ci(desktop): preflight Apple notarization credentials before build (#724 )

Notarization is the last thing tauri-action does, so any credential or account-state fault surfaced ~10 minutes into the macOS job -- after the Rust toolchain, npm install, two Ollama sidecar downloads and a universal cargo build -- as one opaque line: failed to bundle project: failed codesign application: failed to notarize app: Error: HTTP status code: 403. ... That message conflates three unrelated causes, and the signing step succeeds in all of them, so the log actively misleads: the certificate is clearly valid right up until the failure. Add a read-only `notarytool history` call immediately after checkout. It submits nothing and exercises the identical auth path, so all three failures reach us in ~2s with the specific cause and fix named: 401 invalid credentials -> APPLE_PASSWORD is not an app-specific password, or was minted under a different Apple ID than APPLE_ID 403 inaccessible team -> APPLE_ID is not a member of APPLE_TEAM_ID 403 required agreement -> the Program License Agreement lapsed; only the Account Holder can accept it xcrun is preinstalled on macOS runners, hence placement before the toolchain steps rather than beside "Configure Apple signing". Skips cleanly when APPLE_CERTIFICATE is unset (unsigned builds never notarize), mirroring the existing signing step, and errors when a certificate is present but notarization secrets are missing -- previously that combination signed successfully and then failed at the very end. Transient network faults retry 3x; credential errors are deterministic and exit on the first definitive answer. Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
chore: update clone traffic data [skip ci]
2026-08-14 08:52:06 +00:00 · 2026-08-11 17:49:52 -07:00 · 2026-08-11 07:02:06 +00:00 · 2026-08-10 17:22:29 -07:00 · 2026-08-10 17:15:33 -07:00 · 2026-08-10 17:06:48 -07:00
370 changed files with 23522 additions and 6890 deletions
@@ -1,7 +1,7 @@
 {
  "schemaVersion": 1,
  "label": "Git Clones",
-  "message": "107,695",
+  "message": "186,659",
  "color": "green",
  "namedLogo": "git"
 }
@@ -1,6 +1,6 @@
 {
-  "total_clones": 107695,
-  "last_updated": "2026-06-10T07:32:26Z",
+  "total_clones": 186659,
+  "last_updated": "2026-08-11T07:02:05Z",
  "daily": {
    "2026-03-27": 2189,
    "2026-03-28": 1874,
@@ -75,6 +75,69 @@
    "2026-06-05": 2127,
    "2026-06-06": 2204,
    "2026-06-07": 1174,
-    "2026-06-08": 2369
+    "2026-06-08": 2369,
+    "2026-06-09": 1361,
+    "2026-06-10": 1310,
+    "2026-06-11": 2564,
+    "2026-06-12": 1313,
+    "2026-06-13": 2804,
+    "2026-06-14": 1543,
+    "2026-06-15": 1379,
+    "2026-06-16": 1317,
+    "2026-06-17": 1170,
+    "2026-06-18": 1408,
+    "2026-06-19": 1350,
+    "2026-06-20": 1437,
+    "2026-06-21": 1426,
+    "2026-06-22": 1350,
+    "2026-06-23": 1468,
+    "2026-06-24": 1635,
+    "2026-06-25": 1640,
+    "2026-06-26": 1338,
+    "2026-06-27": 1338,
+    "2026-06-28": 1028,
+    "2026-06-29": 765,
+    "2026-06-30": 951,
+    "2026-07-01": 1134,
+    "2026-07-02": 593,
+    "2026-07-03": 537,
+    "2026-07-04": 411,
+    "2026-07-05": 485,
+    "2026-07-06": 555,
+    "2026-07-07": 905,
+    "2026-07-08": 1171,
+    "2026-07-09": 1857,
+    "2026-07-10": 1181,
+    "2026-07-11": 2185,
+    "2026-07-12": 1917,
+    "2026-07-13": 2102,
+    "2026-07-14": 2337,
+    "2026-07-15": 2362,
+    "2026-07-16": 2497,
+    "2026-07-17": 1773,
+    "2026-07-18": 1542,
+    "2026-07-19": 1445,
+    "2026-07-20": 1481,
+    "2026-07-21": 1528,
+    "2026-07-22": 1529,
+    "2026-07-23": 1209,
+    "2026-07-24": 1118,
+    "2026-07-25": 928,
+    "2026-07-26": 740,
+    "2026-07-27": 799,
+    "2026-07-28": 665,
+    "2026-07-29": 745,
+    "2026-07-30": 591,
+    "2026-07-31": 783,
+    "2026-08-01": 567,
+    "2026-08-02": 1248,
+    "2026-08-03": 724,
+    "2026-08-04": 708,
+    "2026-08-05": 647,
+    "2026-08-06": 604,
+    "2026-08-07": 624,
+    "2026-08-08": 706,
+    "2026-08-09": 1076,
+    "2026-08-10": 1060
  }
 }
@@ -21,14 +21,20 @@ jobs:
        id: version
        run: |
          set -euo pipefail
-          # Base version is the next patch above whatever is in pyproject.toml.
+          # Base version is the next patch above the latest plain release tag
+          # (vX.Y.Z) reachable from HEAD. pyproject.toml no longer carries a
+          # static version (#526 switched it to hatch-vcs), so the release tag
+          # is the source of truth. `.devN`/`.rcN`/`desktop-*` tags are excluded
+          # so they can't be mistaken for the release base.
          # Any future manual `X.Y.Z` release will outrank every `X.Y.Z.devN`
          # autotag — PEP 440 sorts dev releases strictly below the final.
-          BASE=$(grep -E '^version = "' pyproject.toml | head -1 | sed -E 's/^version = "([^"]+)"/\1/')
-          if [[ -z "$BASE" ]]; then
-            echo "::error::Could not parse version from pyproject.toml"
+          LATEST_RELEASE=$(git tag --list 'v[0-9]*' --merged HEAD \
+            | grep -E '^v[0-9]+\.[0-9]+\.[0-9]+$' | sort -V | tail -1)
+          if [[ -z "$LATEST_RELEASE" ]]; then
+            echo "::error::No release tag (vX.Y.Z) reachable from HEAD"
            exit 1
          fi
+          BASE="${LATEST_RELEASE#v}"
          MAJOR=$(echo "$BASE" | cut -d. -f1)
          MINOR=$(echo "$BASE" | cut -d. -f2)
          PATCH=$(echo "$BASE" | cut -d. -f3 | sed -E 's/[^0-9].*$//')
@@ -23,6 +23,8 @@ jobs:

      - name: Install uv
        uses: astral-sh/setup-uv@v8.0.0
+        with:
+          enable-cache: true

      - name: Install dependencies
        run: uv sync --extra dev --extra framework-comparison --extra server
@@ -30,6 +32,9 @@ jobs:
      - name: Ruff check
        run: uv run ruff check src/ tests/

+      - name: Ruff format check
+        run: uv run ruff format --check src/ tests/
+
  test:
    runs-on: ubuntu-latest
    steps:
@@ -55,6 +60,8 @@ jobs:

      - name: Install uv
        uses: astral-sh/setup-uv@v8.0.0
+        with:
+          enable-cache: true

      - name: Install dependencies
        run: uv sync --extra dev --extra framework-comparison --extra server
@@ -63,8 +70,13 @@ jobs:
        run: uv run maturin develop --manifest-path rust/crates/openjarvis-python/Cargo.toml

      - name: Run tests
+        # COVERAGE_CORE=sysmon uses CPython 3.12's sys.monitoring backend,
+        # which is dramatically cheaper than the default C trace function.
+        # -n auto fans the suite out across all runner cores via pytest-xdist.
+        env:
+          COVERAGE_CORE: sysmon
        run: |
-          uv run pytest tests/ -v --tb=short -m "not live and not cloud and not hub" \
+          uv run pytest tests/ -n auto -q --tb=short -m "not live and not cloud and not hub" \
            --cov=openjarvis \
            --cov-report=term-missing \
            --cov-report=xml \
@@ -107,6 +119,8 @@ jobs:

      - name: Install uv
        uses: astral-sh/setup-uv@v8.0.0
+        with:
+          enable-cache: true

      - name: Install dependencies
        run: uv sync --extra dev --extra server
@@ -11,22 +11,33 @@ concurrency:
  group: claude-issues-${{ github.event.issue.number || github.run_id }}
  cancel-in-progress: true

+# Least-privilege: only what the issue-fixer job actually needs.
+# id-token (OIDC) is intentionally omitted — claude-code-action@v1 is passed
+# github_token directly, so OIDC is unused here.
 permissions:
  contents: write
  pull-requests: write
  issues: write
-  id-token: write

 jobs:
  fix:
    runs-on: ubuntu-latest
-    timeout-minutes: 60
+    timeout-minutes: 15
+    # Security gate: this job reaches secrets.ANTHROPIC_API_KEY and holds a
+    # write-scoped GITHUB_TOKEN. `issues` / `issue_comment` are public,
+    # attacker-controllable events that run in the base-repo context with full
+    # secret access, so the human-triggered paths are restricted to actors with
+    # write-level association (OWNER / MEMBER / COLLABORATOR). This blocks
+    # external / first-time contributors from draining the API budget or
+    # creating branches/PRs, while leaving maintainer use unaffected.
    if: |
      github.event_name == 'workflow_dispatch' ||
      (github.event_name == 'issues' &&
+       contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.issue.author_association) &&
       (contains(github.event.issue.labels.*.name, 'bug') ||
        contains(github.event.issue.labels.*.name, 'autofix'))) ||
      (github.event_name == 'issue_comment' &&
+       contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.comment.author_association) &&
       !github.event.issue.pull_request &&
       contains(github.event.comment.body, '@claude') &&
       github.actor != 'claude[bot]')
@@ -11,23 +11,33 @@ concurrency:
  group: claude-review-${{ github.event.pull_request.number || github.event.issue.number || github.run_id }}
  cancel-in-progress: true

+# Least-privilege: PR review only needs to post comments on the PR.
+# id-token (OIDC) is omitted — claude-code-action@v1 is passed github_token
+# directly, so OIDC is unused here.
 permissions:
  contents: read
  pull-requests: write
  issues: write
-  id-token: write

 jobs:
  review:
    runs-on: ubuntu-latest
    timeout-minutes: 30
+    # Security gate: this job reaches secrets.ANTHROPIC_API_KEY. Both
+    # issue_comment and pull_request_review_comment are public,
+    # attacker-controllable events that run in the base-repo context with full
+    # secret access, so the @claude paths are restricted to actors with
+    # write-level association (OWNER / MEMBER / COLLABORATOR). External /
+    # first-time contributors cannot trigger the key; maintainers are unaffected.
    if: |
      github.event_name == 'workflow_dispatch' ||
      (github.event_name == 'issue_comment' &&
+       contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.comment.author_association) &&
       github.event.issue.pull_request &&
       contains(github.event.comment.body, '@claude') &&
       github.actor != 'claude[bot]') ||
      (github.event_name == 'pull_request_review_comment' &&
+       contains(fromJSON('["OWNER", "MEMBER", "COLLABORATOR"]'), github.event.comment.author_association) &&
       contains(github.event.comment.body, '@claude') &&
       github.actor != 'claude[bot]')
    steps:
@@ -40,7 +40,8 @@ jobs:
            libappindicator3-dev \
            librsvg2-dev \
            patchelf \
-            libxdo-dev
+            libxdo-dev \
+            libdbus-1-dev

      - name: Setup Node.js
        uses: actions/setup-node@v6
@@ -114,6 +115,109 @@ jobs:

    steps:
      - uses: actions/checkout@v6
+        with:
+          # Full history + tags so the workflow_dispatch fallback in
+          # "Determine release info" can derive the dev version from the
+          # latest release tag (#526).
+          fetch-depth: 0
+
+      # Validate Apple credentials BEFORE the expensive work. Notarization is
+      # the very last thing `tauri-action` does, so a bad credential or a
+      # lapsed account agreement previously surfaced ~10 minutes in — after the
+      # Rust toolchain, npm install, two Ollama sidecar downloads and a
+      # universal cargo build — as a single opaque line:
+      #
+      #   failed to bundle project: failed codesign application: failed to
+      #   notarize app: Error: HTTP status code: 403. ...
+      #
+      # `notarytool history` is a read-only call (it submits nothing) that
+      # exercises the identical auth path, so every credential/account failure
+      # mode reaches us here first, in seconds, with the specific cause named.
+      # `xcrun` is preinstalled on macOS runners, hence placement before the
+      # toolchain steps rather than next to "Configure Apple signing".
+      - name: Preflight Apple notarization credentials
+        if: matrix.platform == 'macos-14'
+        env:
+          CERT: ${{ secrets.APPLE_CERTIFICATE }}
+          A_ID: ${{ secrets.APPLE_ID }}
+          A_PASS: ${{ secrets.APPLE_PASSWORD }}
+          A_TEAM: ${{ secrets.APPLE_TEAM_ID }}
+        shell: bash
+        run: |
+          set -uo pipefail
+
+          # Mirror the skip logic in "Configure Apple signing": without a
+          # certificate the build is unsigned and never notarizes, so there is
+          # nothing to preflight. Tag builds still hard-fail there.
+          if [ -z "$CERT" ]; then
+            echo "No Apple certificate configured; skipping notarization preflight."
+            exit 0
+          fi
+
+          missing=""
+          [ -z "$A_ID" ]   && missing="$missing APPLE_ID"
+          [ -z "$A_PASS" ] && missing="$missing APPLE_PASSWORD"
+          [ -z "$A_TEAM" ] && missing="$missing APPLE_TEAM_ID"
+          if [ -n "$missing" ]; then
+            echo "::error::APPLE_CERTIFICATE is set but notarization secrets are missing:$missing"
+            echo "::error::Signing would succeed and notarization would then fail. Set them or clear APPLE_CERTIFICATE."
+            exit 1
+          fi
+
+          # Retry only to absorb transient network faults. Credential and
+          # account errors are deterministic, so we classify and exit on the
+          # first definitive answer rather than retrying into the same wall.
+          attempt=1
+          while [ "$attempt" -le 3 ]; do
+            out=$(xcrun notarytool history \
+              --apple-id "$A_ID" \
+              --team-id "$A_TEAM" \
+              --password "$A_PASS" \
+              --output-format json 2>&1)
+            rc=$?
+
+            if [ $rc -eq 0 ]; then
+              echo "Apple notarization preflight OK — credentials valid, team reachable, agreements in effect."
+              exit 0
+            fi
+
+            case "$out" in
+              *"Invalid credentials"*|*"401"*)
+                echo "::error::Apple notarization preflight failed: invalid credentials (HTTP 401)."
+                echo "::error::APPLE_PASSWORD must be an app-specific password from appleid.apple.com,"
+                echo "::error::generated while signed in as the SAME Apple ID as APPLE_ID. A regular"
+                echo "::error::Apple ID password will not work, and a password minted under a different"
+                echo "::error::Apple ID authenticates as that other account."
+                exit 1
+                ;;
+              *"Invalid or inaccessible developer team ID"*)
+                echo "::error::Apple notarization preflight failed: APPLE_ID is not a member of team APPLE_TEAM_ID (HTTP 403)."
+                echo "::error::The Team ID must match the signing certificate. Read it from the cert's"
+                echo "::error::subject, where it appears as: Developer ID Application: NAME (TEAMID)."
+                echo "::error::If you belong to several teams, confirm APPLE_ID is a member of this one."
+                exit 1
+                ;;
+              *"required agreement"*|*"agreement"*)
+                echo "::error::Apple notarization preflight failed: the team has no in-effect agreement (HTTP 403)."
+                echo "::error::Apple reissues the Developer Program License Agreement periodically and"
+                echo "::error::notarization is refused until it is accepted. ONLY THE ACCOUNT HOLDER can"
+                echo "::error::accept it — team Admins cannot. Sign in to the account that owns this team:"
+                echo "::error::  1. https://developer.apple.com/account -> review any pending agreement"
+                echo "::error::  2. App Store Connect -> Business -> accept anything pending there too"
+                echo "::error::Certificates stay valid while this is outstanding, so signing still works."
+                exit 1
+                ;;
+            esac
+
+            echo "Preflight attempt ${attempt}/3 failed with a non-credential error."
+            echo "$out" | tail -5
+            attempt=$((attempt + 1))
+            [ "$attempt" -le 3 ] && sleep 10
+          done
+
+          echo "::error::Apple notarization preflight failed after 3 attempts. Last output:"
+          echo "$out" | tail -20
+          exit 1

      - name: Install system dependencies (Linux)
        if: matrix.platform == 'ubuntu-22.04'
@@ -125,7 +229,8 @@ jobs:
            libappindicator3-dev \
            librsvg2-dev \
            patchelf \
-            libxdo-dev
+            libxdo-dev \
+            libdbus-1-dev

      - name: Install Rust stable
        uses: dtolnay/rust-toolchain@stable
@@ -183,7 +288,16 @@ jobs:
            # workflow_dispatch fallback (manual UI dispatch without --ref).
            # Derive a PEP 440 dev version aligned with autotag.yml so we
            # don't burn the X.Y.Z release-version namespace.
-            BASE=$(grep -E '^version = "' pyproject.toml | head -1 | sed -E 's/^version = "([^"]+)"/\1/')
+            # pyproject.toml no longer carries a static version (#526), so the
+            # base comes from the latest plain release tag (vX.Y.Z), matching
+            # autotag.yml. .dev/.rc/desktop-* tags are excluded.
+            LATEST_RELEASE=$(git tag --list 'v[0-9]*' --merged HEAD \
+              | grep -E '^v[0-9]+\.[0-9]+\.[0-9]+$' | sort -V | tail -1)
+            if [[ -z "$LATEST_RELEASE" ]]; then
+              echo "::error::No release tag (vX.Y.Z) reachable from HEAD"
+              exit 1
+            fi
+            BASE="${LATEST_RELEASE#v}"
            MAJOR=$(echo "$BASE" | cut -d. -f1)
            MINOR=$(echo "$BASE" | cut -d. -f2)
            PATCH=$(echo "$BASE" | cut -d. -f3 | sed -E 's/[^0-9].*$//')
@@ -238,6 +352,10 @@ jobs:
          TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
          TAURI_SIGNING_PRIVATE_KEY_PASSWORD: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY_PASSWORD }}
          TAURI_CONFIG: '{"version":"${{ steps.release-info.outputs.tauri_version }}","bundle":{"externalBin":["binaries/ollama"]}}'
+          # tauri-action runs beforeBuildCommand (npm run build:tauri -> vite
+          # build), which requires this at build time (#587). Strict for
+          # releases: a missing/empty secret fails the build by design.
+          VITE_SUPABASE_ANON_KEY: ${{ secrets.VITE_SUPABASE_ANON_KEY }}
        with:
          projectPath: frontend
          tauriScript: npx tauri
@@ -41,6 +41,26 @@ jobs:
      - name: Install dependencies
        run: uv sync --extra docs

+      # Inject the public Supabase anon key so the savings leaderboard works on
+      # the published docs site. Missing/empty (e.g. fork PRs) leaves the
+      # leaderboard gracefully disabled. The key is read from env (not inlined)
+      # and JSON-encoded into a JS string literal to avoid any injection.
+      - name: Inject leaderboard Supabase anon key
+        env:
+          OPENJARVIS_LEADERBOARD_ANON: ${{ secrets.VITE_SUPABASE_ANON_KEY }}
+        run: |
+          python3 - <<'PY'
+          import json, os, pathlib
+
+          key = os.environ.get("OPENJARVIS_LEADERBOARD_ANON", "")
+          pathlib.Path("docs/javascripts/leaderboard-config.js").write_text(
+              "// Generated at docs-build time from the VITE_SUPABASE_ANON_KEY secret.\n"
+              "window.OPENJARVIS_SUPABASE_ANON_KEY = " + json.dumps(key) + ";\n",
+              encoding="utf-8",
+          )
+          print("leaderboard anon key:", "set" if key else "empty (leaderboard disabled)")
+          PY
+
      - name: Build documentation
        run: uv run mkdocs build

@@ -35,3 +35,8 @@ jobs:
      - run: npm ci
      - run: npx tsc --noEmit
      - run: npm run build
+        env:
+          # Optional: when the secret is unset the build still succeeds and the
+          # leaderboard is disabled (see src/lib/supabase.ts). No placeholder,
+          # so a keyless CI build doesn't bake in a bogus anon key.
+          VITE_SUPABASE_ANON_KEY: ${{ secrets.VITE_SUPABASE_ANON_KEY }}
@@ -12,6 +12,11 @@ on:
        description: 'Tag to publish (e.g. v1.0.2.dev500). Overrides github.ref.'
        required: false
        type: string
+      dry_run:
+        description: 'Dry run: build + validate, then publish to TestPyPI instead of PyPI (no production upload).'
+        required: false
+        default: false
+        type: boolean

 permissions:
  contents: read
@@ -50,6 +55,8 @@ jobs:
          cache-dependency-path: frontend/package-lock.json

      - name: Build frontend and bundle into package
+        env:
+          VITE_SUPABASE_ANON_KEY: ${{ secrets.VITE_SUPABASE_ANON_KEY }}
        run: |
          set -euo pipefail
          cd frontend
@@ -67,27 +74,41 @@ jobs:
            exit 1
          }

-      - name: Set version from tag
+      - name: Resolve build version from tag
        env:
          REF: ${{ steps.ref.outputs.ref }}
        run: |
          set -euo pipefail
-          # Strip leading "v" if present (e.g. v1.0.2.dev500 -> 1.0.2.dev500)
+          # Strip leading "v" (e.g. v1.0.3.dev825 -> 1.0.3.dev825).
          VERSION="${REF#v}"
-          if [[ -z "$VERSION" ]]; then
-            echo "::error::Could not resolve version from ref '$REF'"
+          if ! [[ "$VERSION" =~ ^[0-9]+\.[0-9]+\.[0-9]+ ]]; then
+            echo "::error::ref '$REF' is not a version tag (expected vX.Y.Z[.devN]); pass -f tag=vX.Y.Z"
            exit 1
          fi
-          sed -i "s/^version = \".*\"/version = \"${VERSION}\"/" pyproject.toml
-          # Sanity check the substitution actually took
-          grep -q "^version = \"${VERSION}\"" pyproject.toml || {
-            echo "::error::sed failed to update pyproject.toml version"
-            exit 1
-          }
-          echo "Building version $VERSION"
+          # pyproject.toml is now dynamic = ["version"] via hatch-vcs (#526), so
+          # there is no static line to sed. setuptools_scm cannot bump custom
+          # `.devN` tags, so we pin the exact build version explicitly — the
+          # published version always equals the pushed tag.
+          echo "SETUPTOOLS_SCM_PRETEND_VERSION=${VERSION}" >> "$GITHUB_ENV"
+          echo "Building version ${VERSION}"

      - name: Build package
        run: uv build

+      - name: Publish to TestPyPI (dry run)
+        if: ${{ inputs.dry_run }}
+        env:
+          UV_PUBLISH_TOKEN: ${{ secrets.TEST_PYPI_API_TOKEN }}
+        run: |
+          set -euo pipefail
+          if [[ -z "${UV_PUBLISH_TOKEN:-}" ]]; then
+            echo "::warning::TEST_PYPI_API_TOKEN is not set — skipping the TestPyPI upload."
+            echo "Build + twine check passed, which validated version derivation and packaging end to end."
+            echo "To exercise a real upload, add a TEST_PYPI_API_TOKEN secret (or a TestPyPI trusted publisher)."
+            exit 0
+          fi
+          uv publish --publish-url https://test.pypi.org/legacy/
+
      - name: Publish to PyPI
+        if: ${{ !inputs.dry_run }}
        run: uv publish
@@ -8,6 +8,19 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).

 ## [Unreleased]

+### Added
+
+**Vision input for `jarvis ask`** — attach images to a query with
+`-i`/`--image` (repeatable) or capture the current screen with
+`-S`/`--screen`, for vision-capable models such as `gemma3:4b`. Images flow
+through `Message.images` into Ollama's `/api/chat` `images` field; text-only
+requests are unaffected. A privacy guard warns before any image is sent to a
+non-local engine, and the security guardrail now preserves images when it
+sanitizes a flagged prompt. Screen capture uses the built-in Windows .NET
+stack with `mss`/`Pillow` fallbacks on other platforms. Adds the
+`JARVIS_NUM_CTX` environment variable to tune the Ollama context window
+(default `16384`).
+
 ## [1.0.2] - 2026-05-24

 A patch release that fixes a packaging bug which broke the v1.0.1
@@ -0,0 +1,19 @@
+.PHONY: setup build test lint format
+
+# Mirrors .github/workflows/ci.yml so `make test` matches CI locally.
+
+setup:
+	uv sync --extra dev --extra framework-comparison --extra server
+
+build:
+	uv run maturin develop --manifest-path rust/crates/openjarvis-python/Cargo.toml
+
+test: build
+	uv run pytest tests/ -n auto -q --tb=short -m "not live and not cloud and not hub"
+
+lint:
+	uv run ruff check src/ tests/
+	uv run ruff format --check src/ tests/
+
+format:
+	uv run ruff format src/ tests/
@@ -4,7 +4,8 @@
  <p><i>Personal AI, On Personal Devices.</i></p>

  <p>
-    <a href="https://scalingintelligence.stanford.edu/blogs/openjarvis/"><img src="https://img.shields.io/badge/project-OpenJarvis-blue" alt="Project"></a>
+    <a href="https://arxiv.org/abs/2605.17172"><img src="https://img.shields.io/badge/arXiv-2605.17172-b31b1b.svg" alt="arXiv"></a>
+    <a href="https://openjarvis.stanford.edu/"><img src="https://img.shields.io/badge/project-OpenJarvis-blue" alt="Project"></a>
    <a href="https://open-jarvis.github.io/OpenJarvis/"><img src="https://img.shields.io/badge/docs-mkdocs-blue" alt="Docs"></a>
    <img src="https://img.shields.io/badge/python-%3E%3D3.10-blue" alt="Python">
    <img src="https://img.shields.io/badge/license-Apache%202.0-green" alt="License">
@@ -23,7 +24,7 @@

 > **[Documentation](https://open-jarvis.github.io/OpenJarvis/)**
 >
-> **[Project Site](https://scalingintelligence.stanford.edu/blogs/openjarvis/)**
+> **[Project Site](https://openjarvis.stanford.edu/)**
 >
 > **[Paper](https://arxiv.org/abs/2605.17172)**
 >
@@ -106,6 +106,11 @@ enabled = true                            # Record traces for analysis
 db_path = "~/.openjarvis/traces.db"

 [server]
-host = "0.0.0.0"
+# Bind to loopback by default so the API is not exposed to the local network.
+# To serve other devices on your LAN, set host = "0.0.0.0" AND set an API key
+# (OPENJARVIS_API_KEY / `jarvis auth generate-key`) — startup refuses a
+# non-loopback bind without a key. The "server" security profile also flips
+# this to 0.0.0.0 intentionally.
+host = "127.0.0.1"
 port = 8000
 agent = "native_openhands"
@@ -0,0 +1,39 @@
+# Full system access: unrestricted shell and filesystem
+# Copy to ~/.openjarvis/config.toml
+#
+# WARNING: shell_exec runs arbitrary commands as your user. No command
+# allowlist, no denylist, no working-directory restriction. file_read and
+# file_write aren't restricted to any directory either. Only enable what you
+# actually want the agent to have. tools.enabled is the whole permission grant;
+# there's no second allowlist to configure.
+#
+# On macOS, this config alone does not reach TCC-protected data (Messages,
+# Mail, Photos, Safari). That requires Full Disk Access granted to the process
+# hosting the backend. See docs/user-guide/system-access.md.
+#
+# Usage:
+#   jarvis ask "What's using the most disk space in my home directory?"
+#   jarvis chat        # prompts before each shell_exec call
+
+[engine]
+default = "ollama"
+
+[intelligence]
+default_model = "qwen3.5:9b"
+
+[agent]
+default_agent = "orchestrator"
+max_turns = 10
+
+[tools]
+enabled = [
+    "shell_exec",
+    "file_read",
+    "file_write",
+    "apply_patch",
+    "code_interpreter",
+    "git_status",
+    "git_diff",
+    "think",
+    "calculator",
+]
@@ -1,34 +1,83 @@
+# Base images are pinned to an immutable digest (in addition to a human-readable
+# tag) so every build resolves the exact same layers — reproducible builds and
+# safe rollbacks (#563).
+
 # Stage 1: Build frontend SPA
-FROM node:22-slim AS frontend
+FROM node:22.23.0-slim@sha256:d9f850096136edbc402debdd8729579a288aac64574ada0ff4db26b6ae58b0b2 AS frontend
+# Public Supabase anon key for the savings leaderboard; empty by default so
+# the image's leaderboard stays disabled (#589). Pass --build-arg to enable.
+ARG OPENJARVIS_LEADERBOARD_PUBLIC_ANON=

 WORKDIR /frontend
 COPY frontend/package.json frontend/package-lock.json* ./
 RUN npm ci --ignore-scripts 2>/dev/null || npm install
 COPY frontend/ .
-RUN npm run build
+RUN VITE_SUPABASE_ANON_KEY="${OPENJARVIS_LEADERBOARD_PUBLIC_ANON}" npm run build

 # Stage 2: Build Python package
-FROM python:3.12-slim-bookworm AS builder
+FROM python:3.12.13-slim-bookworm@sha256:76d4b7b6305788c6b4c6a19d6a22a3921bf802e9af4d5e1e5bd771208dba74bf AS builder
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends build-essential ca-certificates curl && \
+    rm -rf /var/lib/apt/lists/*
+
+ENV PATH="/root/.cargo/bin:${PATH}"
+
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | \
+        sh -s -- -y --profile minimal --default-toolchain none && \
+    rustup toolchain install 1.88 --profile minimal && \
+    rustup default 1.88

 WORKDIR /app
-COPY pyproject.toml README.md ./
+
+# Install dependencies from the committed lockfile (#567). `uv export --frozen`
+# reads uv.lock as-is (no re-resolution) and emits a fully pinned, hash-verified
+# requirements set; `--no-deps` then installs exactly that set. This is a
+# separate layer from the source copy so dependency installs stay cached when
+# only application code changes.
+COPY pyproject.toml uv.lock README.md ./
+RUN pip install --no-cache-dir uv && \
+    uv export --frozen --no-dev --extra server --no-emit-project > requirements.txt && \
+    uv pip install --system --no-deps -r requirements.txt && \
+    uv pip install --system --no-deps "maturin>=1.12.6,<2"
+
+# Copy the source and the non-src force-include paths (see pyproject
+# [tool.hatch.build.targets.wheel.force-include]) before building the project.
 COPY src/ src/
+COPY rust/ rust/
 COPY scripts/install scripts/install
 COPY deploy/windows deploy/windows

 # Copy built frontend into the server static directory
 COPY --from=frontend /src/openjarvis/server/static src/openjarvis/server/static/

-RUN pip install --no-cache-dir uv && \
-    uv pip install --system ".[server]"
+# Install the project itself without re-resolving dependencies.
+RUN uv pip install --system --no-deps . && \
+    maturin build --release \
+        --manifest-path rust/crates/openjarvis-python/Cargo.toml \
+        --interpreter python3 \
+        --out /tmp/openjarvis-rust-wheel && \
+    uv pip install --system --no-deps /tmp/openjarvis-rust-wheel/*.whl && \
+    python3 -c "import openjarvis_rust; print('openjarvis_rust ok')" && \
+    python3 -m pip uninstall -y maturin && \
+    rm -rf /tmp/openjarvis-rust-wheel rust

 # Stage 3: Runtime
-FROM python:3.12-slim-bookworm
+FROM python:3.12.13-slim-bookworm@sha256:76d4b7b6305788c6b4c6a19d6a22a3921bf802e9af4d5e1e5bd771208dba74bf

 COPY --from=builder /usr/local /usr/local
 COPY --from=builder /app /app
 WORKDIR /app

+# Run as an unprivileged user — the server needs no root privileges, so dropping
+# them limits the blast radius of a compromise (#565). The app writes only to
+# $HOME (config/cache/state), which is owned by this user.
+RUN groupadd --system --gid 10001 openjarvis && \
+    useradd --system --uid 10001 --gid openjarvis \
+        --create-home --home-dir /home/openjarvis openjarvis
+ENV HOME=/home/openjarvis
+USER openjarvis
+
 EXPOSE 8000

 ENTRYPOINT ["jarvis"]
@@ -1,32 +1,69 @@
+# Base images are pinned to an immutable digest (in addition to a human-readable
+# tag) so every build resolves the exact same layers — reproducible builds and
+# safe rollbacks (#563).
+
 # Stage 1: Build frontend SPA
-FROM node:22-slim AS frontend
+FROM node:22.23.0-slim@sha256:d9f850096136edbc402debdd8729579a288aac64574ada0ff4db26b6ae58b0b2 AS frontend
+# Public Supabase anon key for the savings leaderboard; empty by default so
+# the image's leaderboard stays disabled (#589). Pass --build-arg to enable.
+ARG OPENJARVIS_LEADERBOARD_PUBLIC_ANON=

 WORKDIR /frontend
 COPY frontend/package.json frontend/package-lock.json* ./
 RUN npm ci --ignore-scripts 2>/dev/null || npm install
 COPY frontend/ .
-RUN npm run build
+RUN VITE_SUPABASE_ANON_KEY="${OPENJARVIS_LEADERBOARD_PUBLIC_ANON}" npm run build

 # Stage 2: Build Python package (NVIDIA CUDA 12.4)
-FROM nvidia/cuda:12.4.0-runtime-ubuntu22.04 AS builder
+FROM nvidia/cuda:12.4.0-runtime-ubuntu22.04@sha256:af8bd179ed3bf69d4b63b19a763662a6141f0f62ef099283f68d0b14b4bab0e3 AS builder

 RUN apt-get update && \
-    apt-get install -y --no-install-recommends python3 python3-pip python3-venv && \
+    apt-get install -y --no-install-recommends \
+        build-essential \
+        ca-certificates \
+        curl \
+        python3 \
+        python3-dev \
+        python3-pip \
+        python3-venv && \
    rm -rf /var/lib/apt/lists/*

+ENV PATH="/root/.cargo/bin:${PATH}"
+
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | \
+        sh -s -- -y --profile minimal --default-toolchain none && \
+    rustup toolchain install 1.88 --profile minimal && \
+    rustup default 1.88
+
 WORKDIR /app
-COPY pyproject.toml README.md ./
+
+# Install dependencies from the committed lockfile (#567). See deploy/docker/Dockerfile
+# for the rationale behind the frozen export + --no-deps install.
+COPY pyproject.toml uv.lock README.md ./
+RUN pip install --no-cache-dir uv && \
+    uv export --frozen --no-dev --extra server --no-emit-project > requirements.txt && \
+    uv pip install --system --no-deps -r requirements.txt && \
+    uv pip install --system --no-deps "maturin>=1.12.6,<2"
+
 COPY src/ src/
+COPY rust/ rust/
 COPY scripts/install scripts/install
 COPY deploy/windows deploy/windows

 COPY --from=frontend /src/openjarvis/server/static src/openjarvis/server/static/

-RUN pip install --no-cache-dir uv && \
-    uv pip install --system ".[server]"
+RUN uv pip install --system --no-deps . && \
+    maturin build --release \
+        --manifest-path rust/crates/openjarvis-python/Cargo.toml \
+        --interpreter python3 \
+        --out /tmp/openjarvis-rust-wheel && \
+    uv pip install --system --no-deps /tmp/openjarvis-rust-wheel/*.whl && \
+    python3 -c "import openjarvis_rust; print('openjarvis_rust ok')" && \
+    python3 -m pip uninstall -y maturin && \
+    rm -rf /tmp/openjarvis-rust-wheel rust

 # Stage 3: Runtime
-FROM nvidia/cuda:12.4.0-runtime-ubuntu22.04
+FROM nvidia/cuda:12.4.0-runtime-ubuntu22.04@sha256:af8bd179ed3bf69d4b63b19a763662a6141f0f62ef099283f68d0b14b4bab0e3

 RUN apt-get update && \
    apt-get install -y --no-install-recommends python3 python3-pip && \
@@ -36,6 +73,14 @@ COPY --from=builder /usr/local /usr/local
 COPY --from=builder /app /app
 WORKDIR /app

+# Run as an unprivileged user (#565). NVIDIA device nodes (/dev/nvidia*) are
+# world-accessible, so GPU workloads do not require root.
+RUN groupadd --system --gid 10001 openjarvis && \
+    useradd --system --uid 10001 --gid openjarvis \
+        --create-home --home-dir /home/openjarvis openjarvis
+ENV HOME=/home/openjarvis
+USER openjarvis
+
 EXPOSE 8000

 ENTRYPOINT ["jarvis"]
@@ -1,32 +1,69 @@
+# Base images are pinned to an immutable digest (in addition to a human-readable
+# tag) so every build resolves the exact same layers — reproducible builds and
+# safe rollbacks (#563).
+
 # Stage 1: Build frontend SPA
-FROM node:22-slim AS frontend
+FROM node:22.23.0-slim@sha256:d9f850096136edbc402debdd8729579a288aac64574ada0ff4db26b6ae58b0b2 AS frontend
+# Public Supabase anon key for the savings leaderboard; empty by default so
+# the image's leaderboard stays disabled (#589). Pass --build-arg to enable.
+ARG OPENJARVIS_LEADERBOARD_PUBLIC_ANON=

 WORKDIR /frontend
 COPY frontend/package.json frontend/package-lock.json* ./
 RUN npm ci --ignore-scripts 2>/dev/null || npm install
 COPY frontend/ .
-RUN npm run build
+RUN VITE_SUPABASE_ANON_KEY="${OPENJARVIS_LEADERBOARD_PUBLIC_ANON}" npm run build

 # Stage 2: Build Python package (AMD ROCm 7.2)
-FROM rocm/dev-ubuntu-22.04:7.2 AS builder
+FROM rocm/dev-ubuntu-22.04:7.2@sha256:05af5f04a06b04676d4c7438997d0deadaeb7478961ad621376e199bf3aeb644 AS builder

 RUN apt-get update && \
-    apt-get install -y --no-install-recommends python3 python3-pip python3-venv && \
+    apt-get install -y --no-install-recommends \
+        build-essential \
+        ca-certificates \
+        curl \
+        python3 \
+        python3-dev \
+        python3-pip \
+        python3-venv && \
    rm -rf /var/lib/apt/lists/*

+ENV PATH="/root/.cargo/bin:${PATH}"
+
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | \
+        sh -s -- -y --profile minimal --default-toolchain none && \
+    rustup toolchain install 1.88 --profile minimal && \
+    rustup default 1.88
+
 WORKDIR /app
-COPY pyproject.toml README.md ./
+
+# Install dependencies from the committed lockfile (#567). See deploy/docker/Dockerfile
+# for the rationale behind the frozen export + --no-deps install.
+COPY pyproject.toml uv.lock README.md ./
+RUN pip install --no-cache-dir uv && \
+    uv export --frozen --no-dev --extra server --no-emit-project > requirements.txt && \
+    uv pip install --system --no-deps -r requirements.txt && \
+    uv pip install --system --no-deps "maturin>=1.12.6,<2"
+
 COPY src/ src/
+COPY rust/ rust/
 COPY scripts/install scripts/install
 COPY deploy/windows deploy/windows

 COPY --from=frontend /src/openjarvis/server/static src/openjarvis/server/static/

-RUN pip install --no-cache-dir uv && \
-    uv pip install --system ".[server]"
+RUN uv pip install --system --no-deps . && \
+    maturin build --release \
+        --manifest-path rust/crates/openjarvis-python/Cargo.toml \
+        --interpreter python3 \
+        --out /tmp/openjarvis-rust-wheel && \
+    uv pip install --system --no-deps /tmp/openjarvis-rust-wheel/*.whl && \
+    python3 -c "import openjarvis_rust; print('openjarvis_rust ok')" && \
+    python3 -m pip uninstall -y maturin && \
+    rm -rf /tmp/openjarvis-rust-wheel rust

 # Stage 3: Runtime
-FROM rocm/dev-ubuntu-22.04:7.2
+FROM rocm/dev-ubuntu-22.04:7.2@sha256:05af5f04a06b04676d4c7438997d0deadaeb7478961ad621376e199bf3aeb644

 RUN apt-get update && \
    apt-get install -y --no-install-recommends python3 python3-pip && \
@@ -36,6 +73,18 @@ COPY --from=builder /usr/local /usr/local
 COPY --from=builder /app /app
 WORKDIR /app

+# Run as an unprivileged user (#565). ROCm GPU access is gated by the `video` and
+# `render` groups (see group_add in docker-compose.gpu.rocm.yml), so the user is
+# added to both; root is not required.
+RUN groupadd --system --gid 10001 openjarvis && \
+    useradd --system --uid 10001 --gid openjarvis \
+        --create-home --home-dir /home/openjarvis openjarvis && \
+    (getent group video  >/dev/null || groupadd --system video) && \
+    (getent group render >/dev/null || groupadd --system render) && \
+    usermod -aG video,render openjarvis
+ENV HOME=/home/openjarvis
+USER openjarvis
+
 EXPOSE 8000

 ENTRYPOINT ["jarvis"]
@@ -1,15 +1,69 @@
-FROM python:3.12-slim
+# Base images are pinned to an immutable digest (in addition to a human-readable
+# tag) so every build resolves the exact same layers (#563).

-# Install Node.js 22
-RUN apt-get update && apt-get install -y --no-install-recommends \
-    curl ca-certificates && \
-    curl -fsSL https://deb.nodesource.com/setup_22.x | bash - && \
-    apt-get install -y nodejs && \
+# Node.js is sourced from the official, digest-pinned image rather than piping a
+# remote setup script into bash (`curl ... | bash -`), which performed no
+# checksum or signature verification of the downloaded installer (#566). The
+# image digest is the integrity check, and the copy is architecture-agnostic.
+FROM node:22.23.0-slim@sha256:d9f850096136edbc402debdd8729579a288aac64574ada0ff4db26b6ae58b0b2 AS node
+
+FROM python:3.12.13-slim-bookworm@sha256:76d4b7b6305788c6b4c6a19d6a22a3921bf802e9af4d5e1e5bd771208dba74bf AS builder
+
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends build-essential ca-certificates curl && \
    rm -rf /var/lib/apt/lists/*

+ENV PATH="/root/.cargo/bin:${PATH}"
+
+RUN curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | \
+        sh -s -- -y --profile minimal --default-toolchain none && \
+    rustup toolchain install 1.88 --profile minimal && \
+    rustup default 1.88
+
 WORKDIR /app
+
+# Install dependencies from the committed lockfile (#567): `uv export --frozen`
+# reads uv.lock as-is and emits a pinned, hash-verified set installed with
+# --no-deps (no re-resolution). Copied first so this layer caches independently
+# of application source.
+COPY pyproject.toml uv.lock README.md ./
+RUN pip install --no-cache-dir uv && \
+    uv export --frozen --no-dev --extra server --no-emit-project > requirements.txt && \
+    uv pip install --system --no-deps -r requirements.txt && \
+    uv pip install --system --no-deps "maturin>=1.12.6,<2"
+
 COPY . .
-RUN pip install --no-cache-dir ".[server]"
+
+# Install the project itself without re-resolving dependencies.
+RUN uv pip install --system --no-deps . && \
+    maturin build --release \
+        --manifest-path rust/crates/openjarvis-python/Cargo.toml \
+        --interpreter python3 \
+        --out /tmp/openjarvis-rust-wheel && \
+    uv pip install --system --no-deps /tmp/openjarvis-rust-wheel/*.whl && \
+    python3 -c "import openjarvis_rust; print('openjarvis_rust ok')" && \
+    python3 -m pip uninstall -y maturin && \
+    rm -rf /tmp/openjarvis-rust-wheel rust/target
+
+FROM python:3.12.13-slim-bookworm@sha256:76d4b7b6305788c6b4c6a19d6a22a3921bf802e9af4d5e1e5bd771208dba74bf
+
+# libstdc++6 + ca-certificates are the only runtime requirements of the Node
+# binary copied below (the python slim image already provides libc/libgcc).
+RUN apt-get update && \
+    apt-get install -y --no-install-recommends ca-certificates libstdc++6 && \
+    rm -rf /var/lib/apt/lists/*
+
+COPY --from=builder /usr/local /usr/local
+COPY --from=builder /app /app
+
+# Transplant the Node.js runtime from the official image. Both images are Debian
+# bookworm, so the glibc/libstdc++ ABI matches.
+COPY --from=node /usr/local/bin/node /usr/local/bin/node
+COPY --from=node /usr/local/lib/node_modules /usr/local/lib/node_modules
+RUN ln -sf /usr/local/lib/node_modules/npm/bin/npm-cli.js /usr/local/bin/npm && \
+    ln -sf /usr/local/lib/node_modules/npm/bin/npx-cli.js /usr/local/bin/npx
+
+WORKDIR /app

 LABEL openjarvis-sandbox=true

@@ -18,7 +18,9 @@ services:
              capabilities: [gpu]

  ollama:
-    image: ollama/ollama:latest
+    # Pinned to a fixed version + digest for reproducible deployments (#563);
+    # must match the tag in docker-compose.yml.
+    image: ollama/ollama:0.30.10@sha256:bfc9c6d53cc6989aa5131a6fde6b162b2802d4d337657f3253b5f69579bddeee
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
@@ -18,7 +18,9 @@ services:
    restart: unless-stopped

  ollama:
-    image: ollama/ollama:latest
+    # Pinned to a fixed version + digest for reproducible deployments and
+    # predictable rollbacks (#563). Bump deliberately, not implicitly via :latest.
+    image: ollama/ollama:0.30.10@sha256:bfc9c6d53cc6989aa5131a6fde6b162b2802d4d337657f3253b5f69579bddeee
    ports:
      - "11434:11434"
    volumes:
@@ -14,7 +14,27 @@ Environment=HOME=/opt/openjarvis
 #   OPENJARVIS_API_KEY=<key>      (generate one: `jarvis auth generate-key`)
 # It is not prefixed with "-", so the unit fails to start if the file is
 # missing — preventing an accidentally unauthenticated public server.
+# Keep secrets here (mode 0600, owned by root) rather than inline Environment=
+# lines, which leak into `systemctl show` and the journal.
 EnvironmentFile=/etc/openjarvis/env

+# --- Sandboxing / hardening (#564) ---
+# Conservative set: tightens the unit without blocking the server's normal I/O
+# or local GPU inference. ProtectSystem=strict makes the whole filesystem
+# read-only except ReadWritePaths, so $HOME (config/cache/state under
+# /opt/openjarvis) stays writable.
+NoNewPrivileges=true
+ProtectSystem=strict
+ReadWritePaths=/opt/openjarvis
+ProtectHome=true
+PrivateTmp=true
+ProtectControlGroups=true
+ProtectKernelLogs=true
+ProtectKernelModules=true
+ProtectKernelTunables=true
+RestrictRealtime=true
+RestrictSUIDSGID=true
+LockPersonality=true
+
 [Install]
 WantedBy=multi-user.target
@@ -20,8 +20,8 @@ What it does:
 4. Installs `uv` (https://astral.sh/uv) if absent.
 5. Clones the OpenJarvis repository to `%LOCALAPPDATA%\OpenJarvis`
   (override with `$env:OPENJARVIS_HOME`).
-6. Runs `uv sync --extra server` so the FastAPI server entry point is
-   importable.
+6. Runs `uv sync --extra desktop --group desktop-native` so the FastAPI server,
+   speech backend, and native extension are importable.
 7. Optionally prompts to register a scheduled task that auto-starts the
   server at logon.

@@ -105,7 +105,7 @@ To pull the latest:
 ```powershell
 cd "$env:LOCALAPPDATA\OpenJarvis\src"
 git pull --ff-only
-uv sync --extra server
+uv sync --extra desktop --group desktop-native
 ```

 Or re-run the installer with `-Force`:
@@ -5,19 +5,19 @@
 .DESCRIPTION
    Phase-1 of the native-Windows-support RFC (#298). Mirrors the
    behavior of scripts/install/install.sh (the curl-pipe-bash installer
-    for Linux/WSL2/macOS) but for native Windows PowerShell — no WSL,
+    for Linux/WSL2/macOS) but for native Windows PowerShell - no WSL,
    no Docker, no MSYS2.

    Steps:
      1. Refuse non-Windows / Windows < 10.
-      2. Check Python 3.10 — 3.13 on PATH (3.14 has no numpy wheels yet,
+      2. Check Python 3.10 - 3.13 on PATH (3.14 has no numpy wheels yet,
         see #432).
      3. Check git on PATH.
      4. Install uv (https://astral.sh/uv) if absent.
      5. Clone the OpenJarvis repository to $env:LOCALAPPDATA\OpenJarvis
         (override with $env:OPENJARVIS_HOME).
-      6. Run `uv sync --extra server` so the FastAPI server entry point
-         is importable.
+      6. Run `uv sync --extra desktop --group desktop-native` so the FastAPI
+         server, speech backend, and native extension are importable.
      7. Optionally register the scheduled-task service (see
         deploy/windows/jarvis-service.ps1).

@@ -65,7 +65,7 @@ if (-not $Service     -and $env:OPENJARVIS_SERVICE)      { $Service     = $true
 if (-not $Force       -and $env:OPENJARVIS_FORCE)        { $Force       = $true }

 # ---------------------------------------------------------------------------
-# Output helpers — coloured but plain enough for Constrained Language Mode.
+# Output helpers - coloured but plain enough for Constrained Language Mode.
 # ---------------------------------------------------------------------------

 function Write-Info  ($msg) { Write-Host "[info]  $msg" -ForegroundColor Cyan }
@@ -77,13 +77,13 @@ function Write-Fail  ($msg) {
 }

 # ---------------------------------------------------------------------------
-# Shared helpers — winget bootstrap + PATH refresh
+# Shared helpers - winget bootstrap + PATH refresh
 # ---------------------------------------------------------------------------

 # Pull the latest Machine + User PATH from the registry into the current
 # PowerShell session. Tools installed by `winget install` (Python, git,
 # Ollama, etc.) update the User PATH, but the running process inherits
-# the parent shell's environment — so without this refresh the just-
+# the parent shell's environment - so without this refresh the just-
 # installed tool stays invisible to subsequent `Get-Command` calls.
 #
 # CRITICAL: registry PATH entries can be REG_EXPAND_SZ (with literal
@@ -157,7 +157,7 @@ function Get-PythonCommand {
 Write-Info "Checking Python (3.10 - 3.13)..."
 $pythonExe = Get-PythonCommand
 if (-not $pythonExe) {
-    Write-Info "Python not on PATH — attempting auto-install via winget..."
+    Write-Info "Python not on PATH - attempting auto-install via winget..."
    $pythonExe = Install-WithWinget -WingetId 'Python.Python.3.13' -CommandName 'python'
    if (-not $pythonExe) {
        Write-Fail @"
@@ -196,7 +196,7 @@ Write-Ok "Python $pyMajor.$pyMinor ($pythonExe)"
 Write-Info "Checking git..."
 $gitExe = (Get-Command git -ErrorAction SilentlyContinue).Source
 if (-not $gitExe) {
-    Write-Info "git not on PATH — attempting auto-install via winget..."
+    Write-Info "git not on PATH - attempting auto-install via winget..."
    $gitExe = Install-WithWinget -WingetId 'Git.Git' -CommandName 'git'
    if (-not $gitExe) {
        Write-Fail @"
@@ -227,7 +227,7 @@ if (-not $uvExe) {
    }
    # The astral installer puts uv at %USERPROFILE%\.local\bin\uv.exe and
    # adds that dir to the User PATH. The current process's PATH isn't
-    # refreshed automatically — prepend the install dir so the rest of
+    # refreshed automatically - prepend the install dir so the rest of
    # this script picks it up.
    $uvDir = Join-Path $env:USERPROFILE '.local\bin'
    if (Test-Path (Join-Path $uvDir 'uv.exe')) {
@@ -279,13 +279,13 @@ if (Test-Path (Join-Path $srcDir '.git')) {
 }

 # ---------------------------------------------------------------------------
-# 6. uv sync --extra server
+# 6. uv sync --extra desktop --group desktop-native
 # ---------------------------------------------------------------------------

-Write-Info "Running 'uv sync --extra server' in $srcDir (this can take a few minutes)..."
+Write-Info "Running 'uv sync --extra desktop --group desktop-native' in $srcDir (this can take a few minutes)..."
 Push-Location $srcDir
 try {
-    & $uvExe sync --extra server
+    & $uvExe sync --extra desktop --group desktop-native
    if ($LASTEXITCODE -ne 0) {
        Write-Fail "uv sync failed with exit code $LASTEXITCODE. Check the output above."
    }
@@ -295,13 +295,13 @@ try {
 Write-Ok "Dependencies installed"

 # ---------------------------------------------------------------------------
-# 7. Ollama — install + start + wait for daemon
+# 7. Ollama - install + start + wait for daemon
 # ---------------------------------------------------------------------------

 Write-Info "Checking Ollama..."
 $ollamaExe = (Get-Command ollama -ErrorAction SilentlyContinue).Source
 if (-not $ollamaExe) {
-    Write-Info "  Ollama not on PATH — downloading the official installer (~150 MB)..."
+    Write-Info "  Ollama not on PATH - downloading the official installer (~150 MB)..."
    $ollamaSetup = Join-Path $env:TEMP 'OllamaSetup.exe'
    # SilentlyContinue is load-bearing in PS 5.1: the default progress
    # bar renderer slows Invoke-WebRequest down 30x on large downloads
@@ -340,13 +340,18 @@ Write-Ok "Ollama ($ollamaExe)"
 Write-Info "Waiting for Ollama daemon..."
 $ollamaReady = $false
 for ($i = 0; $i -lt 60; $i++) {
-    & $ollamaExe list 2>&1 | Out-Null
+    # 'ollama list' writes to stderr until the daemon is reachable; under
+    # $ErrorActionPreference='Stop' the 2>&1 merge surfaces that as a
+    # terminating NativeCommandError that would abort the whole install on
+    # the very first probe. Swallow it and rely on $LASTEXITCODE so the
+    # Start-Process serve fallback below actually runs (issue #522).
+    try { & $ollamaExe list 2>&1 | Out-Null } catch { }
    if ($LASTEXITCODE -eq 0) {
        $ollamaReady = $true
        break
    }
    if ($i -eq 5) {
-        # Daemon clearly isn't auto-running — start it ourselves. Ollama
+        # Daemon clearly isn't auto-running - start it ourselves. Ollama
        # for Windows uses the tray app `ollama app.exe`; falling back to
        # `ollama serve` works headless.
        Start-Process -FilePath $ollamaExe -ArgumentList 'serve' -WindowStyle Hidden -ErrorAction SilentlyContinue
@@ -354,11 +359,11 @@ for ($i = 0; $i -lt 60; $i++) {
    Start-Sleep -Seconds 1
 }
 if (-not $ollamaReady) {
-    Write-Warn2 "Ollama daemon didn't become ready in 60s. Continuing — bg-orchestrator will retry later."
+    Write-Warn2 "Ollama daemon didn't become ready in 60s. Continuing - bg-orchestrator will retry later."
 }

 # ---------------------------------------------------------------------------
-# 8. Pull a starter model (qwen3.5:2b — ~1.5 GB)
+# 8. Pull a starter model (qwen3.5:2b - ~1.5 GB)
 # ---------------------------------------------------------------------------

 $modelPullOk = $false
@@ -372,11 +377,11 @@ if ($ollamaReady) {
        Write-Warn2 "ollama pull failed; the bg-orchestrator will retry once Ollama is reachable."
    }
 } else {
-    Write-Warn2 "Skipping model pull — daemon wasn't ready."
+    Write-Warn2 "Skipping model pull - daemon wasn't ready."
 }

 # ---------------------------------------------------------------------------
-# 9. jarvis.cmd shim — so bare `jarvis` works in any new PowerShell
+# 9. jarvis.cmd shim - so bare `jarvis` works in any new PowerShell
 # ---------------------------------------------------------------------------

 $binDir = Join-Path $installRoot 'bin'
@@ -387,7 +392,7 @@ if (-not (Test-Path $binDir)) {
 }

 # %~dp0 in a .cmd file resolves to the directory containing the script,
-# so the shim is self-locating — moving %LOCALAPPDATA%\OpenJarvis won't
+# so the shim is self-locating - moving %LOCALAPPDATA%\OpenJarvis won't
 # break it as long as the user moves the whole tree. `uv` is resolved
 # from PATH at runtime (astral installer adds it to User PATH); avoids
 # pinning to the install-time uv.exe path which can shift on uv updates.
@@ -400,7 +405,7 @@ uv run --project "%SRC%" jarvis %*
 Set-Content -Path $shimPath -Value $shimContent -Encoding ASCII

 # Add %LOCALAPPDATA%\OpenJarvis\bin to User PATH if it isn't already
-# there. The current process won't see it until restart — handled in the
+# there. The current process won't see it until restart - handled in the
 # final banner.
 #
 # Compare against the EXPANDED form: a previous install may have written
@@ -430,7 +435,7 @@ Write-Ok "jarvis shim installed at $shimPath"
 $serviceScript = Join-Path $srcDir 'deploy\windows\jarvis-service.ps1'
 $shouldInstallService = $false

-# Pre-check admin if the user wants the service — Register-ScheduledTask
+# Pre-check admin if the user wants the service - Register-ScheduledTask
 # requires elevation. We do this before the prompt so we don't ask "do
 # you want the service?" only to fail with Access Denied after they say
 # yes.
@@ -439,7 +444,7 @@ $isAdmin = ([Security.Principal.WindowsPrincipal] `
 ).IsInRole([Security.Principal.WindowsBuiltInRole]::Administrator)

 if ($Service -and -not $isAdmin) {
-    Write-Fail "-Service was requested, but this PowerShell is not elevated. Register-ScheduledTask needs admin rights — re-run from an elevated PowerShell, or drop -Service."
+    Write-Fail "-Service was requested, but this PowerShell is not elevated. Register-ScheduledTask needs admin rights - re-run from an elevated PowerShell, or drop -Service."
 }
 if ($Service) {
    $shouldInstallService = $true
@@ -448,7 +453,7 @@ if ($Service) {
 } elseif (-not $isAdmin) {
    # Default to skip-with-explanation when we can't elevate, rather
    # than prompting and then failing at Register-ScheduledTask.
-    Write-Warn2 "Skipping scheduled-task setup — this PowerShell is not elevated."
+    Write-Warn2 "Skipping scheduled-task setup - this PowerShell is not elevated."
    Write-Warn2 "  Register-ScheduledTask requires admin. To install the service later:"
    Write-Warn2 "    Right-click PowerShell -> Run as administrator, then run:"
    Write-Warn2 "    powershell -ExecutionPolicy Bypass -File `"$serviceScript`" install"
@@ -464,7 +469,7 @@ if ($Service) {
        $reply = Read-Host "Register OpenJarvis as a Windows scheduled task (auto-start at logon, loopback only)? [y/N]"
        $shouldInstallService = ($reply -match '^[yY]')
    } else {
-        Write-Warn2 "Non-interactive install — skipping scheduled-task setup."
+        Write-Warn2 "Non-interactive install - skipping scheduled-task setup."
        Write-Warn2 "To register the service later, run (from an elevated PowerShell):"
        Write-Warn2 "  powershell -ExecutionPolicy Bypass -File `"$serviceScript`" install"
    }
@@ -487,9 +492,9 @@ if ($shouldInstallService) {
 # ---------------------------------------------------------------------------

 Write-Host ""
-Write-Host "  ┌──────────────────────────────────┐" -ForegroundColor Green
-Write-Host "  │   OpenJarvis install complete    │" -ForegroundColor Green
-Write-Host "  └──────────────────────────────────┘" -ForegroundColor Green
+Write-Host "  +----------------------------------+" -ForegroundColor Green
+Write-Host "  |   OpenJarvis install complete    |" -ForegroundColor Green
+Write-Host "  +----------------------------------+" -ForegroundColor Green
 Write-Host ""
 Write-Host "  Repo:    $srcDir"

@@ -109,7 +109,7 @@ If you prefer to run each step yourself:
    ```bash
    git clone https://github.com/open-jarvis/OpenJarvis.git
    cd OpenJarvis
-    uv sync --extra server
+    uv sync --extra desktop
    cd frontend && npm install && cd ..
    ```

@@ -17,6 +17,46 @@ The configuration file lives at:

 OpenJarvis creates the `~/.openjarvis/` directory and populates it with a default config when you run `jarvis init`.

+## Relocating the OpenJarvis directory
+
+OpenJarvis keeps **all** of its state — config, databases, caches, logs,
+credentials, skills, recipes, connectors — under a **single root** so it never
+clutters your home directory beyond one folder. By default that root is
+`~/.openjarvis`, but you can move it.
+
+The root is resolved in priority order:
+
+1. **`$OPENJARVIS_HOME`** — explicit override. Honored by both the installer
+   and the Python runtime.
+2. **`$XDG_DATA_HOME/openjarvis`** — used when `$XDG_DATA_HOME` is set (a single
+   `openjarvis` directory nested under it, per the XDG Base Directory spec).
+3. **`~/.openjarvis`** — the default. With no environment variables set, the
+   resolved path is exactly this, so existing installs are untouched.
+
+```bash
+# Relocate the whole install + runtime tree at install time:
+OPENJARVIS_HOME=~/apps/openjarvis curl -fsSL https://open-jarvis.github.io/OpenJarvis/install.sh | bash
+
+# Or for a single run / your shell profile:
+export OPENJARVIS_HOME=~/apps/openjarvis
+```
+
+Confirm where your data lives with:
+
+```bash
+jarvis config path
+```
+
+!!! note "Migration"
+    Because the default is unchanged, **no data migration is required** for
+    existing installs. If you set `OPENJARVIS_HOME` (or `XDG_DATA_HOME`) on a
+    machine that already has data in `~/.openjarvis`, OpenJarvis will look in
+    the new location and not see your old data — move it yourself if you want
+    to keep it: `mv ~/.openjarvis "$OPENJARVIS_HOME"`.
+
+`$OPENJARVIS_CONFIG` still points at an explicit `config.toml` file
+independently of the root, if you need to override just the config file path.
+
 ## Generating Configuration

 ### First-Time Setup
@@ -564,7 +604,7 @@ enforce_tool_confirmation = true
 | `scan_output` | bool | `true` | Whether to scan model output. |
 | `secret_scanner` | bool | `true` | Enable secret detection (API keys, tokens, passwords). |
 | `pii_scanner` | bool | `true` | Enable PII detection (emails, SSNs, credit cards). |
-| `enforce_tool_confirmation` | bool | `true` | Require confirmation before executing tools. |
+| `enforce_tool_confirmation` | bool | `true` | Accepted but **not currently enforced**. Whether you get prompts depends on the entry point. See [System Access](../user-guide/system-access.md#confirmation-behaviour). |

 !!! tip "Choosing a security mode"
    Use `"warn"` during development to see what would be flagged without disrupting output.
@@ -41,7 +41,7 @@ If you prefer to run each step yourself:
    ```bash
    git clone https://github.com/open-jarvis/OpenJarvis.git
    cd OpenJarvis
-    uv sync --extra server
+    uv sync --extra desktop
    uv run maturin develop -m rust/crates/openjarvis-python/Cargo.toml
    cd frontend && npm install && cd ..
    ```
@@ -278,6 +278,7 @@ OpenJarvis uses optional extras to keep the base installation lightweight.

 | Extra | Install Command | Description |
 |-------|----------------|-------------|
+| `desktop` | `uv sync --extra desktop` | Desktop/API server plus local speech input |
 | `server` | `uv sync --extra server` | OpenAI-compatible API server (`jarvis serve`) |
 | `dev` | `uv sync --extra dev` | Development and testing tools |
 | `docs` | `uv sync --extra docs` | Documentation build tools |
@@ -285,7 +286,7 @@ OpenJarvis uses optional extras to keep the base installation lightweight.
 Combine extras:

 ```bash
-uv sync --extra server --extra memory-faiss --extra inference-cloud
+uv sync --extra desktop --extra memory-faiss --extra inference-cloud
 ```

 ## Setting Up an Inference Backend
@@ -135,6 +135,19 @@ cd OpenJarvis
 This launches the backend API server and a React frontend at [http://localhost:5173](http://localhost:5173).
 You get a ChatGPT-like interface with streaming responses, tool use, energy monitoring, and a telemetry dashboard — all running locally on your hardware.

+Web search is available through the built-in DuckDuckGo fallback. To use
+Tavily, add `TAVILY_API_KEY` under **Settings → Tools → Web Search** after the
+app starts, or export it before starting quickstart:
+
+```bash
+export TAVILY_API_KEY="tvly-..."
+./scripts/quickstart.sh
+```
+
+The script does not automatically source `.env` files. Run `source .env`
+first if that is where you keep the key. Stop any existing OpenJarvis server
+before restarting so it inherits the updated environment.
+
 To stop all services, press ++ctrl+c++ in the terminal.

 !!! tip "Environment variable"
@@ -8,7 +8,7 @@ avoid a Linux VM; WSL2 remains the smoother experience for most users.
 ## What you get

 - A PowerShell installer that probes prerequisites, installs `uv`,
-  clones the repo, and runs `uv sync --extra server`.
+  clones the repo, and runs `uv sync --extra desktop --group desktop-native`.
 - An optional Windows scheduled-task service equivalent to the systemd
  unit and launchd plist.
 - Loopback default — the service binds `127.0.0.1` so no API key is
@@ -38,7 +38,7 @@ The installer will:
 4. Install `uv` if absent (via the official `astral.sh/uv` PowerShell
   installer).
 5. Clone the repo to `%LOCALAPPDATA%\OpenJarvis\src`.
-6. Run `uv sync --extra server`.
+6. Run `uv sync --extra desktop --group desktop-native`.
 7. Prompt to register the scheduled-task service (skip with
   `-SkipService`).

@@ -183,7 +183,7 @@ OpenJarvis is built around five composable layers. Each has a clean interface an

    ---

-    CLI, Python SDK, and guides for [Morning Digest](user-guide/morning-digest.md), [Deep Research](user-guide/deep-research.md), [Code Assistant](user-guide/code-assistant.md), [Scheduled Monitor](user-guide/scheduled-monitor.md), [Simple Chat](user-guide/chat-simple.md), agents, memory, tools, and telemetry.
+    CLI, Python SDK, and guides for [Morning Digest](user-guide/morning-digest.md), [Deep Research](user-guide/deep-research.md), [Code Assistant](user-guide/code-assistant.md), [Scheduled Monitor](user-guide/scheduled-monitor.md), [Simple Chat](user-guide/chat-simple.md), [Evaluations](user-guide/evaluations.md), agents, memory, tools, and telemetry.

 -   **[Architecture](architecture/overview.md)**

@@ -215,7 +215,7 @@ OpenJarvis is built around five composable layers. Each has a clean interface an

 OpenJarvis is part of [Intelligence Per Watt](https://www.intelligence-per-watt.ai/), a research initiative studying the efficiency of on-device AI systems. Developed at [Hazy Research](https://hazyresearch.stanford.edu/) and the [Scaling Intelligence Lab](https://scalingintelligence.stanford.edu/) at [Stanford SAIL](https://ai.stanford.edu/).

-Read the [blog post](https://scalingintelligence.stanford.edu/blogs/openjarvis/) for the full research motivation, architecture details, and experimental results.
+Read the [blog post](https://openjarvis.stanford.edu/) for the full research motivation, architecture details, and experimental results.

 ## Citation

@@ -0,0 +1,12 @@
+// Public Supabase config for the savings leaderboard.
+//
+// This file is loaded *before* leaderboard.js and supplies the anon key it
+// reads from `window.OPENJARVIS_SUPABASE_ANON_KEY`. The key is injected at
+// docs-build time from the VITE_SUPABASE_ANON_KEY repo secret (see
+// .github/workflows/docs.yml). It is intentionally empty here so that local
+// `mkdocs build` and fork pull requests — which have no secret — render the
+// graceful "Leaderboard not configured yet" message instead of failing.
+//
+// The anon key is public by design: Supabase Row-Level Security protects the
+// data, so shipping it in the public docs bundle is expected.
+window.OPENJARVIS_SUPABASE_ANON_KEY = "";
@@ -1,9 +1,9 @@
 (function () {
  "use strict";

-  var SUPABASE_URL = "https://mtbtgpwzrbostweaanpr.supabase.co";
-  var SUPABASE_ANON_KEY =
-    "eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJzdXBhYmFzZSIsInJlZiI6Im10YnRncHd6cmJvc3R3ZWFhbnByIiwicm9sZSI6ImFub24iLCJpYXQiOjE3NzMxODk0OTQsImV4cCI6MjA4ODc2NTQ5NH0._xMlqCfljtXpwPj54H-ghxfLFO-jiq4W2WhpU8vVL1c";
+  var SUPABASE_URL =
+    window.OPENJARVIS_SUPABASE_URL || "https://mtbtgpwzrbostweaanpr.supabase.co";
+  var SUPABASE_ANON_KEY = window.OPENJARVIS_SUPABASE_ANON_KEY || "";

  var PAGE_SIZE = 50;
  var allRows = [];
@@ -55,5 +55,5 @@ See how the OpenJarvis community saves money, energy, and compute by running AI
 <div id="leaderboard-pagination" class="lb-pagination"></div>

 <p style="font-size:12px;opacity:0.6;margin-top:12px">
-*Dollar savings estimated vs. Claude Opus 4.6 API pricing ($5/1M input, $25/1M output tokens). Assumes local open-source models produce roughly the same number of tokens per request as cloud models.
+*Dollar savings estimated vs. Claude Fable 5 API pricing ($10/1M input, $50/1M output tokens). Assumes local open-source models produce roughly the same number of tokens per request as cloud models.
 </p>
@@ -19,6 +19,71 @@ Agents are the agentic logic layer of OpenJarvis. They determine how a query is

 ---

+## Persistent Persona: SOUL.md, MEMORY.md, USER.md
+
+Every agent's system prompt is assembled at conversation start by the `SystemPromptBuilder`, which injects up to three optional Markdown files -- the **persistent persona**. They are plain text you own and edit, loaded at the start of each conversation. There is no vector database or embedding cache behind them.
+
+| File | What it holds | Example line |
+|------|---------------|--------------|
+| `SOUL.md` | How the agent should behave -- tone, length, what to push back on | `Be concise. Challenge weak assumptions.` |
+| `MEMORY.md` | Facts about you, your projects, your preferences | `I deploy to Postgres, never MySQL.` |
+| `USER.md` | Who you are -- role, team, context | `Backend engineer at Acme, on the payments team.` |
+
+This persona is distinct from the retrieval [memory backend](memory.md): the persona is always-on Markdown context loaded into the prompt, while the memory backend is searchable long-term storage the agent queries on demand.
+
+### Where they live
+
+By default the files are read from the config directory:
+
+```
+~/.openjarvis/SOUL.md
+~/.openjarvis/MEMORY.md
+~/.openjarvis/USER.md
+```
+
+(The config directory honors `$OPENJARVIS_HOME` / `$XDG_DATA_HOME` when set.) The paths are configurable under `[memory_files]`:
+
+```toml
+[memory_files]
+soul_path    = "~/.openjarvis/SOUL.md"
+memory_path  = "~/.openjarvis/MEMORY.md"
+user_path    = "~/.openjarvis/USER.md"
+persona_name = ""    # optional named persona -- see below
+```
+
+### How they're loaded
+
+At the start of each conversation, `SystemPromptBuilder` reads each file as UTF-8 and adds its contents as a section of the system prompt, after the agent template and before the skill catalog:
+
+- **All three are optional.** A missing or empty file is skipped, so any subset works and an install with no persona files behaves exactly as before.
+- **Edits apply to the next conversation.** The files are read once when a conversation's prompt is built, so there is no restart or re-indexing -- edit or delete a line and it takes effect the next time you start a conversation.
+- **Each section is length-capped.** Files are truncated to a per-section character budget so a large `MEMORY.md` cannot crowd out the rest of the prompt.
+
+### Named personas
+
+A single install can answer as different personas without changing global config. A named persona lives in its own directory:
+
+```
+~/.openjarvis/personas/<name>/SOUL.md
+~/.openjarvis/personas/<name>/MEMORY.md
+~/.openjarvis/personas/<name>/USER.md
+```
+
+Select one per invocation, or opt out entirely:
+
+```bash
+jarvis ask --persona work  "summarize my open PRs"
+jarvis ask --persona none  "what is 2 + 2?"     # inject no persona
+```
+
+Set `persona_name` under `[memory_files]` to make a named persona the default. `persona_name = "none"` (equivalently `--persona none`) disables persona injection for that run.
+
+### Editing them
+
+`SOUL.md`, `MEMORY.md`, and `USER.md` are plain Markdown -- open them in any editor. `MEMORY.md` and `USER.md` can also be updated by the agent itself through the `memory_manage` and `user_profile_manage` tools when those are enabled, so the agent can record a new fact mid-conversation. These tools always target the default `MEMORY.md` and `USER.md` (under `~/.openjarvis/`), never a named persona's copies -- edit those by hand.
+
+---
+
 ## BaseAgent ABC

 All agents extend the abstract `BaseAgent` class.
@@ -66,6 +66,8 @@ jarvis ask "What is the capital of France?"
 | `--no-context`                | flag    | off        | Disable memory context injection                       |
 | `-a`, `--agent AGENT`         | string  | none       | Agent to use (`simple`, `orchestrator`)                |
 | `--tools TOOLS`               | string  | none       | Comma-separated tool names to enable                   |
+| `-i`, `--image PATH`          | path    | none       | Image file for a vision model (e.g. `gemma3:4b`); repeatable |
+| `-S`, `--screen`              | flag    | off        | Capture the current screen and send it to the vision model  |

 ### Direct Mode vs Agent Mode

@@ -105,6 +107,39 @@ jarvis ask --no-context "Tell me about Python"
 jarvis ask --max-tokens 2048 "Write a detailed essay about AI"
 ```

+### Vision Input
+
+Vision-capable models (such as `gemma3:4b`) can read images alongside your
+text prompt. Attach one or more image files with `-i`/`--image`, or capture
+the current screen with `-S`/`--screen`:
+
+```bash
+# Ask about a local image
+jarvis ask -i screenshot.png "What is shown in this image?"
+
+# Send multiple images (the flag is repeatable)
+jarvis ask -i chart-a.png -i chart-b.png "Compare these two charts"
+
+# Capture the current screen and ask about it
+jarvis ask --screen "Summarize what's on my screen"
+```
+
+Vision runs in **direct mode** only. If you also pass `--agent`, the image is
+ignored and a note is printed — re-run with `--agent ""` to force direct mode.
+
+The Ollama context window can be tuned for large images or long prompts with
+the `JARVIS_NUM_CTX` environment variable (default `16384`):
+
+```bash
+JARVIS_NUM_CTX=8192 jarvis ask --screen "What's on my screen?"
+```
+
+!!! note "Keep vision on-device"
+    Images are sensitive. OpenJarvis prints a privacy warning before sending
+    an image to a non-local engine, so a screenshot never leaves your machine
+    unnoticed. Use a local engine (e.g. `ollama` with `gemma3:4b`) to keep
+    vision fully local.
+
 ### JSON Output Format

 When using `--json` in **direct mode**, the output includes:
@@ -1,14 +1,14 @@
 # Evaluations

-The OpenJarvis evaluation framework (`openjarvis-evals`) measures model **correctness and accuracy** on academic datasets. It is a separate package from the main OpenJarvis library and is designed specifically for research workflows where you need reproducible, dataset-driven quality assessments.
+The OpenJarvis evaluation framework (`openjarvis.evals`) measures model **correctness and accuracy** on academic datasets. It ships inside the main `openjarvis` package (at `src/openjarvis/evals/`) and is designed specifically for research workflows where you need reproducible, dataset-driven quality assessments.

 !!! info "Evals vs. Benchmarks"
    OpenJarvis has two distinct measurement systems that complement each other:

-    | System | Package | Measures | Entry Point |
-    |--------|---------|----------|-------------|
-    | **Evaluations** | `openjarvis-evals` | Correctness on academic datasets (accuracy, pass rate) | `openjarvis-eval` |
-    | **Benchmarks** | `openjarvis` | Engine performance (latency, throughput) | `jarvis bench` |
+    | System | Module | Measures | Entry Point |
+    |--------|--------|----------|-------------|
+    | **Evaluations** | `openjarvis.evals` | Correctness on academic datasets (accuracy, pass rate) | `jarvis eval` |
+    | **Benchmarks** | `openjarvis.bench` | Engine performance (latency, throughput) | `jarvis bench` |

    Use evaluations to answer "does this model get the right answer?" and benchmarks to answer "how fast does this model respond?". See the [Benchmarks guide](benchmarks.md) for the performance measurement system.

@@ -18,22 +18,38 @@ The OpenJarvis evaluation framework (`openjarvis-evals`) measures model **correc

 ## Installation

-The evaluation framework is a standalone package in the `evals/` directory. Install it alongside OpenJarvis:
+The evaluation framework is part of the main `openjarvis` package — no separate install or extra is required. The standard dev setup is enough:

 ```bash
-uv sync --extra eval
+uv sync --extra dev
 ```

-This installs the `openjarvis-eval` CLI entry point and all required dependencies (`datasets`, `huggingface-hub`, `tqdm`, `rich`).
+The framework's core dependencies (`click`, `datasets`, `rich`) are base dependencies of `openjarvis`. Two optional extras enable experiment tracking integrations:
+
+```bash
+uv sync --extra dev --extra eval-wandb     # Weights & Biases run tracking
+uv sync --extra dev --extra eval-sheets    # Google Sheets results export
+```

 !!! note "Python version requirement"
-    Python 3.10 requires the `tomli` package for TOML config parsing. The `evals/pyproject.toml` includes this as a conditional dependency, so it is installed automatically.
+    Python 3.10 requires the `tomli` package for TOML config parsing. `openjarvis` declares it as a conditional dependency, so it is installed automatically.
+
+## Entry Points
+
+Two equivalent entry points expose the framework:
+
+| Command | Surface |
+|---------|---------|
+| `jarvis eval {list,run,compare,report}` | Canonical CLI. `run` covers the common options; `compare` and `report` post-process result files. |
+| `python -m openjarvis.evals {list,run,run-all,summarize,reparse-judge}` | Full research surface, including judge configuration, the agentic runner, and episode mode. |
+
+The `openjarvis-eval` console script is an alias for `python -m openjarvis.evals` — same commands, same options. This guide uses `jarvis eval` wherever its option set suffices and the module form for research-only options.

 ---

 ## Datasets

-The framework ships with **30+ datasets** covering academic reasoning, agentic tasks, retrieval, conversation quality, and practical use-case benchmarks. Datasets are grouped by category below.
+The framework ships with **40 registered benchmarks** covering academic reasoning, agentic tasks, coding, retrieval, conversation quality, and practical use-case benchmarks. Datasets are grouped by category below; `uv run python -m openjarvis.evals list` prints the authoritative registry.

 ### Use-Case Benchmarks

@@ -64,6 +80,7 @@ These benchmarks measure reasoning and knowledge on established academic dataset
 | **MATH-500** | `math500` | reasoning | Competition-level math problems |
 | **NaturalReasoning** | `natural-reasoning` | reasoning | Natural language reasoning |
 | **HLE** | `hle` | reasoning | Humanity's Last Exam hard challenges |
+| **LiveResearchBench** | `liveresearchbench` | reasoning | Recent research comprehension (Salesforce) |
 | **SimpleQA** | `simpleqa` | chat | Short-form factual question answering |
 | **IPW** | `ipw` | chat | Intelligence Per Watt mixed benchmark |

@@ -79,6 +96,11 @@ These benchmarks test multi-step agent capabilities including tool use, code gen
 | **TerminalBench** | `terminalbench` | agentic | Terminal-based task completion |
 | **TerminalBench Native** | `terminalbench-native` | agentic | TerminalBench with native Docker execution |
 | **TerminalBench V2.1** | `terminalbench-v2.1` | agentic | TB v2.1 Harbor-style Docker tasks |
+| **PinchBench** | `pinchbench` | agentic | Real-world agent tasks |
+| **TauBench** | `taubench` | agentic | Multi-turn customer service |
+| **DeepResearchBench** | `liveresearch` | agentic | Deep research report generation |
+| **DeepResearchBench (alias)** | `deepresearch` | agentic | Same benchmark as `liveresearch` |
+| **ToolCall-15** | `toolcall15` | agentic | Tool calling benchmark |
 | **LifelongAgent** | `lifelong-agent` | agentic | Sequential task learning across sessions |
 | **PaperArena** | `paperarena` | agentic | Scientific paper analysis |
 | **DeepPlanning** | `deepplanning` | agentic | Shopping constraint planning |
@@ -87,6 +109,14 @@ These benchmarks test multi-step agent capabilities including tool use, code gen
 | **WebChoreArena** | `webchorearena` | agentic | Web chore tasks |
 | **WorkArena** | `workarena` | agentic | WorkArena++ enterprise workflows |

+Both `liveresearch` and `deepresearch` are registered keys for the DeepResearchBench report-generation benchmark.
+
+### Coding Benchmarks
+
+| Dataset | Key | Category | Description |
+|---------|-----|----------|-------------|
+| **LiveCodeBench** | `livecodebench` | coding | Competitive programming |
+
 ### Retrieval Benchmarks

 | Dataset | Key | Category | Description |
@@ -123,7 +153,7 @@ The framework includes two pre-built configs for evaluating models on the five c
 ### Cloud models

 ```bash
-uv run python -m openjarvis.evals --config src/openjarvis/evals/configs/use_case_v2_cloud.toml
+uv run jarvis eval run --config src/openjarvis/evals/configs/use_case_v2_cloud.toml
 ```

 This config evaluates **6 cloud models** (Claude Opus 4.6, Claude Haiku 4.5, Gemini 3.1 Pro, Gemini 3.1 Flash Lite, GPT-5.4, GPT-5 Mini) against all 5 use-case benchmarks with 30 samples each, producing a 6x5 = 30-run matrix. Results are written to `results/use-cases-v2-cloud/`.
@@ -131,7 +161,7 @@ This config evaluates **6 cloud models** (Claude Opus 4.6, Claude Haiku 4.5, Gem
 ### Local models

 ```bash
-uv run python -m openjarvis.evals --config src/openjarvis/evals/configs/use_case_v2_local.toml
+uv run jarvis eval run --config src/openjarvis/evals/configs/use_case_v2_local.toml
 ```

 This config evaluates **5 local models** via Ollama (Qwen3.5 122B-A10B, GPT-OSS 120B, GLM4, Qwen3.5 35B-A3B, GLM-4.7-Flash) against the same 5 benchmarks, producing a 5x5 = 25-run matrix. Uses 2 workers (suitable for single-GPU setups). Results are written to `results/use-cases-v2-local/`.
@@ -143,15 +173,22 @@ This config evaluates **5 local models** via Ollama (Qwen3.5 122B-A10B, GPT-OSS

 ## Inference Backends

-Every evaluation run routes model calls through one of two backends:
+Every evaluation run routes model calls through one of four backends:

 | Backend | Key | Description |
 |---------|-----|-------------|
 | **jarvis-direct** | `jarvis-direct` | Engine-level inference via `SystemBuilder`. Works for local (Ollama, vLLM, llama.cpp) and cloud models. |
 | **jarvis-agent** | `jarvis-agent` | Agent-level inference with tool calling. Uses `JarvisSystem.ask()` with the specified agent and tools. |
+| **hermes** | `hermes` | Real Hermes Agent (Nous Research) via subprocess. Requires `--base-url` and `--api-key`. |
+| **openclaw** | `openclaw` | Real OpenClaw via Node subprocess. Requires `--base-url` and `--api-key`. |

 Use `jarvis-direct` for most evaluations. Use `jarvis-agent` when the benchmark requires tool use — for example, GAIA tasks that reference files that must be read with `file_read`, or arithmetic tasks that benefit from `calculator`.

+The `hermes` and `openclaw` backends shell out to external agent frameworks and need an OpenAI-compatible endpoint for their model calls: pass `--base-url`/`--api-key`, set the `JARVIS_BACKEND_BASE_URL`/`JARVIS_BACKEND_API_KEY` environment variables, or add a `[backend.external]` section to your config (see [Config Reference](#backendexternal)).
+
+!!! note "TerminalBench Native"
+    `jarvis eval run --backend` additionally accepts `terminalbench-native`, a Docker-based execution backend used by the TerminalBench Native benchmark.
+
 ---

 ## CLI Usage
@@ -159,73 +196,106 @@ Use `jarvis-direct` for most evaluations. Use `jarvis-agent` when the benchmark
 ### List available benchmarks and backends

 ```bash
-openjarvis-eval list
+uv run python -m openjarvis.evals list
 ```

-Output:
+Abridged output (40 benchmarks, 4 backends):

 ```
-Benchmarks:
-  supergpqa    [reasoning  ] SuperGPQA multiple-choice
-  gaia         [agentic    ] GAIA agentic benchmark
-  frames       [rag        ] FRAMES multi-hop RAG
-  wildchat     [chat       ] WildChat conversation quality
-
-Backends:
-  jarvis-direct    Engine-level inference (local or cloud)
-  jarvis-agent     Agent-level inference with tool calling
+                         Available Benchmarks
+┌──────────────────────┬───────────┬───────────────────────────────────┐
+│ Name                 │ Category  │ Description                       │
+├──────────────────────┼───────────┼───────────────────────────────────┤
+│ supergpqa            │ reasoning │ SuperGPQA multiple-choice         │
+│ gpqa                 │ reasoning │ GPQA graduate-level MCQ           │
+│ ...                  │ ...       │ ...                               │
+│ livecodebench        │ coding    │ LiveCodeBench competitive progr.  │
+│ toolcall15           │ agentic   │ ToolCall-15 tool calling benchmark│
+└──────────────────────┴───────────┴───────────────────────────────────┘
+                         Available Backends
+┌───────────────┬──────────────────────────────────────────────────┐
+│ jarvis-direct │ Engine-level inference (local or cloud)          │
+│ jarvis-agent  │ Agent-level inference with tool calling          │
+│ hermes        │ Real Hermes Agent (Nous Research) via subprocess │
+│ openclaw      │ Real OpenClaw via Node subprocess                │
+└───────────────┴──────────────────────────────────────────────────┘
 ```

+`jarvis eval list` prints a similar table but currently shows a curated subset of the registry; the module form above is the authoritative listing.
+
 ### Run a single benchmark

 ```bash
-# Evaluate qwen3:8b on SuperGPQA (engine-level, 10 samples default)
-openjarvis-eval run -b supergpqa -m qwen3:8b
+# Evaluate qwen3:8b on SuperGPQA (engine-level, 10 samples)
+uv run jarvis eval run -b supergpqa -m qwen3:8b -n 10

-# Evaluate GPT-4o on GAIA using the agent backend with tools
-openjarvis-eval run -b gaia -m gpt-4o --backend jarvis-agent \
+# Evaluate GPT-5 Mini on GAIA using the agent backend with tools
+uv run jarvis eval run -b gaia -m gpt-5-mini --backend jarvis-agent \
    --agent orchestrator --tools calculator,file_read -n 50

-# Run FRAMES with vLLM engine, write output to a file
-openjarvis-eval run -b frames -m llama3:70b -e vllm \
+# Run FRAMES with the vLLM engine, write output to a file
+uv run jarvis eval run -b frames -m llama3:70b -e vllm \
    -o results/frames_llama70b.jsonl

 # Run WildChat with a higher temperature for chat quality
-openjarvis-eval run -b wildchat -m qwen3:8b --temperature 0.7 -n 100
+uv run jarvis eval run -b wildchat -m qwen3:8b --temperature 0.7 -n 100
 ```

-#### Full option reference
+#### `jarvis eval run` option reference

 | Option | Short | Type | Default | Description |
 |--------|-------|------|---------|-------------|
 | `--config` | `-c` | path | — | TOML config file; when provided, `-b` and `-m` are not required |
-| `--benchmark` | `-b` | choice | required* | `supergpqa`, `gaia`, `frames`, or `wildchat` |
-| `--backend` | | choice | `jarvis-direct` | `jarvis-direct` or `jarvis-agent` |
-| `--model` | `-m` | str | required* | Model identifier (e.g., `qwen3:8b`, `gpt-4o`) |
-| `--engine` | `-e` | str | auto | Engine key (`ollama`, `vllm`, `cloud`, ...) |
-| `--agent` | | str | `orchestrator` | Agent name for `jarvis-agent` backend |
-| `--tools` | | str | `""` | Comma-separated tool names (e.g., `calculator,file_read`) |
+| `--benchmark` | `-b` | str | required* | Any registered benchmark key (see `... list`) |
+| `--model` | `-m` | str | required* | Model identifier (e.g., `qwen3:8b`, `gpt-5-mini`) |
 | `--max-samples` | `-n` | int | all | Limit the number of samples evaluated |
-| `--max-workers` | `-w` | int | `4` | Parallel evaluation workers |
-| `--judge-model` | | str | `gpt-4o` | LLM used for judge-based scoring |
-| `--output` | `-o` | path | auto-generated | Output JSONL file path |
+| `--backend` | | choice | `jarvis-direct` | `jarvis-direct`, `jarvis-agent`, `hermes`, `openclaw`, or `terminalbench-native` |
+| `--base-url` | | str | — | OpenAI-compatible endpoint URL (env: `JARVIS_BACKEND_BASE_URL`) |
+| `--api-key` | | str | — | API key for the endpoint (env: `JARVIS_BACKEND_API_KEY`) |
+| `--agent` | | str | — | Agent name for `jarvis-agent` backend (e.g., `orchestrator`) |
+| `--engine` | `-e` | str | auto | Engine key (`ollama`, `vllm`, `cloud`, ...) |
+| `--tools` | | str | `""` | Comma-separated tool names (e.g., `calculator,file_read`) |
+| `--telemetry/--no-telemetry` | | flag | off | Enable telemetry collection during eval |
+| `--gpu-metrics/--no-gpu-metrics` | | flag | off | Enable GPU metric polling |
 | `--seed` | | int | `42` | Random seed for dataset shuffling |
-| `--split` | | str | dataset default | Override the dataset split |
 | `--temperature` | | float | `0.0` | Generation temperature |
 | `--max-tokens` | | int | `2048` | Maximum output tokens |
+| `--model-filter` | | str | — | Filter models by name substring (multi-model configs) |
+| `--output` | `-o` | path | auto-generated | Output JSONL file path |
+| `--wandb-project` / `--wandb-entity` / `--wandb-tags` / `--wandb-group` | | str | `""` | Weights & Biases tracking (requires `eval-wandb` extra) |
+| `--sheets-id` / `--sheets-worksheet` / `--sheets-creds` | | str | `""` | Google Sheets export (requires `eval-sheets` extra) |
 | `--verbose` | `-v` | flag | off | Enable debug logging |

 *Required when `--config` is not provided.

+#### Research-only options (`python -m openjarvis.evals run`)
+
+The module CLI accepts everything above plus research-grade options that `jarvis eval run` does not expose:
+
+| Option | Short | Type | Default | Description |
+|--------|-------|------|---------|-------------|
+| `--max-workers` | `-w` | int | `4` | Parallel evaluation workers |
+| `--judge-model` | | str | `gpt-5-mini-2025-08-07` | LLM used for judge-based scoring (see `--help` for the current default) |
+| `--judge-engine` | | str | `cloud` | Engine key for the LLM judge; use `vllm` to judge locally |
+| `--split` | | str | dataset default | Override the dataset split |
+| `--compact` | | flag | off | Dense single-table output |
+| `--trace-detail` | | flag | off | Full per-step trace listing |
+| `--agentic` | | flag | off | Use `AgenticRunner` for multi-turn agent execution |
+| `--episode-mode` | | flag | off | Sequential episode processing with lifelong learning (required for `lifelong-agent` and similar benchmarks) |
+| `--concurrency` | | int | `1` | Parallel query execution (AgenticRunner only) |
+| `--query-timeout` | | float | — | Per-query wall-clock timeout in seconds (AgenticRunner only) |
+
+Note: the module CLI's `--backend` choice covers `jarvis-direct`, `jarvis-agent`, `hermes`, and `openclaw`; `terminalbench-native` as a backend is available via `jarvis eval run` and TOML configs.
+
 ### Run all benchmarks at once

-The `run-all` command evaluates a single model against all four benchmarks sequentially and writes results to an output directory:
+The `run-all` command (module CLI only) evaluates a single model against **every registered benchmark** sequentially and writes results to an output directory:

 ```bash
-openjarvis-eval run-all -m qwen3:8b
+uv run python -m openjarvis.evals run-all -m qwen3:8b

 # With options
-openjarvis-eval run-all -m gpt-4o -n 100 --output-dir results/gpt4o/
+uv run python -m openjarvis.evals run-all -m gpt-5-mini -n 100 --output-dir results/gpt5mini/
 ```

 Output files are written as `{output_dir}/{benchmark}_{model-slug}.jsonl`. The model slug replaces `/` and `:` with `-`, so `qwen3:8b` becomes `qwen3-8b`.
@@ -235,7 +305,7 @@ Output files are written as `{output_dir}/{benchmark}_{model-slug}.jsonl`. The m
 After a run, inspect a JSONL results file:

 ```bash
-openjarvis-eval summarize results/supergpqa_qwen3-8b.jsonl
+uv run python -m openjarvis.evals summarize results/supergpqa_qwen3-8b.jsonl
 ```

 Output:
@@ -251,6 +321,55 @@ Accuracy:  0.7222
 Errors:    2
 ```

+The module CLI also provides `reparse-judge`, which re-parses stored judge output in a results file and recovers records whose judge verdicts initially failed to parse — useful after improving the judge-output parser without re-running inference.
+
+### Compare and report
+
+`jarvis eval` adds two post-processing commands for result files:
+
+```bash
+# Side-by-side metric comparison across runs
+uv run jarvis eval compare results/supergpqa_qwen3-8b.jsonl results/supergpqa_gpt-5-mini.jsonl
+
+# Detailed report (accuracy, latency, cost, per-subject breakdown) for one run
+uv run jarvis eval report results/supergpqa_qwen3-8b.jsonl
+```
+
+---
+
+## Evaluating an Already-Running Endpoint
+
+If you already have an OpenAI-compatible server running — `jarvis serve`, vLLM, SGLang, llama.cpp's server, or a hosted endpoint — point an eval directly at it with `--base-url` and `--api-key`:
+
+```bash
+# A vLLM server is already serving Qwen/Qwen3-8B on a GPU node:
+#   vllm serve Qwen/Qwen3-8B --port 8000
+uv run jarvis eval run -b supergpqa -m Qwen/Qwen3-8B \
+    --base-url http://gpu-node:8000/v1 \
+    --api-key local-key \
+    -n 50
+```
+
+The `-m` value must match a model id the server reports at `GET /v1/models`. Both flags fall back to the `JARVIS_BACKEND_BASE_URL` and `JARVIS_BACKEND_API_KEY` environment variables, so CI jobs can set them once:
+
+```bash
+export JARVIS_BACKEND_BASE_URL=http://gpu-node:8000/v1
+export JARVIS_BACKEND_API_KEY=local-key
+uv run jarvis eval run -b gaia -m Qwen/Qwen3-8B --backend jarvis-agent -n 25
+```
+
+For the external `hermes` and `openclaw` backends these values are **required** (the foreign frameworks need an endpoint to send model calls to).
+
+!!! tip "Engine-level alternative for vLLM"
+    The vLLM engine also honors the `VLLM_HOST` environment variable (default `http://localhost:8000`):
+
+    ```bash
+    VLLM_HOST=http://gpu-node:8000 uv run python -m openjarvis.evals run \
+        -b supergpqa -m Qwen/Qwen3-8B -e vllm -n 50
+    ```
+
+    `VLLM_HOST` is process-global — if the candidate and the judge both use the `vllm` engine, they share the same endpoint. Prefer `--base-url` when you need them separate.
+
 ---

 ## TOML Config System
@@ -260,7 +379,7 @@ For research workflows that compare multiple models across multiple benchmarks,
 ### Running from a config

 ```bash
-openjarvis-eval run --config src/openjarvis/evals/configs/full-suite.toml
+uv run jarvis eval run --config src/openjarvis/evals/configs/full-suite.toml
 ```

 When `--config` is provided, the `-b`/`--benchmark` and `-m`/`--model` options are not required. All settings come from the config file. The CLI expands the matrix, prints a progress table, and writes results to the configured `output_dir`.
@@ -269,7 +388,7 @@ When `--config` is provided, the `-b`/`--benchmark` and `-m`/`--model` options a

 A config file has six sections: `[meta]`, `[defaults]`, `[judge]`, `[run]`, `[[models]]`, and `[[benchmarks]]`. Only `[[models]]` and `[[benchmarks]]` are required — all other sections are optional and fall back to built-in defaults.

-```toml title="evals/configs/full-suite.toml"
+```toml title="src/openjarvis/evals/configs/full-suite.toml"
 # Suite-level metadata (optional)
 [meta]
 name = "full-suite-v1"
@@ -353,7 +472,7 @@ For example, `temperature` is resolved as: use `[defaults].temperature` (0.0), t

 A config requires only one `[[models]]` and one `[[benchmarks]]` entry:

-```toml title="evals/configs/minimal.toml"
+```toml title="src/openjarvis/evals/configs/minimal.toml"
 [[models]]
 name = "qwen3:8b"

@@ -365,7 +484,7 @@ This runs SuperGPQA against qwen3:8b with all default settings. Use this as a st

 ### Single-run config with full options

-```toml title="evals/configs/single-run.toml"
+```toml title="src/openjarvis/evals/configs/single-run.toml"
 [meta]
 name = "single-run-example"
 description = "Evaluate SuperGPQA with a single model and full configuration"
@@ -425,7 +544,8 @@ Configuration for the LLM used as a judge in GAIA, FRAMES, and WildChat scoring.

 | Field | Type | Default | Description |
 |-------|------|---------|-------------|
-| `model` | str | `"gpt-4o"` | Judge model identifier |
+| `model` | str | `"gpt-5-mini-2025-08-07"` | Judge model identifier |
+| `engine` | str | `None` | Engine key for the judge (e.g., `"vllm"` to judge locally; defaults to cloud) |
 | `provider` | str | `None` | Provider override (e.g., `"openai"`) |
 | `temperature` | float | `0.0` | Judge sampling temperature |
 | `max_tokens` | int | `1024` | Maximum judge output tokens |
@@ -444,6 +564,20 @@ Execution settings that apply to the entire suite.
 | `seed` | int | `42` | Random seed for dataset shuffling |
 | `telemetry` | bool | `false` | Enable GPU telemetry capture (energy, power, utilization, throughput) |
 | `gpu_metrics` | bool | `false` | Enable GPU metric polling via `pynvml` (requires `pynvml` or `nvidia-ml-py`) |
+| `warmup_samples` | int | `0` | Untimed warmup samples before measurement |
+| `energy_vendor` | str | `""` | GPU energy vendor override |
+| `max_turns` | int | `None` | Maximum agent turns per query |
+| `wandb_project` / `wandb_entity` / `wandb_tags` / `wandb_group` | str | `""` | Weights & Biases tracking |
+| `sheets_spreadsheet_id` / `sheets_worksheet` / `sheets_credentials_path` | str | `""` / `"Results"` / `""` | Google Sheets export |
+
+### `[backend.external]`
+
+Endpoint settings for the `hermes` and `openclaw` backends. Environment variables override TOML values.
+
+| Field | Type | Default | Description |
+|-------|------|---------|-------------|
+| `base_url` | str | `None` | OpenAI-compatible endpoint URL (env: `JARVIS_BACKEND_BASE_URL`) |
+| `api_key` | str | `None` | API key for the endpoint (env: `JARVIS_BACKEND_API_KEY`) |

 ### `[[models]]`

@@ -451,7 +585,7 @@ One block per model. The `name` field is required.

 | Field | Type | Default | Description |
 |-------|------|---------|-------------|
-| `name` | str | required | Model identifier (e.g., `"qwen3:8b"`, `"gpt-4o"`) |
+| `name` | str | required | Model identifier (e.g., `"qwen3:8b"`, `"gpt-5-mini"`) |
 | `engine` | str | `None` | Engine key to use (`"ollama"`, `"vllm"`, `"cloud"`, ...) |
 | `provider` | str | `None` | Provider override for cloud models (e.g., `"openai"`) |
 | `temperature` | float | `None` | Override `[defaults].temperature` for this model |
@@ -468,10 +602,12 @@ One block per benchmark. The `name` field is required.

 | Field | Type | Default | Description |
 |-------|------|---------|-------------|
-| `name` | str | required | Benchmark key: `supergpqa`, `gaia`, `frames`, or `wildchat` |
-| `backend` | str | `"jarvis-direct"` | Inference backend: `jarvis-direct` or `jarvis-agent` |
+| `name` | str | required | Any registered benchmark key (see `uv run python -m openjarvis.evals list`) |
+| `backend` | str | `"jarvis-direct"` | `jarvis-direct`, `jarvis-agent`, `hermes`, `openclaw`, or `terminalbench-native` |
 | `max_samples` | int | `None` | Limit number of samples; `None` evaluates the full dataset |
 | `split` | str | `None` | Override the default dataset split |
+| `subset` | str | `None` | Dataset subset/variant (benchmark-specific) |
+| `record_ids` | list[str] | `None` | Evaluate only these record ids |
 | `agent` | str | `None` | Agent name for `jarvis-agent` backend (e.g., `"orchestrator"`) |
 | `tools` | list[str] | `[]` | Tool names for `jarvis-agent` backend |
 | `judge_model` | str | `None` | Override `[judge].model` for this benchmark only |
@@ -647,7 +783,7 @@ The `EvalRunner` processes samples concurrently using a `ThreadPoolExecutor`. Re

 ```bash
 # Use more workers for faster evaluation (if the engine supports concurrent requests)
-openjarvis-eval run -b supergpqa -m qwen3:8b -w 8 -n 500
+uv run python -m openjarvis.evals run -b supergpqa -m qwen3:8b -w 8 -n 500
 ```

 !!! warning "Worker count and engine load"
@@ -392,7 +392,7 @@ enforce_tool_confirmation = true
 | `secret_scanner` | `bool` | `true` | Run `SecretScanner` on all text |
 | `pii_scanner` | `bool` | `true` | Run `PIIScanner` on all text |
 | `audit_log_path` | `str` | `~/.openjarvis/audit.db` | Path to the SQLite audit log |
-| `enforce_tool_confirmation` | `bool` | `true` | Require explicit confirmation before tool execution |
+| `enforce_tool_confirmation` | `bool` | `true` | Accepted by the loader but **not currently enforced**. See [System Access](system-access.md#confirmation-behaviour) for when prompts actually happen |

 !!! tip "Start with warn, tighten later"
    `mode = "warn"` is a good starting point. It lets you observe what patterns are being triggered without disrupting normal usage. Switch to `"redact"` once you are satisfied that the scanner isn't producing too many false positives for your workload.
@@ -0,0 +1,198 @@
+# System Access
+
+How to give an agent access to the machine it runs on, and where the real
+limits are.
+
+!!! warning
+    `shell_exec` runs arbitrary commands as your user. There is no command
+    allowlist, no denylist, and no sandbox unless you turn one on. An agent
+    holding this tool can do anything you can do from a terminal.
+
+---
+
+## Start here: you probably have no tools enabled
+
+If the agent tells you it can't run commands or read files, that's usually not
+a permissions problem. It means no tools were enabled in the first place.
+
+Tools come from `tools.enabled`, falling back to `agent.tools`. Both default to
+empty, and an empty value builds the agent with **zero tools**. Nothing is
+enabled by default.
+
+First check whether you have a config file at all:
+
+```bash
+cat ~/.openjarvis/config.toml
+```
+
+If it isn't there, that's your answer. Create it:
+
+```toml
+[engine]
+default = "ollama"
+
+[intelligence]
+default_model = "qwen3.5:9b"
+
+[agent]
+default_agent = "orchestrator"
+
+[tools]
+enabled = ["shell_exec", "file_read", "file_write", "think"]
+```
+
+There's a fuller version at
+`configs/openjarvis/examples/full-system-access.toml`.
+
+Then confirm the list actually resolved:
+
+```bash
+python -c "from openjarvis.core.config import load_config; print(load_config().tools.enabled)"
+```
+
+---
+
+## What the tools reach
+
+| Tool | Scope |
+|------|-------|
+| `shell_exec` | Any command, as your user. 30s default timeout, 300s max, output capped at 100 KB per stream. |
+| `file_read` | Any readable path. 1 MB cap. |
+| `file_write` | Any writable path. 10 MB cap, can create parent directories. |
+| `apply_patch` | Applies unified diffs to any path. |
+| `code_interpreter` | Python in a subprocess, behind a coarse pattern blocklist. |
+
+`file_read` and `file_write` take an `allowed_dirs` argument that limits them to
+a set of directories, but no config key populates it. When it's empty every path
+is allowed. If you want a filesystem jail today, use the container sandbox
+instead of relying on these tools to enforce one.
+
+### Sensitive filenames
+
+`file_read` and `file_write` refuse names matching a short glob list: `.env`,
+`*.pem`, `id_rsa`, `credentials.*` and a dozen or so others. It matches on the
+filename only, not the path or the contents, and only those two tools consult
+it. `shell_exec`, `apply_patch` and `code_interpreter` skip it entirely, so
+`cat ~/.ssh/id_rsa` through `shell_exec` works fine. Treat it as protection
+against fat fingers, not as a security boundary.
+
+---
+
+## Confirmation behaviour
+
+`shell_exec`, `git_commit` and `agent_kill` are marked `requires_confirmation`.
+What that translates to depends entirely on how you launched the agent:
+
+| Entry point | Behaviour |
+|-------------|-----------|
+| `jarvis chat` | Prompts before each call. |
+| `jarvis ask` | Auto-approves. |
+| `jarvis agent ask` | Auto-approves. Pass `--no-yes` if you want prompts. |
+| HTTP server, desktop app | Auto-approves. Tools you added to an agent's toolkit count as pre-approved. |
+| Embedded via `SystemBuilder` | No callback is wired, so these tools fail closed. |
+
+That last row catches people out. If `shell_exec` returns "requires
+confirmation but no confirmation callback is available", you're constructing the
+agent yourself and need to pass a `confirm_callback`.
+
+!!! note "`enforce_tool_confirmation` doesn't do anything"
+    The config loader accepts `security.enforce_tool_confirmation`, but nothing
+    on the tool execution path reads it. Setting it won't change confirmation
+    behaviour anywhere. Use the table above instead.
+
+---
+
+## macOS: Full Disk Access
+
+On macOS the operating system is the real boundary, not the config. Shell
+access and ordinary file access start working as soon as you enable the tools.
+TCC-protected data does not: Messages, Mail, Photos, Safari history, Contacts
+and Calendar all stay locked, and no config key will change that.
+
+Grant Full Disk Access to whichever process hosts the backend. Child processes
+inherit it:
+
+| How you run OpenJarvis | Grant access to |
+|------------------------|-----------------|
+| CLI (`jarvis ask`, `jarvis chat`) | Your terminal (Terminal, iTerm, Warp) |
+| Desktop app | `OpenJarvis.app`, which spawns `jarvis serve` beneath it |
+| launchd (`deploy/launchd/com.openjarvis.plist`) | The `jarvis` binary, as its own entry |
+
+System Settings, then Privacy & Security, then Full Disk Access, then **+**.
+
+A launchd daemon gets its own TCC context, so granting access to Terminal does
+nothing for it. Add `/usr/local/bin/jarvis` separately.
+
+To check whether the grant took:
+
+```bash
+head -c 16 ~/Library/Messages/chat.db >/dev/null 2>&1 \
+  && echo "granted" || echo "denied"
+```
+
+Restart the host process after you change the setting.
+
+### Driving Mac apps
+
+AppleScript works through `shell_exec`:
+
+```
+osascript -e 'tell application "Music" to play'
+```
+
+macOS asks for Automation permission once per target app, the first time you
+touch it.
+
+---
+
+## What you can't do
+
+There's no computer use. OpenJarvis can't see your screen, move the pointer or
+send keystrokes. No tool for it is registered and no input automation library
+appears anywhere in the codebase, so granting Accessibility or Screen Recording
+buys you nothing on its own.
+
+The `click` and `type` actions you'll find are Playwright, scoped to a browser
+page rather than the desktop.
+
+Some of this is reachable through `shell_exec` if you bring the tooling
+yourself. `screencapture` will take screenshots once you've granted Screen
+Recording, and something like `cliclick` will move the pointer. That gets you
+scripted actions. It doesn't get you an agent that looks at the screen and
+works out where to click.
+
+---
+
+## Narrowing access
+
+Access widens and narrows through `tools.enabled`. Drop entries to take
+capabilities away. That list is the whole grant.
+
+Two stronger isolation options exist. Both are off by default:
+
+```toml
+[sandbox]
+enabled = true          # run tools inside a container
+runtime = "docker"
+
+[security.capabilities]
+enabled = true          # RBAC over declared tool capabilities
+policy_path = "~/.openjarvis/policy.yaml"
+```
+
+!!! note "Capabilities are open by default even once enabled"
+    `CapabilityPolicy` is built with `default_deny=False` and no config key
+    exposes that flag, so an agent with no explicit policy entry gets every
+    capability. Write entries for every agent you mean to restrict.
+
+For anything untrusted, reach for `docker_shell_exec` and
+`code_interpreter_docker` rather than the host-side versions.
+
+---
+
+## See also
+
+- [Security](security.md) for scanners, the audit log and guardrails
+- [Tools](tools.md) for the full registry
+- [Code Assistant](code-assistant.md) for a narrower shell-enabled setup
+- [External MCP Servers](mcp-external-servers.md) for capabilities OpenJarvis doesn't ship
@@ -11,7 +11,7 @@
        "@base-ui/react": "^1.3.0",
        "@fontsource-variable/geist": "^5.2.8",
        "@tailwindcss/vite": "^4.2.1",
-        "@tauri-apps/api": "^2",
+        "@tauri-apps/api": "^2.11.1",
        "@tauri-apps/plugin-autostart": "^2",
        "@tauri-apps/plugin-dialog": "^2.7.0",
        "@tauri-apps/plugin-global-shortcut": "^2",
@@ -42,7 +42,7 @@
        "zustand": "^5.0.11"
      },
      "devDependencies": {
-        "@tauri-apps/cli": "^2",
+        "@tauri-apps/cli": "^2.11.4",
        "@types/react": "^19.0.0",
        "@types/react-dom": "^19.0.0",
        "@vitejs/plugin-react": "^4.3.4",
@@ -3720,9 +3720,9 @@
      }
    },
    "node_modules/@tauri-apps/api": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/api/-/api-2.10.1.tgz",
-      "integrity": "sha512-hKL/jWf293UDSUN09rR69hrToyIXBb8CjGaWC7gfinvnQrBVvnLr08FeFi38gxtugAVyVcTa5/FD/Xnkb1siBw==",
+      "version": "2.11.1",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/api/-/api-2.11.1.tgz",
+      "integrity": "sha512-M2FPuYND2m+wh5hfW9ZpSdxMPdEJovPBWwoHJmwUpysTYNHaOkVFN419m/K0LIgjb/7KU2vBgsUepJWugQCvAA==",
      "license": "Apache-2.0 OR MIT",
      "funding": {
        "type": "opencollective",
@@ -3730,9 +3730,9 @@
      }
    },
    "node_modules/@tauri-apps/cli": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli/-/cli-2.10.1.tgz",
-      "integrity": "sha512-jQNGF/5quwORdZSSLtTluyKQ+o6SMa/AUICfhf4egCGFdMHqWssApVgYSbg+jmrZoc8e1DscNvjTnXtlHLS11g==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli/-/cli-2.11.4.tgz",
+      "integrity": "sha512-R8xGtMpwyetawSqm9kYOuMmEqkhUbvcUy8n0aNXIxollKBLESUu5f4Fx+64hgASYm1H+jSWq6jCW6zqTnH6hqQ==",
      "dev": true,
      "license": "Apache-2.0 OR MIT",
      "bin": {
@@ -3746,23 +3746,23 @@
        "url": "https://opencollective.com/tauri"
      },
      "optionalDependencies": {
-        "@tauri-apps/cli-darwin-arm64": "2.10.1",
-        "@tauri-apps/cli-darwin-x64": "2.10.1",
-        "@tauri-apps/cli-linux-arm-gnueabihf": "2.10.1",
-        "@tauri-apps/cli-linux-arm64-gnu": "2.10.1",
-        "@tauri-apps/cli-linux-arm64-musl": "2.10.1",
-        "@tauri-apps/cli-linux-riscv64-gnu": "2.10.1",
-        "@tauri-apps/cli-linux-x64-gnu": "2.10.1",
-        "@tauri-apps/cli-linux-x64-musl": "2.10.1",
-        "@tauri-apps/cli-win32-arm64-msvc": "2.10.1",
-        "@tauri-apps/cli-win32-ia32-msvc": "2.10.1",
-        "@tauri-apps/cli-win32-x64-msvc": "2.10.1"
+        "@tauri-apps/cli-darwin-arm64": "2.11.4",
+        "@tauri-apps/cli-darwin-x64": "2.11.4",
+        "@tauri-apps/cli-linux-arm-gnueabihf": "2.11.4",
+        "@tauri-apps/cli-linux-arm64-gnu": "2.11.4",
+        "@tauri-apps/cli-linux-arm64-musl": "2.11.4",
+        "@tauri-apps/cli-linux-riscv64-gnu": "2.11.4",
+        "@tauri-apps/cli-linux-x64-gnu": "2.11.4",
+        "@tauri-apps/cli-linux-x64-musl": "2.11.4",
+        "@tauri-apps/cli-win32-arm64-msvc": "2.11.4",
+        "@tauri-apps/cli-win32-ia32-msvc": "2.11.4",
+        "@tauri-apps/cli-win32-x64-msvc": "2.11.4"
      }
    },
    "node_modules/@tauri-apps/cli-darwin-arm64": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-darwin-arm64/-/cli-darwin-arm64-2.10.1.tgz",
-      "integrity": "sha512-Z2OjCXiZ+fbYZy7PmP3WRnOpM9+Fy+oonKDEmUE6MwN4IGaYqgceTjwHucc/kEEYZos5GICve35f7ZiizgqEnQ==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-darwin-arm64/-/cli-darwin-arm64-2.11.4.tgz",
+      "integrity": "sha512-1ryOF3ZhpZ/nemHV5zVwBQBz9jDGKmKPvWPADOhc83ig0P4bMc2iER4NbC6r9sjeIZ6RVQ4g3RZIYvezhcl4TQ==",
      "cpu": [
        "arm64"
      ],
@@ -3777,9 +3777,9 @@
      }
    },
    "node_modules/@tauri-apps/cli-darwin-x64": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-darwin-x64/-/cli-darwin-x64-2.10.1.tgz",
-      "integrity": "sha512-V/irQVvjPMGOTQqNj55PnQPVuH4VJP8vZCN7ajnj+ZS8Kom1tEM2hR3qbbIRoS3dBKs5mbG8yg1WC+97dq17Pw==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-darwin-x64/-/cli-darwin-x64-2.11.4.tgz",
+      "integrity": "sha512-uFsGQAAfuyz1k/yGLmkWfkBlgKAqZfxqlHmLWx81QU27RJWfmbNHCIq8T8w1e+VClleIuZUjpHWfoE4E3DLo3A==",
      "cpu": [
        "x64"
      ],
@@ -3794,9 +3794,9 @@
      }
    },
    "node_modules/@tauri-apps/cli-linux-arm-gnueabihf": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-arm-gnueabihf/-/cli-linux-arm-gnueabihf-2.10.1.tgz",
-      "integrity": "sha512-Hyzwsb4VnCWKGfTw+wSt15Z2pLw2f0JdFBfq2vHBOBhvg7oi6uhKiF87hmbXOBXUZaGkyRDkCHsdzJcIfoJC2w==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-arm-gnueabihf/-/cli-linux-arm-gnueabihf-2.11.4.tgz",
+      "integrity": "sha512-IaHZn5CdBL21oUmjiVOS1ctw6Ip1O0pjp70FwOWmYz1myWe0SY96ZIj2FYf7pT0m8bI2h/hrs5ZbEXXh44/MkQ==",
      "cpu": [
        "arm"
      ],
@@ -3811,13 +3811,16 @@
      }
    },
    "node_modules/@tauri-apps/cli-linux-arm64-gnu": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-arm64-gnu/-/cli-linux-arm64-gnu-2.10.1.tgz",
-      "integrity": "sha512-OyOYs2t5GkBIvyWjA1+h4CZxTcdz1OZPCWAPz5DYEfB0cnWHERTnQ/SLayQzncrT0kwRoSfSz9KxenkyJoTelA==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-arm64-gnu/-/cli-linux-arm64-gnu-2.11.4.tgz",
+      "integrity": "sha512-N41/ukTRVe6XSuUTESuFdGeOW2i7k62tK+6gHK5Kd5/q5RPvvi19GaWAVPPb9u95HSGmTChSolBfzynUsssFaA==",
      "cpu": [
        "arm64"
      ],
      "dev": true,
+      "libc": [
+        "glibc"
+      ],
      "license": "Apache-2.0 OR MIT",
      "optional": true,
      "os": [
@@ -3828,13 +3831,16 @@
      }
    },
    "node_modules/@tauri-apps/cli-linux-arm64-musl": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-arm64-musl/-/cli-linux-arm64-musl-2.10.1.tgz",
-      "integrity": "sha512-MIj78PDDGjkg3NqGptDOGgfXks7SYJwhiMh8SBoZS+vfdz7yP5jN18bNaLnDhsVIPARcAhE1TlsZe/8Yxo2zqg==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-arm64-musl/-/cli-linux-arm64-musl-2.11.4.tgz",
+      "integrity": "sha512-v277UnT/fB64xAfSroL5N3Km3tLmvATWqJJw/wRI+g6o+HkeD0slyE7gOhNs1MbjE41R7bQOTxMVoL3aomUJmw==",
      "cpu": [
        "arm64"
      ],
      "dev": true,
+      "libc": [
+        "musl"
+      ],
      "license": "Apache-2.0 OR MIT",
      "optional": true,
      "os": [
@@ -3845,13 +3851,16 @@
      }
    },
    "node_modules/@tauri-apps/cli-linux-riscv64-gnu": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-riscv64-gnu/-/cli-linux-riscv64-gnu-2.10.1.tgz",
-      "integrity": "sha512-X0lvOVUg8PCVaoEtEAnpxmnkwlE1gcMDTqfhbefICKDnOTJ5Est3qL0SrWxizDackIOKBcvtpejrSiVpuJI1kw==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-riscv64-gnu/-/cli-linux-riscv64-gnu-2.11.4.tgz",
+      "integrity": "sha512-qqgNkQ2u1yZHxjhxsZaxUtRDW8dIqIYm33rx/mzwQv0SfY9x1B+iraj8vWeFiXjjSVVhEMepXSOts1TqPzvXNQ==",
      "cpu": [
        "riscv64"
      ],
      "dev": true,
+      "libc": [
+        "glibc"
+      ],
      "license": "Apache-2.0 OR MIT",
      "optional": true,
      "os": [
@@ -3862,13 +3871,16 @@
      }
    },
    "node_modules/@tauri-apps/cli-linux-x64-gnu": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-x64-gnu/-/cli-linux-x64-gnu-2.10.1.tgz",
-      "integrity": "sha512-2/12bEzsJS9fAKybxgicCDFxYD1WEI9kO+tlDwX5znWG2GwMBaiWcmhGlZ8fi+DMe9CXlcVarMTYc0L3REIRxw==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-x64-gnu/-/cli-linux-x64-gnu-2.11.4.tgz",
+      "integrity": "sha512-2VRNWl84FOH0m2giiDkO2h0QXlcMJeX+zJDpI5kDIQAx6s+geF3v48F4DXfJez4GS/FdoDGnPnw1C2iYGbQ7bQ==",
      "cpu": [
        "x64"
      ],
      "dev": true,
+      "libc": [
+        "glibc"
+      ],
      "license": "Apache-2.0 OR MIT",
      "optional": true,
      "os": [
@@ -3879,13 +3891,16 @@
      }
    },
    "node_modules/@tauri-apps/cli-linux-x64-musl": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-x64-musl/-/cli-linux-x64-musl-2.10.1.tgz",
-      "integrity": "sha512-Y8J0ZzswPz50UcGOFuXGEMrxbjwKSPgXftx5qnkuMs2rmwQB5ssvLb6tn54wDSYxe7S6vlLob9vt0VKuNOaCIQ==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-linux-x64-musl/-/cli-linux-x64-musl-2.11.4.tgz",
+      "integrity": "sha512-o9GyhYor/nc7xarmwDE3ka2szuW3uuZzXjHWh64Q8YX5AtSgxdQkFWzrY4O8KiGtVNvFBI14H3Q49Qj5TOIP/A==",
      "cpu": [
        "x64"
      ],
      "dev": true,
+      "libc": [
+        "musl"
+      ],
      "license": "Apache-2.0 OR MIT",
      "optional": true,
      "os": [
@@ -3896,9 +3911,9 @@
      }
    },
    "node_modules/@tauri-apps/cli-win32-arm64-msvc": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-win32-arm64-msvc/-/cli-win32-arm64-msvc-2.10.1.tgz",
-      "integrity": "sha512-iSt5B86jHYAPJa/IlYw++SXtFPGnWtFJriHn7X0NFBVunF6zu9+/zOn8OgqIWSl8RgzhLGXQEEtGBdR4wzpVgg==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-win32-arm64-msvc/-/cli-win32-arm64-msvc-2.11.4.tgz",
+      "integrity": "sha512-ld5Ehb598m0VkYyylRPNeCFsBe/km0jxis6KgMpl3IGY6I/i1RwQXO05I1AsXUXO2WC6AvB/Lw4qTf/asiuEiQ==",
      "cpu": [
        "arm64"
      ],
@@ -3913,9 +3928,9 @@
      }
    },
    "node_modules/@tauri-apps/cli-win32-ia32-msvc": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-win32-ia32-msvc/-/cli-win32-ia32-msvc-2.10.1.tgz",
-      "integrity": "sha512-gXyxgEzsFegmnWywYU5pEBURkcFN/Oo45EAwvZrHMh+zUSEAvO5E8TXsgPADYm31d1u7OQU3O3HsYfVBf2moHw==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-win32-ia32-msvc/-/cli-win32-ia32-msvc-2.11.4.tgz",
+      "integrity": "sha512-12Hxi0XX/H5VFxO/bGgHkFWhml9VMgEOu9CidjeCeTNQ1l6fpUlbiGgSP7CLI3PFtW9/FfbeHieZ+kyWK5H7CA==",
      "cpu": [
        "ia32"
      ],
@@ -3930,9 +3945,9 @@
      }
    },
    "node_modules/@tauri-apps/cli-win32-x64-msvc": {
-      "version": "2.10.1",
-      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-win32-x64-msvc/-/cli-win32-x64-msvc-2.10.1.tgz",
-      "integrity": "sha512-6Cn7YpPFwzChy0ERz6djKEmUehWrYlM+xTaNzGPgZocw3BD7OfwfWHKVWxXzdjEW2KfKkHddfdxK1XXTYqBRLg==",
+      "version": "2.11.4",
+      "resolved": "https://registry.npmjs.org/@tauri-apps/cli-win32-x64-msvc/-/cli-win32-x64-msvc-2.11.4.tgz",
+      "integrity": "sha512-+vDiqBIU5dMISg/wNvX3sF+ZHfgJGJ5T0AcO+EHNXV9GGAG+P5fzodlDXD3QdKCRgZxMoCm5PPvj3BqLNjBthw==",
      "cpu": [
        "x64"
      ],
@@ -18,7 +18,7 @@
    "@base-ui/react": "^1.3.0",
    "@fontsource-variable/geist": "^5.2.8",
    "@tailwindcss/vite": "^4.2.1",
-    "@tauri-apps/api": "^2",
+    "@tauri-apps/api": "^2.11.1",
    "@tauri-apps/plugin-autostart": "^2",
    "@tauri-apps/plugin-dialog": "^2.7.0",
    "@tauri-apps/plugin-global-shortcut": "^2",
@@ -49,7 +49,7 @@
    "zustand": "^5.0.11"
  },
  "devDependencies": {
-    "@tauri-apps/cli": "^2",
+    "@tauri-apps/cli": "^2.11.4",
    "@types/react": "^19.0.0",
    "@types/react-dom": "^19.0.0",
    "@vitejs/plugin-react": "^4.3.4",
@@ -24,9 +24,22 @@ serde_json = "1"
 reqwest = { version = "0.12", features = ["json", "multipart"] }
 tokio = { version = "1", features = ["full"] }

+# Cloud API keys are stored in the OS credential store via `keyring`. keyring v3
+# enables NO backend by default — without an explicit per-platform feature it
+# silently falls back to a non-persistent in-memory mock, so keys would not
+# survive an app restart. Each desktop target opts into its native store.
 [target.'cfg(target_os = "macos")'.dependencies]
 objc = "0.2"
 dispatch = "0.2"
+keyring = { version = "3", features = ["apple-native"] }
+
+[target.'cfg(target_os = "windows")'.dependencies]
+keyring = { version = "3", features = ["windows-native"] }
+
+[target.'cfg(target_os = "linux")'.dependencies]
+# Blocking Secret Service backend (no internal async runtime, so it is safe to
+# call from the tokio-driven Tauri commands). Needs libdbus-1-dev at build time.
+keyring = { version = "3", features = ["sync-secret-service", "crypto-rust"] }

 [features]
 default = ["custom-protocol"]
@@ -8,8 +8,10 @@ use tokio::sync::Mutex;

 const OLLAMA_PORT: u16 = 11434;
 const JARVIS_PORT: u16 = 8000;
+const DESKTOP_UV_SYNC_COMMAND: &str =
+    "uv sync --extra desktop --extra inference-cloud --extra inference-google --group desktop-native";

-/// Small, fast model pulled at startup so the app opens quickly.
+/// Small, fast model used when startup needs a default Ollama tag.
 const STARTUP_MODEL: &str = "qwen3.5:4b";

 /// Tiny fallback model if even the startup model can't be pulled.
@@ -104,7 +106,7 @@ fn default_local_model(ram_gb: f64) -> &'static str {
 struct BootPlan {
    /// Whether to start and wait for the bundled Ollama.
    launch_ollama: bool,
-    /// The single Ollama model to pull (None for custom endpoints).
+    /// The preferred Ollama model (None for custom endpoints).
    model_to_pull: Option<String>,
    /// Optional `(engine_key, bare_host)` override for a custom endpoint,
    /// e.g. `("lmstudio", "http://localhost:1234")`. Written into
@@ -608,6 +610,69 @@ async fn wait_for_jarvis_health(
 }

 async fn ollama_has_model(model: &str) -> bool {
+    let models = ollama_model_names().await;
+    matching_installed_model(&models, model).is_some()
+}
+
+fn parse_ollama_model_names(body: &serde_json::Value) -> Vec<String> {
+    body.get("models")
+        .and_then(|m| m.as_array())
+        .map(|models| {
+            models
+                .iter()
+                .filter_map(|m| {
+                    m.get("name")
+                        .or_else(|| m.get("model"))
+                        .and_then(|n| n.as_str())
+                })
+                .filter(|name| !name.trim().is_empty())
+                .map(|name| name.to_string())
+                .collect()
+        })
+        .unwrap_or_default()
+}
+
+fn model_names_match(installed: &str, requested: &str) -> bool {
+    installed == requested
+        || installed.strip_suffix(":latest") == Some(requested)
+        || requested.strip_suffix(":latest") == Some(installed)
+}
+
+fn matching_installed_model(models: &[String], requested: &str) -> Option<String> {
+    models
+        .iter()
+        .find(|model| model_names_match(model, requested))
+        .cloned()
+}
+
+fn model_name_looks_embedding_only(model: &str) -> bool {
+    let name = model.to_ascii_lowercase();
+    ["embed", "embedding", "rerank", "minilm", "bge-", "bge_", "e5-", "e5_"]
+        .iter()
+        .any(|marker| name.contains(marker))
+}
+
+fn preferred_installed_model(models: &[String]) -> Option<String> {
+    models
+        .iter()
+        .find(|model| !model.trim().is_empty() && !model_name_looks_embedding_only(model))
+        .or_else(|| models.iter().find(|model| !model.trim().is_empty()))
+        .cloned()
+}
+
+fn startup_installed_model(requested_model: &str, installed_models: &[String]) -> Option<String> {
+    matching_installed_model(installed_models, requested_model)
+        .or_else(|| preferred_installed_model(installed_models))
+}
+
+fn should_persist_resolved_model(cfg: &InferenceConfig) -> bool {
+    cfg.model
+        .as_deref()
+        .map(|model| model.trim().is_empty())
+        .unwrap_or(true)
+}
+
+async fn ollama_model_names() -> Vec<String> {
    let url = format!("http://127.0.0.1:{}/api/tags", OLLAMA_PORT);
    let client = reqwest::Client::builder()
        .timeout(Duration::from_secs(5))
@@ -615,21 +680,10 @@ async fn ollama_has_model(model: &str) -> bool {
        .unwrap();
    if let Ok(resp) = client.get(&url).send().await {
        if let Ok(body) = resp.json::<serde_json::Value>().await {
-            if let Some(models) = body.get("models").and_then(|m| m.as_array()) {
-                return models.iter().any(|m| {
-                    m.get("name")
-                        .and_then(|n| n.as_str())
-                        .map(|n| {
-                            n == model
-                                || n.strip_suffix(":latest") == Some(model)
-                                || model.strip_suffix(":latest") == Some(n)
-                        })
-                        .unwrap_or(false)
-                });
-            }
+            return parse_ollama_model_names(&body);
        }
    }
-    false
+    Vec::new()
 }

 async fn pull_model(model: &str) -> Result<(), String> {
@@ -679,13 +733,21 @@ fn format_uv_sync_failure(
    let code = exit_code
        .map(|c| c.to_string())
        .unwrap_or_else(|| "unknown".to_string());
+    let tail = uv_sync_stderr_tail(stderr, 800);
+    let rust_hint = if looks_like_rust_extension_build_error(stderr) {
+        format!("\n\n{}", rust_toolchain_install_hint())
+    } else {
+        String::new()
+    };
    format!(
        "`uv sync` failed in {} (exit {}). Last output:\n\n{}\n\n\
         Try opening a terminal in that directory and running \
-         `uv sync --extra server` manually for the full output.",
+         `{}` manually for the full output.{}",
        root.display(),
        code,
-        uv_sync_stderr_tail(stderr, 800),
+        tail,
+        DESKTOP_UV_SYNC_COMMAND,
+        rust_hint,
    )
 }

@@ -734,6 +796,122 @@ fn format_uv_sync_spawn_error(root: &std::path::Path, uv_bin: &str, err: &str) -
    )
 }

+fn rust_toolchain_install_hint() -> &'static str {
+    "The desktop app needs the Rust toolchain to build `openjarvis_rust`. \
+     Install Rust from https://rustup.rs. On Windows, also install Visual Studio \
+     Build Tools with the C++ workload, then relaunch."
+}
+
+fn looks_like_rust_extension_build_error(stderr: &str) -> bool {
+    let lower = stderr.to_ascii_lowercase();
+    [
+        "openjarvis-rust",
+        "openjarvis_rust",
+        "maturin",
+        "cargo",
+        "rustc",
+        "link.exe",
+        "visual studio",
+    ]
+    .iter()
+    .any(|marker| lower.contains(marker))
+}
+
+fn format_missing_rust_toolchain() -> String {
+    format!(
+        "Could not find Rust's `cargo` command. {}\n\n\
+         If Rust is already installed, close and relaunch the desktop app so \
+         PATH includes `~/.cargo/bin`.",
+        rust_toolchain_install_hint(),
+    )
+}
+
+fn format_extension_import_failure(root: &std::path::Path, stderr: &str) -> String {
+    let tail = uv_sync_stderr_tail(stderr, 4000);
+    format!(
+        "`openjarvis_rust` is still not importable after building. Last output:\n\n{}\n\n\
+         Run these manually for the full build log:\n\n\
+           cd {}\n\
+           {}\n\
+           uv run python -c \"import openjarvis_rust\"",
+        if tail.is_empty() {
+            "(no stderr output)"
+        } else {
+            &tail
+        },
+        root.display(),
+        DESKTOP_UV_SYNC_COMMAND,
+    )
+}
+
+fn add_cargo_bin_to_path(cmd: &mut tokio::process::Command) {
+    let mut paths: Vec<std::path::PathBuf> = std::env::var_os("PATH")
+        .map(|path| std::env::split_paths(&path).collect())
+        .unwrap_or_default();
+    paths.insert(
+        0,
+        std::path::PathBuf::from(home_dir())
+            .join(".cargo")
+            .join("bin"),
+    );
+    if let Ok(joined) = std::env::join_paths(paths) {
+        cmd.env("PATH", joined);
+    }
+}
+
+async fn verify_openjarvis_rust_extension(
+    root: &std::path::Path,
+    uv_bin: &str,
+) -> Result<(), String> {
+    let mut cmd = tokio::process::Command::new(uv_bin);
+    cmd.args(["run", "python", "-c", "import openjarvis_rust"])
+        .stdout(std::process::Stdio::null())
+        .stderr(std::process::Stdio::piped())
+        .current_dir(root);
+    prepare_subprocess_for_appimage(&mut cmd);
+    add_cargo_bin_to_path(&mut cmd);
+
+    match cmd.output().await {
+        Ok(out) if out.status.success() => Ok(()),
+        Ok(out) => {
+            let stderr = String::from_utf8_lossy(&out.stderr);
+            Err(format_extension_import_failure(root, &stderr))
+        }
+        Err(e) => Err(format!(
+            "Could not verify `openjarvis_rust`: {}. Verify uv is installed at `{}`.",
+            e, uv_bin
+        )),
+    }
+}
+
+fn port_owner_hint() -> String {
+    if cfg!(target_os = "windows") {
+        format!("netstat -ano | findstr :{}", JARVIS_PORT)
+    } else {
+        format!("lsof -i :{}", JARVIS_PORT)
+    }
+}
+
+fn format_port_unavailable(port: u16, reason: &str) -> String {
+    format!(
+        "Port {} is not available: {}. Stop the process using that port or \
+         change the OpenJarvis port, then relaunch.\n\nTo identify it:\n  {}",
+        port,
+        reason,
+        port_owner_hint(),
+    )
+}
+
+fn check_jarvis_port_available() -> Result<(), String> {
+    match std::net::TcpListener::bind(("127.0.0.1", JARVIS_PORT)) {
+        Ok(listener) => {
+            drop(listener);
+            Ok(())
+        }
+        Err(err) => Err(format_port_unavailable(JARVIS_PORT, &err.to_string())),
+    }
+}
+
 // ---------------------------------------------------------------------------
 // Backend boot sequence (runs in background after app launch)
 // ---------------------------------------------------------------------------
@@ -751,7 +929,7 @@ async fn boot_backend(backend: SharedBackend, status: SharedStatus) {
        .into();
    }

-    // For the Ollama path, the model pull may fall back to FALLBACK_MODEL; we
+    // For the Ollama path, model resolution may fall back to FALLBACK_MODEL; we
    // record what is actually available here so the serve command below uses
    // it instead of the originally-planned tag. None on the custom path.
    let mut serve_model_override: Option<String> = None;
@@ -798,8 +976,8 @@ async fn boot_backend(backend: SharedBackend, status: SharedStatus) {
            s.detail = "Inference engine ready.".into();
        }

-        // Phase 2: Pull the single default model (see default_local_model /
-        // boot_plan). We deliberately do NOT pull any others.
+        // Phase 2: Resolve one model to serve. Prefer an installed model on
+        // first run so startup does not depend on a download succeeding.
        let model = plan
            .model_to_pull
            .clone()
@@ -810,41 +988,63 @@ async fn boot_backend(backend: SharedBackend, status: SharedStatus) {
            s.detail = format!("Checking for {}...", model);
        }

-        if !ollama_has_model(&model).await {
+        let installed_models = ollama_model_names().await;
+        let resolved_model = if let Some(installed) = startup_installed_model(&model, &installed_models) {
+            installed
+        } else {
            {
                let mut s = status.lock().await;
                s.detail = format!("Downloading {}... (this may take a minute)", model);
            }
-            if let Err(e) = pull_model(&model).await {
-                // If the chosen model fails, try the tiny fallback
-                eprintln!("Warning: failed to pull {}: {}", model, e);
-                if !ollama_has_model(FALLBACK_MODEL).await {
-                    {
-                        let mut s = status.lock().await;
-                        s.detail = format!("Downloading {}...", FALLBACK_MODEL);
-                    }
-                    if let Err(e2) = pull_model(FALLBACK_MODEL).await {
-                        let mut s = status.lock().await;
-                        s.error = Some(format!("Failed to download model: {}", e2));
-                        return;
+            match pull_model(&model).await {
+                Ok(()) => model.clone(),
+                Err(e) => {
+                    eprintln!("Warning: failed to pull {}: {}", model, e);
+
+                    // If a local model appeared while pulling, use it instead of
+                    // making startup depend on another network pull.
+                    if let Some(installed) = preferred_installed_model(&ollama_model_names().await) {
+                        installed
+                    } else if ollama_has_model(FALLBACK_MODEL).await {
+                        FALLBACK_MODEL.to_string()
+                    } else {
+                        {
+                            let mut s = status.lock().await;
+                            s.detail = format!("Downloading {}...", FALLBACK_MODEL);
+                        }
+                        if let Err(e2) = pull_model(FALLBACK_MODEL).await {
+                            if let Some(installed) =
+                                preferred_installed_model(&ollama_model_names().await)
+                            {
+                                installed
+                            } else {
+                                let mut s = status.lock().await;
+                                s.error = Some(format!("Failed to download model: {}", e2));
+                                return;
+                            }
+                        } else {
+                            FALLBACK_MODEL.to_string()
+                        }
                    }
                }
            }
+        };
+
+        if resolved_model != model {
+            let mut s = status.lock().await;
+            s.detail = format!("Using installed model {}.", resolved_model);
        }

-        // The pull may have fallen back to FALLBACK_MODEL; serve and persist
-        // whatever is actually available now, not the originally-planned tag.
-        let resolved_model = if ollama_has_model(&model).await {
-            model
-        } else {
-            FALLBACK_MODEL.to_string()
-        };
        serve_model_override = Some(resolved_model.clone());

-        // Persist the resolved model so Settings shows it and future boots reuse it.
-        let mut persisted = cfg.clone();
-        persisted.model = Some(resolved_model);
-        let _ = write_inference_config(&persisted);
+        // Persist only first-run/default resolution. If the user explicitly
+        // configured a model, do not overwrite that choice with a temporary
+        // fallback selected just to keep startup nonfatal.
+        if should_persist_resolved_model(&cfg) {
+            let mut persisted = cfg.clone();
+            persisted.model = Some(resolved_model);
+            let _ = write_inference_config(&persisted);
+        }

        {
            let mut s = status.lock().await;
@@ -1097,11 +1297,6 @@ async fn boot_backend(backend: SharedBackend, status: SharedStatus) {
                // Something else (a different web server, a stale process,
                // a 4xx-returning instance) is on our port. Don't kill it —
                // give the user actionable info instead.
-                let lsof_hint = if cfg!(target_os = "windows") {
-                    format!("netstat -ano | findstr :{}", JARVIS_PORT)
-                } else {
-                    format!("lsof -i :{}", JARVIS_PORT)
-                };
                let mut s = status.lock().await;
                s.error = Some(format!(
                    "Port {} is already in use by another service (it answered \
@@ -1109,7 +1304,7 @@ async fn boot_backend(backend: SharedBackend, status: SharedStatus) {
                     OpenJarvis port, then relaunch.\n\nTo identify it:\n  {}",
                    JARVIS_PORT,
                    resp.status(),
-                    lsof_hint,
+                    port_owner_hint(),
                ));
                return;
            }
@@ -1119,8 +1314,21 @@ async fn boot_backend(backend: SharedBackend, status: SharedStatus) {
        }
    }

+    if let Err(err) = check_jarvis_port_available() {
+        let mut s = status.lock().await;
+        s.error = Some(err);
+        return;
+    }
+
    let root = project_root.as_ref().unwrap();

+    let cargo_bin = resolve_bin("cargo");
+    if !std::path::Path::new(&cargo_bin).exists() && cargo_bin == "cargo" {
+        let mut s = status.lock().await;
+        s.error = Some(format_missing_rust_toolchain());
+        return;
+    }
+
    // Install dependencies automatically (handles fresh clones).
    //
    // Previously we ran `uv sync` with both stdout AND stderr piped to
@@ -1143,15 +1351,19 @@ async fn boot_backend(backend: SharedBackend, status: SharedStatus) {
    sync_cmd
        .args([
            "sync",
-            "--extra", "server",
+            "--extra", "desktop",
            "--extra", "inference-cloud",
            "--extra", "inference-google",
+            // openjarvis_rust lives in a uv dependency group (not the published
+            // `desktop` extra) so pip installs from PyPI don't require it (#584).
+            "--group", "desktop-native",
        ])
        .stdout(std::process::Stdio::null())
        .stderr(std::process::Stdio::piped())
        .current_dir(root);
    // Avoid LD_LIBRARY_PATH leak when running inside an AppImage (#455).
    prepare_subprocess_for_appimage(&mut sync_cmd);
+    add_cargo_bin_to_path(&mut sync_cmd);
    let sync_output = sync_cmd.output().await;
    match sync_output {
        Ok(out) if !out.status.success() => {
@@ -1168,6 +1380,16 @@ async fn boot_backend(backend: SharedBackend, status: SharedStatus) {
        Ok(_) => {} // success — fall through
    }

+    {
+        let mut s = status.lock().await;
+        s.detail = "Verifying Rust extension (openjarvis_rust)...".into();
+    }
+    if let Err(err) = verify_openjarvis_rust_extension(root, &uv_bin).await {
+        let mut s = status.lock().await;
+        s.error = Some(err);
+        return;
+    }
+
    {
        let mut s = status.lock().await;
        s.detail = format!("Starting API server from {}...", root.display());
@@ -1204,7 +1426,7 @@ async fn boot_backend(backend: SharedBackend, status: SharedStatus) {
    // additions aren't accidentally stripped.
    prepare_subprocess_for_appimage(&mut cmd);

-    // Inject cloud API keys from ~/.openjarvis/cloud-keys.env
+    // Inject cloud API keys from secure desktop storage.
    for (key, value) in read_cloud_keys() {
        cmd.env(&key, &value);
    }
@@ -1540,19 +1762,89 @@ async fn fetch_models(api_url: String) -> Result<serde_json::Value, String> {

 #[tauri::command]
 async fn run_jarvis_command(args: Vec<String>) -> Result<String, String> {
-    let mut cmd_args = vec!["run".to_string(), "jarvis".to_string()];
-    cmd_args.extend(args);
    let uv_bin = resolve_bin("uv");
-    let output = tokio::process::Command::new(&uv_bin)
-        .args(&cmd_args)
-        .output()
-        .await
-        .map_err(|e| format!("Failed to launch jarvis: {}", e))?;

-    if output.status.success() {
-        Ok(String::from_utf8_lossy(&output.stdout).to_string())
-    } else {
-        Err(String::from_utf8_lossy(&output.stderr).to_string())
+    let mut cmd_args = vec!["run".to_string(), "jarvis".to_string()];
+    cmd_args.extend(args.iter().cloned());
+
+    let mut cmd = tokio::process::Command::new(&uv_bin);
+    cmd.args(&cmd_args);
+    // Run from the project root so `uv run jarvis` resolves the OpenJarvis
+    // project regardless of the app's launch cwd. In a packaged install the
+    // cwd isn't the checkout, so without this `jarvis` isn't found and the
+    // backend never starts — the UI then shows "Failed to get response"
+    // (see #531).
+    if let Some(ref root) = find_project_root() {
+        cmd.current_dir(root);
+    }
+
+    let is_serve = args.first().map(|a| a.as_str() == "serve").unwrap_or(false);
+
+    if !is_serve {
+        // Short-lived command (e.g. `stop`, `status`): wait for it and return
+        // its captured output.
+        let output = cmd
+            .output()
+            .await
+            .map_err(|e| format!("Failed to launch jarvis: {}", e))?;
+        return if output.status.success() {
+            Ok(String::from_utf8_lossy(&output.stdout).to_string())
+        } else {
+            Err(String::from_utf8_lossy(&output.stderr).to_string())
+        };
+    }
+
+    // `jarvis serve` is a long-running server that never exits. The old code
+    // used `.output()`, which waits for the process to exit and so hung this
+    // command forever — the "Start" button never resolved (#531). Spawn it
+    // detached instead, drain stderr (a full 4 KB Windows pipe can otherwise
+    // stall the child mid-startup, #309), and poll /health for readiness.
+    cmd.stdout(std::process::Stdio::null())
+        .stderr(std::process::Stdio::piped());
+    let mut child = cmd
+        .spawn()
+        .map_err(|e| format!("Failed to launch jarvis serve: {}", e))?;
+
+    let tail: StderrTail = Arc::new(Mutex::new(Vec::new()));
+    if let Some(stderr) = child.stderr.take() {
+        spawn_jarvis_stderr_drainer(stderr, tail.clone());
+    }
+
+    let client = reqwest::Client::builder()
+        .timeout(Duration::from_secs(2))
+        .build()
+        .map_err(|e| format!("Failed to build HTTP client: {}", e))?;
+    let url = format!("http://127.0.0.1:{}/health", JARVIS_PORT);
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(120);
+
+    loop {
+        // Surface an early crash (bad venv, missing Rust ext, etc.) right away
+        // instead of waiting out the full readiness timeout.
+        if let Ok(Some(status)) = child.try_wait() {
+            let stderr = String::from_utf8_lossy(tail.lock().await.as_slice()).into_owned();
+            return Err(format!(
+                "jarvis serve exited (code {:?}) before becoming healthy:\n{}",
+                status.code(),
+                stderr.trim()
+            ));
+        }
+        if let Ok(resp) = client.get(&url).send().await {
+            if resp.status().is_success() {
+                // Leave the server running (the Child is detached on drop —
+                // kill_on_drop defaults to false); `stop` tears it down.
+                return Ok(format!(
+                    "jarvis serve is ready on http://127.0.0.1:{}",
+                    JARVIS_PORT
+                ));
+            }
+        }
+        if tokio::time::Instant::now() >= deadline {
+            return Err(format!(
+                "jarvis serve did not become healthy on port {} within 120s.",
+                JARVIS_PORT
+            ));
+        }
+        tokio::time::sleep(Duration::from_millis(500)).await;
    }
 }

@@ -1594,11 +1886,29 @@ async fn transcribe_audio(
        .send()
        .await
        .map_err(|e| format!("Connection failed: {}", e))?;
-    let body: serde_json::Value = resp
-        .json()
+    let status = resp.status();
+    let body = resp
+        .text()
        .await
        .map_err(|e| format!("Invalid response: {}", e))?;
-    Ok(body)
+    if !status.is_success() {
+        let detail = serde_json::from_str::<serde_json::Value>(&body)
+            .ok()
+            .and_then(|value| {
+                value
+                    .get("detail")
+                    .and_then(|detail| detail.as_str())
+                    .map(str::to_string)
+            })
+            .filter(|detail| !detail.is_empty())
+            .unwrap_or(body);
+        return Err(format!(
+            "Transcription failed ({}): {}",
+            status.as_u16(),
+            detail
+        ));
+    }
+    serde_json::from_str(&body).map_err(|e| format!("Invalid response: {}", e))
 }

 /// Submit savings to Supabase leaderboard.
@@ -1632,17 +1942,111 @@ async fn submit_savings(
 // Cloud API key management
 // ---------------------------------------------------------------------------

-/// Path to the cloud keys file (~/.openjarvis/cloud-keys.env).
-fn cloud_keys_path() -> std::path::PathBuf {
+const SECURE_KEY_SERVICE: &str = "OpenJarvis Cloud Keys";
+const MANAGED_CLOUD_KEY_NAMES: &[&str] = &[
+    "OPENAI_API_KEY",
+    "ANTHROPIC_API_KEY",
+    "GEMINI_API_KEY",
+    "GOOGLE_API_KEY",
+    "OPENROUTER_API_KEY",
+    "MINIMAX_API_KEY",
+    "TAVILY_API_KEY",
+];
+
+/// Legacy path used by older desktop builds. New saves never write here.
+fn legacy_cloud_keys_path() -> std::path::PathBuf {
    let home = home_dir();
    std::path::PathBuf::from(home)
        .join(".openjarvis")
        .join("cloud-keys.env")
 }

-/// Read cloud keys from disk and return as key=value pairs.
-fn read_cloud_keys() -> Vec<(String, String)> {
-    let path = cloud_keys_path();
+fn validate_cloud_key_name(key_name: &str) -> Result<(), String> {
+    let valid = !key_name.is_empty()
+        && key_name.len() <= 128
+        && key_name.ends_with("_API_KEY")
+        && key_name
+            .chars()
+            .all(|ch| ch.is_ascii_uppercase() || ch.is_ascii_digit() || ch == '_');
+    if valid {
+        Ok(())
+    } else {
+        Err(format!("Invalid API key name: {}", key_name))
+    }
+}
+
+fn engine_api_key_name(engine: &str) -> String {
+    let normalized: String = engine
+        .chars()
+        .map(|ch| {
+            if ch.is_ascii_alphanumeric() {
+                ch.to_ascii_uppercase()
+            } else {
+                '_'
+            }
+        })
+        .collect();
+    let trimmed = normalized.trim_matches('_');
+    let engine_name = if trimmed.is_empty() {
+        CUSTOM_FALLBACK_ENGINE.to_ascii_uppercase()
+    } else {
+        trimmed.to_string()
+    };
+    format!("{}_API_KEY", engine_name)
+}
+
+fn managed_cloud_key_names() -> Vec<String> {
+    let mut names: Vec<String> = MANAGED_CLOUD_KEY_NAMES
+        .iter()
+        .map(|name| (*name).to_string())
+        .collect();
+
+    let cfg = read_inference_config();
+    if matches!(&cfg.kind, SourceKind::Custom) {
+        let engine = cfg.engine.unwrap_or_else(|| CUSTOM_FALLBACK_ENGINE.to_string());
+        let key_name = engine_api_key_name(&engine);
+        if validate_cloud_key_name(&key_name).is_ok() {
+            names.push(key_name);
+        }
+    }
+
+    names.sort();
+    names.dedup();
+    names
+}
+
+fn secure_store_get(key_name: &str) -> Result<Option<String>, String> {
+    validate_cloud_key_name(key_name)?;
+    let entry = keyring::Entry::new(SECURE_KEY_SERVICE, key_name)
+        .map_err(|err| format!("Failed to open secure key storage for {}: {}", key_name, err))?;
+    match entry.get_password() {
+        Ok(value) => Ok(Some(value)),
+        Err(keyring::Error::NoEntry) => Ok(None),
+        Err(err) => Err(format!("Failed to read {} from secure key storage: {}", key_name, err)),
+    }
+}
+
+fn secure_store_set(key_name: &str, key_value: &str) -> Result<(), String> {
+    validate_cloud_key_name(key_name)?;
+    let entry = keyring::Entry::new(SECURE_KEY_SERVICE, key_name)
+        .map_err(|err| format!("Failed to open secure key storage for {}: {}", key_name, err))?;
+    if key_value.is_empty() {
+        return match entry.delete_credential() {
+            Ok(()) => Ok(()),
+            Err(keyring::Error::NoEntry) => Ok(()),
+            Err(err) => Err(format!(
+                "Failed to remove {} from secure key storage: {}",
+                key_name, err
+            )),
+        };
+    }
+    entry
+        .set_password(key_value)
+        .map_err(|err| format!("Failed to save {} in secure key storage: {}", key_name, err))
+}
+
+fn read_legacy_cloud_keys() -> Vec<(String, String)> {
+    let path = legacy_cloud_keys_path();
    let mut keys = Vec::new();
    if let Ok(contents) = std::fs::read_to_string(&path) {
        for line in contents.lines() {
@@ -1658,47 +2062,68 @@ fn read_cloud_keys() -> Vec<(String, String)> {
    keys
 }

-/// Save a single cloud API key to the keys file.
-#[tauri::command]
-async fn save_cloud_key(key_name: String, key_value: String) -> Result<(), String> {
-    let path = cloud_keys_path();
-    // Ensure directory exists
-    if let Some(parent) = path.parent() {
-        let _ = std::fs::create_dir_all(parent);
+fn migrate_legacy_cloud_keys() {
+    let path = legacy_cloud_keys_path();
+    if !path.exists() {
+        return;
    }

-    // Read existing keys, update/add the one being saved
-    let mut keys: Vec<(String, String)> = read_cloud_keys()
+    let legacy_keys = read_legacy_cloud_keys();
+    if legacy_keys.is_empty() {
+        let _ = std::fs::remove_file(&path);
+        return;
+    }
+
+    let mut migrated_all = true;
+    for (key, value) in legacy_keys {
+        if value.is_empty() {
+            continue;
+        }
+        if secure_store_set(&key, &value).is_err() {
+            migrated_all = false;
+        }
+    }
+
+    if migrated_all {
+        let _ = std::fs::remove_file(path);
+    }
+}
+
+/// Read cloud keys from secure desktop storage and return key=value pairs.
+fn read_cloud_keys() -> Vec<(String, String)> {
+    migrate_legacy_cloud_keys();
+    managed_cloud_key_names()
        .into_iter()
-        .filter(|(k, _)| k != &key_name)
-        .collect();
-    if !key_value.is_empty() {
-        keys.push((key_name, key_value));
-    }
+        .filter_map(|key| match secure_store_get(&key) {
+            Ok(Some(value)) if !value.is_empty() => Some((key, value)),
+            _ => None,
+        })
+        .collect()
+}

-    // Write back
-    let content: String = keys
-        .iter()
-        .map(|(k, v)| format!("{}={}", k, v))
-        .collect::<Vec<_>>()
-        .join("\n");
-    std::fs::write(&path, content + "\n").map_err(|e| format!("Failed to save key: {}", e))?;
-
-    // Set permissions to owner-only (chmod 600)
-    #[cfg(unix)]
-    {
-        use std::os::unix::fs::PermissionsExt;
-        let _ = std::fs::set_permissions(&path, std::fs::Permissions::from_mode(0o600));
-    }
-
-    // Tell the running server to hot-reload its cloud engine so the user
-    // doesn't need to restart the app after entering an API key.
+async fn reload_cloud_keys(keys: Vec<(String, String)>) {
    let reload_url = format!("http://127.0.0.1:{}/v1/cloud/reload", JARVIS_PORT);
+    let key_map: serde_json::Map<String, serde_json::Value> = keys
+        .into_iter()
+        .map(|(key, value)| (key, serde_json::Value::String(value)))
+        .collect();
    let _ = reqwest::Client::new()
        .post(&reload_url)
+        .json(&serde_json::json!({ "keys": key_map }))
        .timeout(std::time::Duration::from_secs(10))
        .send()
        .await;
+}
+
+/// Save a single cloud API key to secure desktop storage.
+#[tauri::command]
+async fn save_cloud_key(key_name: String, key_value: String) -> Result<(), String> {
+    let key_value = key_value.trim().to_string();
+    secure_store_set(&key_name, &key_value)?;
+
+    // Tell the running server to hot-reload its cloud engine so the user
+    // doesn't need to restart the app after entering an API key.
+    reload_cloud_keys(vec![(key_name, key_value)]).await;

    Ok(())
 }
@@ -1706,10 +2131,13 @@ async fn save_cloud_key(key_name: String, key_value: String) -> Result<(), Strin
 /// Get which cloud providers have keys configured (without exposing values).
 #[tauri::command]
 async fn get_cloud_key_status() -> Result<serde_json::Value, String> {
-    let keys = read_cloud_keys();
-    let status: Vec<serde_json::Value> = keys
-        .iter()
-        .map(|(k, v)| serde_json::json!({ "key": k, "set": !v.is_empty() }))
+    migrate_legacy_cloud_keys();
+    let status: Vec<serde_json::Value> = managed_cloud_key_names()
+        .into_iter()
+        .map(|key| {
+            let set = matches!(secure_store_get(&key), Ok(Some(value)) if !value.is_empty());
+            serde_json::json!({ "key": key, "set": set })
+        })
        .collect();
    Ok(serde_json::json!(status))
 }
@@ -1721,8 +2149,8 @@ async fn get_inference_source() -> Result<InferenceConfig, String> {
 }

 /// Persist the chosen inference source. `host` is normalized to a bare base
-/// URL. For custom endpoints, an optional API key is stored in cloud-keys.env
-/// under `<ENGINE>_API_KEY`. Applies on next app launch.
+/// URL. For custom endpoints, an optional API key is stored in secure desktop
+/// storage under `<ENGINE>_API_KEY`. Applies on next app launch.
 #[tauri::command]
 async fn set_inference_source(
    kind: String,
@@ -1754,7 +2182,7 @@ async fn set_inference_source(
                .engine
                .clone()
                .unwrap_or_else(|| CUSTOM_FALLBACK_ENGINE.to_string());
-            let key_name = format!("{}_API_KEY", engine.to_ascii_uppercase());
+            let key_name = engine_api_key_name(&engine);
            // Save the key before persisting the config: if the key can't be
            // written, surface it and DON'T record a custom source whose
            // credential is missing (which would fail confusingly at runtime).
@@ -2440,9 +2868,12 @@ pub fn run() {
 #[cfg(test)]
 mod tests {
    use super::{
-        boot_plan, default_local_model, format_uv_sync_failure, format_uv_sync_spawn_error,
-        normalize_host, parse_inference_config, upsert_engine_host, uv_sync_stderr_tail,
-        InferenceConfig, SourceKind,
+        boot_plan, default_local_model, format_extension_import_failure,
+        format_missing_rust_toolchain, format_port_unavailable, format_uv_sync_failure,
+        format_uv_sync_spawn_error, matching_installed_model, model_names_match, normalize_host,
+        parse_inference_config, parse_ollama_model_names, preferred_installed_model,
+        should_persist_resolved_model, startup_installed_model, upsert_engine_host,
+        uv_sync_stderr_tail, InferenceConfig, SourceKind, DESKTOP_UV_SYNC_COMMAND,
    };
    use std::path::Path;

@@ -2486,7 +2917,7 @@ mod tests {
        assert!(msg.contains("exit 2"));
        assert!(msg.contains("/home/u/.openjarvis/src"));
        assert!(msg.contains("failed to resolve numpy==2.1.3"));
-        assert!(msg.contains("uv sync --extra server")); // actionable next step
+        assert!(msg.contains(DESKTOP_UV_SYNC_COMMAND)); // actionable next step
    }

    #[test]
@@ -2509,6 +2940,49 @@ mod tests {
        assert!(msg.contains("No such file or directory"));
    }

+    #[test]
+    fn missing_rust_toolchain_message_names_cargo_and_installer() {
+        let msg = format_missing_rust_toolchain();
+        assert!(msg.contains("cargo"));
+        assert!(msg.contains("https://rustup.rs"));
+        assert!(msg.contains("openjarvis_rust"));
+        assert!(msg.contains("Visual Studio Build Tools"));
+    }
+
+    #[test]
+    fn uv_sync_rust_failure_mentions_toolchain() {
+        let msg = format_uv_sync_failure(
+            Path::new("C:\\Users\\me\\OpenJarvis"),
+            Some(1),
+            "maturin failed: linker `link.exe` not found while building openjarvis-rust",
+        );
+        assert!(msg.contains("exit 1"));
+        assert!(msg.contains("link.exe"));
+        assert!(msg.contains("https://rustup.rs"));
+        assert!(msg.contains("Visual Studio Build Tools"));
+    }
+
+    #[test]
+    fn extension_import_failure_names_verification_command() {
+        let msg = format_extension_import_failure(
+            Path::new("C:\\Users\\me\\OpenJarvis"),
+            "ModuleNotFoundError: No module named 'openjarvis_rust'",
+        );
+        assert!(msg.contains("openjarvis_rust"));
+        assert!(msg.contains(DESKTOP_UV_SYNC_COMMAND));
+        assert!(msg.contains("uv run python -c \"import openjarvis_rust\""));
+        assert!(msg.contains("ModuleNotFoundError"));
+    }
+
+    #[test]
+    fn port_unavailable_message_names_port_and_owner_hint() {
+        let msg = format_port_unavailable(8000, "address already in use");
+        assert!(msg.contains("Port 8000 is not available"));
+        assert!(msg.contains("address already in use"));
+        assert!(msg.contains("To identify it"));
+        assert!(msg.contains("8000"));
+    }
+
    #[test]
    fn default_local_model_picks_second_largest_that_fits() {
        // QWEN35_MODELS min_ram ladder: 4,6,8,12,24,32,96 GB
@@ -2524,6 +2998,97 @@ mod tests {
        assert_eq!(default_local_model(1.0), super::FALLBACK_MODEL);
    }

+    #[test]
+    fn parse_ollama_model_names_reads_nonempty_names() {
+        let body = serde_json::json!({
+            "models": [
+                {"name": "llama3.2:latest"},
+                {"name": ""},
+                {"name": "qwen3.5:4b"},
+                {"model": "mistral:latest"}
+            ]
+        });
+        assert_eq!(
+            parse_ollama_model_names(&body),
+            vec![
+                "llama3.2:latest".to_string(),
+                "qwen3.5:4b".to_string(),
+                "mistral:latest".to_string()
+            ]
+        );
+    }
+
+    #[test]
+    fn model_names_match_treats_latest_as_optional() {
+        assert!(model_names_match("llama3.2:latest", "llama3.2"));
+        assert!(model_names_match("llama3.2", "llama3.2:latest"));
+        assert!(model_names_match("qwen3.5:4b", "qwen3.5:4b"));
+        assert!(!model_names_match("llama3.2:latest", "qwen3.5:4b"));
+    }
+
+    #[test]
+    fn installed_model_helpers_pick_matching_or_first_model() {
+        let models = vec!["llama3.2:latest".to_string(), "qwen3.5:4b".to_string()];
+        assert_eq!(
+            matching_installed_model(&models, "llama3.2"),
+            Some("llama3.2:latest".to_string())
+        );
+        assert_eq!(
+            preferred_installed_model(&models),
+            Some("llama3.2:latest".to_string())
+        );
+    }
+
+    #[test]
+    fn preferred_installed_model_skips_embedding_names_when_chat_model_exists() {
+        let models = vec![
+            "nomic-embed-text:latest".to_string(),
+            "llama3.2:latest".to_string(),
+        ];
+        assert_eq!(
+            preferred_installed_model(&models),
+            Some("llama3.2:latest".to_string())
+        );
+    }
+
+    #[test]
+    fn startup_installed_model_uses_existing_model_for_defaults() {
+        let models = vec!["llama3.2:latest".to_string()];
+        assert_eq!(
+            startup_installed_model("qwen3.5:4b", &models),
+            Some("llama3.2:latest".to_string())
+        );
+    }
+
+    #[test]
+    fn startup_installed_model_uses_existing_model_when_configured_model_missing() {
+        let models = vec!["llama3.2:latest".to_string()];
+        assert_eq!(
+            startup_installed_model("qwen3.5:4b", &models),
+            Some("llama3.2:latest".to_string())
+        );
+    }
+
+    #[test]
+    fn resolved_model_is_only_persisted_when_no_model_was_configured() {
+        let default_cfg = InferenceConfig { kind: SourceKind::Ollama, ..Default::default() };
+        assert!(should_persist_resolved_model(&default_cfg));
+
+        let empty_cfg = InferenceConfig {
+            kind: SourceKind::Ollama,
+            model: Some(" ".into()),
+            ..Default::default()
+        };
+        assert!(should_persist_resolved_model(&empty_cfg));
+
+        let user_cfg = InferenceConfig {
+            kind: SourceKind::Ollama,
+            model: Some("qwen3.5:9b".into()),
+            ..Default::default()
+        };
+        assert!(!should_persist_resolved_model(&user_cfg));
+    }
+
    #[test]
    fn parse_defaults_to_ollama_when_file_missing_or_garbage() {
        assert!(matches!(parse_inference_config("").kind, SourceKind::Ollama));
@@ -31,7 +31,6 @@ export default function App() {
  const prevModelRef = useRef<string>('');
  const setModels = useAppStore((s) => s.setModels);
  const setModelsLoading = useAppStore((s) => s.setModelsLoading);
-  const setSelectedModel = useAppStore((s) => s.setSelectedModel);
  const selectedModel = useAppStore((s) => s.selectedModel);
  const setServerInfo = useAppStore((s) => s.setServerInfo);
  const setSavings = useAppStore((s) => s.setSavings);
@@ -70,7 +69,6 @@ export default function App() {
    fetchModels()
      .then((m) => {
        setModels(m);
-        if (!selectedModel && m.length > 0) setSelectedModel(m[0].id);
      })
      .catch(() => setModels([]))
      .finally(() => setModelsLoading(false));
@@ -89,7 +87,7 @@ export default function App() {
          setSavings(data);
          if (optInEnabled && optInDisplayName && data) {
            const claudeEntry = data.per_provider.find(
-              (p) => p.provider === 'claude-opus-4.6',
+              (p) => p.provider === 'claude-fable-5',
            );
            const dollarSavings = claudeEntry ? claudeEntry.total_cost : 0;
            const energySaved = data.per_provider.reduce(
@@ -15,6 +15,7 @@ function getGreeting(): string {
 }

 export function ChatArea() {
+  const activeId = useAppStore((s) => s.activeId);
  const messages = useAppStore((s) => s.messages);
  const streamState = useAppStore((s) => s.streamState);
  const systemPanelOpen = useAppStore((s) => s.systemPanelOpen);
@@ -22,6 +23,10 @@ export function ChatArea() {
  const navigate = useNavigate();
  const listRef = useRef<HTMLDivElement>(null);
  const shouldAutoScroll = useRef(true);
+  const wasStreaming = useRef(false);
+  const lastScrollTop = useRef(0);
+  const isCurrentChatStreaming = streamState.isStreaming && streamState.conversationId === activeId;
+  const currentStreamContent = isCurrentChatStreaming ? streamState.content : '';

  // Check if any data sources are connected
  const [hasConnectedSources, setHasConnectedSources] = useState<boolean | null>(null);
@@ -34,18 +39,37 @@ export function ChatArea() {
  }, []);

  useEffect(() => {
+    // Sending a message always pins the view to the bottom, even if the
+    // user had scrolled up to read earlier messages.
+    if (isCurrentChatStreaming && !wasStreaming.current) {
+      shouldAutoScroll.current = true;
+    }
+    wasStreaming.current = isCurrentChatStreaming;
    if (shouldAutoScroll.current && listRef.current) {
      listRef.current.scrollTop = listRef.current.scrollHeight;
    }
-  }, [messages, streamState.content]);
+  }, [messages, currentStreamContent, isCurrentChatStreaming]);

  const handleScroll = () => {
    if (!listRef.current) return;
    const { scrollTop, scrollHeight, clientHeight } = listRef.current;
-    shouldAutoScroll.current = scrollHeight - scrollTop - clientHeight < 100;
+    const distance = scrollHeight - scrollTop - clientHeight;
+    const scrolledUp = scrollTop < lastScrollTop.current;
+    lastScrollTop.current = scrollTop;
+    if (scrolledUp && distance >= 1) {
+      // Any upward scroll away from the bottom stops autoscroll immediately,
+      // so streaming content never fights the user (no jitter). Sub-1px
+      // upward movement (elastic bounce settling at the bottom) is ignored.
+      shouldAutoScroll.current = false;
+    } else if (!scrolledUp) {
+      // Re-engage when scrolled back to the bottom. < 2 rather than < 1:
+      // at fractional zoom levels the at-bottom residual can reach 1px,
+      // which would otherwise leave autoscroll permanently disengaged.
+      shouldAutoScroll.current = distance < 2;
+    }
  };

-  const isEmpty = messages.length === 0 && !streamState.isStreaming;
+  const isEmpty = messages.length === 0 && !isCurrentChatStreaming;

  const PanelIcon = systemPanelOpen ? PanelRightClose : PanelRightOpen;

@@ -153,12 +177,12 @@ export function ChatArea() {
                <MessageBubble
                  key={msg.id}
                  message={msg}
-                  isLive={isLastAssistant && streamState.isStreaming}
+                  isLive={isLastAssistant && isCurrentChatStreaming}
                />
              );
            })}
            {(() => {
-              if (!streamState.isStreaming || streamState.content !== '') return null;
+              if (!isCurrentChatStreaming || streamState.content !== '') return null;
              // For research messages the ResearchTimeline handles its own
              // pre-content loading state — suppress the generic dots.
              const last = messages[messages.length - 1];
@@ -96,8 +96,15 @@ export function InputArea() {
  const deepResearch = useAppStore((s) => s.deepResearch);
  const setDeepResearch = useAppStore((s) => s.setDeepResearch);
  const corpusSync = useResearchCorpusSync(deepResearch);
+  const isCurrentChatStreaming = streamState.isStreaming && streamState.conversationId === activeId;

-  const { state: speechState, available: speechAvailable, startRecording, stopRecording } = useSpeech();
+  const {
+    state: speechState,
+    error: speechError,
+    available: speechAvailable,
+    startRecording,
+    stopRecording,
+  } = useSpeech();

  // Abort in-flight stream when the user switches models mid-generation.
  // This prevents errors from trying to continue a stream with a stale model.
@@ -122,6 +129,12 @@ export function InputArea() {
    : streamState.isStreaming ? 'streaming'
    : undefined;

+  useEffect(() => {
+    if (speechError) {
+      toast.error(speechError, { duration: 8000 });
+    }
+  }, [speechError]);
+
  const handleMicClick = useCallback(async () => {
    if (speechState === 'recording') {
      try {
@@ -214,6 +227,7 @@ export function InputArea() {
    let ttftMs: number | undefined;

    setStreamState({
+      conversationId: convId,
      isStreaming: true,
      phase: deepResearch ? 'Researching...' : 'Generating...',
      elapsedMs: 0,
@@ -231,7 +245,11 @@ export function InputArea() {

    try {
      if (deepResearch) {
-        for await (const ev of streamResearch(content, controller.signal)) {
+        for await (const ev of streamResearch(
+          content,
+          selectedModel,
+          controller.signal,
+        )) {
          if (ev.type === 'search_call') {
            const trace: ResearchSearchTrace = {
              id: generateId(),
@@ -450,7 +468,10 @@ export function InputArea() {
      }
      const totalMs = Date.now() - startTime;
      const _CLOUD_PREFIXES = ['gpt-', 'o1-', 'o3-', 'o4-', 'claude-', 'gemini-', 'openrouter/', 'MiniMax-', 'chatgpt-'];
-      const engineLabel = _CLOUD_PREFIXES.some(p => selectedModel.startsWith(p)) ? 'cloud' : 'ollama';
+      const selectedOwner = useAppStore.getState().models.find((m) => m.id === selectedModel)?.owned_by;
+      const engineLabel = selectedOwner === 'litellm'
+        ? 'litellm'
+        : _CLOUD_PREFIXES.some(p => selectedModel.startsWith(p)) ? 'cloud' : 'ollama';
      const telemetry: MessageTelemetry = {
        engine: engineLabel,
        model_id: selectedModel,
@@ -583,7 +604,7 @@ export function InputArea() {
          style={{ color: 'var(--color-text)', maxHeight: '200px' }}
          disabled={streamState.isStreaming || modelLoading}
        />
-        {streamState.isStreaming ? (
+        {isCurrentChatStreaming ? (
          <button
            onClick={stopStreaming}
            className="p-2 rounded-xl transition-colors shrink-0 cursor-pointer"
@@ -602,7 +623,7 @@ export function InputArea() {
            />
            <button
              onClick={sendMessage}
-              disabled={!input.trim() || modelLoading || !selectedModel}
+              disabled={streamState.isStreaming || !input.trim() || modelLoading || !selectedModel}
              title={selectedModel ? 'Send message' : 'Pick a model first (⌘K)'}
              className="p-2 rounded-xl transition-colors shrink-0 cursor-pointer disabled:opacity-30 disabled:cursor-default"
              style={{
@@ -29,8 +29,8 @@ interface TelemetryStats {
 }

 const CLOUD_PRICING = [
-  { name: 'GPT-5.3', input: 2.00, output: 10.00, primary: true },
-  { name: 'Claude Opus 4.6', input: 5.00, output: 25.00, primary: false },
+  { name: 'GPT-5.6 Sol', input: 5.00, output: 30.00, primary: true },
+  { name: 'Claude Fable 5', input: 10.00, output: 50.00, primary: false },
  { name: 'Gemini 3.1 Pro', input: 2.00, output: 12.00, primary: false },
 ];

@@ -1,7 +1,15 @@
-import { useState, useRef, useEffect } from 'react';
+import { useState, useRef, useEffect, useCallback } from 'react';
 import { Search, Cpu, X, Download, Loader2, Trash2, Check, Cloud, Key, Eye, EyeOff } from 'lucide-react';
 import { useAppStore } from '../lib/store';
-import { pullModel, deleteModel, fetchModels, preloadModel, isTauri } from '../lib/api';
+import {
+  pullModel,
+  deleteModel,
+  fetchModels,
+  preloadModel,
+  isTauri,
+  getCloudKeyStatus,
+  saveCloudKey,
+} from '../lib/api';

 /** Popular models that users can download from the catalogue. */
 const CATALOGUE_MODELS = [
@@ -23,7 +31,6 @@ const CATALOGUE_MODELS = [
 interface CloudProvider {
  name: string;
  envKey: string;
-  storageKey: string;
  models: Array<{ id: string; desc: string }>;
 }

@@ -31,7 +38,6 @@ const CLOUD_PROVIDERS: CloudProvider[] = [
  {
    name: 'OpenAI',
    envKey: 'OPENAI_API_KEY',
-    storageKey: 'openjarvis-openai-key',
    models: [
      { id: 'gpt-4o', desc: 'GPT-4o — fast, multimodal' },
      { id: 'gpt-4o-mini', desc: 'GPT-4o Mini — cheap, fast' },
@@ -41,7 +47,6 @@ const CLOUD_PROVIDERS: CloudProvider[] = [
  {
    name: 'Anthropic',
    envKey: 'ANTHROPIC_API_KEY',
-    storageKey: 'openjarvis-anthropic-key',
    models: [
      { id: 'claude-sonnet-4-6', desc: 'Claude Sonnet 4.6 — balanced' },
      { id: 'claude-opus-4-6', desc: 'Claude Opus 4.6 — most capable' },
@@ -51,7 +56,6 @@ const CLOUD_PROVIDERS: CloudProvider[] = [
  {
    name: 'Google',
    envKey: 'GEMINI_API_KEY',
-    storageKey: 'openjarvis-gemini-key',
    models: [
      { id: 'gemini-2.5-pro', desc: 'Gemini 2.5 Pro — flagship' },
      { id: 'gemini-2.5-flash', desc: 'Gemini 2.5 Flash — fast' },
@@ -61,7 +65,6 @@ const CLOUD_PROVIDERS: CloudProvider[] = [
  {
    name: 'OpenRouter',
    envKey: 'OPENROUTER_API_KEY',
-    storageKey: 'openjarvis-openrouter-key',
    models: [
      { id: 'openrouter/auto', desc: 'Auto — best model for the task' },
      { id: 'openrouter/anthropic/claude-sonnet-4', desc: 'Claude Sonnet 4 via OpenRouter' },
@@ -70,16 +73,6 @@ const CLOUD_PROVIDERS: CloudProvider[] = [
  },
 ];

-function getStoredKey(storageKey: string): string {
-  try { return localStorage.getItem(storageKey) || ''; } catch { return ''; }
-}
-function setStoredKey(storageKey: string, value: string): void {
-  try {
-    if (value) localStorage.setItem(storageKey, value);
-    else localStorage.removeItem(storageKey);
-  } catch {}
-}
-
 type Tab = 'installed' | 'catalogue' | 'cloud';

 export function CommandPalette() {
@@ -92,11 +85,10 @@ export function CommandPalette() {
  const [deleting, setDeleting] = useState<string | null>(null);
  const [customModel, setCustomModel] = useState('');
  const [showKeys, setShowKeys] = useState<Record<string, boolean>>({});
-  const [apiKeys, setApiKeys] = useState<Record<string, string>>(() => {
-    const keys: Record<string, string> = {};
-    for (const p of CLOUD_PROVIDERS) keys[p.storageKey] = getStoredKey(p.storageKey);
-    return keys;
-  });
+  const [apiKeys, setApiKeys] = useState<Record<string, string>>({});
+  const [cloudKeyStatus, setCloudKeyStatus] = useState<Record<string, boolean>>({});
+  const [cloudKeyError, setCloudKeyError] = useState<string | null>(null);
+  const [savingKey, setSavingKey] = useState<string | null>(null);
  const inputRef = useRef<HTMLInputElement>(null);

  const models = useAppStore((s) => s.models);
@@ -106,6 +98,20 @@ export function CommandPalette() {
  const setCommandPaletteOpen = useAppStore((s) => s.setCommandPaletteOpen);

  const installedIds = new Set(models.map((m) => m.id));
+  const desktopKeyStorage = isTauri();
+
+  const refreshCloudKeyStatus = useCallback(async () => {
+    if (!desktopKeyStorage) {
+      setCloudKeyStatus({});
+      return;
+    }
+    try {
+      setCloudKeyStatus(await getCloudKeyStatus());
+      setCloudKeyError(null);
+    } catch (e: any) {
+      setCloudKeyError(e?.message || 'Failed to read cloud key status');
+    }
+  }, [desktopKeyStorage]);

  const filtered = tab === 'installed'
    ? (query
@@ -122,6 +128,10 @@ export function CommandPalette() {
    inputRef.current?.focus();
  }, []);

+  useEffect(() => {
+    void refreshCloudKeyStatus();
+  }, [refreshCloudKeyStatus]);
+
  useEffect(() => {
    setSelectedIdx(0);
  }, [query, tab]);
@@ -133,18 +143,17 @@ export function CommandPalette() {
    }
  }, [pullSuccess]);

-  const handleSelect = async (modelId: string) => {
+  const handleSelect = async (modelId: string, owner?: string) => {
    const previousModel = selectedModel;
    setSelectedModel(modelId);
    setCommandPaletteOpen(false);

    if (modelId !== previousModel) {
-      const { createConversation, setModelLoading, addLogEntry } = useAppStore.getState();
-      createConversation(modelId);
+      const { setModelLoading, addLogEntry } = useAppStore.getState();
      setModelLoading(true);
      addLogEntry({ timestamp: Date.now(), level: 'info', category: 'model', message: `Switching to ${modelId}...` });
      try {
-        await preloadModel(modelId);
+        await preloadModel(modelId, owner);
        addLogEntry({ timestamp: Date.now(), level: 'info', category: 'model', message: `${modelId} loaded` });
      } catch (e: any) {
        addLogEntry({ timestamp: Date.now(), level: 'error', category: 'model', message: `Failed to load ${modelId}: ${e.message}` });
@@ -210,24 +219,29 @@ export function CommandPalette() {
  };

  const handleSaveKey = async (provider: CloudProvider, value: string) => {
-    setStoredKey(provider.storageKey, value);
-    setApiKeys((prev) => ({ ...prev, [provider.storageKey]: value }));
+    const keyValue = value.trim();
+    setSavingKey(provider.envKey);
+    setCloudKeyError(null);

-    // Also save to Tauri backend so the server process picks up the key
-    if (isTauri()) {
-      try {
-        const { invoke } = await import('@tauri-apps/api/core');
-        await invoke('save_cloud_key', { keyName: provider.envKey, keyValue: value });
-      } catch {}
+    try {
+      await saveCloudKey(provider.envKey, keyValue);
+      setApiKeys((prev) => ({ ...prev, [provider.envKey]: '' }));
+      await refreshCloudKeyStatus();
+      useAppStore.getState().addLogEntry({
+        timestamp: Date.now(), level: 'info', category: 'model',
+        message: `${provider.name} API key ${keyValue ? 'saved' : 'removed'}. Refreshing model list...`,
+      });
+      await refreshModels();
+    } catch (e: any) {
+      setCloudKeyError(e?.message || `Failed to save ${provider.name} API key`);
+    } finally {
+      setSavingKey(null);
    }
+  };

-    useAppStore.getState().addLogEntry({
-      timestamp: Date.now(), level: 'info', category: 'model',
-      message: `${provider.name} API key ${value ? 'saved' : 'removed'}. Refreshing model list…`,
-    });
-
-    // Refresh the model list so cloud models appear immediately.
-    await refreshModels();
+  const handleKeyBlur = (provider: CloudProvider) => {
+    const draft = apiKeys[provider.envKey] || '';
+    if (draft.trim()) void handleSaveKey(provider, draft);
  };

  const handleKeyDown = (e: React.KeyboardEvent) => {
@@ -241,7 +255,8 @@ export function CommandPalette() {
      setSelectedIdx((i) => Math.max(i - 1, 0));
    } else if (e.key === 'Enter' && tab === 'installed' && filtered.length > 0) {
      e.preventDefault();
-      handleSelect((filtered[selectedIdx] as any).id);
+      const model = filtered[selectedIdx] as (typeof models)[number];
+      handleSelect(model.id, model.owned_by);
    }
  };

@@ -323,6 +338,11 @@ export function CommandPalette() {
            <Check size={12} /> Downloaded {pullSuccess} successfully
          </div>
        )}
+        {tab === 'cloud' && cloudKeyError && (
+          <div className="px-4 py-2 text-xs" style={{ color: 'var(--color-error)', background: 'rgba(220,38,38,0.05)' }}>
+            {cloudKeyError}
+          </div>
+        )}

        {/* Results */}
        <div className="max-h-[400px] overflow-y-auto py-2">
@@ -346,11 +366,15 @@ export function CommandPalette() {
                    onMouseEnter={() => setSelectedIdx(idx)}
                  >
                    <button
-                      onClick={() => handleSelect(model.id)}
+                      onClick={() => handleSelect(model.id, model.owned_by)}
                      className="flex items-center gap-3 flex-1 min-w-0 text-left cursor-pointer"
                      style={{ background: 'none', border: 'none', padding: 0 }}
                    >
-                      <Cpu size={16} style={{ color: isActive ? 'var(--color-accent)' : 'var(--color-text-tertiary)' }} />
+                      {model.owned_by === 'litellm' ? (
+                        <Cloud size={16} style={{ color: isActive ? 'var(--color-accent)' : 'var(--color-text-tertiary)' }} />
+                      ) : (
+                        <Cpu size={16} style={{ color: isActive ? 'var(--color-accent)' : 'var(--color-text-tertiary)' }} />
+                      )}
                      <div className="flex-1 min-w-0">
                        <div className="text-sm truncate" style={{ color: isActive ? 'var(--color-accent)' : 'var(--color-text)', fontWeight: isActive ? 500 : 400 }}>
                          {model.id}
@@ -362,17 +386,19 @@ export function CommandPalette() {
                        </span>
                      )}
                    </button>
-                    <button
-                      onClick={() => handleDelete(model.id)}
-                      disabled={isDeleting}
-                      className="p-1 rounded transition-colors cursor-pointer"
-                      style={{ color: 'var(--color-text-tertiary)', opacity: 0 }}
-                      title="Delete model"
-                      onMouseEnter={(e) => { e.currentTarget.style.opacity = '1'; e.currentTarget.style.color = 'var(--color-error)'; }}
-                      onMouseLeave={(e) => { e.currentTarget.style.opacity = '0'; e.currentTarget.style.color = 'var(--color-text-tertiary)'; }}
-                    >
-                      {isDeleting ? <Loader2 size={14} className="animate-spin" /> : <Trash2 size={14} />}
-                    </button>
+                    {model.owned_by !== 'litellm' && (
+                      <button
+                        onClick={() => handleDelete(model.id)}
+                        disabled={isDeleting}
+                        className="p-1 rounded transition-colors cursor-pointer"
+                        style={{ color: 'var(--color-text-tertiary)', opacity: 0 }}
+                        title="Delete model"
+                        onMouseEnter={(e) => { e.currentTarget.style.opacity = '1'; e.currentTarget.style.color = 'var(--color-error)'; }}
+                        onMouseLeave={(e) => { e.currentTarget.style.opacity = '0'; e.currentTarget.style.color = 'var(--color-text-tertiary)'; }}
+                      >
+                        {isDeleting ? <Loader2 size={14} className="animate-spin" /> : <Trash2 size={14} />}
+                      </button>
+                    )}
                  </div>
                );
              })
@@ -431,13 +457,17 @@ export function CommandPalette() {
            /* ── Cloud Models tab ── */
            <div className="px-4 py-2">
              <div className="text-[11px] mb-3" style={{ color: 'var(--color-text-tertiary)' }}>
-                Add your API keys to use cloud models. Keys are stored locally on your device only.
+                {desktopKeyStorage
+                  ? 'Add your API keys to use cloud models. Keys are stored in secure desktop storage.'
+                  : 'Configure cloud provider keys in the server environment to use cloud models.'}
              </div>

              {CLOUD_PROVIDERS.map((provider) => {
-                const key = apiKeys[provider.storageKey] || '';
-                const hasKey = !!key;
-                const isVisible = showKeys[provider.storageKey];
+                const key = apiKeys[provider.envKey] || '';
+                const hasSavedKey = !!cloudKeyStatus[provider.envKey];
+                const hasKey = hasSavedKey || !!key.trim();
+                const isVisible = showKeys[provider.envKey];
+                const isSaving = savingKey === provider.envKey;

                return (
                  <div key={provider.name} className="mb-4">
@@ -458,26 +488,28 @@ export function CommandPalette() {
                        <input
                          type={isVisible ? 'text' : 'password'}
                          value={key}
-                          onChange={(e) => setApiKeys((prev) => ({ ...prev, [provider.storageKey]: e.target.value }))}
-                          onBlur={() => handleSaveKey(provider, apiKeys[provider.storageKey] || '')}
-                          placeholder={`${provider.envKey}`}
+                          onChange={(e) => setApiKeys((prev) => ({ ...prev, [provider.envKey]: e.target.value }))}
+                          onBlur={() => handleKeyBlur(provider)}
+                          placeholder={hasSavedKey ? 'Saved in secure storage' : provider.envKey}
+                          disabled={!desktopKeyStorage || isSaving}
                          className="flex-1 text-xs px-2 py-1.5 bg-transparent outline-none font-mono"
                          style={{ color: 'var(--color-text)' }}
                        />
                        <button
-                          onClick={() => setShowKeys((prev) => ({ ...prev, [provider.storageKey]: !prev[provider.storageKey] }))}
+                          onClick={() => setShowKeys((prev) => ({ ...prev, [provider.envKey]: !prev[provider.envKey] }))}
                          className="px-2 cursor-pointer" style={{ color: 'var(--color-text-tertiary)' }}
                        >
                          {isVisible ? <EyeOff size={12} /> : <Eye size={12} />}
                        </button>
                      </div>
-                      {hasKey && (
+                      {hasSavedKey && (
                        <button
                          onClick={() => handleSaveKey(provider, '')}
+                          disabled={isSaving}
                          className="px-2 py-1 rounded-lg text-[10px] cursor-pointer"
-                          style={{ color: 'var(--color-error)', border: '1px solid var(--color-error)' }}
+                          style={{ color: 'var(--color-error)', border: '1px solid var(--color-error)', opacity: isSaving ? 0.5 : 1 }}
                        >
-                          Remove
+                          {isSaving ? 'Saving' : 'Remove'}
                        </button>
                      )}
                    </div>
@@ -2,8 +2,8 @@ import { DollarSign, TrendingDown, Cloud, HardDrive } from 'lucide-react';
 import { useAppStore } from '../../lib/store';

 const CLOUD_PRICING = [
-  { name: 'GPT-5.3', input: 2.00, output: 10.00 },
-  { name: 'Claude Opus 4.6', input: 5.00, output: 25.00 },
+  { name: 'GPT-5.6 Sol', input: 5.00, output: 30.00 },
+  { name: 'Claude Fable 5', input: 10.00, output: 50.00 },
  { name: 'Gemini 3.1 Pro', input: 2.00, output: 12.00 },
 ];

@@ -1,6 +1,7 @@
 import { useState, useEffect, useCallback } from 'react';
 import type React from 'react';
 import { invoke } from '@tauri-apps/api/core';
+import { LEADERBOARD_ENABLED, SUPABASE_ANON_KEY, SUPABASE_URL } from '../../lib/supabase';

 // ---------------------------------------------------------------------------
 // Types
@@ -221,8 +222,8 @@ const styles: Record<string, React.CSSProperties> = {
 };

 const PROVIDER_COLORS: Record<string, string> = {
-  'gpt-5.3': colors.green,
-  'claude-opus-4.6': colors.yellow,
+  'gpt-5.6-sol': colors.green,
+  'claude-fable-5': colors.yellow,
  'gemini-3.1-pro': colors.accent,
 };

@@ -279,9 +280,6 @@ function getOrCreateAnonId(): string {
  return id;
 }

-const SUPABASE_URL = 'https://mtbtgpwzrbostweaanpr.supabase.co';
-const SUPABASE_KEY = 'eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJzdXBhYmFzZSIsInJlZiI6Im10YnRncHd6cmJvc3R3ZWFhbnByIiwicm9sZSI6ImFub24iLCJpYXQiOjE3NzMxODk0OTQsImV4cCI6MjA4ODc2NTQ5NH0._xMlqCfljtXpwPj54H-ghxfLFO-jiq4W2WhpU8vVL1c';
-
 const REFRESH_INTERVAL_MS = 5000;

 export function SavingsDashboard({ apiUrl }: { apiUrl: string }) {
@@ -318,15 +316,16 @@ export function SavingsDashboard({ apiUrl }: { apiUrl: string }) {
    return () => clearInterval(timer);
  }, [fetchData]);

-  // Share savings to Supabase when opted in and data changes
+  // Share savings to Supabase when opted in and data changes. Skipped entirely
+  // when no anon key was built in (leaderboard disabled).
  useEffect(() => {
-    if (!optInEnabled || !displayName || !data) return;
+    if (!LEADERBOARD_ENABLED || !optInEnabled || !displayName || !data) return;
    const dollarSavings = data.per_provider.reduce((s, p) => s + p.total_cost, 0);
    const energySaved = data.per_provider.reduce((s, p) => s + (p.energy_wh || 0), 0);
    const flopsSaved = data.per_provider.reduce((s, p) => s + (p.flops || 0), 0);
    invoke('submit_savings', {
      supabaseUrl: SUPABASE_URL,
-      supabaseKey: SUPABASE_KEY,
+      supabaseKey: SUPABASE_ANON_KEY,
      payload: {
        anon_id: anonId,
        display_name: displayName,
@@ -7,6 +7,7 @@ import {
  type SetupStatus,
 } from '../lib/api';
 import { useAppStore } from '../lib/store';
+import { isEmbedOnlyModel } from '../lib/model-capabilities';

 const STEPS = [
  { key: 'ollama_ready', label: 'Inference Engine', icon: Cpu, detail: 'Starting Ollama...' },
@@ -91,12 +92,14 @@ export function SetupScreen({ onReady }: { onReady: () => void }) {
          fetchRecommendedModel().catch(() => ({ model: '', reason: '' })),
        ]);
        const store = useAppStore.getState();
+        const hadSelection = !!store.selectedModel;
        store.setModels(models);
        store.setModelsLoading(false);
-        const recommended = rec.model && models.some((m) => m.id === rec.model)
+        const chatModels = models.filter((m) => !isEmbedOnlyModel(m.id));
+        const recommended = rec.model && chatModels.some((m) => m.id === rec.model)
          ? rec.model
-          : models[0]?.id || '';
-        if (recommended && !store.selectedModel) {
+          : chatModels[0]?.id || '';
+        if (recommended && !hadSelection) {
          store.setSelectedModel(recommended);
        }
      } catch {
@@ -22,6 +22,9 @@ export function ConversationList({ searchQuery }: Props) {
  const navigate = useNavigate();
  const conversations = useAppStore((s) => s.conversations);
  const activeId = useAppStore((s) => s.activeId);
+  const streamingConversationId = useAppStore((s) =>
+    s.streamState.isStreaming ? s.streamState.conversationId : null,
+  );
  const selectConversation = useAppStore((s) => s.selectConversation);
  const deleteConversation = useAppStore((s) => s.deleteConversation);

@@ -43,6 +46,7 @@ export function ConversationList({ searchQuery }: Props) {
    <div className="flex flex-col gap-0.5 py-1">
      {filtered.map((conv) => {
        const isActive = conv.id === activeId;
+        const isStreaming = conv.id === streamingConversationId;
        return (
          <div
            key={conv.id}
@@ -82,11 +86,18 @@ export function ConversationList({ searchQuery }: Props) {
                e.stopPropagation();
                deleteConversation(conv.id);
              }}
-              className="p-1.5 mr-1 rounded opacity-0 group-hover:opacity-100 transition-opacity cursor-pointer"
+              disabled={isStreaming}
+              className="p-1.5 mr-1 rounded opacity-0 group-hover:opacity-100 transition-opacity cursor-pointer disabled:cursor-not-allowed disabled:opacity-30"
              style={{ color: 'var(--color-text-tertiary)' }}
-              onMouseEnter={(e) => (e.currentTarget.style.color = 'var(--color-error)')}
+              onMouseEnter={(e) => {
+                if (!isStreaming) e.currentTarget.style.color = 'var(--color-error)';
+              }}
              onMouseLeave={(e) => (e.currentTarget.style.color = 'var(--color-text-tertiary)')}
-              title="Delete conversation"
+              title={
+                isStreaming
+                  ? 'Stop generating before deleting this conversation'
+                  : 'Delete conversation'
+              }
            >
              <Trash2 size={14} />
            </button>
@@ -1,4 +1,4 @@
-import { afterEach, beforeEach, describe, expect, it } from 'vitest';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';

 // Regression for #266: the frontend must send the local API key as a Bearer
 // token on /v1 + /api requests, or `jarvis serve` with a key configured 401s
@@ -6,6 +6,7 @@ import { afterEach, beforeEach, describe, expect, it } from 'vitest';
 // authHeaders) that source the key and build the header.

 const SETTINGS_KEY = 'openjarvis-settings';
+const fetchMock = vi.fn<typeof fetch>();

 // Minimal in-memory localStorage stub so the helpers can run under node
 // (no jsdom dependency).
@@ -26,11 +27,16 @@ class MemoryStorage {
 }

 beforeEach(() => {
+  vi.resetModules();
+  vi.stubEnv('VITE_SUPABASE_ANON_KEY', 'test-anon-key');
+  fetchMock.mockReset();
+  globalThis.fetch = fetchMock;
  (globalThis as unknown as { localStorage: MemoryStorage }).localStorage =
    new MemoryStorage();
 });

 afterEach(() => {
+  vi.unstubAllEnvs();
  (globalThis as unknown as { localStorage?: MemoryStorage }).localStorage =
    undefined;
 });
@@ -83,3 +89,50 @@ describe('authHeaders', () => {
    });
  });
 });
+
+describe('tool credentials', () => {
+  it('reads credential status from the local server', async () => {
+    fetchMock.mockResolvedValue(
+      new Response(JSON.stringify({ TAVILY_API_KEY: true }), {
+        status: 200,
+        headers: { 'Content-Type': 'application/json' },
+      }),
+    );
+    const { fetchToolCredentialStatus } = await freshApi();
+
+    await expect(fetchToolCredentialStatus('web_search')).resolves.toEqual({
+      TAVILY_API_KEY: true,
+    });
+    expect(fetchMock).toHaveBeenCalledWith(
+      '/v1/tools/web_search/credentials/status',
+      { headers: {} },
+    );
+  });
+
+  it('saves a tool credential through the local server', async () => {
+    fetchMock.mockResolvedValue(new Response('{}', { status: 200 }));
+    const { saveToolCredentials } = await freshApi();
+
+    await saveToolCredentials('web_search', {
+      TAVILY_API_KEY: 'tvly-test',
+    });
+
+    expect(fetchMock).toHaveBeenCalledWith('/v1/tools/web_search/credentials', {
+      method: 'POST',
+      headers: { 'Content-Type': 'application/json' },
+      body: JSON.stringify({ TAVILY_API_KEY: 'tvly-test' }),
+    });
+  });
+
+  it('deletes a tool credential through the local server', async () => {
+    fetchMock.mockResolvedValue(new Response('{}', { status: 200 }));
+    const { deleteToolCredential } = await freshApi();
+
+    await deleteToolCredential('web_search', 'TAVILY_API_KEY');
+
+    expect(fetchMock).toHaveBeenCalledWith(
+      '/v1/tools/web_search/credentials/TAVILY_API_KEY',
+      { method: 'DELETE', headers: {} },
+    );
+  });
+});
@@ -1,12 +1,10 @@
 import type { ModelInfo, SavingsData, ServerInfo } from '../types';
+import { SUPABASE_ANON_KEY, SUPABASE_URL } from './supabase';

 // ---------------------------------------------------------------------------
-// Supabase config — safe to embed (RLS protects writes)
+// Supabase config
 // ---------------------------------------------------------------------------

-const SUPABASE_URL = import.meta.env.VITE_SUPABASE_URL || 'https://mtbtgpwzrbostweaanpr.supabase.co';
-const SUPABASE_ANON_KEY = import.meta.env.VITE_SUPABASE_ANON_KEY || 'eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJpc3MiOiJzdXBhYmFzZSIsInJlZiI6Im10YnRncHd6cmJvc3R3ZWFhbnByIiwicm9sZSI6ImFub24iLCJpYXQiOjE3NzMxODk0OTQsImV4cCI6MjA4ODc2NTQ5NH0._xMlqCfljtXpwPj54H-ghxfLFO-jiq4W2WhpU8vVL1c';
-
 declare global {
  interface Window {
    __TAURI_INTERNALS__?: unknown;
@@ -15,6 +13,31 @@ declare global {

 export const isTauri = () => typeof window !== 'undefined' && !!window.__TAURI_INTERNALS__;

+export type CloudKeyStatus = Record<string, boolean>;
+
+export async function getCloudKeyStatus(): Promise<CloudKeyStatus> {
+  if (!isTauri()) return {};
+  try {
+    const { invoke } = await import('@tauri-apps/api/core');
+    const rows = await invoke<Array<{ key: string; set: boolean }>>('get_cloud_key_status');
+    return Object.fromEntries(rows.map((row) => [row.key, row.set]));
+  } catch (e: any) {
+    throw new Error(e?.message ?? e ?? 'Failed to read cloud key status');
+  }
+}
+
+export async function saveCloudKey(keyName: string, keyValue: string): Promise<void> {
+  if (!isTauri()) {
+    throw new Error('Cloud API keys can be saved in the desktop app only.');
+  }
+  try {
+    const { invoke } = await import('@tauri-apps/api/core');
+    await invoke('save_cloud_key', { keyName, keyValue });
+  } catch (e: any) {
+    throw new Error(e?.message ?? e ?? 'Failed to save cloud key');
+  }
+}
+
 // Cached API base URL fetched from the Tauri backend at startup.
 // This avoids hardcoding the port — the Rust backend is the single
 // source of truth for JARVIS_PORT.
@@ -195,9 +218,9 @@ export async function deleteModel(modelName: string): Promise<void> {

 const _CLOUD_PREFIXES = ['gpt-', 'o1-', 'o3-', 'o4-', 'claude-', 'gemini-', 'openrouter/'];

-export async function preloadModel(modelName: string): Promise<void> {
+export async function preloadModel(modelName: string, owner?: string): Promise<void> {
  // Cloud models don't need Ollama preloading
-  if (_CLOUD_PREFIXES.some(p => modelName.startsWith(p))) {
+  if (owner === 'litellm' || _CLOUD_PREFIXES.some(p => modelName.startsWith(p))) {
    return;
  }
  // Trigger Ollama to load the model into memory (empty prompt, no generation).
@@ -317,8 +340,9 @@ export async function transcribeAudio(audioBlob: Blob, filename = 'recording.web
        audioData: Array.from(new Uint8Array(buffer)),
        filename,
      });
-    } catch {
-      // Fall through to fetch
+    } catch (err) {
+      const msg = err instanceof Error ? err.message : String(err);
+      throw new Error(msg || 'Transcription failed');
    }
  }
  const formData = new FormData();
@@ -327,7 +351,16 @@ export async function transcribeAudio(audioBlob: Blob, filename = 'recording.web
    method: 'POST',
    body: formData,
  });
-  if (!res.ok) throw new Error(`Transcription failed: ${res.status}`);
+  if (!res.ok) {
+    let detail = "";
+    try {
+      const body = await res.json();
+      detail = typeof body.detail === 'string' ? body.detail : "";
+    } catch {
+      // Keep the status-only message below when the body is not JSON.
+    }
+    throw new Error(detail || `Transcription failed: ${res.status}`);
+  }
  return res.json();
 }

@@ -852,6 +885,25 @@ export async function saveToolCredentials(
  if (!res.ok) throw new Error(`Failed: ${res.status}`);
 }

+export async function fetchToolCredentialStatus(
+  toolName: string,
+): Promise<Record<string, boolean>> {
+  const res = await apiFetch(`/v1/tools/${toolName}/credentials/status`);
+  if (!res.ok) throw new Error(`Failed: ${res.status}`);
+  return await res.json();
+}
+
+export async function deleteToolCredential(
+  toolName: string,
+  keyName: string,
+): Promise<void> {
+  const res = await apiFetch(
+    `/v1/tools/${encodeURIComponent(toolName)}/credentials/${encodeURIComponent(keyName)}`,
+    { method: 'DELETE' },
+  );
+  if (!res.ok) throw new Error(`Failed: ${res.status}`);
+}
+
 export interface AgentTraceDetail {
  id: string;
  agent: string;
@@ -1,5 +1,5 @@
 import { getBase } from './api';
-import type { ConnectorInfo, SyncStatus, ConnectRequest } from '../types/connectors';
+import type { ConnectorInfo, SyncStatus, ConnectRequest, ConnectResponse } from '../types/connectors';

 // ---------------------------------------------------------------------------
 // Connectors API
@@ -18,16 +18,47 @@ export async function getConnector(id: string): Promise<ConnectorInfo> {
  return res.json();
 }

-export async function connectSource(id: string, req: ConnectRequest): Promise<ConnectorInfo> {
+export async function connectSource(id: string, req: ConnectRequest): Promise<ConnectResponse> {
  const res = await fetch(`${getBase()}/v1/connectors/${encodeURIComponent(id)}/connect`, {
    method: 'POST',
    headers: { 'Content-Type': 'application/json' },
    body: JSON.stringify(req),
  });
-  if (!res.ok) throw new Error(`Failed to connect ${id}: ${res.status}`);
+  if (!res.ok) {
+    // Surface the backend's actionable detail (e.g. malformed Client ID /
+    // Secret) instead of a bare status code so the UI can render it.
+    const err = await res.json().catch(() => ({ detail: res.statusText }));
+    throw new Error(err.detail || `Failed to connect ${id}: ${res.status}`);
+  }
  return res.json();
 }

+/** Open the server-side OAuth consent flow in a popup and resolve once the
+ *  connector reports connected (or reject on timeout). Reused for any OAuth
+ *  connector whose /connect returned `oauth_required` (issue #512). */
+export function startServerOAuth(id: string, oauthStartPath?: string): Promise<void> {
+  const path = oauthStartPath || `/v1/connectors/${encodeURIComponent(id)}/oauth/start`;
+  window.open(`${getBase()}${path}`, '_blank', 'width=600,height=700');
+  return new Promise((resolve, reject) => {
+    const interval = setInterval(async () => {
+      try {
+        const info = await getConnector(id);
+        if (info.connected) {
+          clearInterval(interval);
+          clearTimeout(timer);
+          resolve();
+        }
+      } catch {
+        // ignore transient polling errors
+      }
+    }, 2000);
+    const timer = setTimeout(() => {
+      clearInterval(interval);
+      reject(new Error('Authorization timed out — please try again.'));
+    }, 180000);
+  });
+}
+
 export async function disconnectSource(id: string): Promise<void> {
  const res = await fetch(`${getBase()}/v1/connectors/${encodeURIComponent(id)}/disconnect`, {
    method: 'POST',
@@ -0,0 +1,19 @@
+import { describe, expect, it } from 'vitest';
+
+import { isEmbedOnlyModel } from './model-capabilities';
+
+describe('isEmbedOnlyModel', () => {
+  it.each([
+    'nomic-embed-text',
+    'mxbai-embed-large',
+    'text-embedding-3-small',
+    'all-minilm:latest',
+    'hf.co/BAAI/bge-m3:latest',
+  ])('classifies %s as embedding-only', (modelId) => {
+    expect(isEmbedOnlyModel(modelId)).toBe(true);
+  });
+
+  it.each(['qwen3.5:4b', 'codegemma:7b'])('keeps %s available for chat', (modelId) => {
+    expect(isEmbedOnlyModel(modelId)).toBe(false);
+  });
+});
@@ -0,0 +1,22 @@
+const EMBEDDING_MODEL_PREFIXES = [
+  'all-minilm',
+  'bge-',
+  'bge_',
+  'e5-',
+  'e5_',
+  'gte-',
+  'gte_',
+  'jina-embeddings',
+  'nomic-bert',
+  'sentence-transformers',
+];
+
+export function isEmbedOnlyModel(modelId: string): boolean {
+  const name = (modelId || '').trim().toLowerCase();
+  const leaf = name.slice(name.lastIndexOf('/') + 1).split(':')[0];
+  return (
+    leaf.includes('embed') ||
+    leaf.includes('minilm') ||
+    EMBEDDING_MODEL_PREFIXES.some((prefix) => leaf.startsWith(prefix))
+  );
+}
@@ -1,5 +1,5 @@
 import type { ResearchEvent, SSEEvent } from '../types';
-import { getBase } from './api';
+import { getBase, authHeaders } from './api';

 export interface ChatRequest {
  model: string;
@@ -16,7 +16,7 @@ export async function* streamChat(
  const base = getBase();
  const response = await fetch(`${base}/v1/chat/completions`, {
    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
+    headers: authHeaders({ 'Content-Type': 'application/json' }),
    body: JSON.stringify(request),
    signal,
  });
@@ -60,6 +60,7 @@ export async function* streamChat(

 export async function* streamResearch(
  query: string,
+  model?: string,
  signal?: AbortSignal,
 ): AsyncGenerator<ResearchEvent> {
  // /api/research is mounted at the server root — strip any trailing /v1
@@ -67,8 +68,8 @@ export async function* streamResearch(
  const base = getBase().replace(/\/v1\/?$/, '');
  const response = await fetch(`${base}/api/research`, {
    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({ query }),
+    headers: authHeaders({ 'Content-Type': 'application/json' }),
+    body: JSON.stringify({ query, ...(model ? { model } : {}) }),
    signal,
  });

@@ -0,0 +1,64 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import type { ModelInfo } from '../types';
+
+class MemoryStorage {
+  private store = new Map<string, string>();
+
+  getItem(key: string): string | null {
+    return this.store.get(key) ?? null;
+  }
+
+  setItem(key: string, value: string): void {
+    this.store.set(key, String(value));
+  }
+}
+
+const model = (id: string): ModelInfo => ({
+  id,
+  object: 'model',
+  created: 0,
+  owned_by: 'openjarvis',
+});
+
+beforeEach(() => {
+  vi.resetModules();
+  (globalThis as unknown as { localStorage: MemoryStorage }).localStorage =
+    new MemoryStorage();
+});
+
+afterEach(() => {
+  (globalThis as unknown as { localStorage?: MemoryStorage }).localStorage =
+    undefined;
+});
+
+describe('setModels', () => {
+  it('does not select an embedding-only model', async () => {
+    const { useAppStore } = await import('./store');
+
+    useAppStore.getState().setModels([model('nomic-embed-text')]);
+
+    expect(useAppStore.getState().selectedModel).toBe('');
+  });
+
+  it('clears a missing selection when no chat fallback exists', async () => {
+    const { useAppStore } = await import('./store');
+    useAppStore.getState().setSelectedModel('deleted-chat-model');
+
+    useAppStore.getState().setModels([model('nomic-embed-text')]);
+
+    expect(useAppStore.getState().selectedModel).toBe('');
+  });
+
+  it('replaces an embedding selection with an available chat model', async () => {
+    const { useAppStore } = await import('./store');
+    useAppStore.getState().setSelectedModel('all-minilm:latest');
+
+    useAppStore.getState().setModels([
+      model('all-minilm:latest'),
+      model('qwen3.5:4b'),
+    ]);
+
+    expect(useAppStore.getState().selectedModel).toBe('qwen3.5:4b');
+  });
+});
@@ -0,0 +1,82 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+class MemoryStorage {
+  private store = new Map<string, string>();
+
+  getItem(key: string): string | null {
+    return this.store.get(key) ?? null;
+  }
+
+  setItem(key: string, value: string): void {
+    this.store.set(key, String(value));
+  }
+
+  removeItem(key: string): void {
+    this.store.delete(key);
+  }
+}
+
+beforeEach(() => {
+  vi.resetModules();
+  (globalThis as unknown as { localStorage: MemoryStorage }).localStorage =
+    new MemoryStorage();
+});
+
+afterEach(() => {
+  (globalThis as unknown as { localStorage?: MemoryStorage }).localStorage =
+    undefined;
+});
+
+async function freshStore() {
+  return (await import('./store')).useAppStore;
+}
+
+describe('conversation stream ownership', () => {
+  it('persists background stream updates without replacing the active messages', async () => {
+    const store = await freshStore();
+    const sourceId = store.getState().createConversation('test-model');
+    store.getState().addMessage(sourceId, {
+      id: 'assistant',
+      role: 'assistant',
+      content: '',
+      timestamp: 1,
+    });
+
+    const activeId = store.getState().createConversation('test-model');
+    store.getState().setStreamState({
+      conversationId: sourceId,
+      isStreaming: true,
+      content: 'streamed response',
+    });
+    store.getState().updateLastAssistant(sourceId, 'streamed response');
+
+    expect(store.getState().activeId).toBe(activeId);
+    expect(store.getState().messages).toEqual([]);
+
+    store.getState().selectConversation(sourceId);
+    expect(store.getState().messages).toHaveLength(1);
+    expect(store.getState().messages[0].content).toBe('streamed response');
+  });
+
+  it('keeps the stream-owning conversation until generation stops', async () => {
+    const store = await freshStore();
+    const sourceId = store.getState().createConversation('test-model');
+    const activeId = store.getState().createConversation('test-model');
+    store.getState().setStreamState({
+      conversationId: sourceId,
+      isStreaming: true,
+    });
+
+    store.getState().deleteConversation(sourceId);
+    expect(
+      store.getState().conversations.map((conversation) => conversation.id),
+    ).toContain(sourceId);
+    expect(store.getState().activeId).toBe(activeId);
+
+    store.getState().resetStream();
+    store.getState().deleteConversation(sourceId);
+    expect(
+      store.getState().conversations.map((conversation) => conversation.id),
+    ).not.toContain(sourceId);
+  });
+});
@@ -15,6 +15,7 @@ import type {
  TokenUsage,
 } from '../types';
 import type { ManagedAgent } from './api';
+import { isEmbedOnlyModel } from './model-capabilities';

 export interface CachedConnector {
  connector_id: string;
@@ -110,6 +111,7 @@ function saveSettings(settings: Settings): void {
 // ── Store ─────────────────────────────────────────────────────────────

 const INITIAL_STREAM: StreamState = {
+  conversationId: null,
  isStreaming: false,
  phase: '',
  elapsedMs: 0,
@@ -351,6 +353,9 @@ export const useAppStore = create<AppState>((set, get) => {
    },

    deleteConversation: (id: string) => {
+      const streamState = get().streamState;
+      if (streamState.isStreaming && streamState.conversationId === id) return;
+
      const store = loadConversations();
      delete store.conversations[id];
      if (store.activeId === id) {
@@ -393,12 +398,14 @@ export const useAppStore = create<AppState>((set, get) => {
          (message.content.length > 50 ? '...' : '');
      }
      saveConversations(store);
-      set({
-        messages: [...conv.messages],
-        conversations: Object.values(store.conversations).sort(
-          (a, b) => b.updatedAt - a.updatedAt,
-        ),
-      });
+      const conversations = Object.values(store.conversations).sort(
+        (a, b) => b.updatedAt - a.updatedAt,
+      );
+      if (get().activeId === conversationId) {
+        set({ messages: [...conv.messages], conversations });
+      } else {
+        set({ conversations });
+      }
    },

    updateLastAssistant: (
@@ -425,7 +432,9 @@ export const useAppStore = create<AppState>((set, get) => {
        if (researchSources) lastMsg.researchSources = researchSources;
        conv.updatedAt = Date.now();
        saveConversations(store);
-        set({ messages: [...conv.messages] });
+        if (get().activeId === conversationId) {
+          set({ messages: [...conv.messages] });
+        }
      }
    },

@@ -444,11 +453,36 @@ export const useAppStore = create<AppState>((set, get) => {
    // ── Models & server ────────────────────────────────────────────

    setModels: (models: ModelInfo[]) =>
-      set((state) =>
-        !state.selectedModel && models.length > 0
-          ? { models, selectedModel: models[0].id }
-          : { models },
-      ),
+      set((state) => {
+        // Ollama returns embed-only models (e.g. nomic-embed-text) in the
+        // same list as chat models. Auto-picking models[0] selected the
+        // embedder and every chat failed with HTTP 400 "does not support
+        // chat". Prefer a real chat model for selection / fallback.
+        const chatModels = models.filter((m) => !isEmbedOnlyModel(m.id));
+        const preferred =
+          (state.settings.defaultModel &&
+            chatModels.some((m) => m.id === state.settings.defaultModel) &&
+            state.settings.defaultModel) ||
+          chatModels[0]?.id ||
+          models.find((m) => !isEmbedOnlyModel(m.id))?.id ||
+          '';
+
+        const currentIsBad =
+          !!state.selectedModel && isEmbedOnlyModel(state.selectedModel);
+        const currentMissing =
+          !!state.selectedModel &&
+          !models.some((m) => m.id === state.selectedModel);
+
+        if (!state.selectedModel || currentIsBad || currentMissing) {
+          // Prefer a real chat model. If none exist, clear a bad/missing
+          // selection rather than keeping an embed-only id that 400s on chat.
+          return {
+            models,
+            selectedModel: preferred,
+          };
+        }
+        return { models };
+      }),
    setModelsLoading: (loading: boolean) => set({ modelsLoading: loading }),
    setSelectedModel: (model: string) => set({ selectedModel: model }),
    setServerInfo: (info: ServerInfo | null) => set({ serverInfo: info }),
@@ -0,0 +1,11 @@
+export const SUPABASE_URL =
+  import.meta.env.VITE_SUPABASE_URL || 'https://mtbtgpwzrbostweaanpr.supabase.co';
+
+// The Supabase anon key is optional at build time. When it is unset the public
+// savings leaderboard is disabled rather than failing the build — this keeps
+// the `openjarvis` package and desktop app buildable without coupling
+// publishability to a leaderboard credential. Set VITE_SUPABASE_ANON_KEY at
+// build time (from a repo secret) to enable the leaderboard.
+export const SUPABASE_ANON_KEY = import.meta.env.VITE_SUPABASE_ANON_KEY ?? '';
+
+export const LEADERBOARD_ENABLED = SUPABASE_ANON_KEY.length > 0;
@@ -23,7 +23,6 @@ import {
  fetchAgentTrace,
  fetchManagedAgent,
  fetchAvailableTools,
-  saveToolCredentials,
  fetchModels,
  updateManagedAgent,
  fetchRecommendedModel,
@@ -575,7 +574,7 @@ function ToolsPicker({
          </div>
          {/* Live description strip */}
          <div
-            className="flex items-center gap-2 px-2.5 py-1.5"
+            className="flex items-start gap-2 px-2.5 py-1.5"
            style={{
              borderTop: '1px solid var(--color-border)',
              background: 'var(--color-bg)',
@@ -609,10 +608,11 @@ function ToolsPicker({
              </span>
            )}
            <span
-              className="truncate"
+              className="min-w-0 whitespace-normal break-words"
              style={{
                flex: 1,
                color: 'var(--color-text-tertiary)',
+                lineHeight: 1.4,
              }}
            >
              {hovered ? `— ${hint}` : hint}
@@ -3740,8 +3740,8 @@ export function AgentsPage() {
              const paramsB = paramMatch ? parseFloat(paramMatch[1]) : 9;
              const flops = 2 * paramsB * 1e9 * (inTok + outTok);
              const providers = [
-                { label: 'GPT-5.3', inPer1M: 2.0, outPer1M: 10.0 },
-                { label: 'Claude Opus 4.6', inPer1M: 5.0, outPer1M: 25.0 },
+                { label: 'GPT-5.6 Sol', inPer1M: 5.0, outPer1M: 30.0 },
+                { label: 'Claude Fable 5', inPer1M: 10.0, outPer1M: 50.0 },
                { label: 'Gemini 3.1 Pro', inPer1M: 2.0, outPer1M: 12.0 },
              ];
              const energyWh = (inTok + outTok) / 1000 * 0.4;
@@ -24,7 +24,7 @@ import {
 import type { LucideIcon } from 'lucide-react';
 import { SOURCE_CATALOG } from '../types/connectors';
 import type { ConnectRequest } from '../types/connectors';
-import { listConnectors, connectSource, disconnectSource, getSyncStatus, triggerSync } from '../lib/connectors-api';
+import { listConnectors, connectSource, disconnectSource, getSyncStatus, triggerSync, startServerOAuth } from '../lib/connectors-api';
 import type { SyncStatus } from '../types/connectors';

 // ---------------------------------------------------------------------------
@@ -673,7 +673,19 @@ function DataSourcesSection() {
    setConnectStage('Connecting...');
    setConnectError('');
    try {
-      await connectSource(id, req);
+      const resp = await connectSource(id, req);
+
+      // OAuth connectors (Google Drive/Calendar/Contacts/Gmail/Tasks): pasting
+      // a Client ID / Secret only registers the app credentials. The backend
+      // returns `oauth_required` with the path to the in-process consent flow,
+      // which is the only path that actually mints an access token. Open it now
+      // and wait for the callback to flip the connector to connected. Without
+      // this the connector would stay "pending" forever — the exact #512 bug.
+      if (resp.status === 'oauth_required') {
+        setConnectStage('Opening Google sign-in...');
+        await startServerOAuth(id, resp.oauth_start);
+      }
+
      setConnectStage('Connected! Starting sync...');

      // Wait for connector to show as connected
@@ -417,7 +417,7 @@ function SelfHostedView() {
          <p className="text-xs" style={{ color: 'var(--color-text-secondary)' }}>
            Launch the API server to get the full UI in your browser:
          </p>
-          <CodeBlock code={"git clone https://github.com/open-jarvis/OpenJarvis.git\ncd OpenJarvis\nuv sync --extra server\njarvis serve --port 8000"} />
+          <CodeBlock code={"git clone https://github.com/open-jarvis/OpenJarvis.git\ncd OpenJarvis\nuv sync --extra desktop\njarvis serve --port 8000"} />
          <p className="text-xs" style={{ color: 'var(--color-text-tertiary)' }}>
            The chat, dashboard, energy profiling, and cost comparison all run
            locally on your machine.
@@ -19,9 +19,24 @@ import {
  RefreshCw,
 } from 'lucide-react';
 import { useAppStore, type ThemeMode } from '../lib/store';
-import { checkHealth, fetchSpeechHealth, getMemoryStats, getInferenceSource, setInferenceSource, type InferenceSource } from '../lib/api';
+import {
+  checkHealth,
+  fetchSpeechHealth,
+  getMemoryStats,
+  getInferenceSource,
+  setInferenceSource,
+  getCloudKeyStatus,
+  saveCloudKey,
+  fetchToolCredentialStatus,
+  saveToolCredentials,
+  deleteToolCredential,
+  isTauri,
+  type InferenceSource,
+} from '../lib/api';
 import { isAutoUpdateDisabled, setAutoUpdateDisabled } from '../components/Desktop/UpdateChecker';

+const CLOUD_KEY_STATUS_CHANGED = 'openjarvis-cloud-key-status-changed';
+
 function OllamaModelList() {
  const [models, setModels] = useState<Array<{ name: string; size: number }>>([]);
  useEffect(() => {
@@ -44,32 +59,135 @@ function OllamaModelList() {
  );
 }

-function ApiKeyInput({ storageKey, placeholder }: { storageKey: string; placeholder: string }) {
-  const [value, setValue] = useState(() => {
-    try { return localStorage.getItem(storageKey) || ''; } catch { return ''; }
-  });
+function ApiKeyInput({
+  keyName,
+  placeholder,
+  toolName,
+}: {
+  keyName: string;
+  placeholder: string;
+  toolName?: string;
+}) {
+  const [value, setValue] = useState('');
  const [saved, setSaved] = useState(false);
-  const save = (v: string) => {
-    setValue(v);
-    try { if (v) localStorage.setItem(storageKey, v); else localStorage.removeItem(storageKey); } catch {}
-    setSaved(true);
-    setTimeout(() => setSaved(false), 2000);
+  const [hasKey, setHasKey] = useState(false);
+  const [error, setError] = useState('');
+  const desktopKeyStorage = isTauri();
+  const serverToolStorage = !desktopKeyStorage && !!toolName;
+  const canManage = desktopKeyStorage || serverToolStorage;
+
+  const refresh = useCallback(async () => {
+    if (!canManage) {
+      setHasKey(false);
+      return;
+    }
+    try {
+      const status = desktopKeyStorage
+        ? await getCloudKeyStatus()
+        : await fetchToolCredentialStatus(toolName!);
+      setHasKey(!!status[keyName]);
+    } catch {
+      setHasKey(false);
+    }
+  }, [canManage, desktopKeyStorage, keyName, toolName]);
+
+  useEffect(() => {
+    void refresh();
+    window.addEventListener(CLOUD_KEY_STATUS_CHANGED, refresh);
+    return () => window.removeEventListener(CLOUD_KEY_STATUS_CHANGED, refresh);
+  }, [refresh]);
+
+  const save = async (v: string) => {
+    const next = v.trim();
+    if (!next) return;
+    setError('');
+    try {
+      if (desktopKeyStorage) {
+        await saveCloudKey(keyName, next);
+      } else if (toolName) {
+        await saveToolCredentials(toolName, { [keyName]: next });
+      } else {
+        return;
+      }
+      setValue('');
+      setHasKey(true);
+      setSaved(true);
+      window.dispatchEvent(new Event(CLOUD_KEY_STATUS_CHANGED));
+      setTimeout(() => setSaved(false), 2000);
+    } catch (e: any) {
+      setError(e?.message || 'Failed to save API key');
+    }
  };
+
+  const remove = async () => {
+    setError('');
+    try {
+      if (desktopKeyStorage) {
+        await saveCloudKey(keyName, '');
+      } else if (toolName) {
+        await deleteToolCredential(toolName, keyName);
+      } else {
+        return;
+      }
+      setValue('');
+      setHasKey(false);
+      setSaved(true);
+      window.dispatchEvent(new Event(CLOUD_KEY_STATUS_CHANGED));
+      setTimeout(() => setSaved(false), 2000);
+    } catch (e: any) {
+      setError(e?.message || 'Failed to remove API key');
+    }
+  };
+
  return (
    <div className="flex items-center gap-2">
-      <input type="password" value={value} onChange={e => save(e.target.value)} placeholder={placeholder}
+      <input
+        type="password"
+        value={value}
+        onChange={e => setValue(e.target.value)}
+        onBlur={() => { if (value.trim()) void save(value); }}
+        placeholder={hasKey ? (desktopKeyStorage ? 'Saved in secure storage' : 'Saved by local server') : placeholder}
+        disabled={!canManage}
        className="w-48 px-2 py-1 rounded text-xs"
        style={{ background: 'var(--color-bg)', border: '1px solid var(--color-border)', color: 'var(--color-text)' }} />
+      {hasKey && (
+        <button
+          onClick={() => void remove()}
+          className="px-2 py-1 rounded text-[10px] cursor-pointer"
+          style={{ color: 'var(--color-error)', border: '1px solid var(--color-error)' }}
+        >
+          Remove
+        </button>
+      )}
      {saved && <span className="text-[10px]" style={{ color: 'var(--color-success)' }}>Saved</span>}
+      {error && <span className="text-[10px]" style={{ color: 'var(--color-error)' }}>{error}</span>}
    </div>
  );
 }

-function CloudProviderStatus({ label, storageKey }: { label: string; storageKey: string }) {
+function CloudProviderStatus({ label, keyName }: { label: string; keyName: string }) {
  const [hasKey, setHasKey] = useState(false);
+  const desktopKeyStorage = isTauri();
+
+  const refresh = useCallback(async () => {
+    if (!desktopKeyStorage) {
+      setHasKey(false);
+      return;
+    }
+    try {
+      const status = await getCloudKeyStatus();
+      setHasKey(!!status[keyName]);
+    } catch {
+      setHasKey(false);
+    }
+  }, [desktopKeyStorage, keyName]);
+
  useEffect(() => {
-    try { setHasKey(!!localStorage.getItem(storageKey)); } catch { setHasKey(false); }
-  }, [storageKey]);
+    void refresh();
+    window.addEventListener(CLOUD_KEY_STATUS_CHANGED, refresh);
+    return () => window.removeEventListener(CLOUD_KEY_STATUS_CHANGED, refresh);
+  }, [refresh]);
+
  return (
    <span className="flex items-center gap-1 text-xs" style={{ color: 'var(--color-text-secondary)' }}>
      <span style={{
@@ -424,10 +542,10 @@ export function SettingsPage() {
            </div>
            <SettingRow label="Cloud providers" description="Green dot means API key is configured">
              <div className="flex flex-wrap gap-3">
-                <CloudProviderStatus label="OpenAI" storageKey="openjarvis-openai-key" />
-                <CloudProviderStatus label="Anthropic" storageKey="openjarvis-anthropic-key" />
-                <CloudProviderStatus label="Google" storageKey="openjarvis-gemini-key" />
-                <CloudProviderStatus label="OpenRouter" storageKey="openjarvis-openrouter-key" />
+                <CloudProviderStatus label="OpenAI" keyName="OPENAI_API_KEY" />
+                <CloudProviderStatus label="Anthropic" keyName="ANTHROPIC_API_KEY" />
+                <CloudProviderStatus label="Google" keyName="GEMINI_API_KEY" />
+                <CloudProviderStatus label="OpenRouter" keyName="OPENROUTER_API_KEY" />
              </div>
            </SettingRow>
          </Section>
@@ -435,23 +553,23 @@ export function SettingsPage() {
          {/* API Keys */}
          <Section title="API Keys">
            <SettingRow label="OpenAI" description="GPT-4, GPT-3.5, etc.">
-              <ApiKeyInput storageKey="openjarvis-openai-key" placeholder="sk-..." />
+              <ApiKeyInput keyName="OPENAI_API_KEY" placeholder="sk-..." />
            </SettingRow>
            <SettingRow label="Anthropic" description="Claude models">
-              <ApiKeyInput storageKey="openjarvis-anthropic-key" placeholder="sk-ant-..." />
+              <ApiKeyInput keyName="ANTHROPIC_API_KEY" placeholder="sk-ant-..." />
            </SettingRow>
            <SettingRow label="Google" description="Gemini models">
-              <ApiKeyInput storageKey="openjarvis-gemini-key" placeholder="AI..." />
+              <ApiKeyInput keyName="GEMINI_API_KEY" placeholder="AI..." />
            </SettingRow>
            <SettingRow label="OpenRouter" description="Multi-provider routing">
-              <ApiKeyInput storageKey="openjarvis-openrouter-key" placeholder="sk-or-..." />
+              <ApiKeyInput keyName="OPENROUTER_API_KEY" placeholder="sk-or-..." />
            </SettingRow>
          </Section>

          {/* Tools */}
          <Section title="Tools">
-            <SettingRow label="Web Search" description="SerpAPI or Tavily key for web search tool">
-              <ApiKeyInput storageKey="openjarvis-search-key" placeholder="API key..." />
+            <SettingRow label="Web Search" description="Tavily key for web search tool">
+              <ApiKeyInput keyName="TAVILY_API_KEY" placeholder="tvly-..." toolName="web_search" />
            </SettingRow>
          </Section>

@@ -714,7 +832,7 @@ export function SettingsPage() {
              </p>
              <div className="flex gap-3 mt-3 text-xs">
                <a
-                  href="https://scalingintelligence.stanford.edu/blogs/openjarvis/"
+                  href="https://openjarvis.stanford.edu/"
                  target="_blank"
                  rel="noopener noreferrer"
                  style={{ color: 'var(--color-accent)' }}
@@ -55,6 +55,19 @@ export interface ConnectRequest {
  password?: string;
 }

+/** Response from POST /v1/connectors/{id}/connect.
+ *  For OAuth connectors, pasting a Client ID / Secret pair only registers the
+ *  app credentials; the backend returns `status: "oauth_required"` plus an
+ *  `oauth_start` path the UI must open to run the browser consent flow that
+ *  actually mints an access token (see issue #512). */
+export interface ConnectResponse {
+  connector_id: string;
+  connected: boolean;
+  status: "connected" | "pending" | "oauth_required" | "disconnected";
+  oauth_start?: string;
+  sync_status?: string | null;
+}
+
 export type WizardStep = "pick" | "connect" | "ingest" | "ready";

 // Backward-compatible alias
@@ -257,12 +270,12 @@ export const SOURCE_CATALOG: ConnectorMeta[] = [
        urlLabel: 'Enable Drive API',
      },
      {
-        label: 'Create OAuth credentials: go to Credentials (link below) → click "+ Create Credentials" → choose "OAuth client ID" → Application type: "Desktop app" → click "Create"',
+        label: 'Create OAuth credentials: go to Credentials (link below) → click "+ Create Credentials" → choose "OAuth client ID" → Application type: "Web application". Under "Authorized redirect URIs" add this server\'s callback (e.g. http://localhost:1313/v1/connectors/gdrive/oauth/callback — match the host/port your OpenJarvis server is bound to) → click "Create".',
        url: 'https://console.cloud.google.com/apis/credentials',
        urlLabel: 'Open Credentials',
      },
      {
-        label: 'A dialog will show your Client ID and Client Secret. Copy both and paste them below. (If you miss it, click the download icon next to your OAuth client to see them again)',
+        label: 'A dialog will show your Client ID and Client Secret. Copy both and paste them below, then click Connect — a Google sign-in window opens to finish authorization. (If you miss the dialog, click the download icon next to your OAuth client to see them again.)',
      },
    ],
    inputFields: [
@@ -147,6 +147,7 @@ export interface ConversationStore {
 // --- Stream State ---

 export interface StreamState {
+  conversationId: string | null;
  isStreaming: boolean;
  phase: string;
  elapsedMs: number;
@@ -1,7 +1,9 @@
 /// <reference types="vite/client" />

 interface ImportMetaEnv {
-  readonly VITE_API_URL: string;
+  readonly VITE_API_URL?: string;
+  readonly VITE_SUPABASE_URL?: string;
+  readonly VITE_SUPABASE_ANON_KEY?: string;
 }

 interface ImportMeta {
@@ -4,6 +4,9 @@ import react from '@vitejs/plugin-react';
 import tailwindcss from '@tailwindcss/vite';
 import { VitePWA } from 'vite-plugin-pwa';

+// VITE_SUPABASE_ANON_KEY is intentionally NOT required here: a missing key
+// disables the savings leaderboard at runtime (see src/lib/supabase.ts) rather
+// than failing the build, so the package/app stays publishable without it.
 export default defineConfig({
  resolve: {
    alias: {
@@ -51,7 +54,15 @@ export default defineConfig({
  server: {
    port: 5173,
    proxy: {
-      '/v1': process.env.VITE_API_URL || 'http://localhost:8000',
+      // ws: true is required for the /v1/agents/events WebSocket. Without it
+      // Vite proxies the HTTP request but not the upgrade, so the socket never
+      // opens — no error, no close event, just silence — and every live agent
+      // view sits empty in dev while working in a production build.
+      '/v1': {
+        target: process.env.VITE_API_URL || 'http://localhost:8000',
+        changeOrigin: true,
+        ws: true,
+      },
      '/health': process.env.VITE_API_URL || 'http://localhost:8000',
      '/api': process.env.VITE_API_URL || 'http://localhost:8000',
    },
@@ -127,6 +127,7 @@ markdown_extensions:
  - pymdownx.tilde

 extra_javascript:
+  - javascripts/leaderboard-config.js
  - javascripts/leaderboard.js
  - https://cdn.jsdelivr.net/npm/@docsearch/js@3
  - javascripts/docsearch-init.js
@@ -193,6 +194,9 @@ nav:
      - External MCP Servers: user-guide/mcp-external-servers.md
      - Scheduler: user-guide/scheduler.md
      - Telemetry: user-guide/telemetry.md
+      - Evaluations: user-guide/evaluations.md
+      - Benchmarks: user-guide/benchmarks.md
+      - System Access: user-guide/system-access.md
      - Security: user-guide/security.md
      - LLM-guided spec search: user-guide/llm-guided-spec-search.md
  - Leaderboard: leaderboard.md
@@ -1,10 +1,10 @@
 [build-system]
-requires = ["hatchling"]
+requires = ["hatchling", "hatch-vcs"]
 build-backend = "hatchling.build"

 [project]
 name = "OpenJarvis"
-version = "1.0.2"
+dynamic = ["version"]
 description = "OpenJarvis — modular AI assistant backend with composable intelligence primitives"
 readme = "README.md"
 # Upper bound: numpy 2.2.x (pinned transitively via datasets/pandas) ships no
@@ -48,6 +48,7 @@ dev = [
    "pytest>=8",
    "pytest-asyncio>=0.24",
    "pytest-cov>=5",
+    "pytest-xdist>=3",
    "respx>=0.22",
    "ruff>=0.4",
    "pre-commit>=3.0",
@@ -84,6 +85,13 @@ server = [
    "pydantic>=2.0",
    "python-multipart>=0.0.9",
 ]
+desktop = [
+    "fastapi>=0.110",
+    "uvicorn>=0.30",
+    "pydantic>=2.0",
+    "python-multipart>=0.0.9",
+    "faster-whisper>=1.0",
+]
 openhands = ["openhands-sdk>=1.0; python_version >= '3.12'"]
 gpu-metrics = ["pynvml>=12.0"]
 energy-amd = ["amdsmi>=6.1"]
@@ -152,6 +160,34 @@ Issues = "https://github.com/open-jarvis/OpenJarvis/issues"

 [project.scripts]
 jarvis = "openjarvis.cli:main"
+openjarvis-eval = "openjarvis.evals.cli:main"
+
+# Version is derived from git tags by hatch-vcs (see #526). For source/editable
+# checkouts this yields the true `git describe` version (e.g. 1.0.3.dev109+g<sha>)
+# rather than a stale static string. CI release builds override this with
+# SETUPTOOLS_SCM_PRETEND_VERSION so the published version equals the pushed tag.
+#
+# setuptools_scm cannot bump custom `.devN` tags (only `.dev0`), so the autotag
+# `vX.Y.Z.devN` tags are deliberately EXCLUDED from version derivation here; the
+# base is taken from the latest plain release tag (vX.Y.Z) and the dev distance
+# is computed from commit count since that release.
+[tool.hatch.version]
+source = "vcs"
+
+[tool.hatch.version.raw-options]
+tag_regex = '^v(?P<version>[0-9]+\.[0-9]+\.[0-9]+)$'
+git_describe_command = [
+    "git", "describe", "--dirty", "--tags", "--long",
+    "--match", "v[0-9]*", "--exclude", "*dev*", "--exclude", "*rc*", "--exclude", "desktop-*",
+]
+# Builds without a git checkout (e.g. the `COPY src/ src/` Docker stages, which
+# never include .git) can't run `git describe`. Without a fallback that would
+# hard-fail the build. Mirror the runtime sentinel in src/openjarvis/__init__.py.
+# Such builds can inject the real version via SETUPTOOLS_SCM_PRETEND_VERSION.
+fallback_version = "0.0.0+unknown"
+
+[tool.uv.sources]
+openjarvis-rust = { path = "rust/crates/openjarvis-python" }

 [tool.hatch.build.targets.wheel]
 packages = ["src/openjarvis"]
@@ -202,3 +238,12 @@ select = ["E", "F", "I", "W"]
 dev = [
    "maturin>=1.12.6",
 ]
+# openjarvis_rust is the native PyO3 extension, built from the local Rust
+# workspace. It lives in a uv dependency group (PEP 735) — not the published
+# `desktop` extra — so `uv sync --group desktop-native` builds it from source
+# for the desktop app, while `pip install openjarvis[desktop]` from PyPI does
+# NOT try to resolve openjarvis-rust from PyPI, where it isn't published
+# (dependency groups are excluded from wheel metadata). See #584 / #615.
+desktop-native = [
+    "openjarvis-rust",
+]
@@ -144,16 +144,21 @@ impl MemoryBackend for SQLiteMemory {
    ) -> Result<Vec<RetrievalResult>, OpenJarvisError> {
        let conn = self.conn.lock();

+        // Split on any non-alphanumeric character (not just whitespace) so
+        // internal punctuation — apostrophes in particular ("user's") — never
+        // reaches the FTS5 MATCH string. FTS5's query grammar treats an
+        // unescaped `'` as a string delimiter, so passing a raw token like
+        // `user's` through silently fails to parse and yields zero rows with
+        // no visible error. Splitting fully avoids needing to escape anything.
        let words: Vec<String> = query
-            .split_whitespace()
-            .map(|w| w.trim_matches(|c: char| "?.,!;:'\"()[]{}/ ".contains(c)).to_string())
+            .split(|c: char| !c.is_alphanumeric())
+            .map(|w| w.to_string())
            .filter(|w| !w.is_empty())
            .collect();
-        let fts_query = if words.len() == 1 {
-            words[0].clone()
-        } else {
-            words.join(" OR ")
-        };
+        if words.is_empty() {
+            return Ok(Vec::new());
+        }
+        let fts_query = words.join(" OR ");

        let mut stmt = conn
            .prepare(
@@ -320,6 +325,27 @@ mod tests {
        assert_eq!(mixed.len(), 2, "mixed-case query should find both documents");
    }

+    #[test]
+    fn test_sqlite_apostrophe_in_query() {
+        let mem = SQLiteMemory::in_memory().unwrap();
+        mem.store("The user's name is Trev.", "identity", None).unwrap();
+
+        // A query containing an internal apostrophe must not break FTS5's
+        // MATCH syntax (an unescaped `'` is a string delimiter in FTS5's
+        // query grammar), which previously caused this to silently return
+        // zero results instead of matching or erroring.
+        let multi_word = mem.retrieve("what is the user's name", 5).unwrap();
+        assert!(
+            !multi_word.is_empty(),
+            "query with an internal apostrophe should not silently return zero results"
+        );
+
+        // Bare single-word possessive: exercises the (former) single-word
+        // bypass path that skipped the OR-join entirely.
+        let bare = mem.retrieve("user's", 5).unwrap();
+        assert!(!bare.is_empty(), "single-word possessive query should still match");
+    }
+
    #[test]
    fn test_sqlite_scores_are_positive() {
        let mem = SQLiteMemory::in_memory().unwrap();
@@ -210,6 +210,14 @@ if ! command -v python3 >/dev/null 2>&1; then
 fi

 # ---- env ----
+# OpenJarvis keeps ALL of its state (install tree + runtime data, configs,
+# databases, caches, logs) under a single root so it never clutters $HOME
+# beyond one directory. Relocate it by exporting OPENJARVIS_HOME before
+# running the installer, e.g.:
+#     OPENJARVIS_HOME=~/apps/openjarvis curl ... | bash
+# The Python runtime honors the same override (and, when OPENJARVIS_HOME is
+# unset, $XDG_DATA_HOME/openjarvis if XDG_DATA_HOME is set). With nothing set
+# the root is ~/.openjarvis, so existing installs are untouched.
 OPENJARVIS_HOME="${OPENJARVIS_HOME:-$HOME/.openjarvis}"
 OPENJARVIS_REPO_URL="${OPENJARVIS_REPO_URL:-https://github.com/open-jarvis/OpenJarvis.git}"
 SRC_DIR="$OPENJARVIS_HOME/src"
@@ -148,7 +148,8 @@ fi

 # ── 7. Install Python dependencies ──────────────────────────────────
 info "Installing Python dependencies..."
-uv sync --extra server --quiet 2>/dev/null || uv sync --extra server
+uv sync --extra desktop --extra tools-search --quiet 2>/dev/null \
+  || uv sync --extra desktop --extra tools-search
 ok "Python dependencies installed"

 # ── 7b. Build Rust extension ──────────────────────────────────────
@@ -164,11 +165,17 @@ ok "Frontend dependencies installed"

 # ── 9. Start backend ────────────────────────────────────────────────
 info "Starting backend API server on port 8000..."
+if curl -sf http://localhost:8000/health &>/dev/null; then
+  fail "An OpenJarvis server is already running on port 8000. Stop it before re-running quickstart so updated environment variables are applied."
+fi
 uv run jarvis serve --port 8000 &>/dev/null &
-CLEANUP_PIDS+=($!)
+BACKEND_PID=$!
+CLEANUP_PIDS+=("$BACKEND_PID")
 sleep 3

-if curl -sf http://localhost:8000/health &>/dev/null; then
+if ! kill -0 "$BACKEND_PID" 2>/dev/null; then
+  fail "Backend exited during startup. Run 'uv run jarvis serve --port 8000' to see the error."
+elif curl -sf http://localhost:8000/health &>/dev/null; then
  ok "Backend running at http://localhost:8000"
 else
  warn "Backend may still be starting..."
@@ -57,6 +57,10 @@ class BaseAgent(ABC):

    agent_id: str
    accepts_tools: bool = False
+    # Plain conversational agents may opt into the managed runtime's generic
+    # function-calling loop.  Specialized agents keep their own execution
+    # class even when process-wide MCP tools are available.
+    supports_managed_tool_fallback: bool = False

    def __init__(
        self,
@@ -19,6 +19,7 @@ from typing import Any, List, Optional

 from openjarvis.agents._stubs import AgentContext, AgentResult, BaseAgent
 from openjarvis.core.events import EventBus
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.registry import AgentRegistry
 from openjarvis.core.types import ToolResult
 from openjarvis.engine._stubs import InferenceEngine
@@ -103,7 +104,7 @@ class ClaudeCodeAgent(BaseAgent):
                "Install it from https://nodejs.org/ or via your package manager."
            )

-        dest = Path.home() / ".openjarvis" / "claude_code_runner"
+        dest = get_config_dir() / "claude_code_runner"
        dest.mkdir(parents=True, exist_ok=True)

        # Copy runner files if missing or outdated
@@ -9,6 +9,8 @@ from datetime import datetime
 from pathlib import Path
 from typing import Dict, List, Optional

+from openjarvis.core.paths import get_config_dir
+

@dataclass
 class DigestArtifact:
@@ -30,7 +32,7 @@ class DigestStore:

    def __init__(self, db_path: str = "") -> None:
        if not db_path:
-            db_path = str(Path.home() / ".openjarvis" / "digest.db")
+            db_path = str(get_config_dir() / "digest.db")
        self._db_path = db_path
        self._conn = sqlite3.connect(db_path, check_same_thread=False)
        self._conn.execute("PRAGMA journal_mode=WAL")
@@ -2,6 +2,8 @@

 from __future__ import annotations

+from openjarvis.engine._base import looks_like_context_length_error
+

 class AgentTickError(Exception):
    """Base class for agent tick errors."""
@@ -64,6 +66,14 @@ def classify_error(exc: Exception) -> AgentTickError:

    msg = str(exc).lower()

+    # A context-window overflow is deterministic: retrying the identical
+    # over-length request can never succeed, so fail fast instead of burning
+    # the retry budget on it.
+    if getattr(exc, "is_context_length_error", False) or (
+        looks_like_context_length_error(msg)
+    ):
+        return FatalError(str(exc))
+
    # Check fatal patterns first (more specific)
    if isinstance(exc, PermissionError):
        return FatalError(str(exc))
@@ -90,6 +100,11 @@ def retry_delay(attempt: int) -> int:
 def suggest_action(error: AgentTickError) -> str:
    """Return a human-readable suggested action for the given error."""
    msg = str(error).lower()
+    if looks_like_context_length_error(msg):
+        return (
+            "Conversation too long for the model's context window \u2014 "
+            "start a new chat or shorten the conversation"
+        )
    if any(p in msg for p in ("rate limit", "rate_limit", "429", "too many requests")):
        return "Rate limited \u2014 agent will auto-retry on next tick"
    if any(p in msg for p in ("timeout", "timed out", "connection", "unavailable")):
@@ -2,7 +2,9 @@

 from __future__ import annotations

+import json
 import logging
+import threading
 import time
 from typing import TYPE_CHECKING, Any

@@ -14,6 +16,7 @@ from openjarvis.agents.errors import (
    classify_error,
    retry_delay,
 )
+from openjarvis.agents.tool_resolver import resolve_agent_tools
 from openjarvis.core.events import EventBus, EventType

 if TYPE_CHECKING:
@@ -33,6 +36,32 @@ _MAX_RETRIES = 3
 _AGENT_TICK_DEFAULT_MODEL = "gemma4:31b"


+def _tool_calls_for_storage(result: AgentResult) -> list[dict[str, Any]] | None:
+    """Convert executor tool results to the managed-message storage contract."""
+
+    calls: list[dict[str, Any]] = []
+    for tool_result in result.tool_results:
+        metadata = getattr(tool_result, "metadata", {}) or {}
+        arguments = metadata.get("arguments", "")
+        if not isinstance(arguments, str):
+            try:
+                arguments = json.dumps(arguments, sort_keys=True)
+            except (TypeError, ValueError):
+                arguments = json.dumps(str(arguments))
+        calls.append(
+            {
+                "tool": getattr(tool_result, "tool_name", ""),
+                "arguments": arguments,
+                "result": getattr(tool_result, "content", "") or "",
+                "success": bool(getattr(tool_result, "success", False)),
+                # SSE and the frontend persist/display latency in milliseconds.
+                "latency": float(getattr(tool_result, "latency_seconds", 0.0) or 0.0)
+                * 1000.0,
+            }
+        )
+    return calls or None
+
+
 class AgentExecutor:
    """Executes a single tick for a managed agent.

@@ -51,6 +80,7 @@ class AgentExecutor:
        self._manager = manager
        self._bus = event_bus
        self._trace_store = trace_store
+        self._toolkit_local = threading.local()

    def set_system(self, system: Any) -> None:
        """Deferred system injection — called after JarvisSystem is constructed."""
@@ -63,27 +93,6 @@ class AgentExecutor:
        except Exception:
            pass  # Non-critical

-    def _inject_tool_deps(self, tool: Any) -> None:
-        """Inject runtime dependencies into a tool instance.
-
-        Mirrors SystemBuilder._inject_tool_deps (system.py:920-945)
-        but uses the lightweight system's references.
-        """
-        if self._system is None:
-            return
-        name = getattr(getattr(tool, "spec", None), "name", "")
-        if name == "llm":
-            if hasattr(tool, "_engine"):
-                tool._engine = self._system.engine
-            if hasattr(tool, "_model"):
-                tool._model = self._system.model
-        elif name == "retrieval" or name.startswith("memory_"):
-            if hasattr(tool, "_backend"):
-                tool._backend = getattr(self._system, "memory_backend", None)
-        elif name.startswith("channel_"):
-            if hasattr(tool, "_channel"):
-                tool._channel = getattr(self._system, "channel_backend", None)
-
    def run_ephemeral(
        self,
        agent_type: str,
@@ -102,9 +111,7 @@ class AgentExecutor:
        )
        return agent.run(input_text)

-    def execute_tick(
-        self, agent_id: str, *, lock_already_held: bool = False
-    ) -> None:
+    def execute_tick(self, agent_id: str, *, lock_already_held: bool = False) -> None:
        """Run one tick for the given agent.

        1. Acquire concurrency guard (start_tick)
@@ -126,9 +133,7 @@ class AgentExecutor:
                self._manager.start_tick(agent_id)
                self._set_activity(agent_id, "Preparing tick...")
            except ValueError:
-                logger.warning(
-                    "Agent %s already running, skipping tick", agent_id
-                )
+                logger.warning("Agent %s already running, skipping tick", agent_id)
                return

        agent = self._manager.get_agent(agent_id)
@@ -252,7 +257,20 @@ class AgentExecutor:
        raise last_error or FatalError("max retries exhausted")

    def _invoke_agent(self, agent: dict) -> AgentResult:
-        """Invoke the actual agent run. Tests mock this method."""
+        """Invoke one agent while owning every resource its resolver opens."""
+
+        previous = getattr(self._toolkit_local, "current", None)
+        self._toolkit_local.current = None
+        try:
+            return self._invoke_agent_impl(agent)
+        finally:
+            current = getattr(self._toolkit_local, "current", None)
+            if current is not None:
+                current.close()
+            self._toolkit_local.current = previous
+
+    def _invoke_agent_impl(self, agent: dict) -> AgentResult:
+        """Implementation split out so the wrapper owns resolver lifetime."""
        from openjarvis.agents import AgentRegistry

        agent_type = agent.get("agent_type", "monitor_operative")
@@ -261,6 +279,10 @@ class AgentExecutor:
            raise FatalError(f"Unknown agent type: {agent_type}")

        config = agent.get("config", {})
+        agent_accepts_tools = bool(getattr(agent_cls, "accepts_tools", False))
+        supports_tool_fallback = bool(
+            getattr(agent_cls, "supports_managed_tool_fallback", False)
+        )

        # Resolve engine + model from JarvisSystem
        engine = self._system.engine if self._system else None
@@ -304,64 +326,88 @@ class AgentExecutor:
            except Exception:
                pass  # Fall back to configured model

-        # Resolve tools from config via ToolRegistry
-        tool_names = config.get("tools", [])
-        if isinstance(tool_names, str):
-            tool_names = [t.strip() for t in tool_names.split(",") if t.strip()]
+        mcp_tools: list[Any] = []
+        mcp_clients: list[Any] = []
+        if (
+            config.get("mcp_tools", True) is not False
+            and self._system is not None
+            and (agent_accepts_tools or supports_tool_fallback)
+        ):
+            provider = getattr(
+                self._system,
+                "get_managed_agent_mcp_tools",
+                None,
+            )
+            if callable(provider):
+                try:
+                    mcp_tools, mcp_clients = provider()
+                except Exception as exc:
+                    logger.warning("Managed-agent MCP discovery failed: %s", exc)
+            else:
+                mcp_tools = list(getattr(self._system, "mcp_tools", []) or [])
+                mcp_clients = list(getattr(self._system, "_mcp_clients", []) or [])

-        tool_instances: list[Any] = []
-        if tool_names:
-            try:
-                from openjarvis.server.agent_manager_routes import (
-                    _ensure_registries_populated,
-                )
+            if not mcp_tools:
+                try:
+                    from openjarvis.tools.mcp_adapter import MCPToolAdapter

-                _ensure_registries_populated()
-            except ImportError:
-                pass
-            from openjarvis.core.registry import ToolRegistry
+                    pool = (
+                        getattr(
+                            getattr(self._system, "tool_executor", None),
+                            "_tools",
+                            {},
+                        )
+                        or {}
+                    )
+                    mcp_tools = [
+                        tool
+                        for tool in pool.values()
+                        if isinstance(tool, MCPToolAdapter)
+                    ]
+                except Exception:
+                    mcp_tools = []

-            for tname in tool_names:
-                if ToolRegistry.contains(tname):
-                    try:
-                        tool_cls = ToolRegistry.get(tname)
-                        tool = tool_cls()
-                        self._inject_tool_deps(tool)
-                        tool_instances.append(tool)
-                    except Exception:
-                        logger.warning("Failed to instantiate tool %s", tname)
+        resolved_toolkit = resolve_agent_tools(
+            agent,
+            engine=engine,
+            model=model,
+            memory_backend=getattr(self._system, "memory_backend", None),
+            channel_backend=getattr(self._system, "channel_backend", None),
+            mcp_tools=mcp_tools,
+            mcp_clients=mcp_clients,
+            knowledge_db_path=getattr(self._system, "knowledge_db_path", None),
+        )
+        self._toolkit_local.current = resolved_toolkit
+        tool_instances = resolved_toolkit.instances
+        logger.info(
+            "Agent %s: resolved %d tools (%s)",
+            agent["name"],
+            len(tool_instances),
+            ", ".join(resolved_toolkit.by_name) or "none",
+        )

-            # Pull tools already discovered by SystemBuilder (e.g. external MCP
-            # adapters) that aren't in the static ToolRegistry. Without this,
-            # agents declaring MCP-discovered tools in their template would
-            # silently fall back to natives only.
-            if (
-                self._system is not None
-                and getattr(self._system, "tool_executor", None) is not None
-            ):
-                mcp_pool = getattr(self._system.tool_executor, "_tools", {}) or {}
-                existing = {t.spec.name for t in tool_instances}
-                for tname in tool_names:
-                    if tname in existing:
-                        continue
-                    pooled = mcp_pool.get(tname)
-                    if pooled is not None:
-                        tool_instances.append(pooled)
+        execution_agent_cls = agent_cls
+        if tool_instances and not agent_accepts_tools and supports_tool_fallback:
+            # Managed SSE already runs configured tools through a native
+            # function-calling loop regardless of the selected class. Use the
+            # same capability for immediate/scheduled ticks instead of
+            # silently discarding the resolved toolkit for SimpleAgent and
+            # other explicitly compatible non-tool classes.
+            from openjarvis.agents.orchestrator import OrchestratorAgent

-            if tool_instances:
-                logger.info(
-                    "Agent %s: resolved %d/%d tools",
-                    agent["name"],
-                    len(tool_instances),
-                    len(tool_names),
-                )
+            execution_agent_cls = OrchestratorAgent
+            logger.info(
+                "Agent %s: %s does not accept tools; using %s for this "
+                "tool-enabled tick",
+                agent["name"],
+                agent_cls.__name__,
+                execution_agent_cls.__name__,
+            )

        # Construct agent instance
        agent_kwargs: dict[str, Any] = {}
        sys_prompt = config.get("system_prompt")
-        if sys_prompt is not None:
-            agent_kwargs["system_prompt"] = sys_prompt
-        if getattr(agent_cls, "accepts_tools", False) and tool_instances:
+        if getattr(execution_agent_cls, "accepts_tools", False) and tool_instances:
            agent_kwargs["tools"] = tool_instances
        # Hand the agent our EventBus so its ToolExecutor can publish
        # TOOL_CALL_START/END — without this, ToolExecutor's ``self._bus``
@@ -383,7 +429,7 @@ class AgentExecutor:
        # recall / persistence paths.
        import inspect

-        init_sig = inspect.signature(agent_cls.__init__)
+        init_sig = inspect.signature(execution_agent_cls.__init__)
        accepts_var_kw = any(
            p.kind == inspect.Parameter.VAR_KEYWORD
            for p in init_sig.parameters.values()
@@ -392,6 +438,16 @@ class AgentExecutor:
        def _accepts(name: str) -> bool:
            return accepts_var_kw or name in init_sig.parameters

+        # Unsupported kwargs used to trigger the broad TypeError fallback
+        # below, which retried with a bare constructor and silently discarded
+        # valid prompt/state wiring. Filter by the selected class's signature
+        # before construction instead.
+        if sys_prompt is not None and _accepts("system_prompt"):
+            agent_kwargs["system_prompt"] = sys_prompt
+        agent_kwargs = {
+            name: value for name, value in agent_kwargs.items() if _accepts(name)
+        }
+
        state_kwargs: dict[str, Any] = {}
        if _accepts("operator_id"):
            state_kwargs["operator_id"] = agent["id"]
@@ -408,27 +464,49 @@ class AgentExecutor:
            # agents, mirroring the one-shot `jarvis ask` path so they no
            # longer apply to CLI calls only (#376).
            cfg = getattr(self._system, "config", None)
-            if cfg is not None and _accepts("prompt_builder"):
+            if _accepts("prompt_builder") and (
+                cfg is not None or sys_prompt is not None
+            ):
                from openjarvis.prompt.builder import SystemPromptBuilder

                state_kwargs["prompt_builder"] = SystemPromptBuilder(
-                    agent_template=getattr(
-                        cfg.agent, "default_system_prompt", ""
-                    )
-                    or "",
-                    memory_files_config=cfg.memory_files,
-                    system_prompt_config=cfg.system_prompt,
+                    agent_template=(
+                        sys_prompt
+                        if sys_prompt is not None
+                        else getattr(
+                            getattr(cfg, "agent", None),
+                            "default_system_prompt",
+                            "",
+                        )
+                        or ""
+                    ),
+                    memory_files_config=getattr(cfg, "memory_files", None),
+                    system_prompt_config=getattr(cfg, "system_prompt", None),
                )

        try:
-            agent_instance = agent_cls(
-                engine, model, **agent_kwargs, **state_kwargs
-            )
-        except TypeError:
            try:
-                agent_instance = agent_cls(engine, model, **agent_kwargs)
+                agent_instance = execution_agent_cls(
+                    engine,
+                    model,
+                    **agent_kwargs,
+                    **state_kwargs,
+                )
            except TypeError:
-                agent_instance = agent_cls(engine, model)
+                try:
+                    agent_instance = execution_agent_cls(
+                        engine,
+                        model,
+                        **agent_kwargs,
+                    )
+                except TypeError:
+                    agent_instance = execution_agent_cls(engine, model)
+        except Exception:
+            resolved_toolkit.close()
+            raise
+
+        if resolved_toolkit.mcp_clients:
+            agent_instance._mcp_clients = resolved_toolkit.mcp_clients

        # Inject the managed-agent UUID into the agent's ToolExecutor so
        # emitted TOOL_CALL_START/END events carry it; the trace subscriber
@@ -444,7 +522,7 @@ class AgentExecutor:
            agent["name"],
            len(tool_instances),
            ", ".join(t.spec.name for t in tool_instances) or "none",
-            agent_cls.__name__,
+            execution_agent_cls.__name__,
        )

        # Build input from instruction + summary_memory + pending messages.
@@ -474,9 +552,7 @@ class AgentExecutor:
                tick_note = f"Previous tick: {first_sentence}"

        if instruction:
-            input_text = (
-                f"Current date: {today}\n\nStanding instruction: {instruction}"
-            )
+            input_text = f"Current date: {today}\n\nStanding instruction: {instruction}"
            if tick_note:
                input_text += f"\n\n{tick_note}"
        else:
@@ -561,21 +637,24 @@ class AgentExecutor:
            len(input_text),
        )
        _t0 = time.time()
-        result = agent_instance.run(input_text, context=agent_ctx)
-
-        # Retry once if the model returned empty content (common with
-        # Qwen3.5 thinking mode consuming all tokens).
-        if not (result.content or "").strip():
-            self._set_activity(
-                agent["id"],
-                "Retrying (empty response)...",
-            )
-            logger.warning(
-                "Agent %s: empty content, retrying once",
-                agent["name"],
-            )
+        try:
            result = agent_instance.run(input_text, context=agent_ctx)

+            # Retry once if the model returned empty content (common with
+            # Qwen3.5 thinking mode consuming all tokens).
+            if not (result.content or "").strip():
+                self._set_activity(
+                    agent["id"],
+                    "Retrying (empty response)...",
+                )
+                logger.warning(
+                    "Agent %s: empty content, retrying once",
+                    agent["name"],
+                )
+                result = agent_instance.run(input_text, context=agent_ctx)
+        finally:
+            resolved_toolkit.close()
+
        _elapsed = time.time() - _t0
        logger.info(
            "Agent %s: agent.run() completed in %.1fs, "
@@ -665,7 +744,11 @@ class AgentExecutor:
                # message keeps the complete report. The old [:2000] slices
                # double-truncated and cut findings off mid-sentence.
                self._manager.update_summary_memory(agent_id, result.content)
-                self._manager.store_agent_response(agent_id, result.content)
+                self._manager.store_agent_response(
+                    agent_id,
+                    result.content,
+                    tool_calls=_tool_calls_for_storage(result),
+                )

            # Budget enforcement (post-tick check)
            agent_data = self._manager.get_agent(agent_id)
@@ -73,12 +73,19 @@ WEB_SEARCH_COST_PER_CALL = 0.01
 # $0.01/call number — kept as a separate constant so it can drift.
 OPENAI_WEB_SEARCH_COST_PER_CALL = 0.01

-# Gemini Google-Search grounding: billed at $35 per 1000 grounded
-# *requests* (2025-12 public list price for the Grounding-with-Google-Search
-# tool, charged once per request that uses the tool regardless of how many
-# internal queries it issues). We charge per grounded request, not per
-# `web_search_queries` entry.
-GEMINI_SEARCH_COST_PER_CALL = 0.035
+# Gemini 3 Google-Search grounding: billed at $14 per 1000 search queries.
+# `_call_gemini_agent` reports the model's `web_search_queries`, so this is
+# charged per query, not per outer generate_content request.
+GEMINI_SEARCH_COST_PER_CALL = 0.014
+
+# Tavily Search, advanced depth: 2 API credits per search request at $0.008
+# per credit on the public pay-as-you-go plan. WebSearchTool captures actual
+# credits when Tavily returns usage metadata; this is the fallback estimate.
+TAVILY_SEARCH_COST_PER_CREDIT = 0.008
+TAVILY_ADVANCED_SEARCH_CREDITS = 2
+TAVILY_SEARCH_COST_PER_CALL = (
+    TAVILY_SEARCH_COST_PER_CREDIT * TAVILY_ADVANCED_SEARCH_CREDITS
+)

 ANTHROPIC_WEB_SEARCH_TOOL = {
    "type": "web_search_20250305",
@@ -101,6 +108,40 @@ def build_web_search_tool(max_uses: int = 8) -> Dict[str, Any]:
    }


+def tavily_search_context(
+    query: str,
+    *,
+    max_results: int = 5,
+) -> Dict[str, Any]:
+    """Run OpenJarvis WebSearchTool and return accounting-friendly metadata."""
+    from openjarvis.tools.web_search import WebSearchTool
+
+    tool = WebSearchTool(max_results=max_results)
+    res = tool.execute(query=query, max_results=max_results)
+    meta = dict(res.metadata or {})
+    engine = str(meta.get("engine") or "unknown")
+    credits = 0
+    cost_usd = 0.0
+    if engine == "tavily":
+        try:
+            credits = int(meta.get("credits") or TAVILY_ADVANCED_SEARCH_CREDITS)
+        except (TypeError, ValueError):
+            credits = TAVILY_ADVANCED_SEARCH_CREDITS
+        cost_usd = credits * TAVILY_SEARCH_COST_PER_CREDIT
+    text = res.content or ""
+    if not res.success and not text:
+        text = "(no search results)"
+    return {
+        "text": text,
+        "success": bool(res.success),
+        "engine": engine,
+        "credits": credits,
+        "cost_usd": cost_usd,
+        "n_searches": 1 if (query or "").strip() else 0,
+        "error": None if res.success else text,
+    }
+
+
 def web_search_cfg(method_cfg: Optional[Dict[str, Any]]) -> Tuple[bool, int]:
    """Parse ``method_cfg.web_search = { enabled, max_uses }``.

@@ -262,7 +303,9 @@ def _openrouter_limiter() -> _OpenRouterLimiter:
    if _OPENROUTER_LIMITER is None:
        with _OPENROUTER_LIMITER_LOCK:
            if _OPENROUTER_LIMITER is None:
-                max_concurrent = int(os.environ.get("OJ_OPENROUTER_MAX_CONCURRENT", "20") or 20)
+                max_concurrent = int(
+                    os.environ.get("OJ_OPENROUTER_MAX_CONCURRENT", "20") or 20
+                )
                rpm = int(os.environ.get("OJ_OPENROUTER_RPM", "60") or 60)
                _OPENROUTER_LIMITER = _OpenRouterLimiter(max_concurrent, rpm)
    return _OPENROUTER_LIMITER
@@ -278,8 +321,14 @@ def _serialize_block(block: Any) -> Dict[str, Any]:
    """
    out: Dict[str, Any] = {"type": getattr(block, "type", type(block).__name__)}
    for attr in (
-        "id", "name", "input", "text", "thinking", "signature",
-        "tool_use_id", "content",
+        "id",
+        "name",
+        "input",
+        "text",
+        "thinking",
+        "signature",
+        "tool_use_id",
+        "content",
    ):
        if hasattr(block, attr):
            val = getattr(block, attr)
@@ -300,14 +349,16 @@ def _serialize_openai_tool_calls(tool_calls: Any) -> List[Dict[str, Any]]:
        return out
    for tc in tool_calls:
        fn = getattr(tc, "function", None)
-        out.append({
-            "id": getattr(tc, "id", None),
-            "type": getattr(tc, "type", "function"),
-            "function": {
-                "name": getattr(fn, "name", None) if fn else None,
-                "arguments": getattr(fn, "arguments", None) if fn else None,
-            },
-        })
+        out.append(
+            {
+                "id": getattr(tc, "id", None),
+                "type": getattr(tc, "type", "function"),
+                "function": {
+                    "name": getattr(fn, "name", None) if fn else None,
+                    "arguments": getattr(fn, "arguments", None) if fn else None,
+                },
+            }
+        )
    return out


@@ -432,32 +483,46 @@ class LocalCloudAgent(BaseAgent):
        srv = getattr(msg.usage, "server_tool_use", None)
        n_searches = getattr(srv, "web_search_requests", 0) if srv else 0
        content_blocks = [_serialize_block(b) for b in msg.content]
-        tool_use_blocks = [b for b in content_blocks if b.get("type") in (
-            "tool_use", "server_tool_use",
-        )]
-        tool_result_blocks = [b for b in content_blocks if b.get("type") in (
-            "web_search_tool_result", "tool_result",
-        )]
-        _record_event({
-            "kind": "anthropic",
-            "role": trace_role,
-            "model": model,
-            "system": system,
-            "user": user,
-            "response": text,
-            "content_blocks": content_blocks,
-            "tool_calls": tool_use_blocks,
-            "tool_results": tool_result_blocks,
-            "tokens_in": msg.usage.input_tokens,
-            "tokens_out": msg.usage.output_tokens,
-            "n_web_searches": n_searches,
-            "tools_declared": tools,
-            "tool_choice": tool_choice,
-            "output_config": output_config,
-            "stop_reason": getattr(msg, "stop_reason", None),
-            "latency_s": latency,
-            "ts": time.time(),
-        })
+        tool_use_blocks = [
+            b
+            for b in content_blocks
+            if b.get("type")
+            in (
+                "tool_use",
+                "server_tool_use",
+            )
+        ]
+        tool_result_blocks = [
+            b
+            for b in content_blocks
+            if b.get("type")
+            in (
+                "web_search_tool_result",
+                "tool_result",
+            )
+        ]
+        _record_event(
+            {
+                "kind": "anthropic",
+                "role": trace_role,
+                "model": model,
+                "system": system,
+                "user": user,
+                "response": text,
+                "content_blocks": content_blocks,
+                "tool_calls": tool_use_blocks,
+                "tool_results": tool_result_blocks,
+                "tokens_in": msg.usage.input_tokens,
+                "tokens_out": msg.usage.output_tokens,
+                "n_web_searches": n_searches,
+                "tools_declared": tools,
+                "tool_choice": tool_choice,
+                "output_config": output_config,
+                "stop_reason": getattr(msg, "stop_reason", None),
+                "latency_s": latency,
+                "ts": time.time(),
+            }
+        )
        return text, msg.usage.input_tokens, msg.usage.output_tokens, n_searches

    @staticmethod
@@ -509,24 +574,26 @@ class LocalCloudAgent(BaseAgent):
        u = resp.usage
        p = getattr(u, "prompt_tokens", 0) if u else 0
        c = getattr(u, "completion_tokens", 0) if u else 0
-        _record_event({
-            "kind": "openai",
-            "role": trace_role,
-            "model": model,
-            "system": system,
-            "user": user,
-            "response": text,
-            "tool_calls": tool_calls,
-            "reasoning_content": reasoning,
-            "tokens_in": p,
-            "tokens_out": c,
-            "response_format": response_format,
-            "tools_declared": tools,
-            "tool_choice": tool_choice,
-            "finish_reason": getattr(choice, "finish_reason", None),
-            "latency_s": latency,
-            "ts": time.time(),
-        })
+        _record_event(
+            {
+                "kind": "openai",
+                "role": trace_role,
+                "model": model,
+                "system": system,
+                "user": user,
+                "response": text,
+                "tool_calls": tool_calls,
+                "reasoning_content": reasoning,
+                "tokens_in": p,
+                "tokens_out": c,
+                "response_format": response_format,
+                "tools_declared": tools,
+                "tool_choice": tool_choice,
+                "finish_reason": getattr(choice, "finish_reason", None),
+                "latency_s": latency,
+                "ts": time.time(),
+            }
+        )
        return text, p, c

    @staticmethod
@@ -573,12 +640,10 @@ class LocalCloudAgent(BaseAgent):
        from openai import OpenAI

        if model.startswith("openrouter/"):
-            model = model[len("openrouter/"):]
+            model = model[len("openrouter/") :]
        api_key = os.environ.get("OPENROUTER_API_KEY")
        if not api_key:
-            raise RuntimeError(
-                "OPENROUTER_API_KEY is not set; cannot call OpenRouter."
-            )
+            raise RuntimeError("OPENROUTER_API_KEY is not set; cannot call OpenRouter.")
        client = OpenAI(
            base_url="https://openrouter.ai/api/v1",
            api_key=api_key,
@@ -617,21 +682,23 @@ class LocalCloudAgent(BaseAgent):
        u = resp.usage
        p = getattr(u, "prompt_tokens", 0) if u else 0
        c = getattr(u, "completion_tokens", 0) if u else 0
-        _record_event({
-            "kind": "openrouter",
-            "role": trace_role,
-            "model": model,
-            "system": system,
-            "user": user,
-            "response": text,
-            "tool_calls": tool_calls,
-            "reasoning_content": reasoning,
-            "tokens_in": p,
-            "tokens_out": c,
-            "finish_reason": getattr(choice, "finish_reason", None),
-            "latency_s": latency,
-            "ts": time.time(),
-        })
+        _record_event(
+            {
+                "kind": "openrouter",
+                "role": trace_role,
+                "model": model,
+                "system": system,
+                "user": user,
+                "response": text,
+                "tool_calls": tool_calls,
+                "reasoning_content": reasoning,
+                "tokens_in": p,
+                "tokens_out": c,
+                "finish_reason": getattr(choice, "finish_reason", None),
+                "latency_s": latency,
+                "ts": time.time(),
+            }
+        )
        return text, p, c

    @staticmethod
@@ -660,7 +727,9 @@ class LocalCloudAgent(BaseAgent):
        from google import genai
        from google.genai import types

-        client = genai.Client(http_options=types.HttpOptions(timeout=int(timeout * 1000)))
+        client = genai.Client(
+            http_options=types.HttpOptions(timeout=int(timeout * 1000))
+        )
        cfg = types.GenerateContentConfig(
            temperature=temperature,
            max_output_tokens=max_tokens,
@@ -687,21 +756,23 @@ class LocalCloudAgent(BaseAgent):
            finish_reason = str(resp.candidates[0].finish_reason)
        except Exception:
            pass
-        _record_event({
-            "kind": "gemini",
-            "role": trace_role,
-            "model": model,
-            "system": system,
-            "user": user,
-            "response": text,
-            "tokens_in": p,
-            "tokens_out": c,
-            "temperature": temperature,
-            "max_tokens": max_tokens,
-            "finish_reason": finish_reason,
-            "latency_s": latency,
-            "ts": time.time(),
-        })
+        _record_event(
+            {
+                "kind": "gemini",
+                "role": trace_role,
+                "model": model,
+                "system": system,
+                "user": user,
+                "response": text,
+                "tokens_in": p,
+                "tokens_out": c,
+                "temperature": temperature,
+                "max_tokens": max_tokens,
+                "finish_reason": finish_reason,
+                "latency_s": latency,
+                "ts": time.time(),
+            }
+        )
        return text, p, c

    @staticmethod
@@ -756,27 +827,29 @@ class LocalCloudAgent(BaseAgent):
        u = resp.usage
        p = getattr(u, "prompt_tokens", 0) if u else 0
        c = getattr(u, "completion_tokens", 0) if u else 0
-        _record_event({
-            "kind": "vllm",
-            "role": trace_role,
-            "model": model,
-            "endpoint": endpoint,
-            "system": system,
-            "user": user,
-            "response": text,
-            "tool_calls": tool_calls,
-            "reasoning_content": reasoning,
-            "tokens_in": p,
-            "tokens_out": c,
-            "temperature": temperature,
-            "max_tokens": max_tokens,
-            "enable_thinking": enable_thinking,
-            "tools_declared": tools,
-            "tool_choice": tool_choice,
-            "finish_reason": getattr(choice, "finish_reason", None),
-            "latency_s": latency,
-            "ts": time.time(),
-        })
+        _record_event(
+            {
+                "kind": "vllm",
+                "role": trace_role,
+                "model": model,
+                "endpoint": endpoint,
+                "system": system,
+                "user": user,
+                "response": text,
+                "tool_calls": tool_calls,
+                "reasoning_content": reasoning,
+                "tokens_in": p,
+                "tokens_out": c,
+                "temperature": temperature,
+                "max_tokens": max_tokens,
+                "enable_thinking": enable_thinking,
+                "tools_declared": tools,
+                "tool_choice": tool_choice,
+                "finish_reason": getattr(choice, "finish_reason", None),
+                "latency_s": latency,
+                "ts": time.time(),
+            }
+        )
        return text, p, c

    @staticmethod
@@ -843,33 +916,37 @@ class LocalCloudAgent(BaseAgent):
            n_searches = getattr(srv, "web_search_requests", 0) if srv else 0
            content_blocks = [_serialize_block(b) for b in msg.content]
            tool_use_blocks = [
-                b for b in content_blocks
+                b
+                for b in content_blocks
                if b.get("type") in ("tool_use", "server_tool_use")
            ]
            tool_result_blocks = [
-                b for b in content_blocks
+                b
+                for b in content_blocks
                if b.get("type") in ("web_search_tool_result", "tool_result")
            ]
            stop_reason = getattr(msg, "stop_reason", None)
-            _record_event({
-                "kind": "anthropic",
-                "role": trace_role,
-                "model": model,
-                "system": system if turn == 0 else None,
-                "user": user if turn == 0 else None,
-                "turn": turn,
-                "response": text,
-                "content_blocks": content_blocks,
-                "tool_calls": tool_use_blocks,
-                "tool_results": tool_result_blocks,
-                "tokens_in": msg.usage.input_tokens,
-                "tokens_out": msg.usage.output_tokens,
-                "n_web_searches": n_searches,
-                "tools_declared": tools,
-                "stop_reason": stop_reason,
-                "latency_s": latency,
-                "ts": time.time(),
-            })
+            _record_event(
+                {
+                    "kind": "anthropic",
+                    "role": trace_role,
+                    "model": model,
+                    "system": system if turn == 0 else None,
+                    "user": user if turn == 0 else None,
+                    "turn": turn,
+                    "response": text,
+                    "content_blocks": content_blocks,
+                    "tool_calls": tool_use_blocks,
+                    "tool_results": tool_result_blocks,
+                    "tokens_in": msg.usage.input_tokens,
+                    "tokens_out": msg.usage.output_tokens,
+                    "n_web_searches": n_searches,
+                    "tools_declared": tools,
+                    "stop_reason": stop_reason,
+                    "latency_s": latency,
+                    "ts": time.time(),
+                }
+            )
            p_total += msg.usage.input_tokens
            c_total += msg.usage.output_tokens
            n_searches_total += n_searches
@@ -879,9 +956,7 @@ class LocalCloudAgent(BaseAgent):
            # here — break and let the caller (or future loop variant)
            # handle it. Only ``server_tool_use`` blocks (web_search)
            # are auto-continued by Anthropic itself.
-            client_tool_use = any(
-                b.get("type") == "tool_use" for b in content_blocks
-            )
+            client_tool_use = any(b.get("type") == "tool_use" for b in content_blocks)
            if client_tool_use:
                break
            if stop_reason == "end_turn" or stop_reason is None:
@@ -890,10 +965,12 @@ class LocalCloudAgent(BaseAgent):
            # (server side) — Anthropic returned mid-thought. Append the
            # assistant turn and ask it to continue.
            messages.append({"role": "assistant", "content": msg.content})
-            messages.append({
-                "role": "user",
-                "content": "Continue.",
-            })
+            messages.append(
+                {
+                    "role": "user",
+                    "content": "Continue.",
+                }
+            )
        return last_text, p_total, c_total, n_searches_total, turns

    @staticmethod
@@ -967,8 +1044,12 @@ class LocalCloudAgent(BaseAgent):
                    continue
                raise
        if resp is None:
-            raise last_exc if last_exc is not None else RuntimeError(
-                "openai responses.create failed for all web_search tool names"
+            raise (
+                last_exc
+                if last_exc is not None
+                else RuntimeError(
+                    "openai responses.create failed for all web_search tool names"
+                )
            )
        _bump_cloud_calls()
        latency = time.time() - t0
@@ -992,30 +1073,35 @@ class LocalCloudAgent(BaseAgent):
            text = "".join(chunks)

        n_searches = sum(
-            1 for item in output_items
-            if getattr(item, "type", None) in (
-                "web_search_call", "web_search_tool_call",
+            1
+            for item in output_items
+            if getattr(item, "type", None)
+            in (
+                "web_search_call",
+                "web_search_tool_call",
            )
        )
        u = getattr(resp, "usage", None)
        p = int(getattr(u, "input_tokens", 0) or 0) if u else 0
        c = int(getattr(u, "output_tokens", 0) or 0) if u else 0
-        _record_event({
-            "kind": "openai_agent",
-            "role": trace_role,
-            "model": model,
-            "system": system,
-            "user": user,
-            "response": text,
-            "output_items": _jsonable(output_items),
-            "tokens_in": p,
-            "tokens_out": c,
-            "n_web_searches": n_searches,
-            "tools_declared": [{"type": used_tool_name}],
-            "stop_reason": getattr(resp, "status", None),
-            "latency_s": latency,
-            "ts": time.time(),
-        })
+        _record_event(
+            {
+                "kind": "openai_agent",
+                "role": trace_role,
+                "model": model,
+                "system": system,
+                "user": user,
+                "response": text,
+                "output_items": _jsonable(output_items),
+                "tokens_in": p,
+                "tokens_out": c,
+                "n_web_searches": n_searches,
+                "tools_declared": [{"type": used_tool_name}],
+                "stop_reason": getattr(resp, "status", None),
+                "latency_s": latency,
+                "ts": time.time(),
+            }
+        )
        return text, p, c, n_searches, 1

    @staticmethod
@@ -1088,23 +1174,25 @@ class LocalCloudAgent(BaseAgent):
                n_searches = len(web_search_queries)
        except Exception:  # noqa: BLE001
            pass
-        _record_event({
-            "kind": "gemini_agent",
-            "role": trace_role,
-            "model": model,
-            "system": system,
-            "user": user,
-            "response": text,
-            "tokens_in": p,
-            "tokens_out": c,
-            "n_web_searches": n_searches,
-            "web_search_queries": web_search_queries,
-            "temperature": temperature,
-            "max_tokens": max_tokens,
-            "finish_reason": finish_reason,
-            "latency_s": latency,
-            "ts": time.time(),
-        })
+        _record_event(
+            {
+                "kind": "gemini_agent",
+                "role": trace_role,
+                "model": model,
+                "system": system,
+                "user": user,
+                "response": text,
+                "tokens_in": p,
+                "tokens_out": c,
+                "n_web_searches": n_searches,
+                "web_search_queries": web_search_queries,
+                "temperature": temperature,
+                "max_tokens": max_tokens,
+                "finish_reason": finish_reason,
+                "latency_s": latency,
+                "ts": time.time(),
+            }
+        )
        return text, p, c, n_searches, 1

    def _call_cloud(
@@ -1228,8 +1316,13 @@ class LocalCloudAgent(BaseAgent):
            # Persist the trace before the trace state is closed (and even on
            # hard failure, so we get a record of what we did before it broke).
            self._write_trace_log(
-                context, input, answer, meta if "meta" in locals() else {},
-                events, soft_reason, exc_obj,
+                context,
+                input,
+                answer,
+                meta if "meta" in locals() else {},
+                events,
+                soft_reason,
+                exc_obj,
            )
            _close_trace()
            _close_call_counts()
@@ -1285,9 +1378,7 @@ class LocalCloudAgent(BaseAgent):
                "metadata": meta,
                "events": events,
                "soft_error": soft_reason,
-                "error": (
-                    f"{type(exc).__name__}: {exc}" if exc is not None else None
-                ),
+                "error": (f"{type(exc).__name__}: {exc}" if exc is not None else None),
            }
            (out_dir / f"{task_id}.json").write_text(
                json.dumps(blob, indent=2, default=str)
@@ -1322,6 +1413,9 @@ __all__ = [
    "LocalCloudAgent",
    "NO_TEMP_PREFIXES",
    "OPENAI_WEB_SEARCH_COST_PER_CALL",
+    "TAVILY_ADVANCED_SEARCH_CREDITS",
+    "TAVILY_SEARCH_COST_PER_CALL",
+    "TAVILY_SEARCH_COST_PER_CREDIT",
    "WEB_SEARCH_COST_PER_CALL",
    "_bump_cloud_calls",
    "_bump_local_calls",
@@ -1329,5 +1423,6 @@ __all__ = [
    "estimate_cost",
    "is_gpt5_family",
    "supports_temperature",
+    "tavily_search_context",
    "web_search_cfg",
 ]
@@ -96,6 +96,7 @@ class EnergyCollector:
            return self
        try:
            import pynvml  # type: ignore[import-not-found]
+
            pynvml.nvmlInit()
            total = pynvml.nvmlDeviceGetCount()
            self.gpu_indices = _resolve_gpu_indices(total)
@@ -174,12 +174,15 @@ def _is_retryable(exc: BaseException) -> bool:
        import openai
    except ImportError:
        return False
-    if isinstance(exc, (
-        openai.RateLimitError,
-        openai.APITimeoutError,
-        openai.APIConnectionError,
-        openai.InternalServerError,
-    )):
+    if isinstance(
+        exc,
+        (
+            openai.RateLimitError,
+            openai.APITimeoutError,
+            openai.APIConnectionError,
+            openai.InternalServerError,
+        ),
+    ):
        return True
    if isinstance(exc, openai.APIStatusError):
        status = getattr(exc, "status_code", None)
@@ -196,7 +199,7 @@ def _sleep_for(attempt: int, exc: BaseException) -> float:
        # Respect a server-provided hint, but clamp to our cap so a
        # pathological header can't stall the run for hours.
        return min(_RETRY_CAP, hinted) + random.uniform(0, 0.5)
-    base = min(_RETRY_CAP, _RETRY_BASE * (2 ** attempt))
+    base = min(_RETRY_CAP, _RETRY_BASE * (2**attempt))
    # Full jitter — better tail behavior than equal jitter when many
    # workers wake at the same moment.
    return random.uniform(0.0, base)
@@ -237,16 +240,19 @@ def _wrap_create(orig: Callable[..., Any]) -> Callable[..., Any]:
                        import openai
                    except ImportError:
                        raise
-                    if not isinstance(exc, (
-                        openai.APIConnectionError,
-                        openai.APITimeoutError,
-                        openai.InternalServerError,
-                    )):
+                    if not isinstance(
+                        exc,
+                        (
+                            openai.APIConnectionError,
+                            openai.APITimeoutError,
+                            openai.InternalServerError,
+                        ),
+                    ):
                        raise
                    local_last_exc = exc
                    if attempt >= 2:
                        break
-                    time.sleep(2 ** attempt)
+                    time.sleep(2**attempt)
            assert local_last_exc is not None
            raise local_last_exc

@@ -266,6 +272,7 @@ def _wrap_create(orig: Callable[..., Any]) -> Callable[..., Any]:
                # stay parseable in the runner log.
                try:
                    import sys
+
                    print(
                        f"[openai-retry] attempt {attempt + 1}/{_MAX_RETRIES} "
                        f"{type(exc).__name__}: {str(exc)[:120]} — "
@@ -333,9 +340,7 @@ def patch_openai_globally() -> None:
            from openai.resources.chat import completions as _comp_mod_async

            cls = getattr(_comp_mod_async, "AsyncCompletions", None)
-            if cls is not None and not getattr(
-                cls.create, "_hybrid_patched", False
-            ):
+            if cls is not None and not getattr(cls.create, "_hybrid_patched", False):
                # Async wrapper is structurally different — only patch
                # the bumped defaults via __init__; full retry loop on
                # async would need an async wrapper. Leave that for the
@@ -11,24 +11,27 @@ from __future__ import annotations

 # USD per million tokens, (input, output). Local models = 0.
 PRICES: dict[str, tuple[float, float]] = {
-    "claude-opus-4-7":             (5.00, 25.0),
-    "claude-sonnet-4-6":           (3.00, 15.0),
-    "claude-haiku-4-5":            (1.00, 5.00),
-    "claude-haiku-4-5-20251001":   (1.00, 5.00),
-    "gpt-5":                       (1.25, 10.0),
-    "gpt-5-mini":                  (0.25, 2.00),
-    "gpt-5-mini-2025-08-07":       (0.25, 2.00),
-    "gpt-4o":                      (0.15, 0.60),
-    # Gemini Developer API prices (USD per 1M tokens), 2025-12 list price.
-    # 2.5 Pro uses tiered pricing (>200K context = $2.50/$15); we charge the
-    # low-context tier since GAIA / SWE-bench prompts stay well under 200K.
-    "gemini-2.5-pro":              (1.25, 10.0),
-    "gemini-2.5-flash":            (0.30, 2.50),
-    "gemini-2.5-flash-lite":       (0.10, 0.40),
+    "claude-opus-4-7": (5.00, 25.0),
+    "claude-sonnet-4-6": (3.00, 15.0),
+    "claude-haiku-4-5": (1.00, 5.00),
+    "claude-haiku-4-5-20251001": (1.00, 5.00),
+    "gpt-5.5": (5.00, 30.0),
+    "gpt-5": (1.25, 10.0),
+    "gpt-5-mini": (0.25, 2.00),
+    "gpt-5-mini-2025-08-07": (0.25, 2.00),
+    "gpt-4o": (0.15, 0.60),
+    # Gemini Developer API prices (USD per 1M tokens). Pro models use tiered
+    # pricing above 200K prompt tokens; GAIA prompts stay under that tier, so
+    # charge the low-context standard rate.
+    "gemini-3.1-pro-preview": (2.00, 12.0),
+    "gemini-3.1-pro-preview-customtools": (2.00, 12.0),
+    "gemini-2.5-pro": (1.25, 10.0),
+    "gemini-2.5-flash": (0.30, 2.50),
+    "gemini-2.5-flash-lite": (0.10, 0.40),
    # OpenRouter slugs (used by toolorchestra paper-match pool).
    # Prices are OpenRouter list (USD/1M tokens), 2026-05 snapshot.
-    "qwen/qwen-2.5-coder-32b-instruct":  (0.08, 0.18),
-    "qwen/qwen3-32b":                    (0.10, 0.30),
+    "qwen/qwen-2.5-coder-32b-instruct": (0.08, 0.18),
+    "qwen/qwen3-32b": (0.10, 0.30),
    "meta-llama/llama-3.3-70b-instruct": (0.13, 0.39),
 }

@@ -61,7 +64,7 @@ def is_reasoning_model(model: str) -> bool:
    before emitting visible answer text. At max_tokens=4096 these silently
    truncate with empty answers on GAIA (26/100 GPT-5, 18/100 Gemini Pro)."""
    m = (model or "").lower()
-    return is_gpt5_family(model) or "gemini-2.5-pro" in m
+    return is_gpt5_family(model) or "gemini-2.5-pro" in m or "gemini-3.1-pro" in m


 def default_max_output_tokens(model: str) -> int:
@@ -34,6 +34,7 @@ from openjarvis.agents.hybrid._base import (
    WEB_SEARCH_COST_PER_CALL,
    LocalCloudAgent,
    build_web_search_tool,
+    tavily_search_context,
    web_search_cfg,
 )
 from openjarvis.agents.hybrid.mini_swe_agent import (
@@ -79,9 +80,7 @@ def _resolve_local_model(endpoint: str, registry_model: str) -> str:
    a model id (e.g. ``Qwen3.5-9B``) that's different from what's loaded.
    """
    try:
-        with urllib.request.urlopen(
-            endpoint.rstrip("/") + "/models", timeout=5
-        ) as r:
+        with urllib.request.urlopen(endpoint.rstrip("/") + "/models", timeout=5) as r:
            data = json.loads(r.read())
        served = [m["id"] for m in data.get("data", [])]
    except Exception:
@@ -135,7 +134,12 @@ class AdvisorsAgent(LocalCloudAgent):
        advisor_temperature = float(cfg.get("advisor_temperature", 0.2))

        ws_enabled, ws_max_uses = web_search_cfg(cfg)
-        if ws_enabled and self._cloud_endpoint not in _SEARCH_CAPABLE_ENDPOINTS:
+        search_backend = str(cfg.get("search_backend", "provider")).lower()
+        if (
+            ws_enabled
+            and search_backend != "tavily"
+            and self._cloud_endpoint not in _SEARCH_CAPABLE_ENDPOINTS
+        ):
            raise ValueError(
                f"web_search.enabled=true but cloud_endpoint={self._cloud_endpoint!r}; "
                "server-side web_search is wired for anthropic / openai / gemini "
@@ -146,19 +150,24 @@ class AdvisorsAgent(LocalCloudAgent):
        use_ws = ws_enabled
        gaia_max_turns = int(cfg.get("gaia_max_turns", 8))
        n_searches_total = 0
+        search_cost_total = 0.0

        # 1. Initial executor pass — advisor (Qwen) doesn't get tools;
        # only the cloud executor passes do. With web_search on, dispatch
        # to the search-capable agent loop for the configured provider.
        if use_ws:
-            initial_resp, e1_in, e1_out, n_s1, e1_turns = self._executor_search(
-                user=f"Question:\n{question}",
-                system=EXECUTOR_INITIAL_SYS,
-                max_tokens=executor_max_tokens,
-                ws_max_uses=ws_max_uses,
-                max_turns=gaia_max_turns,
+            (initial_resp, e1_in, e1_out, n_s1, e1_turns, e1_search_cost) = (
+                self._executor_search(
+                    user=f"Question:\n{question}",
+                    system=EXECUTOR_INITIAL_SYS,
+                    max_tokens=executor_max_tokens,
+                    ws_max_uses=ws_max_uses,
+                    max_turns=gaia_max_turns,
+                    query=question,
+                )
            )
            n_searches_total += n_s1
+            search_cost_total += e1_search_cost
        else:
            initial_resp, e1_in, e1_out = self._call_cloud(
                user=f"Question:\n{question}",
@@ -176,7 +185,8 @@ class AdvisorsAgent(LocalCloudAgent):
            )
        local_model = _resolve_local_model(self._local_endpoint, self._local_model)
        advisor_prompt = ADVISOR_TEMPLATE.format(
-            question=question, initial_response=initial_resp,
+            question=question,
+            initial_response=initial_resp,
        )
        advisor_text, adv_in, adv_out = self._call_vllm(
            local_model,
@@ -196,14 +206,18 @@ class AdvisorsAgent(LocalCloudAgent):
            f"answer-format rules."
        )
        if use_ws:
-            final_answer, e2_in, e2_out, n_s2, e2_turns = self._executor_search(
-                user=final_user,
-                system=EXECUTOR_FINAL_SYS,
-                max_tokens=executor_max_tokens,
-                ws_max_uses=ws_max_uses,
-                max_turns=gaia_max_turns,
+            (final_answer, e2_in, e2_out, n_s2, e2_turns, e2_search_cost) = (
+                self._executor_search(
+                    user=final_user,
+                    system=EXECUTOR_FINAL_SYS,
+                    max_tokens=executor_max_tokens,
+                    ws_max_uses=ws_max_uses,
+                    max_turns=gaia_max_turns,
+                    query=question,
+                )
            )
            n_searches_total += n_s2
+            search_cost_total += e2_search_cost
        else:
            final_answer, e2_in, e2_out = self._call_cloud(
                user=final_user,
@@ -216,7 +230,10 @@ class AdvisorsAgent(LocalCloudAgent):
        tokens_local = adv_in + adv_out
        tokens_cloud = e1_in + e1_out + e2_in + e2_out
        cost = self.cost_usd(self._cloud_model, e1_in + e2_in, e1_out + e2_out)
-        cost += n_searches_total * _search_cost_per_call(self._cloud_endpoint)
+        if search_backend == "tavily":
+            cost += search_cost_total
+        else:
+            cost += n_searches_total * _search_cost_per_call(self._cloud_endpoint)

        meta: Dict[str, Any] = {
            "tokens_local": tokens_local,
@@ -233,7 +250,9 @@ class AdvisorsAgent(LocalCloudAgent):
                "initial_response": initial_resp,
                "advisor_feedback": advisor_text,
                "web_search_enabled": use_ws,
+                "search_backend": search_backend,
                "n_web_searches": n_searches_total,
+                "search_cost_usd": search_cost_total,
                "note": "inference-only advisor (untrained); lower bound on the technique.",
            },
        }
@@ -251,16 +270,40 @@ class AdvisorsAgent(LocalCloudAgent):
        max_tokens: int,
        ws_max_uses: int,
        max_turns: int,
-    ) -> Tuple[str, int, int, int, int]:
+        query: Optional[str] = None,
+    ) -> Tuple[str, int, int, int, int, float]:
        """Run a search-capable executor pass for the configured cloud.

        Dispatches by ``self._cloud_endpoint`` to the matching ``_base``
-        agent loop. Returns the shared 5-tuple ``(text, p_tok, c_tok,
-        n_searches, turns)``. The endpoint is assumed already validated
-        against ``_SEARCH_CAPABLE_ENDPOINTS`` by the caller.
+        agent loop, or through Tavily when ``method_cfg.search_backend`` is
+        ``"tavily"``. Returns ``(text, p_tok, c_tok, n_searches, turns,
+        search_cost_usd)``.
        """
+        if str(self._cfg.get("search_backend", "provider")).lower() == "tavily":
+            res = tavily_search_context(
+                query or user,
+                max_results=int(self._cfg.get("tavily_max_results", 5)),
+            )
+            grounded_user = (
+                f"Web search results:\n{res['text']}\n\n"
+                f"Using the search results above, answer this request:\n{user}"
+            )
+            text, p, c = self._call_cloud(
+                user=grounded_user,
+                system=system,
+                max_tokens=max_tokens,
+                temperature=0.0,
+            )
+            return (
+                text,
+                p,
+                c,
+                int(res["n_searches"]),
+                1,
+                float(res["cost_usd"]),
+            )
        if self._cloud_endpoint == "anthropic":
-            return self._call_anthropic_agent(
+            text, p, c, n_searches, turns = self._call_anthropic_agent(
                self._cloud_model,
                user=user,
                system=system,
@@ -269,8 +312,9 @@ class AdvisorsAgent(LocalCloudAgent):
                tools=[build_web_search_tool(ws_max_uses)],
                max_turns=max_turns,
            )
+            return text, p, c, n_searches, turns, 0.0
        if self._cloud_endpoint == "openai":
-            return self._call_openai_agent(
+            text, p, c, n_searches, turns = self._call_openai_agent(
                self._cloud_model,
                user=user,
                system=system,
@@ -278,8 +322,9 @@ class AdvisorsAgent(LocalCloudAgent):
                temperature=0.0,
                max_turns=max_turns,
            )
+            return text, p, c, n_searches, turns, 0.0
        if self._cloud_endpoint == "gemini":
-            return self._call_gemini_agent(
+            text, p, c, n_searches, turns = self._call_gemini_agent(
                self._cloud_model,
                user=user,
                system=system,
@@ -287,6 +332,7 @@ class AdvisorsAgent(LocalCloudAgent):
                temperature=0.0,
                max_turns=max_turns,
            )
+            return text, p, c, n_searches, turns, 0.0
        # Genuinely unsupported (openrouter / vllm / unknown). The caller
        # guard should have caught this; raise defensively.
        raise ValueError(
@@ -379,8 +425,10 @@ class AdvisorsAgent(LocalCloudAgent):

        tokens_local = adv_in + adv_out
        tokens_cloud = (
-            initial_out["tokens_in"] + initial_out["tokens_out"]
-            + final_out["tokens_in"] + final_out["tokens_out"]
+            initial_out["tokens_in"]
+            + initial_out["tokens_out"]
+            + final_out["tokens_in"]
+            + final_out["tokens_out"]
        )
        cost = initial_out["cost_usd"] + final_out["cost_usd"]
        meta: Dict[str, Any] = {
@@ -73,6 +73,7 @@ ARCHON_SWE_RANKER_SYS = (

 # ---------- Stubs for Archon's eager-imported heavy deps we don't need ----------

+
 def _stub_archon_imports() -> None:
    """``utils.py`` imports groq/google/litellm/dotenv at module load. Stub
    the ones we don't use so the import chain doesn't fail when those
@@ -97,6 +98,7 @@ def _add_archon_to_path() -> None:

 # ---------- Anthropic patch for Opus 4.7 ----------

+
 def _patch_anthropic_for_opus() -> None:
    from anthropic.resources.messages import messages as _msgs_mod

@@ -129,8 +131,10 @@ def _tally() -> Dict[str, int]:
    counts = getattr(_TALLY_LOCAL, "counts", None)
    if counts is None:
        counts = {
-            "cloud_prompt": 0, "cloud_completion": 0,
-            "local_prompt": 0, "local_completion": 0,
+            "cloud_prompt": 0,
+            "cloud_completion": 0,
+            "local_prompt": 0,
+            "local_completion": 0,
            "n_web_searches": 0,
        }
        _TALLY_LOCAL.counts = counts
@@ -141,8 +145,10 @@ def _tally() -> Dict[str, int]:

 def _reset_tally() -> None:
    _TALLY_LOCAL.counts = {
-        "cloud_prompt": 0, "cloud_completion": 0,
-        "local_prompt": 0, "local_completion": 0,
+        "cloud_prompt": 0,
+        "cloud_completion": 0,
+        "local_prompt": 0,
+        "local_completion": 0,
        "n_web_searches": 0,
    }

@@ -169,6 +175,7 @@ def _make_local_generator(local_endpoint: str, local_model: str):

    def local_gen(model, messages, max_tokens=2048, temperature=0.7, **_kw):  # type: ignore[no-untyped-def]
        import time as _time
+
        t0 = _time.time()
        try:
            resp = client.chat.completions.create(
@@ -179,31 +186,35 @@ def _make_local_generator(local_endpoint: str, local_model: str):
            )
            _bump_local_calls()
        except Exception as e:
-            _record_event({
-                "kind": "archon_local_gen_error",
-                "model": local_model,
-                "messages": messages,
-                "error": f"{type(e).__name__}: {e}",
-                "ts": _time.time(),
-            })
+            _record_event(
+                {
+                    "kind": "archon_local_gen_error",
+                    "model": local_model,
+                    "messages": messages,
+                    "error": f"{type(e).__name__}: {e}",
+                    "ts": _time.time(),
+                }
+            )
            return f"[local-vllm error: {e!r}]"
        u = resp.usage
        if u:
            _tally()["local_prompt"] += getattr(u, "prompt_tokens", 0) or 0
            _tally()["local_completion"] += getattr(u, "completion_tokens", 0) or 0
        text = (resp.choices[0].message.content or "").strip()
-        _record_event({
-            "kind": "archon_local_gen",
-            "model": local_model,
-            "messages": messages,
-            "response": text,
-            "tokens_in": getattr(u, "prompt_tokens", 0) if u else 0,
-            "tokens_out": getattr(u, "completion_tokens", 0) if u else 0,
-            "temperature": temperature,
-            "max_tokens": max_tokens,
-            "latency_s": _time.time() - t0,
-            "ts": _time.time(),
-        })
+        _record_event(
+            {
+                "kind": "archon_local_gen",
+                "model": local_model,
+                "messages": messages,
+                "response": text,
+                "tokens_in": getattr(u, "prompt_tokens", 0) if u else 0,
+                "tokens_out": getattr(u, "completion_tokens", 0) if u else 0,
+                "temperature": temperature,
+                "max_tokens": max_tokens,
+                "latency_s": _time.time() - t0,
+                "ts": _time.time(),
+            }
+        )
        return text

    return local_gen
@@ -218,10 +229,13 @@ def _wrap_archon_cloud_generators() -> None:

    def gen_openai(model, messages, max_tokens=2048, temperature=0.7, **_kw):  # type: ignore[no-untyped-def]
        import time as _time
+
        client = _OAI()
        kwargs: Dict[str, Any] = dict(
-            model=model, messages=messages,
-            max_tokens=max_tokens, temperature=temperature,
+            model=model,
+            messages=messages,
+            max_tokens=max_tokens,
+            temperature=temperature,
        )
        # GPT-5/o1/o3 reject non-default temperature and use max_completion_tokens.
        if model.startswith(("gpt-5", "o1", "o3")):
@@ -236,20 +250,23 @@ def _wrap_archon_cloud_generators() -> None:
            _tally()["cloud_prompt"] += getattr(u, "prompt_tokens", 0) or 0
            _tally()["cloud_completion"] += getattr(u, "completion_tokens", 0) or 0
        text = (resp.choices[0].message.content or "").strip()
-        _record_event({
-            "kind": "archon_cloud_openai",
-            "model": model,
-            "messages": messages,
-            "response": text,
-            "tokens_in": getattr(u, "prompt_tokens", 0) if u else 0,
-            "tokens_out": getattr(u, "completion_tokens", 0) if u else 0,
-            "latency_s": _time.time() - t0,
-            "ts": _time.time(),
-        })
+        _record_event(
+            {
+                "kind": "archon_cloud_openai",
+                "model": model,
+                "messages": messages,
+                "response": text,
+                "tokens_in": getattr(u, "prompt_tokens", 0) if u else 0,
+                "tokens_out": getattr(u, "completion_tokens", 0) if u else 0,
+                "latency_s": _time.time() - t0,
+                "ts": _time.time(),
+            }
+        )
        return text

    def gen_anthropic(model, messages, max_tokens=2048, temperature=0.7, **_kw):  # type: ignore[no-untyped-def]
        import time as _time
+
        client = _anth.Anthropic(timeout=600.0)
        system = ""
        msgs = []
@@ -259,7 +276,10 @@ def _wrap_archon_cloud_generators() -> None:
            else:
                msgs.append(m)
        kwargs: Dict[str, Any] = dict(
-            model=model, system=system, messages=msgs, max_tokens=max_tokens,
+            model=model,
+            system=system,
+            messages=msgs,
+            max_tokens=max_tokens,
        )
        if not model.startswith(NO_TEMP_PREFIXES):
            kwargs["temperature"] = temperature
@@ -277,24 +297,27 @@ def _wrap_archon_cloud_generators() -> None:
        srv = getattr(u, "server_tool_use", None) if u else None
        n_searches = getattr(srv, "web_search_requests", 0) if srv else 0
        _tally()["n_web_searches"] += int(n_searches)
-        _record_event({
-            "kind": "archon_cloud_anthropic",
-            "model": model,
-            "system": system,
-            "messages": msgs,
-            "response": text.strip(),
-            "tokens_in": getattr(u, "input_tokens", 0) if u else 0,
-            "tokens_out": getattr(u, "output_tokens", 0) if u else 0,
-            "n_web_searches": int(n_searches),
-            "tools_declared": kwargs.get("tools"),
-            "latency_s": _time.time() - t0,
-            "ts": _time.time(),
-        })
+        _record_event(
+            {
+                "kind": "archon_cloud_anthropic",
+                "model": model,
+                "system": system,
+                "messages": msgs,
+                "response": text.strip(),
+                "tokens_in": getattr(u, "input_tokens", 0) if u else 0,
+                "tokens_out": getattr(u, "output_tokens", 0) if u else 0,
+                "n_web_searches": int(n_searches),
+                "tools_declared": kwargs.get("tools"),
+                "latency_s": _time.time() - t0,
+                "ts": _time.time(),
+            }
+        )
        return text.strip()

    from archon.completions.components.Generator import (
        GENERATE_MAP as _GMAP,  # type: ignore[import-not-found]
    )
+
    _GMAP["OpenAI_API"] = gen_openai
    _GMAP["Anthropic_API"] = gen_anthropic

@@ -330,7 +353,9 @@ def _patch_archon_prompts() -> None:
    orig = _p.make_fuser_prompt

    def patched(conv, references, critiques=None, length_control=False):  # type: ignore[no-untyped-def]
-        base = orig(conv, references, critiques=critiques, length_control=length_control)
+        base = orig(
+            conv, references, critiques=critiques, length_control=length_control
+        )
        return base + _FUSER_FORMAT_REMINDER

    patched._hybrid_format_patched = True  # type: ignore[attr-defined]
@@ -339,6 +364,7 @@ def _patch_archon_prompts() -> None:
    from archon.completions.components import (
        Fuser as _F,  # type: ignore[import-not-found]
    )
+
    _F.make_fuser_prompt = patched


@@ -354,6 +380,7 @@ def _apply_patches_once() -> None:
    _patch_anthropic_for_opus()
    # Trigger Archon imports so GENERATE_MAP exists.
    import archon.completions.components.Generator  # type: ignore[import-not-found]  # noqa: F401
+
    _wrap_archon_cloud_generators()
    _patch_archon_prompts()
    _PATCHES_APPLIED = True
@@ -361,49 +388,62 @@ def _apply_patches_once() -> None:

 # ---------- Architecture presets ----------

+
 def _presets():
    return {
        "ensemble_rank_fuse": lambda K, local_model, ranker_model, fuser_model, max_tokens, temperature: [
-            [{
-                "type": "generator",
-                "model": local_model,
-                "model_type": "vllm_local",
-                "top_k": 1,
-                "temperature": temperature,
-                "max_tokens": max_tokens,
-                "samples": K,
-            }],
-            [{
-                "type": "ranker",
-                "model": ranker_model,
-                "model_type": "Anthropic_API" if ranker_model.startswith("claude") else "OpenAI_API",
-                "top_k": min(K, 5),
-                "temperature": 0.0,
-                "max_tokens": max_tokens,
-            }],
-            [{
-                "type": "fuser",
-                "model": fuser_model,
-                "model_type": "Anthropic_API" if fuser_model.startswith("claude") else "OpenAI_API",
-                "temperature": 0.0,
-                "max_tokens": max_tokens,
-                "samples": 1,
-            }],
+            [
+                {
+                    "type": "generator",
+                    "model": local_model,
+                    "model_type": "vllm_local",
+                    "top_k": 1,
+                    "temperature": temperature,
+                    "max_tokens": max_tokens,
+                    "samples": K,
+                }
+            ],
+            [
+                {
+                    "type": "ranker",
+                    "model": ranker_model,
+                    "model_type": "Anthropic_API"
+                    if ranker_model.startswith("claude")
+                    else "OpenAI_API",
+                    "top_k": min(K, 5),
+                    "temperature": 0.0,
+                    "max_tokens": max_tokens,
+                }
+            ],
+            [
+                {
+                    "type": "fuser",
+                    "model": fuser_model,
+                    "model_type": "Anthropic_API"
+                    if fuser_model.startswith("claude")
+                    else "OpenAI_API",
+                    "temperature": 0.0,
+                    "max_tokens": max_tokens,
+                    "samples": 1,
+                }
+            ],
        ],
        # ``single_local`` honors the cfg ``max_tokens`` (passed positionally
        # like ``ensemble_rank_fuse``). Previously it hard-coded 2048, which
        # cut Qwen off mid-reasoning before it could emit the GAIA
        # ``FINAL ANSWER:`` line — the scorer then had nothing to extract.
        "single_local": lambda K, local_model, ranker_model, fuser_model, max_tokens, temperature: [
-            [{
-                "type": "generator",
-                "model": local_model,
-                "model_type": "vllm_local",
-                "top_k": 1,
-                "temperature": 0.0,
-                "max_tokens": max_tokens,
-                "samples": 1,
-            }],
+            [
+                {
+                    "type": "generator",
+                    "model": local_model,
+                    "model_type": "vllm_local",
+                    "top_k": 1,
+                    "temperature": 0.0,
+                    "max_tokens": max_tokens,
+                    "samples": 1,
+                }
+            ],
        ],
    }

@@ -464,7 +504,12 @@ class ArchonAgent(LocalCloudAgent):
        )

        layers = presets[arch](
-            K, self._local_model, ranker_model, fuser_model, max_tokens, temperature,
+            K,
+            self._local_model,
+            ranker_model,
+            fuser_model,
+            max_tokens,
+            temperature,
        )
        archon_cfg = {"name": f"hybrid-archon-{arch}", "layers": layers}

@@ -480,10 +525,12 @@ class ArchonAgent(LocalCloudAgent):
        archon = Archon(archon_cfg)

        try:
-            answer = archon.generate([
-                {"role": "system", "content": "You are a helpful assistant."},
-                {"role": "user", "content": input},
-            ])
+            answer = archon.generate(
+                [
+                    {"role": "system", "content": "You are a helpful assistant."},
+                    {"role": "user", "content": input},
+                ]
+            )
        except Exception:
            # Re-raise so the base ``run()`` / runner's ``_run_one_inner``
            # records this in the row's ``error`` field instead of stashing
@@ -518,10 +565,10 @@ class ArchonAgent(LocalCloudAgent):
            "tool_calls": int(n_searches),
            "traces": {
                "architecture": arch,
-                "n_samples":    K,
+                "n_samples": K,
                "ranker_model": ranker_model,
-                "fuser_model":  fuser_model,
-                "local_model":  self._local_model,
+                "fuser_model": fuser_model,
+                "local_model": self._local_model,
                "tokens_breakdown": dict(_tally()),
                "web_search_enabled": ws_enabled,
                "n_web_searches": n_searches,
@@ -568,26 +615,30 @@ class ArchonAgent(LocalCloudAgent):
                turn_max_tokens=turn_max_tokens,
                trace_prefix=f"archon_gen{k}",
            )
-            candidates.append({
-                "idx": k,
-                "summary": out["final_summary"],
-                "patch": out["patch"],
-                "framed": out["answer"],
-                "tokens_in": out["tokens_in"],
-                "tokens_out": out["tokens_out"],
-                "turns": out["turns"],
-            })
+            candidates.append(
+                {
+                    "idx": k,
+                    "summary": out["final_summary"],
+                    "patch": out["patch"],
+                    "framed": out["answer"],
+                    "tokens_in": out["tokens_in"],
+                    "tokens_out": out["tokens_out"],
+                    "turns": out["turns"],
+                }
+            )
            total_tokens_local += out["tokens_in"] + out["tokens_out"]
-            self.record_trace_event({
-                "kind": "archon_swe_candidate",
-                "idx": k,
-                "patch_chars": len(out["patch"]),
-                "summary": out["final_summary"],
-            })
+            self.record_trace_event(
+                {
+                    "kind": "archon_swe_candidate",
+                    "idx": k,
+                    "patch_chars": len(out["patch"]),
+                    "summary": out["final_summary"],
+                }
+            )

        # Ranker: cloud picks the best candidate.
        ranker_user = (
-            f"Issue:\n{task.get('problem_statement','')}\n\n"
+            f"Issue:\n{task.get('problem_statement', '')}\n\n"
            f"K = {K} candidate patches:\n\n"
            + "\n\n".join(
                f"=== Candidate {c['idx']} ===\nSummary: {c['summary']}\n"
@@ -615,13 +666,15 @@ class ArchonAgent(LocalCloudAgent):
            chosen_idx = 0
        chosen = candidates[chosen_idx]

-        self.record_trace_event({
-            "kind": "archon_swe_rank",
-            "chosen_idx": chosen_idx,
-            "ranker_raw": ranker_text,
-            "tokens_in": r_in,
-            "tokens_out": r_out,
-        })
+        self.record_trace_event(
+            {
+                "kind": "archon_swe_rank",
+                "chosen_idx": chosen_idx,
+                "ranker_raw": ranker_text,
+                "tokens_in": r_in,
+                "tokens_out": r_out,
+            }
+        )

        meta = {
            "tokens_local": total_tokens_local,
@@ -635,8 +688,12 @@ class ArchonAgent(LocalCloudAgent):
                "swe_mode": True,
                "K": K,
                "candidates": [
-                    {"idx": c["idx"], "summary": c["summary"],
-                     "patch_chars": len(c["patch"]), "turns": c["turns"]}
+                    {
+                        "idx": c["idx"],
+                        "summary": c["summary"],
+                        "patch_chars": len(c["patch"]),
+                        "turns": c["turns"],
+                    }
                    for c in candidates
                ],
                "chosen_idx": chosen_idx,
@@ -82,7 +82,11 @@ class BaselineCloudAgent(LocalCloudAgent):
                max_turns=int(cfg.get("swe_max_turns", 30)),
                bash_timeout=int(cfg.get("swe_bash_timeout_s", 120)),
                output_cap=int(cfg.get("swe_output_cap", 10_000)),
-                turn_max_tokens=int(cfg.get("cloud_max_tokens", default_max_output_tokens(self._cloud_model))),
+                turn_max_tokens=int(
+                    cfg.get(
+                        "cloud_max_tokens", default_max_output_tokens(self._cloud_model)
+                    )
+                ),
                trace_prefix="baseline_cloud",
            )
            meta = {
@@ -112,7 +116,11 @@ class BaselineCloudAgent(LocalCloudAgent):
            text, p_tok, c_tok, n_searches, turns = self._call_anthropic_agent(
                self._cloud_model,
                user=input,
-                max_tokens=int(cfg.get("cloud_max_tokens", default_max_output_tokens(self._cloud_model))),
+                max_tokens=int(
+                    cfg.get(
+                        "cloud_max_tokens", default_max_output_tokens(self._cloud_model)
+                    )
+                ),
                temperature=0.0,
                tools=[build_web_search_tool(ws_max_uses)],
                max_turns=gaia_max_turns,
@@ -145,17 +153,23 @@ class BaselineCloudAgent(LocalCloudAgent):
            # wired here. Skip cleanly rather than fake one. Cells that
            # want web_search must run on Anthropic until those backends
            # are wired.
-            self.record_trace_event({
-                "kind": "web_search_skipped",
-                "reason": "non_anthropic_endpoint",
-                "endpoint": self._cloud_endpoint,
-            })
+            self.record_trace_event(
+                {
+                    "kind": "web_search_skipped",
+                    "reason": "non_anthropic_endpoint",
+                    "endpoint": self._cloud_endpoint,
+                }
+            )

        # One-shot direct cloud call. GAIA only — SWE goes through the
        # mini-SWE-agent loop above (now supports anthropic/openai/gemini).
        text, p_tok, c_tok = self._call_cloud(
            user=input,
-            max_tokens=int(cfg.get("cloud_max_tokens", default_max_output_tokens(self._cloud_model))),
+            max_tokens=int(
+                cfg.get(
+                    "cloud_max_tokens", default_max_output_tokens(self._cloud_model)
+                )
+            ),
            temperature=0.0,
        )
        meta = {
@@ -46,6 +46,7 @@ from openjarvis.agents.hybrid._base import (
    WEB_SEARCH_COST_PER_CALL,
    LocalCloudAgent,
    build_web_search_tool,
+    tavily_search_context,
    web_search_cfg,
 )
 from openjarvis.agents.hybrid._prices import (
@@ -88,19 +89,20 @@ CONDUCTOR_STRICTER = (
    "Your previous response was not valid JSON or was missing required fields. "
    "Reply with ONLY a single JSON object — no prose, no code fences, no commentary "
    "— containing exactly the three keys model_id (list[int]), subtasks (list[str]), "
-    "and access_list (list[list[int] or \"all\"]) of equal length, at most 5 entries, "
+    'and access_list (list[list[int] or "all"]) of equal length, at most 5 entries, '
    "and access_list[0] must be [] (an empty list)."
 )


 # ---------- Plan parsing ----------

+
 def _strip_fences(s: str) -> str:
    s = s.strip()
    if s.startswith("```"):
        first_nl = s.find("\n")
        if first_nl != -1:
-            s = s[first_nl + 1:]
+            s = s[first_nl + 1 :]
        if s.endswith("```"):
            s = s[:-3]
        s = s.strip()
@@ -118,9 +120,7 @@ def _try_literal(s: str):
    """Fallback for the paper's literal Python-list output style."""
    out = {}
    for key in ("model_id", "subtasks", "access_list"):
-        m = re.search(
-            rf"{key}\s*=\s*(\[[^\]]*\](?:\s*\+\s*\[[^\]]*\])*)", s, re.DOTALL
-        )
+        m = re.search(rf"{key}\s*=\s*(\[[^\]]*\](?:\s*\+\s*\[[^\]]*\])*)", s, re.DOTALL)
        if not m:
            return None
        try:
@@ -153,7 +153,7 @@ def _validate_plan(plan: Any, n_workers: int) -> Optional[str]:
        if a == "all":
            continue
        if not isinstance(a, list):
-            return f"access_list[{i}] must be list or \"all\""
+            return f'access_list[{i}] must be list or "all"'
        for j in a:
            if not isinstance(j, int) or not (0 <= j < i):
                return f"access_list[{i}] has bad ref {j!r}"
@@ -173,17 +173,18 @@ def _parse_plan(text: str, n_workers: int):

 # ---------- Worker pool ----------

+
 def _vllm_alive(base_url: str) -> bool:
    try:
-        with urllib.request.urlopen(
-            base_url.rstrip("/") + "/models", timeout=3
-        ) as r:
+        with urllib.request.urlopen(base_url.rstrip("/") + "/models", timeout=3) as r:
            return r.status == 200
    except Exception:
        return False


-def _default_pool(local_model: Optional[str], local_endpoint: Optional[str]) -> List[Dict[str, Any]]:
+def _default_pool(
+    local_model: Optional[str], local_endpoint: Optional[str]
+) -> List[Dict[str, Any]]:
    """Default worker pool — faithful to the Sakana Conductor paper (arXiv 2512.04388).

    The paper composes a heterogeneous 7-worker pool spanning three frontier
@@ -205,98 +206,112 @@ def _default_pool(local_model: Optional[str], local_endpoint: Optional[str]) ->
    del local_model, local_endpoint  # paper default carries no local worker
    pool: List[Dict[str, Any]] = []
    if not os.environ.get("OJ_CONDUCTOR_DISABLE_GEMINI"):
-        pool.append({
-            "id": len(pool),
-            "name": "gemini-pro",
-            "endpoint": "gemini",
-            "model": "gemini-2.5-pro",
-            "description": (
-                "Google Gemini 2.5 Pro. Frontier multimodal reasoner with a "
-                "very large context window. Strong at long-document synthesis, "
-                "multi-hop factual reasoning, and tasks that benefit from "
-                "wide retrieval. Slower and pricier than mid-tier workers."
-            ),
-        })
+        pool.append(
+            {
+                "id": len(pool),
+                "name": "gemini-pro",
+                "endpoint": "gemini",
+                "model": "gemini-2.5-pro",
+                "description": (
+                    "Google Gemini 2.5 Pro. Frontier multimodal reasoner with a "
+                    "very large context window. Strong at long-document synthesis, "
+                    "multi-hop factual reasoning, and tasks that benefit from "
+                    "wide retrieval. Slower and pricier than mid-tier workers."
+                ),
+            }
+        )
    if not os.environ.get("OJ_CONDUCTOR_DISABLE_ANTHROPIC"):
-        pool.append({
-            "id": len(pool),
-            "name": "claude-sonnet-4",
-            "endpoint": "anthropic",
-            "model": "claude-sonnet-4-6",
-            "description": (
-                "Anthropic Claude Sonnet 4. Strong general-purpose reasoner "
-                "with careful instruction following and reliable formatting. "
-                "Good default for code, structured writing, and decisive "
-                "steps where accuracy matters more than raw throughput."
-            ),
-        })
+        pool.append(
+            {
+                "id": len(pool),
+                "name": "claude-sonnet-4",
+                "endpoint": "anthropic",
+                "model": "claude-sonnet-4-6",
+                "description": (
+                    "Anthropic Claude Sonnet 4. Strong general-purpose reasoner "
+                    "with careful instruction following and reliable formatting. "
+                    "Good default for code, structured writing, and decisive "
+                    "steps where accuracy matters more than raw throughput."
+                ),
+            }
+        )
    if not os.environ.get("OJ_CONDUCTOR_DISABLE_OPENAI"):
-        pool.append({
-            "id": len(pool),
-            "name": "gpt-5",
-            "endpoint": "openai",
-            "model": "gpt-5",
-            "description": (
-                "OpenAI GPT-5. Frontier-tier broad-knowledge model. Best for "
-                "open-domain factual recall, creative generation, and "
-                "ambiguous questions where coverage matters. Expensive; use "
-                "for steps where breadth of world knowledge is the bottleneck."
-            ),
-        })
+        pool.append(
+            {
+                "id": len(pool),
+                "name": "gpt-5",
+                "endpoint": "openai",
+                "model": "gpt-5",
+                "description": (
+                    "OpenAI GPT-5. Frontier-tier broad-knowledge model. Best for "
+                    "open-domain factual recall, creative generation, and "
+                    "ambiguous questions where coverage matters. Expensive; use "
+                    "for steps where breadth of world knowledge is the bottleneck."
+                ),
+            }
+        )
    if not os.environ.get("OJ_CONDUCTOR_DISABLE_OPENROUTER"):
-        pool.append({
-            "id": len(pool),
-            "name": "deepseek-r1-distill-qwen-32b",
-            "endpoint": "openrouter",
-            "model": "deepseek/deepseek-r1-distill-qwen-32b",
-            "description": (
-                "DeepSeek R1 distilled into Qwen-32B (open weights via "
-                "OpenRouter). Specialized for chain-of-thought math, logic, "
-                "and competitive-programming-style problems. Verbose; "
-                "produces extensive reasoning traces before the final answer."
-            ),
-        })
-        pool.append({
-            "id": len(pool),
-            "name": "gemma3-27b-it",
-            "endpoint": "openrouter",
-            "model": "google/gemma-3-27b-it",
-            "description": (
-                "Google Gemma 3 27B Instruct (open weights via OpenRouter). "
-                "Mid-size instruction-tuned model. Cheap and fast; solid at "
-                "concise summarization, extraction, and short-form Q&A on "
-                "given context. Weaker than the frontier workers on multi-step "
-                "reasoning."
-            ),
-        })
-        pool.append({
-            "id": len(pool),
-            "name": "qwen3-32b",
-            "endpoint": "openrouter",
-            "model": "qwen/qwen3-32b",
-            "description": (
-                "Qwen3-32B in non-thinking mode (open weights via OpenRouter). "
-                "Fast general-purpose dialogue and instruction following. "
-                "Use when the step is straightforward generation, "
-                "summarization, or formatting — does NOT spend tokens on "
-                "internal reasoning."
-            ),
-        })
-        pool.append({
-            "id": len(pool),
-            "name": "qwen3-32b-thinking",
-            "endpoint": "openrouter",
-            "model": "qwen/qwen3-32b",
-            "extra_body": {"reasoning": {"effort": "medium"}},
-            "description": (
-                "Qwen3-32B with reasoning enabled (open weights via "
-                "OpenRouter). Same backbone as 'qwen3-32b' but spends tokens "
-                "on an internal chain of thought before answering. Stronger "
-                "on math, code, and multi-step logic; slower and consumes "
-                "more completion tokens. Prefer this for hard reasoning "
-                "steps; prefer the non-thinking variant for plain dialogue."
-            ),
-        })
+        pool.append(
+            {
+                "id": len(pool),
+                "name": "deepseek-r1-distill-qwen-32b",
+                "endpoint": "openrouter",
+                "model": "deepseek/deepseek-r1-distill-qwen-32b",
+                "description": (
+                    "DeepSeek R1 distilled into Qwen-32B (open weights via "
+                    "OpenRouter). Specialized for chain-of-thought math, logic, "
+                    "and competitive-programming-style problems. Verbose; "
+                    "produces extensive reasoning traces before the final answer."
+                ),
+            }
+        )
+        pool.append(
+            {
+                "id": len(pool),
+                "name": "gemma3-27b-it",
+                "endpoint": "openrouter",
+                "model": "google/gemma-3-27b-it",
+                "description": (
+                    "Google Gemma 3 27B Instruct (open weights via OpenRouter). "
+                    "Mid-size instruction-tuned model. Cheap and fast; solid at "
+                    "concise summarization, extraction, and short-form Q&A on "
+                    "given context. Weaker than the frontier workers on multi-step "
+                    "reasoning."
+                ),
+            }
+        )
+        pool.append(
+            {
+                "id": len(pool),
+                "name": "qwen3-32b",
+                "endpoint": "openrouter",
+                "model": "qwen/qwen3-32b",
+                "description": (
+                    "Qwen3-32B in non-thinking mode (open weights via OpenRouter). "
+                    "Fast general-purpose dialogue and instruction following. "
+                    "Use when the step is straightforward generation, "
+                    "summarization, or formatting — does NOT spend tokens on "
+                    "internal reasoning."
+                ),
+            }
+        )
+        pool.append(
+            {
+                "id": len(pool),
+                "name": "qwen3-32b-thinking",
+                "endpoint": "openrouter",
+                "model": "qwen/qwen3-32b",
+                "extra_body": {"reasoning": {"effort": "medium"}},
+                "description": (
+                    "Qwen3-32B with reasoning enabled (open weights via "
+                    "OpenRouter). Same backbone as 'qwen3-32b' but spends tokens "
+                    "on an internal chain of thought before answering. Stronger "
+                    "on math, code, and multi-step logic; slower and consumes "
+                    "more completion tokens. Prefer this for hard reasoning "
+                    "steps; prefer the non-thinking variant for plain dialogue."
+                ),
+            }
+        )
    # Reassign ids contiguously in case env-gates skipped some entries.
    for new_id, entry in enumerate(pool):
        entry["id"] = new_id
@@ -362,16 +377,17 @@ def _resolve_worker_pool(
                f"Invalid worker_pool entry [{wid_repr}]: 'id' must be an int"
            )
        if wid in seen_ids:
-            raise ValueError(
-                f"Invalid worker_pool entry [{wid}]: duplicate id"
-            )
+            raise ValueError(f"Invalid worker_pool entry [{wid}]: duplicate id")
        seen_ids.add(wid)
        if not entry.get("name") or not isinstance(entry["name"], str):
            raise ValueError(
                f"Invalid worker_pool entry [{wid}]: 'name' must be a non-empty string"
            )
        endpoint = entry.get("endpoint") or entry.get("type")
-        if not isinstance(endpoint, str) or endpoint.lower() not in _CONDUCTOR_VALID_ENDPOINTS:
+        if (
+            not isinstance(endpoint, str)
+            or endpoint.lower() not in _CONDUCTOR_VALID_ENDPOINTS
+        ):
            raise ValueError(
                f"Invalid worker_pool entry [{wid}]: 'endpoint' must be one of "
                f"{_CONDUCTOR_VALID_ENDPOINTS} (got {endpoint!r})"
@@ -456,12 +472,18 @@ def _format_worker_pool(workers: List[Dict[str, Any]]) -> str:
    )


-def _search_capable_indices(workers: List[Dict[str, Any]]) -> List[int]:
+def _search_capable_indices(
+    workers: List[Dict[str, Any]],
+    *,
+    search_backend: str = "provider",
+) -> List[int]:
    """Indices of workers whose endpoint can run server-side web search."""
+    if search_backend == "tavily":
+        return [w["id"] for w in workers]
    return [
-        w["id"] for w in workers
-        if (w.get("endpoint") or "openai").lower()
-        in _SEARCH_CAPABLE_WORKER_ENDPOINTS
+        w["id"]
+        for w in workers
+        if (w.get("endpoint") or "openai").lower() in _SEARCH_CAPABLE_WORKER_ENDPOINTS
    ]


@@ -470,6 +492,7 @@ def _build_conductor_prompt(
    workers: List[Dict[str, Any]],
    *,
    web_search_enabled: bool = False,
+    search_backend: str = "provider",
 ) -> str:
    """Build the planner prompt.

@@ -485,12 +508,18 @@ def _build_conductor_prompt(
    )
    if not web_search_enabled:
        return base
-    capable = _search_capable_indices(workers)
+    capable = _search_capable_indices(workers, search_backend=search_backend)
    if capable:
        cap_str = ", ".join(str(i) for i in capable)
+        if search_backend == "tavily":
+            capability = (
+                "External Tavily search results will be prepended to worker prompts"
+            )
+        else:
+            capability = "Only these model indices can perform live web search"
        constraint = (
            "\n\nWEB SEARCH CONSTRAINT:\n"
-            f"Only these model indices can perform live web search: [{cap_str}]. "
+            f"{capability}: [{cap_str}]. "
            "Any step that needs to look up facts, current events, or other "
            "information not reliably known from memory MUST be routed to one "
            "of those indices. Steps routed to any other model can only use "
@@ -550,8 +579,8 @@ def _call_worker(
    *,
    web_search_tool: Optional[Dict[str, Any]] = None,
    web_search_max_uses: int = 8,
-) -> Tuple[str, int, int, bool, int]:
-    """Returns (text, p_tok, c_tok, is_local, n_web_searches).
+) -> Tuple[str, int, int, bool, int, float]:
+    """Returns (text, p_tok, c_tok, is_local, n_web_searches, extra_cost).

    ``web_search_tool``: a truthy marker that web_search is enabled for
    this run. When set AND the worker endpoint is search-capable
@@ -565,6 +594,22 @@ def _call_worker(
    max_tok = int(cfg.get("worker_max_tokens", 4096))
    temp = float(cfg.get("worker_temperature", 0.2))
    use_ws = web_search_tool is not None
+    search_backend = str(cfg.get("search_backend", "provider")).lower()
+    extra_cost = 0.0
+    if use_ws and search_backend == "tavily":
+        res = tavily_search_context(
+            prompt,
+            max_results=int(cfg.get("tavily_max_results", 5)),
+        )
+        prompt = (
+            f"Web search results:\n{res['text']}\n\n"
+            f"Using the search results above, answer this request:\n{prompt}"
+        )
+        extra_cost = float(res["cost_usd"])
+        use_ws = False
+        tavily_searches = int(res["n_searches"])
+    else:
+        tavily_searches = 0

    if ep == "vllm":
        text, p, c = LocalCloudAgent._call_vllm(
@@ -575,7 +620,7 @@ def _call_worker(
            temperature=temp,
            enable_thinking=False,
        )
-        return text, p, c, True, 0
+        return text, p, c, True, tavily_searches, extra_cost
    if ep == "openai":
        if use_ws:
            text, p, c, n_searches, _ = LocalCloudAgent._call_openai_agent(
@@ -584,14 +629,14 @@ def _call_worker(
                max_tokens=max_tok,
                temperature=(1.0 if is_gpt5_family(worker["model"]) else temp),
            )
-            return text, p, c, False, n_searches
+            return text, p, c, False, n_searches, 0.0
        text, p, c = LocalCloudAgent._call_openai(
            worker["model"],
            user=prompt,
            max_tokens=max_tok,
            temperature=(1.0 if is_gpt5_family(worker["model"]) else temp),
        )
-        return text, p, c, False, 0
+        return text, p, c, False, tavily_searches, extra_cost
    if ep == "openrouter":
        # OpenRouter is OpenAI-compatible; the helper handles the
        # base_url + OPENROUTER_API_KEY plumbing. No server-side web
@@ -607,7 +652,7 @@ def _call_worker(
            temperature=temp,
            extra_body=extra_body if isinstance(extra_body, dict) else None,
        )
-        return text, p, c, False, 0
+        return text, p, c, False, tavily_searches, extra_cost
    if ep == "anthropic":
        eff_temp = temp if supports_temperature(worker["model"]) else 0.0
        anthropic_kwargs: Dict[str, Any] = dict(
@@ -620,7 +665,7 @@ def _call_worker(
        text, p, c, n_searches = LocalCloudAgent._call_anthropic(
            worker["model"], **anthropic_kwargs
        )
-        return text, p, c, False, n_searches
+        return text, p, c, False, n_searches or tavily_searches, extra_cost
    if ep == "gemini":
        # Gemini Developer API via google-genai. With web_search on, route
        # through the Google-Search-grounded agent loop; otherwise plain
@@ -632,14 +677,14 @@ def _call_worker(
                max_tokens=max_tok,
                temperature=temp,
            )
-            return text, p, c, False, n_searches
+            return text, p, c, False, n_searches, 0.0
        text, p, c = LocalCloudAgent._call_gemini(
            worker["model"],
            user=prompt,
            max_tokens=max_tok,
            temperature=temp,
        )
-        return text, p, c, False, 0
+        return text, p, c, False, tavily_searches, extra_cost
    raise ValueError(f"unsupported worker endpoint: {ep!r}")


@@ -659,12 +704,18 @@ def _swe_worker_step(
    ep = (worker.get("endpoint") or "openai").lower()
    if ep == "vllm":
        backbone, model, endpoint, is_local = (
-            "local", worker["model"], worker.get("base_url"), True,
+            "local",
+            worker["model"],
+            worker.get("base_url"),
+            True,
        )
        cloud_endpoint = "anthropic"  # unused on the local path
    elif ep == "anthropic":
        backbone, model, endpoint, is_local = (
-            "cloud", worker["model"], None, False,
+            "cloud",
+            worker["model"],
+            None,
+            False,
        )
        cloud_endpoint = "anthropic"
    else:
@@ -672,7 +723,7 @@ def _swe_worker_step(
        # backbones today (the loop's tool-call format is Anthropic- or
        # OpenAI-via-vllm-shaped only). Fall back to one-shot for those —
        # SWE-bench-wise they were already weak; this preserves behavior.
-        text, p, c, is_local, n_searches = _call_worker(worker, prompt, cfg)
+        text, p, c, is_local, n_searches, _extra = _call_worker(worker, prompt, cfg)
        return text, p, c, is_local, n_searches, 0
    out = run_swe_agent_loop(
        task,
@@ -690,7 +741,11 @@ def _swe_worker_step(
    )
    return (
        out["final_summary"] or out["answer"],
-        out["tokens_in"], out["tokens_out"], is_local, 0, int(out["turns"]),
+        out["tokens_in"],
+        out["tokens_out"],
+        is_local,
+        0,
+        int(out["turns"]),
    )


@@ -753,13 +808,17 @@ class ConductorAgent(LocalCloudAgent):
            and bool(task_meta_early.get("base_commit"))
        )
        ws_enabled, ws_max_uses = web_search_cfg(cfg)
+        search_backend = str(cfg.get("search_backend", "provider")).lower()
        planner_ws = ws_enabled and not swe_mode_early

        # 1. Plan — when web_search is on (GAIA), the prompt names which
        # worker indices can actually search, so the planner routes
        # research steps to a search-capable worker.
        user = _build_conductor_prompt(
-            question, workers, web_search_enabled=planner_ws,
+            question,
+            workers,
+            web_search_enabled=planner_ws,
+            search_backend=search_backend,
        )
        plan_text, p_in, p_out = self._call_cloud(
            user=user,
@@ -787,21 +846,22 @@ class ConductorAgent(LocalCloudAgent):
        if plan is None:
            fallback_used = True
            plan = {
-                "model_id":    [len(workers) - 1],
-                "subtasks":    [question],
+                "model_id": [len(workers) - 1],
+                "subtasks": [question],
                "access_list": [[]],
            }

-        self.record_trace_event({
-            "kind": "conductor_plan",
-            "plan": plan,
-            "fallback_used": fallback_used,
-            "parse_attempts": parse_attempts,
-            "workers": [
-                {k: v for k, v in w.items() if k != "api_key"}
-                for w in workers
-            ],
-        })
+        self.record_trace_event(
+            {
+                "kind": "conductor_plan",
+                "plan": plan,
+                "fallback_used": fallback_used,
+                "parse_attempts": parse_attempts,
+                "workers": [
+                    {k: v for k, v in w.items() if k != "api_key"} for w in workers
+                ],
+            }
+        )

        # 2. Execute
        # If we're on a SWE-bench task AND cfg["swe_use_agent_loop"] is on,
@@ -833,16 +893,17 @@ class ConductorAgent(LocalCloudAgent):
        # memory. Fail loud instead of degrading silently.
        # ``ws_enabled`` / ``ws_max_uses`` computed up front for the planner
        # constraint — reuse them here.
-        if ws_enabled and not swe_mode:
+        if ws_enabled and search_backend != "tavily" and not swe_mode:
            search_workers = [
-                w for w in workers
+                w
+                for w in workers
                if (w.get("endpoint") or "openai").lower()
                in _SEARCH_CAPABLE_WORKER_ENDPOINTS
            ]
            if not search_workers:
-                endpoints = sorted({
-                    (w.get("endpoint") or "openai").lower() for w in workers
-                })
+                endpoints = sorted(
+                    {(w.get("endpoint") or "openai").lower() for w in workers}
+                )
                raise ValueError(
                    f"web_search.enabled=true but the worker pool has no "
                    f"search-capable worker (endpoints present: {endpoints}); "
@@ -853,39 +914,43 @@ class ConductorAgent(LocalCloudAgent):
                )
        # ``ws_tool`` doubles as the enable marker passed to `_call_worker`
        # (truthy => route search-capable workers through their agent loop).
-        ws_tool = (
-            build_web_search_tool(ws_max_uses) if ws_enabled else None
-        )
+        ws_tool = build_web_search_tool(ws_max_uses) if ws_enabled else None

        try:
            if swe_mode:
-                shared_workdir = Path(tempfile.mkdtemp(
-                    prefix=f"conductor-swe-{task_meta.get('task_id','x')}-"
-                ))
+                shared_workdir = Path(
+                    tempfile.mkdtemp(
+                        prefix=f"conductor-swe-{task_meta.get('task_id', 'x')}-"
+                    )
+                )
                _clone_repo(task_meta["repo"], task_meta["base_commit"], shared_workdir)
-                self.record_trace_event({
-                    "kind": "conductor_swe_workdir",
-                    "workdir": str(shared_workdir),
-                    "repo": task_meta["repo"],
-                    "base_commit": task_meta["base_commit"],
-                })
+                self.record_trace_event(
+                    {
+                        "kind": "conductor_swe_workdir",
+                        "workdir": str(shared_workdir),
+                        "repo": task_meta["repo"],
+                        "base_commit": task_meta["base_commit"],
+                    }
+                )

            for i, (mid, subtask, access) in enumerate(
                zip(plan["model_id"], plan["subtasks"], plan["access_list"])
            ):
                worker = workers[mid]
                prompt = _build_step_prompt(question, subtask, steps, access)
-                self.record_trace_event({
-                    "kind": "conductor_step_dispatch",
-                    "step_idx": i,
-                    "worker_id": mid,
-                    "worker_name": worker["name"],
-                    "worker_model": worker["model"],
-                    "subtask": subtask,
-                    "access": access,
-                    "prompt": prompt,
-                    "swe_mode": swe_mode,
-                })
+                self.record_trace_event(
+                    {
+                        "kind": "conductor_step_dispatch",
+                        "step_idx": i,
+                        "worker_id": mid,
+                        "worker_name": worker["name"],
+                        "worker_model": worker["model"],
+                        "subtask": subtask,
+                        "access": access,
+                        "prompt": prompt,
+                        "swe_mode": swe_mode,
+                    }
+                )

                worker_ep = (worker.get("endpoint") or "openai").lower()
                # Post-hoc routing check: if web_search is on but the
@@ -894,35 +959,49 @@ class ConductorAgent(LocalCloudAgent):
                # may legitimately not need search; see Task-3 planner
                # constraint that tries to prevent this upfront).
                if (
-                    ws_enabled and not swe_mode
+                    ws_enabled
+                    and search_backend != "tavily"
+                    and not swe_mode
                    and worker_ep not in _SEARCH_CAPABLE_WORKER_ENDPOINTS
                ):
-                    self.record_trace_event({
-                        "kind": "conductor_search_routing_warning",
-                        "step_idx": i,
-                        "worker_id": mid,
-                        "worker_name": worker["name"],
-                        "worker_endpoint": worker_ep,
-                        "warning": (
-                            f"web_search enabled but step {i} routed to "
-                            f"search-incapable worker {worker['name']!r} "
-                            f"(endpoint {worker_ep!r}); this step cannot "
-                            "ground and may answer blind."
-                        ),
-                    })
+                    self.record_trace_event(
+                        {
+                            "kind": "conductor_search_routing_warning",
+                            "step_idx": i,
+                            "worker_id": mid,
+                            "worker_name": worker["name"],
+                            "worker_endpoint": worker_ep,
+                            "warning": (
+                                f"web_search enabled but step {i} routed to "
+                                f"search-incapable worker {worker['name']!r} "
+                                f"(endpoint {worker_ep!r}); this step cannot "
+                                "ground and may answer blind."
+                            ),
+                        }
+                    )

+                extra_cost = 0.0
                if swe_mode:
                    text, w_in, w_out, is_local, n_searches, bash_turns = (
                        _swe_worker_step(
-                            worker, task_meta, prompt, cfg, shared_workdir, i,
+                            worker,
+                            task_meta,
+                            prompt,
+                            cfg,
+                            shared_workdir,
+                            i,
                        )
                    )
                    tool_calls += bash_turns
                else:
-                    text, w_in, w_out, is_local, n_searches = _call_worker(
-                        worker, prompt, cfg,
-                        web_search_tool=ws_tool,
-                        web_search_max_uses=ws_max_uses,
+                    (text, w_in, w_out, is_local, n_searches, extra_cost) = (
+                        _call_worker(
+                            worker,
+                            prompt,
+                            cfg,
+                            web_search_tool=ws_tool,
+                            web_search_max_uses=ws_max_uses,
+                        )
                    )

                if is_local:
@@ -930,20 +1009,25 @@ class ConductorAgent(LocalCloudAgent):
                else:
                    tokens_cloud += w_in + w_out
                    cost += self.cost_usd(worker["model"], w_in, w_out)
-                    cost += n_searches * _worker_search_cost_per_call(worker_ep)
+                    if search_backend != "tavily":
+                        cost += n_searches * _worker_search_cost_per_call(worker_ep)
+                if search_backend == "tavily":
+                    cost += extra_cost
                n_web_searches_total += n_searches
                tool_calls += n_searches
-                steps.append({
-                    "step_idx": i,
-                    "model_id": mid,
-                    "worker_name": worker["name"],
-                    "worker_model": worker["model"],
-                    "subtask": subtask,
-                    "access": access,
-                    "output": text,
-                    "tokens_in": w_in,
-                    "tokens_out": w_out,
-                })
+                steps.append(
+                    {
+                        "step_idx": i,
+                        "model_id": mid,
+                        "worker_name": worker["name"],
+                        "worker_model": worker["model"],
+                        "subtask": subtask,
+                        "access": access,
+                        "output": text,
+                        "tokens_in": w_in,
+                        "tokens_out": w_out,
+                    }
+                )
                final_answer = text

            # For SWE mode, the authoritative patch is whatever lives in
@@ -954,7 +1038,8 @@ class ConductorAgent(LocalCloudAgent):
                if patch.strip():
                    final_answer = (
                        f"{final_answer}\n\n```diff\n{patch}```"
-                        if final_answer else f"```diff\n{patch}```"
+                        if final_answer
+                        else f"```diff\n{patch}```"
                    )
        finally:
            if shared_workdir is not None:
@@ -965,8 +1050,7 @@ class ConductorAgent(LocalCloudAgent):
        tokens_cloud += conductor_p_in + conductor_p_out

        traces = [
-            (s["step_idx"], s["model_id"], s["subtask"], s["output"])
-            for s in steps
+            (s["step_idx"], s["model_id"], s["subtask"], s["output"]) for s in steps
        ]

        meta = {
@@ -981,11 +1065,11 @@ class ConductorAgent(LocalCloudAgent):
                "plan": plan,
                "fallback_used": fallback_used,
                "web_search_enabled": ws_enabled,
+                "search_backend": search_backend,
                "n_web_searches": n_web_searches_total,
                "parse_attempts": parse_attempts,
                "workers": [
-                    {k: v for k, v in w.items() if k != "api_key"}
-                    for w in workers
+                    {k: v for k, v in w.items() if k != "api_key"} for w in workers
                ],
            },
        }
@@ -172,11 +172,16 @@ def _clone_repo(repo: str, base_commit: str, dest: Path) -> None:
    url = f"https://github.com/{repo}.git"
    subprocess.run(
        ["git", "clone", "--quiet", url, str(dest)],
-        check=True, timeout=300, capture_output=True,
+        check=True,
+        timeout=300,
+        capture_output=True,
    )
    subprocess.run(
        ["git", "checkout", "--quiet", base_commit],
-        cwd=str(dest), check=True, timeout=120, capture_output=True,
+        cwd=str(dest),
+        check=True,
+        timeout=120,
+        capture_output=True,
    )


@@ -265,10 +270,14 @@ def _run_bash(
    stderr = _decode_bash_output(stderr_b, exit_code)
    truncated = False
    if len(stdout) > output_cap:
-        stdout = stdout[:output_cap] + f"\n…[+{len(stdout) - output_cap} chars truncated]"
+        stdout = (
+            stdout[:output_cap] + f"\n…[+{len(stdout) - output_cap} chars truncated]"
+        )
        truncated = True
    if len(stderr) > output_cap:
-        stderr = stderr[:output_cap] + f"\n…[+{len(stderr) - output_cap} chars truncated]"
+        stderr = (
+            stderr[:output_cap] + f"\n…[+{len(stderr) - output_cap} chars truncated]"
+        )
        truncated = True
    return {
        "stdout": stdout,
@@ -297,7 +306,10 @@ def _extract_diff(workdir: Path) -> str:
    """``git diff`` against the base commit — the final SWE-bench patch."""
    proc = subprocess.run(
        ["git", "diff", "--no-color"],
-        cwd=str(workdir), capture_output=True, text=True, timeout=60,
+        cwd=str(workdir),
+        capture_output=True,
+        text=True,
+        timeout=60,
    )
    return proc.stdout

@@ -320,10 +332,11 @@ def _anthropic_assistant_block(block: Any) -> Dict[str, Any]:

 # ---------- Reusable agent-loop entry point ----------

+
 def run_swe_agent_loop(
    task: Dict[str, Any],
    *,
-    backbone: str,                          # "cloud" or "local"
+    backbone: str,  # "cloud" or "local"
    backbone_model: str,
    cloud_endpoint: str = "anthropic",
    local_endpoint: Optional[str] = None,
@@ -387,32 +400,33 @@ def run_swe_agent_loop(

    own_workdir = workdir is None
    if own_workdir:
-        workdir = Path(tempfile.mkdtemp(
-            prefix=f"mini-swe-{task.get('task_id','x')}-"
-        ))
+        workdir = Path(tempfile.mkdtemp(prefix=f"mini-swe-{task.get('task_id', 'x')}-"))
        try:
            _clone_repo(repo, base_commit, workdir)
        except Exception:
            shutil.rmtree(workdir, ignore_errors=True)
            raise

-    _record_event({
-        "kind": f"{trace_prefix}_setup",
-        "repo": repo,
-        "base_commit": base_commit,
-        "workdir": str(workdir),
-        "owns_workdir": own_workdir,
-        "backbone": backbone,
-        "backbone_model": backbone_model,
-        "ts": time.time(),
-    })
+    _record_event(
+        {
+            "kind": f"{trace_prefix}_setup",
+            "repo": repo,
+            "base_commit": base_commit,
+            "workdir": str(workdir),
+            "owns_workdir": own_workdir,
+            "backbone": backbone,
+            "backbone_model": backbone_model,
+            "ts": time.time(),
+        }
+    )

    user_prompt = initial_prompt or task.get("problem_statement") or ""

    try:
        if backbone == "cloud":
            result = _loop_cloud(
-                user_prompt, workdir,
+                user_prompt,
+                workdir,
                model=backbone_model,
                cloud_endpoint=cloud_endpoint,
                max_turns=max_turns,
@@ -423,9 +437,12 @@ def run_swe_agent_loop(
            )
        elif backbone == "local":
            if not local_endpoint:
-                raise ValueError("run_swe_agent_loop(backbone='local') needs local_endpoint")
+                raise ValueError(
+                    "run_swe_agent_loop(backbone='local') needs local_endpoint"
+                )
            result = _loop_local(
-                user_prompt, workdir,
+                user_prompt,
+                workdir,
                model=backbone_model,
                endpoint=local_endpoint,
                max_turns=max_turns,
@@ -440,7 +457,7 @@ def run_swe_agent_loop(
            raise ValueError(f"unsupported backbone: {backbone!r}")

        patch = _extract_diff(workdir)
-        framed = (result["final_summary"] or "[mini-swe-agent produced no summary text]")
+        framed = result["final_summary"] or "[mini-swe-agent produced no summary text]"
        if patch.strip():
            framed = f"{framed}\n\n```diff\n{patch}```"

@@ -450,11 +467,16 @@ def run_swe_agent_loop(
            "final_summary": result["final_summary"],
            "tokens_in": result["tokens_in"],
            "tokens_out": result["tokens_out"],
-            "tokens_local": result["tokens_in"] + result["tokens_out"] if backbone == "local" else 0,
-            "tokens_cloud": result["tokens_in"] + result["tokens_out"] if backbone == "cloud" else 0,
+            "tokens_local": result["tokens_in"] + result["tokens_out"]
+            if backbone == "local"
+            else 0,
+            "tokens_cloud": result["tokens_in"] + result["tokens_out"]
+            if backbone == "cloud"
+            else 0,
            "cost_usd": (
                estimate_cost(backbone_model, result["tokens_in"], result["tokens_out"])
-                if backbone == "cloud" else 0.0
+                if backbone == "cloud"
+                else 0.0
            ),
            "turns": result["turns"],
            "max_turns_hit": result["max_turns_hit"],
@@ -467,6 +489,7 @@ def run_swe_agent_loop(

 # ---------- Cloud loop (dispatcher → per-endpoint multi-turn tool loops) ----------

+
 def _loop_cloud(
    problem: str,
    workdir: Path,
@@ -485,24 +508,36 @@ def _loop_cloud(
    to unblock the 8 SWE cells that were stuck on Anthropic-only support."""
    if cloud_endpoint == "anthropic":
        return _loop_cloud_anthropic(
-            problem, workdir,
-            model=model, max_turns=max_turns,
-            bash_timeout=bash_timeout, output_cap=output_cap,
-            turn_max_tokens=turn_max_tokens, trace_prefix=trace_prefix,
+            problem,
+            workdir,
+            model=model,
+            max_turns=max_turns,
+            bash_timeout=bash_timeout,
+            output_cap=output_cap,
+            turn_max_tokens=turn_max_tokens,
+            trace_prefix=trace_prefix,
        )
    if cloud_endpoint == "openai":
        return _loop_cloud_openai(
-            problem, workdir,
-            model=model, max_turns=max_turns,
-            bash_timeout=bash_timeout, output_cap=output_cap,
-            turn_max_tokens=turn_max_tokens, trace_prefix=trace_prefix,
+            problem,
+            workdir,
+            model=model,
+            max_turns=max_turns,
+            bash_timeout=bash_timeout,
+            output_cap=output_cap,
+            turn_max_tokens=turn_max_tokens,
+            trace_prefix=trace_prefix,
        )
    if cloud_endpoint == "gemini":
        return _loop_cloud_gemini(
-            problem, workdir,
-            model=model, max_turns=max_turns,
-            bash_timeout=bash_timeout, output_cap=output_cap,
-            turn_max_tokens=turn_max_tokens, trace_prefix=trace_prefix,
+            problem,
+            workdir,
+            model=model,
+            max_turns=max_turns,
+            bash_timeout=bash_timeout,
+            output_cap=output_cap,
+            turn_max_tokens=turn_max_tokens,
+            trace_prefix=trace_prefix,
        )
    raise ValueError(
        f"mini-SWE-agent cloud backbone unsupported endpoint: {cloud_endpoint!r}"
@@ -521,6 +556,7 @@ def _loop_cloud_anthropic(
    trace_prefix: str,
 ) -> Dict[str, Any]:
    import anthropic
+
    client = anthropic.Anthropic(timeout=600.0, max_retries=5)
    messages: List[Dict[str, Any]] = [{"role": "user", "content": problem}]

@@ -553,30 +589,39 @@ def _loop_cloud_anthropic(
            btype = getattr(block, "type", None)
            if btype == "tool_use":
                tool_uses.append((block.id, block.name, dict(block.input or {})))
-                content_blocks.append({
-                    "type": "tool_use", "id": block.id, "name": block.name,
-                    "input": dict(block.input or {}),
-                })
+                content_blocks.append(
+                    {
+                        "type": "tool_use",
+                        "id": block.id,
+                        "name": block.name,
+                        "input": dict(block.input or {}),
+                    }
+                )
            elif hasattr(block, "text"):
                text_parts.append(block.text)
                content_blocks.append({"type": "text", "text": block.text})
            else:
                content_blocks.append({"type": btype or "unknown"})

-        _record_event({
-            "kind": f"{trace_prefix}_turn",
-            "turn": turn,
-            "stop_reason": msg.stop_reason,
-            "tokens_in": msg.usage.input_tokens,
-            "tokens_out": msg.usage.output_tokens,
-            "latency_s": latency,
-            "content_blocks": content_blocks,
-            "ts": time.time(),
-        })
+        _record_event(
+            {
+                "kind": f"{trace_prefix}_turn",
+                "turn": turn,
+                "stop_reason": msg.stop_reason,
+                "tokens_in": msg.usage.input_tokens,
+                "tokens_out": msg.usage.output_tokens,
+                "latency_s": latency,
+                "content_blocks": content_blocks,
+                "ts": time.time(),
+            }
+        )

-        messages.append({"role": "assistant", "content": [
-            _anthropic_assistant_block(b) for b in msg.content
-        ]})
+        messages.append(
+            {
+                "role": "assistant",
+                "content": [_anthropic_assistant_block(b) for b in msg.content],
+            }
+        )

        if not tool_uses:
            final_text = "\n".join(text_parts).strip()
@@ -586,28 +631,40 @@ def _loop_cloud_anthropic(
        for tu_id, tu_name, tu_input in tool_uses:
            if tu_name != "bash":
                obs = f"unknown tool: {tu_name!r}"
-                _record_event({
-                    "kind": f"{trace_prefix}_unknown_tool",
-                    "turn": turn, "name": tu_name, "input": tu_input,
-                    "ts": time.time(),
-                })
+                _record_event(
+                    {
+                        "kind": f"{trace_prefix}_unknown_tool",
+                        "turn": turn,
+                        "name": tu_name,
+                        "input": tu_input,
+                        "ts": time.time(),
+                    }
+                )
            else:
                command = str(tu_input.get("command", ""))
                result = _run_bash(
-                    command, workdir,
-                    timeout=bash_timeout, output_cap=output_cap,
+                    command,
+                    workdir,
+                    timeout=bash_timeout,
+                    output_cap=output_cap,
+                )
+                _record_event(
+                    {
+                        "kind": f"{trace_prefix}_bash",
+                        "turn": turn,
+                        "command": command,
+                        **result,
+                        "ts": time.time(),
+                    }
                )
-                _record_event({
-                    "kind": f"{trace_prefix}_bash",
-                    "turn": turn, "command": command,
-                    **result, "ts": time.time(),
-                })
                obs = _format_observation(result)
-            tool_result_blocks.append({
-                "type": "tool_result",
-                "tool_use_id": tu_id,
-                "content": obs,
-            })
+            tool_result_blocks.append(
+                {
+                    "type": "tool_result",
+                    "tool_use_id": tu_id,
+                    "content": obs,
+                }
+            )
        messages.append({"role": "user", "content": tool_result_blocks})

    return {
@@ -621,6 +678,7 @@ def _loop_cloud_anthropic(

 # ---------- Cloud loop (OpenAI multi-turn with function tools) ----------

+
 def _loop_cloud_openai(
    problem: str,
    workdir: Path,
@@ -646,6 +704,7 @@ def _loop_cloud_openai(
      ``_loop_local`` behavior).
    """
    from openai import OpenAI
+
    client = OpenAI(timeout=600.0)

    messages: List[Dict[str, Any]] = [
@@ -681,21 +740,27 @@ def _loop_cloud_openai(
        tool_calls = list(getattr(message, "tool_calls", None) or [])
        text = message.content or ""

-        _record_event({
-            "kind": f"{trace_prefix}_turn",
-            "turn": turn,
-            "endpoint": "openai",
-            "finish_reason": choice.finish_reason,
-            "tokens_in": getattr(u, "prompt_tokens", 0) if u else 0,
-            "tokens_out": getattr(u, "completion_tokens", 0) if u else 0,
-            "latency_s": latency,
-            "text": text,
-            "tool_calls": [
-                {"id": tc.id, "name": tc.function.name, "arguments": tc.function.arguments}
-                for tc in tool_calls
-            ],
-            "ts": time.time(),
-        })
+        _record_event(
+            {
+                "kind": f"{trace_prefix}_turn",
+                "turn": turn,
+                "endpoint": "openai",
+                "finish_reason": choice.finish_reason,
+                "tokens_in": getattr(u, "prompt_tokens", 0) if u else 0,
+                "tokens_out": getattr(u, "completion_tokens", 0) if u else 0,
+                "latency_s": latency,
+                "text": text,
+                "tool_calls": [
+                    {
+                        "id": tc.id,
+                        "name": tc.function.name,
+                        "arguments": tc.function.arguments,
+                    }
+                    for tc in tool_calls
+                ],
+                "ts": time.time(),
+            }
+        )

        # Append the assistant turn (including any tool_calls) so the
        # follow-up tool messages have the right call ids to reference.
@@ -711,7 +776,8 @@ def _loop_cloud_openai(
        if tool_calls:
            assistant_msg["tool_calls"] = [
                {
-                    "id": tc.id, "type": "function",
+                    "id": tc.id,
+                    "type": "function",
                    "function": {
                        "name": tc.function.name,
                        "arguments": tc.function.arguments,
@@ -734,21 +800,26 @@ def _loop_cloud_openai(
                and not text.strip()
                and turn < max_turns
            ):
-                messages.append({
-                    "role": "user",
-                    "content": (
-                        "Your previous response was truncated by the token limit "
-                        "before producing a tool call or final summary. Retry: "
-                        "either issue ONE bash tool call (short command, no large "
-                        "output) or send a brief one-line final summary with no "
-                        "tool calls to end the loop."
-                    ),
-                })
-                _record_event({
-                    "kind": f"{trace_prefix}_recover",
-                    "turn": turn, "reason": "length_truncation_no_tool_call",
-                    "ts": time.time(),
-                })
+                messages.append(
+                    {
+                        "role": "user",
+                        "content": (
+                            "Your previous response was truncated by the token limit "
+                            "before producing a tool call or final summary. Retry: "
+                            "either issue ONE bash tool call (short command, no large "
+                            "output) or send a brief one-line final summary with no "
+                            "tool calls to end the loop."
+                        ),
+                    }
+                )
+                _record_event(
+                    {
+                        "kind": f"{trace_prefix}_recover",
+                        "turn": turn,
+                        "reason": "length_truncation_no_tool_call",
+                        "ts": time.time(),
+                    }
+                )
                continue
            # No tool call → the model is done. Same termination rule as
            # the Anthropic branch.
@@ -762,28 +833,40 @@ def _loop_cloud_openai(
                args = {}
            if tc.function.name != "bash":
                obs = f"unknown tool: {tc.function.name!r}"
-                _record_event({
-                    "kind": f"{trace_prefix}_unknown_tool",
-                    "turn": turn, "name": tc.function.name, "input": args,
-                    "ts": time.time(),
-                })
+                _record_event(
+                    {
+                        "kind": f"{trace_prefix}_unknown_tool",
+                        "turn": turn,
+                        "name": tc.function.name,
+                        "input": args,
+                        "ts": time.time(),
+                    }
+                )
            else:
                command = str(args.get("command", ""))
                result = _run_bash(
-                    command, workdir,
-                    timeout=bash_timeout, output_cap=output_cap,
+                    command,
+                    workdir,
+                    timeout=bash_timeout,
+                    output_cap=output_cap,
+                )
+                _record_event(
+                    {
+                        "kind": f"{trace_prefix}_bash",
+                        "turn": turn,
+                        "command": command,
+                        **result,
+                        "ts": time.time(),
+                    }
                )
-                _record_event({
-                    "kind": f"{trace_prefix}_bash",
-                    "turn": turn, "command": command,
-                    **result, "ts": time.time(),
-                })
                obs = _format_observation(result)
-            messages.append({
-                "role": "tool",
-                "tool_call_id": tc.id,
-                "content": obs,
-            })
+            messages.append(
+                {
+                    "role": "tool",
+                    "tool_call_id": tc.id,
+                    "content": obs,
+                }
+            )

    return {
        "tokens_in": tokens_in,
@@ -796,6 +879,7 @@ def _loop_cloud_openai(

 # ---------- Cloud loop (Gemini multi-turn with function tools) ----------

+
 def _loop_cloud_gemini(
    problem: str,
    workdir: Path,
@@ -831,13 +915,15 @@ def _loop_cloud_gemini(
    from google.genai import types

    client = genai.Client(http_options=types.HttpOptions(timeout=600_000))
-    bash_tool = types.Tool(function_declarations=[
-        types.FunctionDeclaration(
-            name="bash",
-            description=BASH_TOOL_ANTHROPIC["description"],
-            parameters=BASH_TOOL_GEMINI_PARAMETERS,
-        ),
-    ])
+    bash_tool = types.Tool(
+        function_declarations=[
+            types.FunctionDeclaration(
+                name="bash",
+                description=BASH_TOOL_ANTHROPIC["description"],
+                parameters=BASH_TOOL_GEMINI_PARAMETERS,
+            ),
+        ]
+    )

    contents: List[types.Content] = [
        types.Content(role="user", parts=[types.Part(text=problem)]),
@@ -859,7 +945,9 @@ def _loop_cloud_gemini(
        )
        t0 = time.time()
        resp = client.models.generate_content(
-            model=model, contents=contents, config=cfg,
+            model=model,
+            contents=contents,
+            config=cfg,
        )
        _bump_cloud_calls()
        latency = time.time() - t0
@@ -896,21 +984,22 @@ def _loop_cloud_gemini(
        except Exception:
            pass

-        _record_event({
-            "kind": f"{trace_prefix}_turn",
-            "turn": turn,
-            "endpoint": "gemini",
-            "finish_reason": finish_reason,
-            "tokens_in": p,
-            "tokens_out": c,
-            "latency_s": latency,
-            "text": "\n".join(text_parts),
-            "tool_calls": [
-                {"name": name, "arguments": args}
-                for name, args in function_calls
-            ],
-            "ts": time.time(),
-        })
+        _record_event(
+            {
+                "kind": f"{trace_prefix}_turn",
+                "turn": turn,
+                "endpoint": "gemini",
+                "finish_reason": finish_reason,
+                "tokens_in": p,
+                "tokens_out": c,
+                "latency_s": latency,
+                "text": "\n".join(text_parts),
+                "tool_calls": [
+                    {"name": name, "arguments": args} for name, args in function_calls
+                ],
+                "ts": time.time(),
+            }
+        )

        # Append the model's content as-is so the next turn sees its own
        # prior function_call parts (Gemini requires this for the
@@ -934,28 +1023,37 @@ def _loop_cloud_gemini(
            # treat genuine ``STOP`` with text as a final answer.
            fr_str = str(finish_reason or "")
            empty_text = not any(t.strip() for t in text_parts)
-            recoverable = empty_text and turn < max_turns and (
-                "MALFORMED_FUNCTION_CALL" in fr_str
-                or "MAX_TOKENS" in fr_str
+            recoverable = (
+                empty_text
+                and turn < max_turns
+                and ("MALFORMED_FUNCTION_CALL" in fr_str or "MAX_TOKENS" in fr_str)
            )
            if recoverable:
-                contents.append(types.Content(
-                    role="user",
-                    parts=[types.Part(text=(
-                        "Your previous response had no parsable function call "
-                        "and no final text (finish_reason="
-                        f"{fr_str}). Retry: either issue ONE well-formed "
-                        "`bash` function call (short command, valid JSON-ish "
-                        "args) or send a brief final text message with no "
-                        "function call to end the loop."
-                    ))],
-                ))
-                _record_event({
-                    "kind": f"{trace_prefix}_recover",
-                    "turn": turn,
-                    "reason": f"empty_response_{fr_str}",
-                    "ts": time.time(),
-                })
+                contents.append(
+                    types.Content(
+                        role="user",
+                        parts=[
+                            types.Part(
+                                text=(
+                                    "Your previous response had no parsable function call "
+                                    "and no final text (finish_reason="
+                                    f"{fr_str}). Retry: either issue ONE well-formed "
+                                    "`bash` function call (short command, valid JSON-ish "
+                                    "args) or send a brief final text message with no "
+                                    "function call to end the loop."
+                                )
+                            )
+                        ],
+                    )
+                )
+                _record_event(
+                    {
+                        "kind": f"{trace_prefix}_recover",
+                        "turn": turn,
+                        "reason": f"empty_response_{fr_str}",
+                        "ts": time.time(),
+                    }
+                )
                continue
            final_text = "\n".join(text_parts).strip()
            break
@@ -964,26 +1062,39 @@ def _loop_cloud_gemini(
        for name, args in function_calls:
            if name != "bash":
                obs = f"unknown tool: {name!r}"
-                _record_event({
-                    "kind": f"{trace_prefix}_unknown_tool",
-                    "turn": turn, "name": name, "input": args,
-                    "ts": time.time(),
-                })
+                _record_event(
+                    {
+                        "kind": f"{trace_prefix}_unknown_tool",
+                        "turn": turn,
+                        "name": name,
+                        "input": args,
+                        "ts": time.time(),
+                    }
+                )
            else:
                command = str(args.get("command", ""))
                result = _run_bash(
-                    command, workdir,
-                    timeout=bash_timeout, output_cap=output_cap,
+                    command,
+                    workdir,
+                    timeout=bash_timeout,
+                    output_cap=output_cap,
+                )
+                _record_event(
+                    {
+                        "kind": f"{trace_prefix}_bash",
+                        "turn": turn,
+                        "command": command,
+                        **result,
+                        "ts": time.time(),
+                    }
                )
-                _record_event({
-                    "kind": f"{trace_prefix}_bash",
-                    "turn": turn, "command": command,
-                    **result, "ts": time.time(),
-                })
                obs = _format_observation(result)
-            response_parts.append(types.Part.from_function_response(
-                name=name, response={"output": obs},
-            ))
+            response_parts.append(
+                types.Part.from_function_response(
+                    name=name,
+                    response={"output": obs},
+                )
+            )
        contents.append(types.Content(role="user", parts=response_parts))

    return {
@@ -1013,10 +1124,14 @@ def _get_tiktoken_enc() -> Any:
        return _TIKTOKEN_ENC
    try:
        import tiktoken
+
        _TIKTOKEN_ENC = tiktoken.get_encoding("cl100k_base")
    except Exception as exc:
        if not _TIKTOKEN_WARNED:
-            print(f"[mini_swe_agent] tiktoken unavailable ({exc!r}); falling back to len(s)//4", flush=True)
+            print(
+                f"[mini_swe_agent] tiktoken unavailable ({exc!r}); falling back to len(s)//4",
+                flush=True,
+            )
            _TIKTOKEN_WARNED = True
        _TIKTOKEN_ENC = False
    return _TIKTOKEN_ENC
@@ -1038,7 +1153,7 @@ def _estimate_prompt_tokens(messages: List[Dict[str, Any]]) -> int:
            s = "\n".join(parts)
        else:
            s = ""
-        for tc in (m.get("tool_calls") or []):
+        for tc in m.get("tool_calls") or []:
            try:
                s += "\n" + (tc["function"]["arguments"] or "")
                s += "\n" + (tc["function"].get("name") or "")
@@ -1119,7 +1234,7 @@ def _compact_local_messages(
    before_tokens = _estimate_prompt_tokens(messages)
    new_messages: List[Dict[str, Any]] = list(messages)
    n_tool_elided = 0
-    for (s, e) in old_turns:
+    for s, e in old_turns:
        for k in range(s, e):
            m = new_messages[k]
            if m.get("role") != "tool":
@@ -1140,17 +1255,19 @@ def _compact_local_messages(
            n_tool_elided += 1

    after_stage1_tokens = _estimate_prompt_tokens(new_messages)
-    _record_event({
-        "kind": f"{trace_prefix}_compact",
-        "stage": "1",
-        "msgs_before": len(messages),
-        "msgs_after": len(new_messages),
-        "before_tokens": before_tokens,
-        "after_tokens": after_stage1_tokens,
-        "n_tool_elided": n_tool_elided,
-        "n_turns_folded": 0,
-        "ts": time.time(),
-    })
+    _record_event(
+        {
+            "kind": f"{trace_prefix}_compact",
+            "stage": "1",
+            "msgs_before": len(messages),
+            "msgs_after": len(new_messages),
+            "before_tokens": before_tokens,
+            "after_tokens": after_stage1_tokens,
+            "n_tool_elided": n_tool_elided,
+            "n_turns_folded": 0,
+            "ts": time.time(),
+        }
+    )

    if after_stage1_tokens <= compact_at_tokens:
        return new_messages
@@ -1163,12 +1280,21 @@ def _compact_local_messages(

    summary_input = [
        {"role": "system", "content": _COMPACT_PROMPT},
-        {"role": "user", "content": json.dumps(
-            [{"role": m.get("role"),
-              "content": m.get("content") if isinstance(m.get("content"), str) else str(m.get("content"))[:4000]}
-             for m in middle],
-            default=str,
-        )[:60_000]},
+        {
+            "role": "user",
+            "content": json.dumps(
+                [
+                    {
+                        "role": m.get("role"),
+                        "content": m.get("content")
+                        if isinstance(m.get("content"), str)
+                        else str(m.get("content"))[:4000],
+                    }
+                    for m in middle
+                ],
+                default=str,
+            )[:60_000],
+        },
    ]
    summary = ""
    try:
@@ -1194,18 +1320,20 @@ def _compact_local_messages(
    }
    folded = [system_msg, initial_user, synthetic, *tail]
    after_stage2_tokens = _estimate_prompt_tokens(folded)
-    _record_event({
-        "kind": f"{trace_prefix}_compact",
-        "stage": "2",
-        "msgs_before": len(new_messages),
-        "msgs_after": len(folded),
-        "before_tokens": after_stage1_tokens,
-        "after_tokens": after_stage2_tokens,
-        "n_tool_elided": n_tool_elided,
-        "n_turns_folded": n_turns_folded,
-        "summary_chars": len(summary),
-        "ts": time.time(),
-    })
+    _record_event(
+        {
+            "kind": f"{trace_prefix}_compact",
+            "stage": "2",
+            "msgs_before": len(new_messages),
+            "msgs_after": len(folded),
+            "before_tokens": after_stage1_tokens,
+            "after_tokens": after_stage2_tokens,
+            "n_tool_elided": n_tool_elided,
+            "n_turns_folded": n_turns_folded,
+            "summary_chars": len(summary),
+            "ts": time.time(),
+        }
+    )
    return folded


@@ -1231,6 +1359,7 @@ def _loop_local(
    # but still saw 28k-input 400s on the n=100 SWE sweep (the keep window
    # alone routinely exceeded the budget once bash outputs piled up).
    from openai import OpenAI
+
    client = OpenAI(base_url=endpoint, api_key="EMPTY", timeout=600.0)

    messages: List[Dict[str, Any]] = [
@@ -1243,10 +1372,16 @@ def _loop_local(
    turns = 0
    for turn in range(1, max_turns + 1):
        turns = turn
-        if compact_at_tokens > 0 and _estimate_prompt_tokens(messages) > compact_at_tokens:
+        if (
+            compact_at_tokens > 0
+            and _estimate_prompt_tokens(messages) > compact_at_tokens
+        ):
            messages = _compact_local_messages(
-                messages, client=client, model=model,
-                keep_last=compact_keep_last, trace_prefix=trace_prefix,
+                messages,
+                client=client,
+                model=model,
+                keep_last=compact_keep_last,
+                trace_prefix=trace_prefix,
                compact_at_tokens=compact_at_tokens,
            )
        t0 = time.time()
@@ -1268,23 +1403,27 @@ def _loop_local(
            # server walled the call. Compact aggressively (keep_last=1)
            # and retry once. Re-raise on anything else or on a second
            # failure — the runner records the row as errored.
+            from openjarvis.engine._base import looks_like_context_length_error
+
            msg = str(exc)
-            is_ctx = (
-                "maximum context length" in msg
-                or "context length" in msg.lower() and "exceed" in msg.lower()
-            )
+            is_ctx = looks_like_context_length_error(msg)
            if not is_ctx:
                raise
-            _record_event({
-                "kind": f"{trace_prefix}_emergency_compact",
-                "turn": turn,
-                "error": msg[:300],
-                "tokens_before": _estimate_prompt_tokens(messages),
-                "ts": time.time(),
-            })
+            _record_event(
+                {
+                    "kind": f"{trace_prefix}_emergency_compact",
+                    "turn": turn,
+                    "error": msg[:300],
+                    "tokens_before": _estimate_prompt_tokens(messages),
+                    "ts": time.time(),
+                }
+            )
            messages = _compact_local_messages(
-                messages, client=client, model=model,
-                keep_last=1, trace_prefix=trace_prefix,
+                messages,
+                client=client,
+                model=model,
+                keep_last=1,
+                trace_prefix=trace_prefix,
                compact_at_tokens=max(8_000, compact_at_tokens // 2),
            )
            resp = client.chat.completions.create(
@@ -1306,20 +1445,26 @@ def _loop_local(
        tool_calls = list(getattr(message, "tool_calls", None) or [])
        text = message.content or ""

-        _record_event({
-            "kind": f"{trace_prefix}_turn",
-            "turn": turn,
-            "finish_reason": choice.finish_reason,
-            "tokens_in": getattr(u, "prompt_tokens", 0) if u else 0,
-            "tokens_out": getattr(u, "completion_tokens", 0) if u else 0,
-            "latency_s": latency,
-            "text": text,
-            "tool_calls": [
-                {"id": tc.id, "name": tc.function.name, "arguments": tc.function.arguments}
-                for tc in tool_calls
-            ],
-            "ts": time.time(),
-        })
+        _record_event(
+            {
+                "kind": f"{trace_prefix}_turn",
+                "turn": turn,
+                "finish_reason": choice.finish_reason,
+                "tokens_in": getattr(u, "prompt_tokens", 0) if u else 0,
+                "tokens_out": getattr(u, "completion_tokens", 0) if u else 0,
+                "latency_s": latency,
+                "text": text,
+                "tool_calls": [
+                    {
+                        "id": tc.id,
+                        "name": tc.function.name,
+                        "arguments": tc.function.arguments,
+                    }
+                    for tc in tool_calls
+                ],
+                "ts": time.time(),
+            }
+        )

        # Match the OpenAI cloud branch: content="" (not None) when only
        # tool_calls are present; omit ``tool_calls`` entirely when there
@@ -1333,7 +1478,8 @@ def _loop_local(
        if tool_calls:
            assistant_local_msg["tool_calls"] = [
                {
-                    "id": tc.id, "type": "function",
+                    "id": tc.id,
+                    "type": "function",
                    "function": {
                        "name": tc.function.name,
                        "arguments": tc.function.arguments,
@@ -1357,20 +1503,28 @@ def _loop_local(
            else:
                command = str(args.get("command", ""))
                result = _run_bash(
-                    command, workdir,
-                    timeout=bash_timeout, output_cap=output_cap,
+                    command,
+                    workdir,
+                    timeout=bash_timeout,
+                    output_cap=output_cap,
+                )
+                _record_event(
+                    {
+                        "kind": f"{trace_prefix}_bash",
+                        "turn": turn,
+                        "command": command,
+                        **result,
+                        "ts": time.time(),
+                    }
                )
-                _record_event({
-                    "kind": f"{trace_prefix}_bash",
-                    "turn": turn, "command": command,
-                    **result, "ts": time.time(),
-                })
                obs = _format_observation(result)
-            messages.append({
-                "role": "tool",
-                "tool_call_id": tc.id,
-                "content": obs,
-            })
+            messages.append(
+                {
+                    "role": "tool",
+                    "tool_call_id": tc.id,
+                    "content": obs,
+                }
+            )

    return {
        "tokens_in": tokens_in,
@@ -1383,6 +1537,7 @@ def _loop_local(

 # ---------- Standalone agent ----------

+
@AgentRegistry.register("mini_swe_agent")
 class MiniSWEAgent(LocalCloudAgent):
    """Single-model bash-loop agent for SWE-bench-shaped tasks.
@@ -1410,10 +1565,7 @@ class MiniSWEAgent(LocalCloudAgent):
            task = context.metadata.get("task") or {}

        backbone = cfg.get("backbone", "cloud")
-        model = (
-            self._cloud_model if backbone == "cloud"
-            else (self._local_model or "")
-        )
+        model = self._cloud_model if backbone == "cloud" else (self._local_model or "")

        out = run_swe_agent_loop(
            task,
@@ -48,12 +48,13 @@ from openjarvis.agents.hybrid._base import (
    WEB_SEARCH_COST_PER_CALL,
    LocalCloudAgent,
    build_web_search_tool,
+    tavily_search_context,
    web_search_cfg,
 )
 from openjarvis.agents.hybrid._openai_retry import (
    patch_openai_globally as _patch_openai_globally,
 )
-from openjarvis.agents.hybrid._prices import NO_TEMP_PREFIXES
+from openjarvis.agents.hybrid._prices import NO_TEMP_PREFIXES, default_max_output_tokens
 from openjarvis.agents.hybrid.mini_swe_agent import run_swe_agent_loop
 from openjarvis.core.registry import AgentRegistry

@@ -86,7 +87,7 @@ MINIONS_FIRST_TURN_SCHEMA = {
            "type": "object",
            "properties": {
                "reasoning": {"type": "string"},
-                "message":   {"type": "string"},
+                "message": {"type": "string"},
            },
            "required": ["reasoning", "message"],
            "additionalProperties": False,
@@ -103,7 +104,7 @@ MINIONS_CONVERSATION_SCHEMA = {
                    "type": "object",
                    "properties": {
                        "decision": {"const": "request_additional_info"},
-                        "message":  {"type": "string"},
+                        "message": {"type": "string"},
                    },
                    "required": ["decision", "message"],
                    "additionalProperties": False,
@@ -112,7 +113,7 @@ MINIONS_CONVERSATION_SCHEMA = {
                    "type": "object",
                    "properties": {
                        "decision": {"const": "provide_final_answer"},
-                        "answer":   {"type": "string"},
+                        "answer": {"type": "string"},
                    },
                    "required": ["decision", "answer"],
                    "additionalProperties": False,
@@ -125,8 +126,8 @@ MINIONS_CONVERSATION_SCHEMA = {
 # Markers from Minions's supervisor prompts (prompts/minion.py). Any one
 # being present in the call's messages/system is a strong Minions signal.
 MINIONS_PROMPT_MARKERS = (
-    "small language model that has read",   # SUPERVISOR_INITIAL_PROMPT
-    "provide_final_answer",                  # SUPERVISOR_CONVERSATION_PROMPT
+    "small language model that has read",  # SUPERVISOR_INITIAL_PROMPT
+    "provide_final_answer",  # SUPERVISOR_CONVERSATION_PROMPT
    "request_additional_info",
 )

@@ -170,6 +171,7 @@ def _stub_missing_imports() -> None:
    """
    try:
        import mistralai
+
        if not hasattr(mistralai, "Mistral"):
            mistralai.Mistral = type("Mistral", (), {})  # type: ignore[attr-defined]
    except ImportError:
@@ -214,12 +216,12 @@ def _patch_anthropic_globally() -> None:
                model = kwargs.get("model", "")
                if model.startswith(NO_TEMP_PREFIXES):
                    kwargs.pop("temperature", None)
-                    if (
-                        "output_config" not in kwargs
-                        and _looks_like_minions_call(kwargs)
+                    if "output_config" not in kwargs and _looks_like_minions_call(
+                        kwargs
                    ):
                        kwargs["output_config"] = _minions_turn_schema(kwargs)
                return orig(self, **kwargs)
+
            patched._hybrid_patched = True  # type: ignore[attr-defined]
            return patched

@@ -357,11 +359,14 @@ def _apply_patches_once() -> None:

 # ---------- Pre-fetch helper (GAIA only) ----------

+
 def _prefetch_context(
    question: str,
    cloud_endpoint: str,
    cloud_model: str,
    max_uses: int = 8,
+    search_backend: str = "provider",
+    tavily_max_results: int = 5,
 ) -> Dict[str, Any]:
    """Use Anthropic web_search to fetch real source material the worker can read.

@@ -373,8 +378,27 @@ def _prefetch_context(
    and zeros — the protocol still runs.
    """
    out: Dict[str, Any] = {
-        "text": "", "tokens": 0, "cost_usd": 0.0, "n_searches": 0,
+        "text": "",
+        "tokens": 0,
+        "cost_usd": 0.0,
+        "n_searches": 0,
    }
+    if search_backend == "tavily":
+        try:
+            res = tavily_search_context(question, max_results=tavily_max_results)
+            out.update(
+                text=res["text"],
+                cost_usd=float(res["cost_usd"]),
+                n_searches=int(res["n_searches"]),
+                tokens=0,
+                engine=res.get("engine"),
+                credits=res.get("credits"),
+            )
+            if res.get("error"):
+                out["error"] = res["error"]
+        except Exception as e:
+            out["error"] = f"{type(e).__name__}: {e}"
+        return out
    if cloud_endpoint != "anthropic" or not (question or "").strip():
        return out
    try:
@@ -393,10 +417,12 @@ def _prefetch_context(
            tool_choice={"type": "any"},
        )
        from openjarvis.agents.hybrid._prices import cost as _cost_usd
+
        out.update(
            text=text,
            tokens=p + c,
-            cost_usd=_cost_usd(cloud_model, p, c) + n_searches * WEB_SEARCH_COST_PER_CALL,
+            cost_usd=_cost_usd(cloud_model, p, c)
+            + n_searches * WEB_SEARCH_COST_PER_CALL,
            n_searches=n_searches,
        )
    except Exception as e:
@@ -404,9 +430,7 @@ def _prefetch_context(
    return out


-def _context_for(
-    task: Optional[Dict[str, Any]], prefetched: str = ""
-) -> List[str]:
+def _context_for(task: Optional[Dict[str, Any]], prefetched: str = "") -> List[str]:
    """Minions wants a context list."""
    bits: List[str] = []
    task = task or {}
@@ -421,6 +445,7 @@ def _context_for(

 # ---------- Main agent ----------

+
@AgentRegistry.register("minions")
 class MinionsAgent(LocalCloudAgent):
    """HazyResearch Minions supervisor/worker protocol. See module docstring."""
@@ -432,6 +457,7 @@ class MinionsAgent(LocalCloudAgent):
        # 400/529, KeyError on missing schema fields.
        try:
            import anthropic
+
            if isinstance(exc, anthropic.BadRequestError):
                return f"{type(exc).__name__}: {str(exc)[:120]}"
        except Exception:
@@ -498,18 +524,21 @@ class MinionsAgent(LocalCloudAgent):
            max_tokens=cfg.get("worker_max_tokens", 4096),
            local=True,
        )
+        cloud_max_tokens = int(
+            cfg.get("cloud_max_tokens") or default_max_output_tokens(self._cloud_model)
+        )
        if self._cloud_endpoint == "openai":
            cloud_client = OpenAIClient(
                model_name=self._cloud_model,
                temperature=0.0,
-                max_tokens=4096,
+                max_tokens=cloud_max_tokens,
            )
        elif self._cloud_endpoint == "anthropic":
            # Temperature stripping is handled by the global patch above for Opus 4.7+.
            cloud_client = AnthropicClient(
                model_name=self._cloud_model,
                temperature=0.0,
-                max_tokens=4096,
+                max_tokens=cloud_max_tokens,
            )
        elif self._cloud_endpoint == "gemini":
            # The vendored Minion library already special-cases GeminiClient
@@ -520,7 +549,7 @@ class MinionsAgent(LocalCloudAgent):
            cloud_client = GeminiClient(
                model_name=self._cloud_model,
                temperature=0.0,
-                max_tokens=4096,
+                max_tokens=cloud_max_tokens,
            )
        else:
            raise ValueError(f"unsupported cloud endpoint: {self._cloud_endpoint!r}")
@@ -543,9 +572,14 @@ class MinionsAgent(LocalCloudAgent):
        #   - enabled = false  → prefetch OFF
        #   - enabled = true   → prefetch ON (honors max_uses)
        prefetch: Dict[str, Any] = {
-            "text": "", "tokens": 0, "cost_usd": 0.0, "n_searches": 0,
+            "text": "",
+            "tokens": 0,
+            "cost_usd": 0.0,
+            "n_searches": 0,
        }
-        ws_block = cfg.get("web_search") if isinstance(cfg.get("web_search"), dict) else None
+        ws_block = (
+            cfg.get("web_search") if isinstance(cfg.get("web_search"), dict) else None
+        )
        ws_enabled, ws_max_uses = web_search_cfg(cfg)
        # If the cell explicitly set web_search.enabled = false, honor that.
        # If it set web_search.enabled = true, honor max_uses. If it didn't
@@ -560,17 +594,21 @@ class MinionsAgent(LocalCloudAgent):
                self._cloud_endpoint,
                self._cloud_model,
                max_uses=ws_max_uses,
+                search_backend=str(cfg.get("search_backend", "provider")).lower(),
+                tavily_max_results=int(cfg.get("tavily_max_results", 5)),
            )

        if prefetch.get("text"):
-            self.record_trace_event({
-                "kind": "minions_prefetch",
-                "n_searches": prefetch["n_searches"],
-                "tokens": prefetch["tokens"],
-                "cost_usd": prefetch["cost_usd"],
-                "text": prefetch["text"],
-                "error": prefetch.get("error"),
-            })
+            self.record_trace_event(
+                {
+                    "kind": "minions_prefetch",
+                    "n_searches": prefetch["n_searches"],
+                    "tokens": prefetch["tokens"],
+                    "cost_usd": prefetch["cost_usd"],
+                    "text": prefetch["text"],
+                    "error": prefetch.get("error"),
+                }
+            )

        out = protocol(
            task=input,  # full formatted prompt (with bench instruction)
@@ -582,15 +620,17 @@ class MinionsAgent(LocalCloudAgent):
        # The Minions library doesn't go through our SDK helpers, so the
        # auto-trace missed every turn. Record the protocol output directly —
        # supervisor_messages + worker_messages contain the full conversation.
-        self.record_trace_event({
-            "kind": "minions_protocol",
-            "mode": mode,
-            "supervisor_messages": out.get("supervisor_messages"),
-            "worker_messages": out.get("worker_messages"),
-            "timing": out.get("timing"),
-            "log_file": out.get("log_file"),
-            "final_answer": out.get("final_answer", ""),
-        })
+        self.record_trace_event(
+            {
+                "kind": "minions_protocol",
+                "mode": mode,
+                "supervisor_messages": out.get("supervisor_messages"),
+                "worker_messages": out.get("worker_messages"),
+                "timing": out.get("timing"),
+                "log_file": out.get("log_file"),
+                "final_answer": out.get("final_answer", ""),
+            }
+        )

        local_usage = out.get("local_usage")
        remote_usage = out.get("remote_usage")
@@ -625,7 +665,6 @@ class MinionsAgent(LocalCloudAgent):
        }
        return out.get("final_answer", ""), meta

-
    # ------------------------------------------------------------------
    # SWE-bench variant
    # ------------------------------------------------------------------
@@ -643,21 +682,23 @@ class MinionsAgent(LocalCloudAgent):
        # 1. Cloud supervisor writes a high-level plan (no tools).
        plan_text, p_in, p_out = self._call_cloud(
            user=(
-                f"Issue:\n{task.get('problem_statement','')}\n\n"
-                f"Repo: {task.get('repo','')}\n"
-                f"Base commit: {task.get('base_commit','')}\n\n"
-                f"{task.get('hints_text','')}"
+                f"Issue:\n{task.get('problem_statement', '')}\n\n"
+                f"Repo: {task.get('repo', '')}\n"
+                f"Base commit: {task.get('base_commit', '')}\n\n"
+                f"{task.get('hints_text', '')}"
            ),
            system=MINIONS_SWE_PLANNER_SYS,
            max_tokens=int(cfg.get("supervisor_max_tokens", 1024)),
            temperature=0.0,
        )
-        self.record_trace_event({
-            "kind": "minions_swe_plan",
-            "plan": plan_text,
-            "tokens_in": p_in,
-            "tokens_out": p_out,
-        })
+        self.record_trace_event(
+            {
+                "kind": "minions_swe_plan",
+                "plan": plan_text,
+                "tokens_in": p_in,
+                "tokens_out": p_out,
+            }
+        )
        supervisor_cost = self.cost_usd(self._cloud_model, p_in, p_out)

        # 2. Local worker runs mini-SWE-agent with the plan as context.
@@ -38,13 +38,14 @@ except ModuleNotFoundError:
 from openjarvis.agents._stubs import AgentContext, AgentResult
 from openjarvis.agents.hybrid._energy import EnergyCollector
 from openjarvis.agents.hybrid._prompts import format_prompt as _format_prompt
+from openjarvis.core.paths import get_config_dir

 PACKAGE_DIR = Path(__file__).parent
 DEFAULT_REGISTRY_DIR = PACKAGE_DIR / "registry"
 DEFAULT_EXPERIMENTS_DIR = Path(
    os.environ.get(
        "OPENJARVIS_HYBRID_EXPERIMENTS_DIR",
-        Path.home() / ".openjarvis" / "experiments" / "hybrid",
+        get_config_dir() / "experiments" / "hybrid",
    )
 )
 DEFAULT_SUBSETS_DIR = DEFAULT_EXPERIMENTS_DIR / "subsets"
@@ -122,6 +123,7 @@ def load_registry(registry_dir: Optional[Path] = None) -> Dict[str, Dict[str, An

 # ---------- Bench dispatch ----------

+
 def _load_gaia_tasks(n: Optional[int]) -> List[Dict[str, Any]]:
    """GAIA validation. Each task is a dict with `task_id` + `question`."""
    from openjarvis.evals.datasets.gaia import GAIADataset
@@ -137,12 +139,14 @@ def _load_gaia_tasks(n: Optional[int]) -> List[Dict[str, Any]]:
        # id round-trip.
        md = rec.metadata or {}
        task_id = md.get("task_id") or rec.record_id
-        out.append({
-            "task_id": task_id,
-            "question": md.get("question", rec.problem),
-            "reference": rec.reference,
-            "metadata": dict(md),
-        })
+        out.append(
+            {
+                "task_id": task_id,
+                "question": md.get("question", rec.problem),
+                "reference": rec.reference,
+                "metadata": dict(md),
+            }
+        )
    return out


@@ -155,19 +159,21 @@ def _load_swebench_tasks(n: Optional[int]) -> List[Dict[str, Any]]:
    out: List[Dict[str, Any]] = []
    for rec in ds.iter_records():
        md = rec.metadata or {}
-        out.append({
-            "task_id": md.get("instance_id", rec.record_id),
-            "repo": md.get("repo", ""),
-            "base_commit": md.get("base_commit", ""),
-            "problem_statement": md.get("problem_statement", rec.problem),
-            "hints_text": md.get("hints_text", ""),
-            "test_patch": md.get("test_patch", ""),
-            "FAIL_TO_PASS": md.get("FAIL_TO_PASS", []),
-            "PASS_TO_PASS": md.get("PASS_TO_PASS", []),
-            "version": md.get("version"),
-            "reference": rec.reference,
-            "metadata": dict(md),
-        })
+        out.append(
+            {
+                "task_id": md.get("instance_id", rec.record_id),
+                "repo": md.get("repo", ""),
+                "base_commit": md.get("base_commit", ""),
+                "problem_statement": md.get("problem_statement", rec.problem),
+                "hints_text": md.get("hints_text", ""),
+                "test_patch": md.get("test_patch", ""),
+                "FAIL_TO_PASS": md.get("FAIL_TO_PASS", []),
+                "PASS_TO_PASS": md.get("PASS_TO_PASS", []),
+                "version": md.get("version"),
+                "reference": rec.reference,
+                "metadata": dict(md),
+            }
+        )
    return out


@@ -206,7 +212,9 @@ def _load_subset_file(subset_path: str) -> Dict[str, Any]:
                f"subset {p.name} has no 'task_ids' field; got keys {list(data.keys())}"
            )
        return data
-    raise ValueError(f"subset {p.name} must be a list or dict; got {type(data).__name__}")
+    raise ValueError(
+        f"subset {p.name} must be a list or dict; got {type(data).__name__}"
+    )


 def _apply_subset(
@@ -332,7 +340,11 @@ def _score_swebench(

    patch = extract_patch(answer)
    if patch is None:
-        return {"success": False, "score": 0.0, "details": {"reason": "no_patch_extracted"}}
+        return {
+            "success": False,
+            "score": 0.0,
+            "details": {"reason": "no_patch_extracted"},
+        }

    record = EvalRecord(
        record_id=task["task_id"],
@@ -368,6 +380,7 @@ def score(

 # ---------- Cell run ----------

+
 def _cell_dir(cell_name: str, root: Path) -> Path:
    d = root / cell_name
    d.mkdir(parents=True, exist_ok=True)
@@ -440,8 +453,10 @@ def _error_row(task: Dict[str, Any], t0: float, error: str) -> Dict[str, Any]:
    return {
        "task_id": task["task_id"],
        "answer": "",
-        "tokens_local": 0, "tokens_cloud": 0,
-        "cost_usd": 0.0, "latency_s": time.time() - t0,
+        "tokens_local": 0,
+        "tokens_cloud": 0,
+        "cost_usd": 0.0,
+        "latency_s": time.time() - t0,
        "web_search_uses": 0,
        "tool_calls": 0,
        "n_cloud_calls": 0,
@@ -456,11 +471,13 @@ def _run_one_inner(
 ) -> Dict[str, Any]:
    """Run the agent on one task. Returns a hybrid-shape row."""
    prompt = _format_prompt(task)
-    ctx = AgentContext(metadata={
-        "task": task,
-        "task_id": task["task_id"],
-        "log_dir": log_dir,
-    })
+    ctx = AgentContext(
+        metadata={
+            "task": task,
+            "task_id": task["task_id"],
+            "log_dir": log_dir,
+        }
+    )
    t0 = time.time()
    try:
        result: AgentResult = agent.run(prompt, ctx)
@@ -534,11 +551,12 @@ def _run_one(
    if worker.is_alive():
        print(
            f"[timeout] task={task['task_id']} exceeded "
-            f"{task_timeout_s/60:.1f}m — abandoning worker, recording error row",
+            f"{task_timeout_s / 60:.1f}m — abandoning worker, recording error row",
            flush=True,
        )
        return _error_row(
-            task, t0,
+            task,
+            t0,
            f"TaskTimeout: task exceeded the {task_timeout_s:.0f}s hybrid "
            "per-task wall-clock cap (likely a hung network or Modal-harness "
            "call); worker thread abandoned, task left for resume.",
@@ -558,7 +576,7 @@ def _heartbeat(done: int, total: int, row: Dict[str, Any], t_start: float) -> No
    print(
        f"[{done}/{total}] {ok} task={row['task_id']} score={sc_str} "
        f"local={row['tokens_local']} cloud={row['tokens_cloud']} "
-        f"${row['cost_usd']:.3f} {row['latency_s']:.1f}s eta={eta/60:.1f}m",
+        f"${row['cost_usd']:.3f} {row['latency_s']:.1f}s eta={eta / 60:.1f}m",
        flush=True,
    )

@@ -644,9 +662,9 @@ def _write_summary(
    summary_path.write_text(json.dumps(summary, indent=2))
    print(
        f"[summary] {cell_name}: n={n_done}/{cell['n']} err={n_err} "
-        f"acc={acc:.3f} cost=${total_cost:.2f} time={wall/60:.1f}m "
-        f"energy={energy_j/1000:.1f}kJ "
-        f"(session +{elapsed/60:.1f}m +{energy_j_session/1000:.1f}kJ, "
+        f"acc={acc:.3f} cost=${total_cost:.2f} time={wall / 60:.1f}m "
+        f"energy={energy_j / 1000:.1f}kJ "
+        f"(session +{elapsed / 60:.1f}m +{energy_j_session / 1000:.1f}kJ, "
        f"processed={n_processed})",
        flush=True,
    )
@@ -664,8 +682,11 @@ def run_cell(
    out_dir = _cell_dir(cell_name, out_root)
    with _cell_lock(out_dir, cell_name):
        _run_cell_locked(
-            cell_name, cell, out_dir,
-            do_score=do_score, resume=resume,
+            cell_name,
+            cell,
+            out_dir,
+            do_score=do_score,
+            resume=resume,
        )


@@ -727,7 +748,7 @@ def _run_cell_locked(
    mcfg = cell.get("method_cfg") or {}
    task_timeout_s = float(mcfg.get("task_timeout_s", DEFAULT_TASK_TIMEOUT_S))
    if task_timeout_s > 0:
-        print(f"[task-timeout] {task_timeout_s/60:.1f}m per task", flush=True)
+        print(f"[task-timeout] {task_timeout_s / 60:.1f}m per task", flush=True)

    agent = _build_agent(cell)

@@ -739,18 +760,25 @@ def _run_cell_locked(

    def _process(task: Dict[str, Any]) -> None:
        row = _run_one(
-            agent, cell["bench"], task, log_dir,
+            agent,
+            cell["bench"],
+            task,
+            log_dir,
            task_timeout_s=task_timeout_s,
        )
        scored: Optional[Dict[str, Any]] = None
        if do_score and row.get("error") is None:
            try:
                scored = score(
-                    cell["bench"], task, row["answer"], cell_name=cell_name,
+                    cell["bench"],
+                    task,
+                    row["answer"],
+                    cell_name=cell_name,
                )
            except Exception as e:
                scored = {
-                    "success": False, "score": 0.0,
+                    "success": False,
+                    "score": 0.0,
                    "details": {"score_error": str(e)},
                }
        full_row = {**row, "score": scored}
@@ -789,9 +817,7 @@ def _run_cell_locked(
        while not watchdog_stop.wait(60.0):
            try:
                cur = (
-                    results_path.stat().st_mtime
-                    if results_path.exists()
-                    else last_seen
+                    results_path.stat().st_mtime if results_path.exists() else last_seen
                )
            except Exception:
                cur = last_seen
@@ -830,7 +856,11 @@ def _run_cell_locked(
    watchdog_stop.set()

    _write_summary(
-        out_dir, cell_name, cell, tasks, t_start,
+        out_dir,
+        cell_name,
+        cell,
+        tasks,
+        t_start,
        n_processed=len(pending),
        energy_j_session=energy.energy_j_total,
    )
@@ -838,6 +868,7 @@ def _run_cell_locked(

 # ---------- CLI ----------

+
 def main(argv: Optional[List[str]] = None) -> int:
    p = argparse.ArgumentParser(
        prog="python -m openjarvis.agents.hybrid.runner",
@@ -855,13 +886,18 @@ def main(argv: Optional[List[str]] = None) -> int:
        help="Override experiments output root.",
    )
    p.add_argument("--no-score", action="store_true", help="Skip scoring.")
-    p.add_argument("--no-resume", action="store_true", help="Don't resume from results.jsonl.")
+    p.add_argument(
+        "--no-resume", action="store_true", help="Don't resume from results.jsonl."
+    )
    args = p.parse_args(argv)

    reg_dir = Path(args.registry_dir) if args.registry_dir else None
    cells = load_registry(reg_dir)
    if not cells:
-        print(f"[error] no cells found in {reg_dir or DEFAULT_REGISTRY_DIR}", file=sys.stderr)
+        print(
+            f"[error] no cells found in {reg_dir or DEFAULT_REGISTRY_DIR}",
+            file=sys.stderr,
+        )
        return 2
    if args.cell not in cells:
        print(
@@ -871,7 +907,8 @@ def main(argv: Optional[List[str]] = None) -> int:
        return 2
    root = Path(args.root) if args.root else None
    run_cell(
-        args.cell, cells[args.cell],
+        args.cell,
+        cells[args.cell],
        do_score=not args.no_score,
        resume=not args.no_resume,
        root=root,
@@ -47,33 +47,33 @@ from openjarvis.core.registry import AgentRegistry
 # would seed before any oracle update.

 SKILL_CATALOG: Dict[str, str] = {
-    "factual_recall":     "Recall named entities, dates, places, well-known facts from training data without external lookup.",
+    "factual_recall": "Recall named entities, dates, places, well-known facts from training data without external lookup.",
    "multi_step_reasoning": "Chain several inference steps together (e.g. compose dates, traverse relationships, decompose then aggregate).",
-    "arithmetic":         "Exact numeric computation on values already given in the question.",
-    "web_grounding":      "Question needs information likely NOT in a small model's parametric memory (rare facts, recent events, niche sources).",
+    "arithmetic": "Exact numeric computation on values already given in the question.",
+    "web_grounding": "Question needs information likely NOT in a small model's parametric memory (rare facts, recent events, niche sources).",
    "long_text_extraction": "Read a long supplied document/context and extract a specific piece.",
-    "format_compliance":  "Strict output formatting (e.g. GAIA's `FINAL ANSWER: <answer>` rule, comma-separated lists with no units).",
-    "code_or_logic":      "Write or trace code, or apply logical/symbolic constraints precisely.",
+    "format_compliance": "Strict output formatting (e.g. GAIA's `FINAL ANSWER: <answer>` rule, comma-separated lists with no units).",
+    "code_or_logic": "Write or trace code, or apply logical/symbolic constraints precisely.",
 }

 DEFAULT_AGENT_COMPETENCE: Dict[str, Dict[str, float]] = {
    "local-qwen-27b": {
-        "factual_recall":       0.25,
+        "factual_recall": 0.25,
        "multi_step_reasoning": 0.30,
-        "arithmetic":           0.55,
-        "web_grounding":        0.10,
+        "arithmetic": 0.55,
+        "web_grounding": 0.10,
        "long_text_extraction": 0.55,
-        "format_compliance":    0.65,
-        "code_or_logic":        0.45,
+        "format_compliance": 0.65,
+        "code_or_logic": 0.45,
    },
    "cloud-opus-4-7": {
-        "factual_recall":       0.85,
+        "factual_recall": 0.85,
        "multi_step_reasoning": 0.88,
-        "arithmetic":           0.85,
-        "web_grounding":        0.70,
+        "arithmetic": 0.85,
+        "web_grounding": 0.70,
        "long_text_extraction": 0.90,
-        "format_compliance":    0.92,
-        "code_or_logic":        0.90,
+        "format_compliance": 0.92,
+        "code_or_logic": 0.90,
    },
 }

@@ -149,6 +149,17 @@ def _build_router_schema(agent_ids: List[str]) -> Dict[str, Any]:
    }


+def _openai_response_format(schema: Dict[str, Any]) -> Dict[str, Any]:
+    return {
+        "type": "json_schema",
+        "json_schema": {
+            "name": "skillorchestra_route",
+            "schema": schema["format"]["schema"],
+            "strict": True,
+        },
+    }
+
+
 def _parse_router_json(text: str) -> Dict[str, Any]:
    s = (text or "").strip()
    try:
@@ -178,9 +189,7 @@ def _score_agents(
    lam = 0.5
    scores: Dict[str, Dict[str, float]] = {}
    for aid, comps in competence.items():
-        comp = sum(
-            skill_weights.get(sid, 0.0) * comps[sid] for sid in SKILL_CATALOG
-        )
+        comp = sum(skill_weights.get(sid, 0.0) * comps[sid] for sid in SKILL_CATALOG)
        cost_pen = lam * cost.get(aid, 0.0)
        scores[aid] = {
            "competence": comp,
@@ -196,6 +205,70 @@ class SkillOrchestraAgent(LocalCloudAgent):

    agent_id = "skillorchestra"

+    def _route_call(
+        self,
+        *,
+        question: str,
+        router_sys: str,
+        router_schema: Dict[str, Any],
+        router_max: int,
+    ) -> Tuple[str, int, int]:
+        user = f"Question:\n{question}"
+        if self._cloud_endpoint == "anthropic":
+            kwargs: Dict[str, Any] = {
+                "user": user,
+                "system": router_sys,
+                "max_tokens": router_max,
+                "output_config": router_schema,
+            }
+            if supports_temperature(self._cloud_model):
+                kwargs["temperature"] = 0.0
+            text, r_in, r_out, _ = self._call_anthropic(
+                self._cloud_model,
+                **kwargs,
+            )
+            return text, r_in, r_out
+        if self._cloud_endpoint == "openai":
+            return self._call_openai(
+                self._cloud_model,
+                user=user,
+                system=router_sys,
+                max_tokens=router_max,
+                temperature=0.0,
+                response_format=_openai_response_format(router_schema),
+            )
+        if self._cloud_endpoint == "gemini":
+            return self._call_gemini(
+                self._cloud_model,
+                user=user,
+                system=router_sys,
+                max_tokens=router_max,
+                temperature=0.0,
+            )
+        raise ValueError(
+            f"SkillOrchestra router unsupported cloud_endpoint={self._cloud_endpoint!r}"
+        )
+
+    def _executor_call(
+        self,
+        *,
+        question: str,
+        max_tokens: int,
+    ) -> Tuple[str, int, int]:
+        if self._cloud_endpoint == "anthropic":
+            text, w_in, w_out, _ = self._call_anthropic(
+                self._cloud_model,
+                user=question,
+                max_tokens=max_tokens,
+                temperature=0.0,
+            )
+            return text, w_in, w_out
+        return self._call_cloud(
+            user=question,
+            max_tokens=max_tokens,
+            temperature=0.0,
+        )
+
    def _is_soft_failure(self, exc: BaseException) -> Optional[str]:
        # Empty/unbalanced router JSON — treat as soft failure to match the
        # hybrid adapter's behavior (matches `err=1` rows in the n=30 cell).
@@ -220,33 +293,13 @@ class SkillOrchestraAgent(LocalCloudAgent):
        router_sys = _build_router_sys(competence, cost)
        router_schema = _build_router_schema(agent_ids)

-        # 1. Route — Anthropic only (output_config schema is Anthropic-specific
-        # in the hybrid adapter). If you need OpenAI routing, swap the prompt
-        # to JSON-mode and bypass output_config.
-        if self._cloud_endpoint != "anthropic":
-            raise ValueError(
-                "SkillOrchestra router requires cloud_endpoint='anthropic'; "
-                f"got {self._cloud_endpoint!r}"
-            )
        router_max = int(cfg.get("router_max_tokens", 1024))
-        # Strip temperature for Opus 4.7+; Anthropic's output_config does the schema.
-        if supports_temperature(self._cloud_model):
-            router_text, r_in, r_out, _ = self._call_anthropic(
-                self._cloud_model,
-                user=f"Question:\n{question}",
-                system=router_sys,
-                max_tokens=router_max,
-                temperature=0.0,
-                output_config=router_schema,
-            )
-        else:
-            router_text, r_in, r_out, _ = self._call_anthropic(
-                self._cloud_model,
-                user=f"Question:\n{question}",
-                system=router_sys,
-                max_tokens=router_max,
-                output_config=router_schema,
-            )
+        router_text, r_in, r_out = self._route_call(
+            question=question,
+            router_sys=router_sys,
+            router_schema=router_schema,
+            router_max=router_max,
+        )

        decision = _parse_router_json(router_text)
        skill_weights: Dict[str, float] = decision.get("skill_weights") or {}
@@ -258,14 +311,16 @@ class SkillOrchestraAgent(LocalCloudAgent):
        if chosen not in competence:
            chosen = max(scored, key=lambda a: scored[a]["final_score"])

-        self.record_trace_event({
-            "kind": "skillorchestra_route",
-            "chosen_agent": chosen,
-            "skill_weights": skill_weights,
-            "agent_scores": scored,
-            "reasoning": decision.get("reasoning", ""),
-            "router_raw": router_text,
-        })
+        self.record_trace_event(
+            {
+                "kind": "skillorchestra_route",
+                "chosen_agent": chosen,
+                "skill_weights": skill_weights,
+                "agent_scores": scored,
+                "reasoning": decision.get("reasoning", ""),
+                "router_raw": router_text,
+            }
+        )

        tokens_local = 0
        tokens_cloud = r_in + r_out
@@ -329,11 +384,9 @@ class SkillOrchestraAgent(LocalCloudAgent):
                tokens_cloud += out["tokens_in"] + out["tokens_out"]
                run_cost += out["cost_usd"]
            else:
-                ans, w_in, w_out, _ = self._call_anthropic(
-                    self._cloud_model,
-                    user=question,
+                ans, w_in, w_out = self._executor_call(
+                    question=question,
                    max_tokens=int(cfg.get("cloud_max_tokens", 4096)),
-                    temperature=0.0,
                )
                tokens_cloud += w_in + w_out
                run_cost += self.cost_usd(self._cloud_model, w_in, w_out)
@@ -41,8 +41,12 @@ from .orchestrator import run_orchestrator
 from .stage_router import StageSkillHandbook

 _VALID_STRATEGIES = {
-    "none", "router_decides", "analyze_model_decide",
-    "weighted_avg", "weakest_skill", "strongest_skill",
+    "none",
+    "router_decides",
+    "analyze_model_decide",
+    "weighted_avg",
+    "weakest_skill",
+    "strongest_skill",
 }


@@ -146,7 +150,11 @@ class SkillOrchestraAgent(LocalCloudAgent):
            strategy = "none"

        return run_orchestrator(
-            self, input, cfg=cfg, handbook=handbook, strategy=strategy,
+            self,
+            input,
+            cfg=cfg,
+            handbook=handbook,
+            strategy=strategy,
        )


@@ -31,7 +31,14 @@ from .stage_router import (
    get_routing_strategy,
    parse_skill_analysis,
 )
-from .tools import anthropic_tools, openai_tools, run_answer, run_code, run_search
+from .tools import (
+    anthropic_tools,
+    gemini_tools,
+    openai_tools,
+    run_answer,
+    run_code,
+    run_search,
+)

 # tool name -> routing stage (stage_router uses "reasoning" for code).
 _TOOL_STAGE = {
@@ -51,6 +58,7 @@ _STAGE_DEFAULT_ALIAS = {
 # Orchestrator decision step (raw SDK — needs tool_use blocks back)
 # ---------------------------------------------------------------------------

+
 def _orchestrate_step(
    agent: Any,
    *,
@@ -115,24 +123,63 @@ def _orchestrate_step(
        u = resp.usage
        p = getattr(u, "prompt_tokens", 0) if u else 0
        c = getattr(u, "completion_tokens", 0) if u else 0
+    elif endpoint == "gemini":
+        from google import genai
+        from google.genai import types
+
+        client = genai.Client(http_options=types.HttpOptions(timeout=600_000))
+        cfg = types.GenerateContentConfig(
+            temperature=1.0,
+            max_output_tokens=max_tokens,
+            tools=[types.Tool(function_declarations=gemini_tools())],
+        )
+        resp = client.models.generate_content(
+            model=model,
+            contents=user,
+            config=cfg,
+        )
+        text = (resp.text or "") if hasattr(resp, "text") else ""
+        tool_calls = []
+        try:
+            parts = resp.candidates[0].content.parts or []
+        except Exception:  # noqa: BLE001
+            parts = []
+        for part in parts:
+            fc = getattr(part, "function_call", None)
+            if fc is None:
+                continue
+            name = getattr(fc, "name", None)
+            if not isinstance(name, str) or not name:
+                continue
+            args = getattr(fc, "args", None) or {}
+            try:
+                args = dict(args)
+            except Exception:  # noqa: BLE001
+                args = {}
+            tool_calls.append({"name": name, "input": args})
+        um = getattr(resp, "usage_metadata", None)
+        p = int(getattr(um, "prompt_token_count", 0) or 0) if um else 0
+        c = int(getattr(um, "candidates_token_count", 0) or 0) if um else 0
    else:
        raise ValueError(
            f"orchestrator endpoint {endpoint!r} unsupported — route the "
-            "orchestrator through anthropic/openai (set method_cfg."
+            "orchestrator through anthropic/openai/gemini (set method_cfg."
            "orchestrator_endpoint)."
        )

    cost = agent.cost_usd(model, p, c)
-    agent.record_trace_event({
-        "kind": "skillorchestra_orchestrate",
-        "model": model,
-        "endpoint": endpoint,
-        "prompt": user,
-        "response": text,
-        "tool_calls": tool_calls,
-        "tokens_in": p,
-        "tokens_out": c,
-    })
+    agent.record_trace_event(
+        {
+            "kind": "skillorchestra_orchestrate",
+            "model": model,
+            "endpoint": endpoint,
+            "prompt": user,
+            "response": text,
+            "tool_calls": tool_calls,
+            "tokens_in": p,
+            "tokens_out": c,
+        }
+    )
    return text, tool_calls, p, c, cost


@@ -140,6 +187,7 @@ def _orchestrate_step(
 # Context assembly  —  eval_frames.py:1305-1351
 # ---------------------------------------------------------------------------

+
 def _build_context(
    doc_list: List[Tuple[str, str]],
    code_list: List[Tuple[str, str]],
@@ -177,6 +225,7 @@ def _build_context(
 # Main loop  —  eval_frames.py:run_single
 # ---------------------------------------------------------------------------

+
 def run_orchestrator(
    agent: Any,
    problem: str,
@@ -192,18 +241,22 @@ def run_orchestrator(
    code_timeout = int(cfg.get("code_timeout_s", 60))
    answer_max_tokens = int(cfg.get("answer_max_tokens", 40000))
    ws_max_uses = int(cfg.get("web_search_max_uses", 5))
+    search_backend = str(cfg.get("search_backend", "provider")).lower()
+    tavily_max_results = int(cfg.get("tavily_max_results", 5))

    # The orchestrator model: a fixed model per run (the original's
    # MODEL_NAME). Defaults to the cell's cloud model when that endpoint
    # supports tool calls, else Opus. ``router_model`` / ``router_endpoint``
    # are accepted as back-compat aliases (pre-restructure cfg key names).
-    orch_endpoint = (cfg.get("orchestrator_endpoint")
-                     or cfg.get("router_endpoint")
-                     or agent._cloud_endpoint).lower()
-    orch_model = (cfg.get("orchestrator_model")
-                  or cfg.get("router_model")
-                  or agent._cloud_model)
-    if orch_endpoint not in ("anthropic", "openai"):
+    orch_endpoint = (
+        cfg.get("orchestrator_endpoint")
+        or cfg.get("router_endpoint")
+        or agent._cloud_endpoint
+    ).lower()
+    orch_model = (
+        cfg.get("orchestrator_model") or cfg.get("router_model") or agent._cloud_model
+    )
+    if orch_endpoint not in ("anthropic", "openai", "gemini"):
        orch_endpoint, orch_model = "anthropic", "claude-opus-4-7"
    orch_max_tokens = int(cfg.get("orchestrator_max_tokens", 4096))

@@ -233,7 +286,9 @@ def run_orchestrator(
        if handbook is not None and strategy != "none":
            sa = parse_skill_analysis(orch_text)
            rr = get_routing_strategy(strategy, handbook).select_model(
-                stage, sa, tool_call_model=tool_alias,
+                stage,
+                sa,
+                tool_call_model=tool_alias,
            )
            return rr.model_alias
        return tool_alias or _STAGE_DEFAULT_ALIAS[stage]
@@ -242,7 +297,10 @@ def run_orchestrator(
        used_rounds = step + 1
        is_last = step == max_rounds - 1
        context_str = _build_context(
-            doc_list, code_list, attempt_list, char_cap=char_cap,
+            doc_list,
+            code_list,
+            attempt_list,
+            char_cap=char_cap,
        )

        if handbook is not None and strategy != "none":
@@ -253,14 +311,14 @@ def run_orchestrator(
                handbook=handbook,
            )
        else:
-            user = (
-                f"Problem: {problem}\n\n{context_str}\n\n"
-                "Choose an appropriate tool."
-            )
+            user = f"Problem: {problem}\n\n{context_str}\n\nChoose an appropriate tool."

        text, tcalls, p, c, ocost = _orchestrate_step(
-            agent, user=user, model=orch_model,
-            endpoint=orch_endpoint, max_tokens=orch_max_tokens,
+            agent,
+            user=user,
+            model=orch_model,
+            endpoint=orch_endpoint,
+            max_tokens=orch_max_tokens,
        )
        tokens_cloud += p + c
        cost_usd += ocost
@@ -289,23 +347,31 @@ def run_orchestrator(
            tool_alias = (tc.get("input") or {}).get("model")
            stage = _TOOL_STAGE.get(tool, "answer")
            chosen_alias = _route(stage, tool_alias, text)
-            spec: ModelSpec = pool.get(chosen_alias) or pool[
-                _STAGE_DEFAULT_ALIAS[stage]
-            ]
-            route_log.append({
-                "step": step,
-                "tool": tool,
-                "orchestrator_alias": tool_alias,
-                "routed_alias": chosen_alias,
-                "routed_model": spec.model,
-                "is_local": spec.is_local,
-            })
+            spec: ModelSpec = (
+                pool.get(chosen_alias) or pool[_STAGE_DEFAULT_ALIAS[stage]]
+            )
+            route_log.append(
+                {
+                    "step": step,
+                    "tool": tool,
+                    "orchestrator_alias": tool_alias,
+                    "routed_alias": chosen_alias,
+                    "routed_model": spec.model,
+                    "is_local": spec.is_local,
+                }
+            )
            tool_calls_n += 1

            if tool == "search":
                res = run_search(
-                    agent, spec, context_str=context_str, problem=problem,
-                    retriever_url=retriever_url, web_search_max_uses=ws_max_uses,
+                    agent,
+                    spec,
+                    context_str=context_str,
+                    problem=problem,
+                    retriever_url=retriever_url,
+                    web_search_max_uses=ws_max_uses,
+                    search_backend=search_backend,
+                    tavily_max_results=tavily_max_results,
                )
                docs = res["search_results_data"]
                joined = "\n---\n".join(d for d in docs if d)[:char_cap]
@@ -313,13 +379,19 @@ def run_orchestrator(
                web_uses += res.get("web_search_uses", 0)
            elif tool in ("enhance_reasoning", "code"):
                res = run_code(
-                    agent, spec, context_str=context_str, problem=problem,
+                    agent,
+                    spec,
+                    context_str=context_str,
+                    problem=problem,
                    bash_timeout_s=code_timeout,
                )
                code_list.append((res["generated_code"], res["exec_result"]))
            else:  # answer
                res = run_answer(
-                    agent, spec, context_str=context_str, problem=problem,
+                    agent,
+                    spec,
+                    context_str=context_str,
+                    problem=problem,
                    max_tokens=answer_max_tokens,
                )
                final_pred = res["pred"]
@@ -335,12 +407,14 @@ def run_orchestrator(
        if finish:
            break

-    agent.record_trace_event({
-        "kind": "skillorchestra_route_log",
-        "strategy": strategy,
-        "rounds_used": used_rounds,
-        "routes": route_log,
-    })
+    agent.record_trace_event(
+        {
+            "kind": "skillorchestra_route_log",
+            "strategy": strategy,
+            "rounds_used": used_rounds,
+            "routes": route_log,
+        }
+    )

    meta = {
        "tokens_local": tokens_local,
@@ -32,8 +32,14 @@ from typing import Any, Dict, List, Optional, Tuple
 STAGE_ALIASES: Dict[str, List[str]] = {
    "search": ["search-1", "search-2", "search-3"],
    "reasoning": ["reasoner-1", "reasoner-2", "reasoner-3"],
-    "answer": ["answer-1", "answer-2", "answer-3", "answer-4",
-               "answer-math-1", "answer-math-2"],
+    "answer": [
+        "answer-1",
+        "answer-2",
+        "answer-3",
+        "answer-4",
+        "answer-math-1",
+        "answer-math-2",
+    ],
 }

 # Every alias the orchestrator can emit, flat.
@@ -42,9 +48,13 @@ ALL_ALIASES: List[str] = [a for aliases in STAGE_ALIASES.values() for a in alias
 # Default tier: which aliases collapse onto the cloud model vs the local
 # model. Dearer ``-1``/``-2`` (+ answer-math-1) -> cloud; cheaper -> local.
 _CLOUD_ALIASES = {
-    "search-1", "search-2",
-    "reasoner-1", "reasoner-2",
-    "answer-1", "answer-2", "answer-math-1",
+    "search-1",
+    "search-2",
+    "reasoner-1",
+    "reasoner-2",
+    "answer-1",
+    "answer-2",
+    "answer-math-1",
 }


@@ -54,8 +64,8 @@ class ModelSpec:

    alias: str
    model: str
-    endpoint: str          # "anthropic" | "openai" | "gemini" | "http://..."
-    kind: str              # "cloud" | "local"
+    endpoint: str  # "anthropic" | "openai" | "gemini" | "http://..."
+    kind: str  # "cloud" | "local"

    @property
    def is_local(self) -> bool:
@@ -87,7 +97,10 @@ def build_pool(
            pool[alias] = ModelSpec(alias, cloud_model, cloud_endpoint, "cloud")
        else:
            pool[alias] = ModelSpec(
-                alias, local_model, local_endpoint, "local"  # type: ignore[arg-type]
+                alias,
+                local_model,
+                local_endpoint,
+                "local",  # type: ignore[arg-type]
            )

    for alias, spec in (overrides or {}).items():
@@ -133,18 +146,30 @@ def call_alias(
    ep = spec.endpoint.lower()
    if ep == "anthropic":
        text, p, c, _ = agent._call_anthropic(
-            spec.model, user=user, system=system,
-            max_tokens=max_tokens, temperature=temperature, trace_role="cloud",
+            spec.model,
+            user=user,
+            system=system,
+            max_tokens=max_tokens,
+            temperature=temperature,
+            trace_role="cloud",
        )
    elif ep == "openai":
        text, p, c = agent._call_openai(
-            spec.model, user=user, system=system,
-            max_tokens=max_tokens, temperature=temperature, trace_role="cloud",
+            spec.model,
+            user=user,
+            system=system,
+            max_tokens=max_tokens,
+            temperature=temperature,
+            trace_role="cloud",
        )
    elif ep == "gemini":
        text, p, c = agent._call_gemini(
-            spec.model, user=user, system=system,
-            max_tokens=max_tokens, temperature=temperature, trace_role="cloud",
+            spec.model,
+            user=user,
+            system=system,
+            max_tokens=max_tokens,
+            temperature=temperature,
+            trace_role="cloud",
        )
    else:
        raise ValueError(f"unsupported pool endpoint: {spec.endpoint!r}")
@@ -94,7 +94,12 @@ class StageSkillHandbook:
            "answer": {},
        }
        self.model_profiles: Dict[str, ModelProfile] = {}
-        self.usage_patterns: Dict[str, Any] = {"stages": {}, "guidelines": {}, "models": {}, "raw": {}}
+        self.usage_patterns: Dict[str, Any] = {
+            "stages": {},
+            "guidelines": {},
+            "models": {},
+            "raw": {},
+        }
        self.routing_insights: List[str] = []
        self.learning_history: List[Dict[str, Any]] = []
        self.version = "1.0.0"
@@ -102,7 +107,10 @@ class StageSkillHandbook:
        self.updated_at = ""

    def get_model_skill_scores(self) -> Dict[str, Dict[str, float]]:
-        return {alias: profile.skill_scores for alias, profile in self.model_profiles.items()}
+        return {
+            alias: profile.skill_scores
+            for alias, profile in self.model_profiles.items()
+        }

    def get_models_for_stage(self, stage: str) -> List[ModelProfile]:
        return [p for p in self.model_profiles.values() if p.stage == stage]
@@ -135,19 +143,29 @@ class StageSkillHandbook:

    def format_model_performance(self, stage: str) -> str:
        profiles = self.get_models_for_stage(stage)
-        valid_prefixes = {"search": ["search-"], "code": ["reasoner-", "code-"], "answer": ["answer-"]}
+        valid_prefixes = {
+            "search": ["search-"],
+            "code": ["reasoner-", "code-"],
+            "answer": ["answer-"],
+        }
        prefixes = valid_prefixes.get(stage, [])

        lines = []
        for p in profiles:
            if not any(p.model_alias.startswith(prefix) for prefix in prefixes):
                continue
-            has_data = (p.skill_scores and len(p.skill_scores) > 0) or p.strengths or p.weaknesses
+            has_data = (
+                (p.skill_scores and len(p.skill_scores) > 0)
+                or p.strengths
+                or p.weaknesses
+            )
            if p.total_attempts > 0 or has_data:
                lines.append(f"\n### {p.model_alias} ({p.actual_model})")
                if p.total_attempts > 0:
                    rate = p.total_successes / p.total_attempts
-                    lines.append(f"Overall: {rate:.0%} success ({p.total_successes}/{p.total_attempts})")
+                    lines.append(
+                        f"Overall: {rate:.0%} success ({p.total_successes}/{p.total_attempts})"
+                    )
                else:
                    lines.append("Overall: 0% overall")
                if p.skill_scores:
@@ -157,7 +175,9 @@ class StageSkillHandbook:
                        for sid, s in p.skill_scores.items()
                        if (sid.split(".")[0] if "." in sid else sid) in ("code", stage)
                    }
-                    for skill_id, score in sorted(stage_skill_scores.items(), key=lambda x: x[1], reverse=True):
+                    for skill_id, score in sorted(
+                        stage_skill_scores.items(), key=lambda x: x[1], reverse=True
+                    ):
                        lines.append(f"  - {skill_id}: {score:.0%}")
                if p.strengths:
                    lines.append(f"Strengths: {', '.join(p.strengths[:3])}")
@@ -220,7 +240,9 @@ def parse_skill_analysis(output: str) -> Optional[SkillAnalysis]:
    try:
        data = json.loads(match.group(1).strip())
        required_skills = [
-            SkillWeight(skill_id=s.get("skill_id", ""), percentage=float(s.get("percentage", 0)))
+            SkillWeight(
+                skill_id=s.get("skill_id", ""), percentage=float(s.get("percentage", 0))
+            )
            for s in data.get("required_skills", [])
        ]
        return SkillAnalysis(
@@ -272,7 +294,14 @@ class RoutingStrategy:
        if stage == "reasoning":
            return ["reasoner-1", "reasoner-2", "reasoner-3"]
        if stage == "answer":
-            return ["answer-1", "answer-2", "answer-3", "answer-4", "answer-math-1", "answer-math-2"]
+            return [
+                "answer-1",
+                "answer-2",
+                "answer-3",
+                "answer-4",
+                "answer-math-1",
+                "answer-math-2",
+            ]
        return []

    def select_model(
@@ -292,9 +321,17 @@ class RouterDecidesStrategy(RoutingStrategy):
        tool_call_model: Optional[str] = None,
    ) -> ModelRoutingResult:
        if tool_call_model:
-            return ModelRoutingResult(tool_call_model, "router_decides_from_tool_call", 1.0)
-        defaults = {"search": "search-1", "reasoning": "reasoner-1", "answer": "answer-1"}
-        return ModelRoutingResult(defaults.get(stage, "answer-1"), "router_decides_fallback", 0.5)
+            return ModelRoutingResult(
+                tool_call_model, "router_decides_from_tool_call", 1.0
+            )
+        defaults = {
+            "search": "search-1",
+            "reasoning": "reasoner-1",
+            "answer": "answer-1",
+        }
+        return ModelRoutingResult(
+            defaults.get(stage, "answer-1"), "router_decides_fallback", 0.5
+        )


 class AnalyzeModelDecideStrategy(RoutingStrategy):
@@ -305,17 +342,33 @@ class AnalyzeModelDecideStrategy(RoutingStrategy):
        tool_call_model: Optional[str] = None,
    ) -> ModelRoutingResult:
        if tool_call_model:
-            return ModelRoutingResult(tool_call_model, "analyze_model_decide_with_skill_analysis", 1.0)
-        defaults = {"search": "search-1", "reasoning": "reasoner-1", "answer": "answer-1"}
-        return ModelRoutingResult(defaults.get(stage, "answer-1"), "analyze_model_decide_fallback", 0.5)
+            return ModelRoutingResult(
+                tool_call_model, "analyze_model_decide_with_skill_analysis", 1.0
+            )
+        defaults = {
+            "search": "search-1",
+            "reasoning": "reasoner-1",
+            "answer": "answer-1",
+        }
+        return ModelRoutingResult(
+            defaults.get(stage, "answer-1"), "analyze_model_decide_fallback", 0.5
+        )


 class WeightedAverageStrategy(RoutingStrategy):
    COST_TIERS = {
-        "search-3": 1, "search-2": 2, "search-1": 3,
-        "reasoner-3": 1, "reasoner-2": 2, "reasoner-1": 3,
-        "answer-math-2": 1, "answer-4": 1, "answer-3": 2,
-        "answer-math-1": 2, "answer-2": 3, "answer-1": 4,
+        "search-3": 1,
+        "search-2": 2,
+        "search-1": 3,
+        "reasoner-3": 1,
+        "reasoner-2": 2,
+        "reasoner-1": 3,
+        "answer-math-2": 1,
+        "answer-4": 1,
+        "answer-3": 2,
+        "answer-math-1": 2,
+        "answer-2": 3,
+        "answer-1": 4,
    }

    def select_model(
@@ -326,9 +379,17 @@ class WeightedAverageStrategy(RoutingStrategy):
    ) -> ModelRoutingResult:
        if not skill_analysis or not skill_analysis.required_skills:
            if tool_call_model:
-                return ModelRoutingResult(tool_call_model, "weighted_avg_no_skills_use_tool_call", 0.7)
-            defaults = {"search": "search-1", "reasoning": "reasoner-1", "answer": "answer-1"}
-            return ModelRoutingResult(defaults.get(stage, "answer-1"), "weighted_avg_no_skills_fallback", 0.5)
+                return ModelRoutingResult(
+                    tool_call_model, "weighted_avg_no_skills_use_tool_call", 0.7
+                )
+            defaults = {
+                "search": "search-1",
+                "reasoning": "reasoner-1",
+                "answer": "answer-1",
+            }
+            return ModelRoutingResult(
+                defaults.get(stage, "answer-1"), "weighted_avg_no_skills_fallback", 0.5
+            )

        models = self._get_models_for_stage(stage)
        model_scores = {}
@@ -341,15 +402,25 @@ class WeightedAverageStrategy(RoutingStrategy):
                score = scores.get(sid, 0.0)
                weighted_sum += weight * score
                total_weight += weight
-            model_scores[model] = weighted_sum / total_weight if total_weight > 0 else 0.5
+            model_scores[model] = (
+                weighted_sum / total_weight if total_weight > 0 else 0.5
+            )

        if not model_scores:
-            defaults = {"search": "search-1", "reasoning": "reasoner-1", "answer": "answer-1"}
-            return ModelRoutingResult(defaults.get(stage, "answer-1"), "weighted_avg_no_model_scores", 0.5)
+            defaults = {
+                "search": "search-1",
+                "reasoning": "reasoner-1",
+                "answer": "answer-1",
+            }
+            return ModelRoutingResult(
+                defaults.get(stage, "answer-1"), "weighted_avg_no_model_scores", 0.5
+            )
        max_score = max(model_scores.values())
        best = [m for m, s in model_scores.items() if abs(s - max_score) < 0.001]
        best.sort(key=lambda m: self.COST_TIERS.get(m, 999))
-        return ModelRoutingResult(best[0], "weighted_avg_from_skill_analysis", max_score, model_scores)
+        return ModelRoutingResult(
+            best[0], "weighted_avg_from_skill_analysis", max_score, model_scores
+        )


 class WeakestSkillStrategy(RoutingStrategy):
@@ -361,18 +432,36 @@ class WeakestSkillStrategy(RoutingStrategy):
    ) -> ModelRoutingResult:
        if not skill_analysis or not skill_analysis.required_skills:
            if tool_call_model:
-                return ModelRoutingResult(tool_call_model, "weakest_skill_no_skills_use_tool_call", 0.7)
-            defaults = {"search": "search-1", "reasoning": "reasoner-1", "answer": "answer-1"}
-            return ModelRoutingResult(defaults.get(stage, "answer-1"), "weakest_skill_no_skills_fallback", 0.5)
+                return ModelRoutingResult(
+                    tool_call_model, "weakest_skill_no_skills_use_tool_call", 0.7
+                )
+            defaults = {
+                "search": "search-1",
+                "reasoning": "reasoner-1",
+                "answer": "answer-1",
+            }
+            return ModelRoutingResult(
+                defaults.get(stage, "answer-1"), "weakest_skill_no_skills_fallback", 0.5
+            )
        weakest = min(skill_analysis.required_skills, key=lambda s: s.percentage)
        sid = self._find_skill_id(stage, weakest.skill_id) or weakest.skill_id
        models = self._get_models_for_stage(stage)
-        model_scores = {m: self._model_skill_scores.get(m, {}).get(sid, 0.5) for m in models}
+        model_scores = {
+            m: self._model_skill_scores.get(m, {}).get(sid, 0.5) for m in models
+        }
        if not model_scores:
-            defaults = {"search": "search-1", "reasoning": "reasoner-1", "answer": "answer-1"}
-            return ModelRoutingResult(defaults.get(stage, "answer-1"), "weakest_skill_no_model_scores", 0.5)
+            defaults = {
+                "search": "search-1",
+                "reasoning": "reasoner-1",
+                "answer": "answer-1",
+            }
+            return ModelRoutingResult(
+                defaults.get(stage, "answer-1"), "weakest_skill_no_model_scores", 0.5
+            )
        best = max(model_scores, key=model_scores.get)
-        return ModelRoutingResult(best, f"weakest_skill_{weakest.skill_id}", model_scores[best], model_scores)
+        return ModelRoutingResult(
+            best, f"weakest_skill_{weakest.skill_id}", model_scores[best], model_scores
+        )


 class StrongestSkillStrategy(RoutingStrategy):
@@ -384,18 +473,41 @@ class StrongestSkillStrategy(RoutingStrategy):
    ) -> ModelRoutingResult:
        if not skill_analysis or not skill_analysis.required_skills:
            if tool_call_model:
-                return ModelRoutingResult(tool_call_model, "strongest_skill_no_skills_use_tool_call", 0.7)
-            defaults = {"search": "search-1", "reasoning": "reasoner-1", "answer": "answer-1"}
-            return ModelRoutingResult(defaults.get(stage, "answer-1"), "strongest_skill_no_skills_fallback", 0.5)
+                return ModelRoutingResult(
+                    tool_call_model, "strongest_skill_no_skills_use_tool_call", 0.7
+                )
+            defaults = {
+                "search": "search-1",
+                "reasoning": "reasoner-1",
+                "answer": "answer-1",
+            }
+            return ModelRoutingResult(
+                defaults.get(stage, "answer-1"),
+                "strongest_skill_no_skills_fallback",
+                0.5,
+            )
        strongest = max(skill_analysis.required_skills, key=lambda s: s.percentage)
        sid = self._find_skill_id(stage, strongest.skill_id) or strongest.skill_id
        models = self._get_models_for_stage(stage)
-        model_scores = {m: self._model_skill_scores.get(m, {}).get(sid, 0.5) for m in models}
+        model_scores = {
+            m: self._model_skill_scores.get(m, {}).get(sid, 0.5) for m in models
+        }
        if not model_scores:
-            defaults = {"search": "search-1", "reasoning": "reasoner-1", "answer": "answer-1"}
-            return ModelRoutingResult(defaults.get(stage, "answer-1"), "strongest_skill_no_model_scores", 0.5)
+            defaults = {
+                "search": "search-1",
+                "reasoning": "reasoner-1",
+                "answer": "answer-1",
+            }
+            return ModelRoutingResult(
+                defaults.get(stage, "answer-1"), "strongest_skill_no_model_scores", 0.5
+            )
        best = max(model_scores, key=model_scores.get)
-        return ModelRoutingResult(best, f"strongest_skill_{strongest.skill_id}", model_scores[best], model_scores)
+        return ModelRoutingResult(
+            best,
+            f"strongest_skill_{strongest.skill_id}",
+            model_scores[best],
+            model_scores,
+        )


 ROUTING_STRATEGIES = {
@@ -27,6 +27,7 @@ from .._base import (
    OPENAI_WEB_SEARCH_COST_PER_CALL,
    WEB_SEARCH_COST_PER_CALL,
    build_web_search_tool,
+    tavily_search_context,
 )
 from .pool import ModelSpec, call_alias

@@ -40,8 +41,7 @@ _SEARCH_CAPABLE_ENDPOINTS = ("anthropic", "openai", "gemini")

 _SEARCH_DESC = "Search for missing information."
 _CODE_DESC = (
-    "Write and execute Python code to compute intermediate results for "
-    "the problem."
+    "Write and execute Python code to compute intermediate results for the problem."
 )
 _ANSWER_DESC = (
    "Extract the final answer when you have gathered enough information "
@@ -51,8 +51,14 @@ _ANSWER_DESC = (
 _ENUMS = {
    "search": ["search-1", "search-2", "search-3"],
    "enhance_reasoning": ["reasoner-1", "reasoner-2", "reasoner-3"],
-    "answer": ["answer-1", "answer-2", "answer-3", "answer-4",
-               "answer-math-1", "answer-math-2"],
+    "answer": [
+        "answer-1",
+        "answer-2",
+        "answer-3",
+        "answer-4",
+        "answer-math-1",
+        "answer-math-2",
+    ],
 }


@@ -75,15 +81,17 @@ def anthropic_tools() -> List[Dict[str, Any]]:
        ("enhance_reasoning", _CODE_DESC),
        ("answer", _ANSWER_DESC),
    ):
-        out.append({
-            "name": name,
-            "description": desc,
-            "input_schema": {
-                "type": "object",
-                "properties": {"model": _model_prop(name)},
-                "required": ["model"],
-            },
-        })
+        out.append(
+            {
+                "name": name,
+                "description": desc,
+                "input_schema": {
+                    "type": "object",
+                    "properties": {"model": _model_prop(name)},
+                    "required": ["model"],
+                },
+            }
+        )
    return out


@@ -95,9 +103,33 @@ def openai_tools() -> List[Dict[str, Any]]:
        ("enhance_reasoning", _CODE_DESC),
        ("answer", _ANSWER_DESC),
    ):
-        out.append({
-            "type": "function",
-            "function": {
+        out.append(
+            {
+                "type": "function",
+                "function": {
+                    "name": name,
+                    "description": desc,
+                    "parameters": {
+                        "type": "object",
+                        "properties": {"model": _model_prop(name)},
+                        "required": ["model"],
+                    },
+                },
+            }
+        )
+    return out
+
+
+def gemini_tools() -> List[Dict[str, Any]]:
+    """The 3 orchestrator tools in Gemini function-declaration shape."""
+    out = []
+    for name, desc in (
+        ("search", _SEARCH_DESC),
+        ("enhance_reasoning", _CODE_DESC),
+        ("answer", _ANSWER_DESC),
+    ):
+        out.append(
+            {
                "name": name,
                "description": desc,
                "parameters": {
@@ -105,8 +137,8 @@ def openai_tools() -> List[Dict[str, Any]]:
                    "properties": {"model": _model_prop(name)},
                    "required": ["model"],
                },
-            },
-        })
+            }
+        )
    return out


@@ -114,6 +146,7 @@ def openai_tools() -> List[Dict[str, Any]]:
 # enhance_reasoning / code  —  eval_frames.py:659-812
 # ---------------------------------------------------------------------------

+
 def run_code(
    agent: Any,
    spec: ModelSpec,
@@ -129,7 +162,8 @@ def run_code(
    rather than raising — the orchestrator learns the model can't code.
    """
    prompt = (
-        context_str.strip() + "\n\n"
+        context_str.strip()
+        + "\n\n"
        + f"Question: {problem}\nInstead of directly answering the question, "
        "please write additional python code that will give intermidiate "
        "results after execution. Wrap the code within ```python and ```. "
@@ -137,7 +171,11 @@ def run_code(
        "initialization."
    )
    text, p, c, cost = call_alias(
-        agent, spec, user=prompt, max_tokens=8000, temperature=1.0,
+        agent,
+        spec,
+        user=prompt,
+        max_tokens=8000,
+        temperature=1.0,
    )
    generated_code = ""
    if "```python" in text:
@@ -176,6 +214,7 @@ def run_code(
 # answer  —  eval_frames.py:814-997
 # ---------------------------------------------------------------------------

+
 def run_answer(
    agent: Any,
    spec: ModelSpec,
@@ -198,11 +237,15 @@ def run_answer(
    boxed = False

    if "qwen3" in model_l and "235" not in model_l:
-        system = "Please reason step by step, and put your final answer within \\boxed{}."
+        system = (
+            "Please reason step by step, and put your final answer within \\boxed{}."
+        )
        user = base
        boxed = True
    elif "qwen2.5-math" in model_l or "qwen-2.5-math" in model_l:
-        system = "Please reason step by step, and put your final answer within \\boxed{}."
+        system = (
+            "Please reason step by step, and put your final answer within \\boxed{}."
+        )
        user = base
        boxed = True
    else:
@@ -215,8 +258,12 @@ def run_answer(
        )

    text, p, c, cost = call_alias(
-        agent, spec, user=user, system=system,
-        max_tokens=max_tokens, temperature=1.0,
+        agent,
+        spec,
+        user=user,
+        system=system,
+        max_tokens=max_tokens,
+        temperature=1.0,
    )

    pred = ""
@@ -247,6 +294,7 @@ def run_answer(
 # search  —  eval_frames.py:999-1096
 # ---------------------------------------------------------------------------

+
 def run_search(
    agent: Any,
    spec: ModelSpec,
@@ -256,6 +304,8 @@ def run_search(
    retriever_url: Optional[str] = None,
    topk: int = 150,
    web_search_max_uses: int = 5,
+    search_backend: str = "provider",
+    tavily_max_results: int = 5,
 ) -> Dict[str, Any]:
    """Write a search query with ``spec``, then retrieve documents.

@@ -265,13 +315,18 @@ def run_search(
    OpenJarvis substitution for the missing FAISS wiki index).
    """
    prompt = (
-        context_str.strip() + "\n\n"
+        context_str.strip()
+        + "\n\n"
        + f"Question: {problem}\nInstead of directly answering the question, "
        "please think hard and write a concise query to search Wikipedia. "
        "Wrap the query within <query> and </query>."
    )
    text, p, c, cost = call_alias(
-        agent, spec, user=prompt, max_tokens=8000, temperature=1.0,
+        agent,
+        spec,
+        user=prompt,
+        max_tokens=8000,
+        temperature=1.0,
    )
    if "<query>" in text:
        query = text.split("<query>")[-1].split("</query>")[0].strip()
@@ -283,7 +338,12 @@ def run_search(
    contents: List[str] = []
    search_uses = 0

-    if retriever_url:
+    if search_backend == "tavily":
+        res = tavily_search_context(query, max_results=tavily_max_results)
+        contents.append(res["text"])
+        search_uses = int(res["n_searches"])
+        cost += float(res["cost_usd"])
+    elif retriever_url:
        # Faithful path — the original FAISS retriever service.
        import requests

@@ -294,7 +354,9 @@ def run_search(
        }
        try:
            results = requests.post(
-                f"{retriever_url.rstrip('/')}/retrieve", json=payload, timeout=120,
+                f"{retriever_url.rstrip('/')}/retrieve",
+                json=payload,
+                timeout=120,
            ).json()
            for r in results[0]:
                doc = r.get("document", {})
@@ -20,6 +20,7 @@ from typing import Any, Dict, List, Optional
 # BetaCompetence
 # ---------------------------------------------------------------------------

+
@dataclass
 class BetaCompetence:
    """Bayesian competence estimate for an agent on a specific skill.
@@ -80,6 +81,7 @@ class BetaCompetence:
 # CostStats
 # ---------------------------------------------------------------------------

+
@dataclass
 class CostStats:
    """Execution cost statistics for an agent under a specific mode.
@@ -110,11 +112,17 @@ class CostStats:
        """Incremental running-average update."""
        n = self.total_executions
        self.avg_prompt_tokens = (self.avg_prompt_tokens * n + prompt_tokens) / (n + 1)
-        self.avg_completion_tokens = (self.avg_completion_tokens * n + completion_tokens) / (n + 1)
+        self.avg_completion_tokens = (
+            self.avg_completion_tokens * n + completion_tokens
+        ) / (n + 1)
        self.avg_latency_s = (self.avg_latency_s * n + latency_s) / (n + 1)
        self.avg_cost_usd = (self.avg_cost_usd * n + cost_usd) / (n + 1)
-        self.avg_completion_cost_usd = (self.avg_completion_cost_usd * n + completion_cost_usd) / (n + 1)
-        self.avg_prompt_cost_usd = (self.avg_prompt_cost_usd * n + prompt_cost_usd) / (n + 1)
+        self.avg_completion_cost_usd = (
+            self.avg_completion_cost_usd * n + completion_cost_usd
+        ) / (n + 1)
+        self.avg_prompt_cost_usd = (self.avg_prompt_cost_usd * n + prompt_cost_usd) / (
+            n + 1
+        )
        self.total_executions = n + 1

    def to_dict(self) -> Dict[str, Any]:
@@ -137,6 +145,7 @@ class CostStats:
 # RoutingInsight
 # ---------------------------------------------------------------------------

+
@dataclass
 class RoutingInsight:
    """A single routing insight learned from execution traces"""
@@ -165,6 +174,7 @@ class RoutingInsight:
 # ModeMetadata
 # ---------------------------------------------------------------------------

+
@dataclass
 class ModeMetadata:
    """Mode-level routing metadata."""
@@ -197,6 +207,7 @@ class ModeMetadata:
 # Skill
 # ---------------------------------------------------------------------------

+
@dataclass
 class SkillProvenance:
    """Tracks how and why a skill was discovered."""
@@ -231,7 +242,7 @@ class Skill:
    indicators: List[str] = field(default_factory=list)
    examples: List[str] = field(default_factory=list)
    mode: str = ""
-    parent_skill_id: Optional[str] = None # for hierarchical skills
+    parent_skill_id: Optional[str] = None  # for hierarchical skills
    provenance: SkillProvenance = field(default_factory=SkillProvenance)

    def to_dict(self) -> Dict[str, Any]:
@@ -273,6 +284,7 @@ class Skill:
 # AgentProfile
 # ---------------------------------------------------------------------------

+
@dataclass
 class AgentProfile:
    """Agent profile for skill-aware orchestration."""
@@ -309,9 +321,7 @@ class AgentProfile:
        """Update competence estimate for a skill."""
        self.get_competence_dist(skill_id).update(success)

-    def weighted_competence(
-        self, skill_weights: Dict[str, float]
-    ) -> float:
+    def weighted_competence(self, skill_weights: Dict[str, float]) -> float:
        """Compute weighted competence: sum w_{t,sigma} * alpha/(alpha+beta)."""
        if not skill_weights:
            return 0.5
@@ -334,9 +344,7 @@ class AgentProfile:
        ]
        return sum(scores) / len(scores) if scores else 0.0

-    def category_competence_for_skills(
-        self, active_skill_ids: List[str]
-    ) -> float:
+    def category_competence_for_skills(self, active_skill_ids: List[str]) -> float:
        """Category-level competence for hierarchical tie-breaking.

        Extracts parent categories from active_skill_ids (e.g. 'entertainment_knowledge'
@@ -349,7 +357,9 @@ class AgentProfile:
            categories.add(cat)
        if not categories:
            return 0.0
-        return sum(self.category_competence(cat) for cat in categories) / len(categories)
+        return sum(self.category_competence(cat) for cat in categories) / len(
+            categories
+        )

    @property
    def overall_success_rate(self) -> float:
@@ -390,8 +400,12 @@ class AgentProfile:
            "skill_scores": skill_scores,
            "skill_attempts": skill_attempts,
            "skill_successes": skill_successes,
-            "total_attempts": self.total_attempts if self.total_attempts > 0 else skill_total_attempts,
-            "total_successes": self.total_successes if self.total_attempts > 0 else skill_total_successes,
+            "total_attempts": self.total_attempts
+            if self.total_attempts > 0
+            else skill_total_attempts,
+            "total_successes": self.total_successes
+            if self.total_attempts > 0
+            else skill_total_successes,
            "cost_stats": self.cost_stats.to_dict(),
            "routing_signals": self.routing_signals,
            "strengths": self.strengths,
@@ -11,10 +11,11 @@ import logging
 import sqlite3
 import time
 import uuid
-from pathlib import Path
 from typing import Any, Dict, List, Optional
 from uuid import uuid4

+from openjarvis.core.paths import get_config_dir
+
 logger = logging.getLogger(__name__)

 _CREATE_AGENTS = """\
@@ -537,7 +538,7 @@ class AgentManager:
            pass

        # User templates
-        user_dir = Path("~/.openjarvis/templates").expanduser()
+        user_dir = get_config_dir() / "templates"
        if user_dir.is_dir():
            for f in user_dir.glob("*.toml"):
                try:
--- a/Show More
+++ b/Show More