ci: parallelize the test suite and cut install/coverage overhead (#580)

Run pytest with -n auto (pytest-xdist) and COVERAGE_CORE=sysmon, enable the uv cache, and switch test output to -q. Cuts the test job from ~40min to ~4min without changing what's tested or the 60% coverage gate.

.github/clone-stats/badge.json

@@ -1,7 +1,7 @@
 {
   "schemaVersion": 1,
   "label": "Git Clones",
-  "message": "110,366",
+  "message": "128,077",
   "color": "green",
   "namedLogo": "git"
 }

.github/clone-stats/clone-data.json

@@ -1,6 +1,6 @@
 {
-  "total_clones": 110366,
-  "last_updated": "2026-06-11T07:44:18Z",
+  "total_clones": 128077,
+  "last_updated": "2026-06-22T08:05:29Z",
   "daily": {
     "2026-03-27": 2189,
     "2026-03-28": 1874,
@@ -77,6 +77,17 @@
     "2026-06-07": 1174,
     "2026-06-08": 2369,
     "2026-06-09": 1361,
-    "2026-06-10": 1310
+    "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
   }
 }

.github/workflows/autotag.yml

@@ -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].*$//')

.github/workflows/ci.yml

@@ -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
@@ -55,6 +57,8 @@ 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 +67,13 @@ 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 +116,8 @@ 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

.github/workflows/claude-issues.yml

@@ -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]')

.github/workflows/claude-review.yml

@@ -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:

.github/workflows/desktop.yml

@@ -114,6 +114,11 @@ 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
 
       - name: Install system dependencies (Linux)
         if: matrix.platform == 'ubuntu-22.04'
@@ -183,7 +188,16 @@ 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 "$BASE" | cut -d. -f3 | sed -E 's/[^0-9].*$//')

.github/workflows/pypi-publish.yml

@@ -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
@@ -67,27 +72,41 @@ 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

CHANGELOG.md

@@ -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

deploy/windows/README.md

@@ -20,7 +20,7 @@ What it does:
 4. Installs `uv` (https://astral.sh/uv) if absent.
 5. Clones the OpenJarvis repository to `%LOCALAPPDATA%\OpenJarvis`
    (override with `$env:OPENJARVIS_HOME`).
-6. Runs `uv sync --extra server` so the FastAPI server entry point is
+6. Runs `uv sync --extra desktop` so the FastAPI server and speech backend are
    importable.
 7. Optionally prompts to register a scheduled task that auto-starts the
    server at logon.
@@ -105,7 +105,7 @@ To pull the latest:
 ```powershell
 cd "$env:LOCALAPPDATA\OpenJarvis\src"
 git pull --ff-only
-uv sync --extra server
+uv sync --extra desktop
 ```
 
 Or re-run the installer with `-Force`:

deploy/windows/install.ps1

@@ -16,8 +16,8 @@
       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` so the FastAPI server and speech
+         backend are importable.
       7. Optionally register the scheduled-task service (see
          deploy/windows/jarvis-service.ps1).
 
@@ -279,13 +279,13 @@ if (Test-Path (Join-Path $srcDir '.git')) {
 }
 
 # ---------------------------------------------------------------------------
-# 6. uv sync --extra server
+# 6. uv sync --extra desktop
 # ---------------------------------------------------------------------------
 
-Write-Info "Running 'uv sync --extra server' in $srcDir (this can take a few minutes)..."
+Write-Info "Running 'uv sync --extra desktop' in $srcDir (this can take a few minutes)..."
 Push-Location $srcDir
 try {
-    & $uvExe sync --extra server
+    & $uvExe sync --extra desktop
     if ($LASTEXITCODE -ne 0) {
         Write-Fail "uv sync failed with exit code $LASTEXITCODE. Check the output above."
     }

docs/downloads.md

@@ -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 ..
     ```
 

docs/getting-started/configuration.md

@@ -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

docs/getting-started/installation.md

@@ -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

docs/getting-started/windows-native.md

@@ -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`.
 - 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`.
 7. Prompt to register the scheduled-task service (skip with
    `-SkipService`).
 

docs/user-guide/cli.md

@@ -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:

frontend/src-tauri/src/lib.rs

@@ -682,7 +682,7 @@ fn format_uv_sync_failure(
     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.",
+         `uv sync --extra desktop` manually for the full output.",
         root.display(),
         code,
         uv_sync_stderr_tail(stderr, 800),
@@ -1143,7 +1143,7 @@ async fn boot_backend(backend: SharedBackend, status: SharedStatus) {
     sync_cmd
         .args([
             "sync",
-            "--extra", "server",
+            "--extra", "desktop",
             "--extra", "inference-cloud",
             "--extra", "inference-google",
         ])
@@ -1664,11 +1664,29 @@ 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.
@@ -2556,7 +2574,7 @@ 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("uv sync --extra desktop")); // actionable next step
     }
 
     #[test]

frontend/src/components/Chat/InputArea.tsx

@@ -97,7 +97,13 @@ export function InputArea() {
   const setDeepResearch = useAppStore((s) => s.setDeepResearch);
   const corpusSync = useResearchCorpusSync(deepResearch);
 
-  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 +128,12 @@ export function InputArea() {
     : streamState.isStreaming ? 'streaming'
     : undefined;
 
+  useEffect(() => {
+    if (speechError) {
+      toast.error(speechError, { duration: 8000 });
+    }
+  }, [speechError]);
+
   const handleMicClick = useCallback(async () => {
     if (speechState === 'recording') {
       try {

frontend/src/lib/api.ts

@@ -317,8 +317,9 @@ 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 +328,16 @@ 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();
 }
 

frontend/src/lib/connectors-api.ts

@@ -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',

frontend/src/lib/sse.ts

@@ -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(`${base}/v1/chat/completions`, {
     method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
+    headers: authHeaders({ 'Content-Type': 'application/json' }),
     body: JSON.stringify(request),
     signal,
   });
@@ -67,7 +67,7 @@ export async function* streamResearch(
   const base = getBase().replace(/\/v1\/?$/, '');
   const response = await fetch(`${base}/api/research`, {
     method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
+    headers: authHeaders({ 'Content-Type': 'application/json' }),
     body: JSON.stringify({ query }),
     signal,
   });
@@ -106,3 +106,4 @@ export async function* streamResearch(
     reader.releaseLock();
   }
 }
+

frontend/src/pages/DataSourcesPage.tsx

@@ -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

frontend/src/pages/GetStartedPage.tsx

@@ -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.

frontend/src/types/connectors.ts

@@ -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: [

pyproject.toml

@@ -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"]
@@ -154,6 +162,30 @@ Issues = "https://github.com/open-jarvis/OpenJarvis/issues"
 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.hatch.build.targets.wheel]
 packages = ["src/openjarvis"]
 

scripts/install/install.sh

@@ -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"

scripts/quickstart.sh

@@ -148,7 +148,7 @@ 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 --quiet 2>/dev/null || uv sync --extra desktop
 ok "Python dependencies installed"
 
 # ── 7b. Build Rust extension ──────────────────────────────────────

src/openjarvis/agents/claude_code.py

@@ -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

src/openjarvis/agents/digest_store.py

@@ -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")

src/openjarvis/agents/hybrid/_base.py

@@ -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 }``.
 
@@ -1322,6 +1363,9 @@ __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 +1373,6 @@ __all__ = [
     "estimate_cost",
     "is_gpt5_family",
     "supports_temperature",
+    "tavily_search_context",
     "web_search_cfg",
 ]

src/openjarvis/agents/hybrid/_prices.py

@@ -15,13 +15,16 @@ PRICES: dict[str, tuple[float, float]] = {
     "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), 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 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),
@@ -61,7 +64,11 @@ 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:

src/openjarvis/agents/hybrid/advisors.py

@@ -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 (
@@ -135,7 +136,12 @@ 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 +152,23 @@ 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(
+            (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}",
@@ -196,14 +206,17 @@ class AdvisorsAgent(LocalCloudAgent):
             f"answer-format rules."
         )
         if use_ws:
-            final_answer, e2_in, e2_out, n_s2, e2_turns = self._executor_search(
+            (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 +229,10 @@ 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 +249,9 @@ 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 +269,40 @@ 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 +311,9 @@ 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 +321,9 @@ 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 +331,7 @@ 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(

src/openjarvis/agents/hybrid/conductor.py

@@ -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 (
@@ -456,8 +457,14 @@ 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()
@@ -470,6 +477,7 @@ 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 +493,16 @@ 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 +562,8 @@ 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 +577,22 @@ 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 +603,7 @@ 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 +612,14 @@ 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 +635,7 @@ 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 +648,7 @@ 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 +660,14 @@ 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}")
 
 
@@ -672,7 +700,7 @@ 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,
@@ -753,13 +781,17 @@ 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,
@@ -833,7 +865,7 @@ 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
                 if (w.get("endpoint") or "openai").lower()
@@ -894,7 +926,7 @@ 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({
@@ -911,6 +943,7 @@ class ConductorAgent(LocalCloudAgent):
                         ),
                     })
 
+                extra_cost = 0.0
                 if swe_mode:
                     text, w_in, w_out, is_local, n_searches, bash_turns = (
                         _swe_worker_step(
@@ -919,7 +952,9 @@ class ConductorAgent(LocalCloudAgent):
                     )
                     tool_calls += bash_turns
                 else:
-                    text, w_in, w_out, is_local, n_searches = _call_worker(
+                    (
+                        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,
@@ -930,7 +965,10 @@ 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({
@@ -981,6 +1019,7 @@ 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": [

src/openjarvis/agents/hybrid/minions.py

@@ -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
 
@@ -362,6 +363,8 @@ def _prefetch_context(
     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.
 
@@ -375,6 +378,22 @@ def _prefetch_context(
     out: Dict[str, Any] = {
         "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:
@@ -498,18 +517,22 @@ 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 +543,7 @@ 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}")
@@ -560,6 +583,8 @@ 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"):

src/openjarvis/agents/hybrid/runner.py

@@ -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,

src/openjarvis/agents/hybrid/skillorchestra.py

@@ -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:
@@ -196,6 +207,70 @@ 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 +295,13 @@ 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 {}
@@ -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)

src/openjarvis/agents/hybrid/skillorchestra/orchestrator.py

@@ -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 = {
@@ -115,10 +122,49 @@ 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)."
         )
 
@@ -192,6 +238,8 @@ 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
@@ -203,7 +251,7 @@ def run_orchestrator(
     orch_model = (cfg.get("orchestrator_model")
                   or cfg.get("router_model")
                   or agent._cloud_model)
-    if orch_endpoint not in ("anthropic", "openai"):
+    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))
 
@@ -306,6 +354,8 @@ def run_orchestrator(
                 res = run_search(
                     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]

src/openjarvis/agents/hybrid/skillorchestra/tools.py

@@ -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
 
@@ -110,6 +111,26 @@ def openai_tools() -> List[Dict[str, Any]]:
     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": {
+                "type": "object",
+                "properties": {"model": _model_prop(name)},
+                "required": ["model"],
+            },
+        })
+    return out
+
+
 # ---------------------------------------------------------------------------
 # enhance_reasoning / code  —  eval_frames.py:659-812
 # ---------------------------------------------------------------------------
@@ -256,6 +277,8 @@ 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.
 
@@ -283,7 +306,12 @@ 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
 

src/openjarvis/agents/hybrid/toolorchestra.py

@@ -16,7 +16,8 @@ Two modes, gated by ``method_cfg.orchestrator_mode``:
   (``answer-1``, ``reasoner-2``, ``search-3``, …) is mapped to a real
   backend through ``EXPERT_MODEL_MAPPING`` — by default the frontier
   Anthropic worker for `*-1` slots, gpt-5-mini for `*-2`, local Qwen
-  for `*-3`. Search routes to the Anthropic server-side web_search.
+  for `*-3`. Search routes to the configured provider's server-side
+  web-search helper when available.
 
   We do NOT reproduce the upstream Tavily / FAISS-wiki retriever, the
   code-interpreter sandbox, or the multi-vLLM mix (Llama-3.3-70B,
@@ -43,8 +44,8 @@ Prompted-mode pipeline:
    prompt; fallback to strongest worker on parse failure.
 
 Workers come from ``cfg["workers"]`` or a sensible default pool (local
-Qwen if vLLM up, plus a web-search tool via Anthropic, Opus 4.7,
-gpt-5-mini).
+Qwen if vLLM up, plus provider-native web search, the configured frontier
+cloud model, and gpt-5-mini).
 """
 
 from __future__ import annotations
@@ -59,8 +60,11 @@ from typing import Any, Dict, List, Optional, Tuple
 from openjarvis.agents._stubs import AgentContext
 from openjarvis.agents.hybrid._base import (
     ANTHROPIC_WEB_SEARCH_TOOL,
+    GEMINI_SEARCH_COST_PER_CALL,
+    OPENAI_WEB_SEARCH_COST_PER_CALL,
     WEB_SEARCH_COST_PER_CALL,
     LocalCloudAgent,
+    tavily_search_context,
 )
 from openjarvis.agents.hybrid._prices import (
     PRICES,
@@ -197,10 +201,23 @@ def _expert_for(slot: str, local_model: Optional[str],
                                   cost tier for mid OpenAI calls)
       - `*-3` (local tier)     -> local vLLM (`local_model`)
       - `answer-math-*`        -> same tiers as the numeric suffix
-      - `search-*`             -> always the Anthropic web_search tool (the
-                                  upstream uses Tavily; we have web_search)
+      - `search-*`             -> provider-native web search when the cloud
+                                  endpoint supports it; otherwise Anthropic
     """
     if slot.startswith("search"):
+        ep = (cloud_endpoint or "anthropic").lower()
+        if ep == "openai":
+            return {
+                "name": f"search:{slot}",
+                "type": "openai-web-search",
+                "model": cloud_model,
+            }
+        if ep == "gemini":
+            return {
+                "name": f"search:{slot}",
+                "type": "gemini-web-search",
+                "model": cloud_model,
+            }
         return {
             "name": f"search:{slot}",
             "type": "anthropic-web-search",
@@ -340,21 +357,18 @@ def _paper_expert_for(
 
 # ---- Tavily + Modal helpers -------------------------------------------------
 
-def _call_tavily_search(query: str, max_results: int = 5) -> Tuple[str, int, int]:
-    """One-shot Tavily search. Returns (text, p_tok=0, c_tok=0).
+def _call_tavily_search(
+    query: str,
+    max_results: int = 5,
+) -> Tuple[str, int, int, float, int]:
+    """One-shot Tavily search. Returns (text, p_tok=0, c_tok=0, cost, uses).
 
     Token counts are reported as zero (no LLM was billed); the OpenJarvis
     accounting layer separately tallies tool-call counts. Falls back to
     DuckDuckGo if Tavily is unreachable (see ``WebSearchTool``).
     """
-    from openjarvis.tools.web_search import WebSearchTool
-
-    tool = WebSearchTool(max_results=max_results)
-    res = tool.execute(query=query, max_results=max_results)
-    text = res.content or ""
-    if not res.success and not text:
-        text = "(no results)"
-    return text, 0, 0
+    res = tavily_search_context(query, max_results=max_results)
+    return res["text"], 0, 0, float(res["cost_usd"]), int(res["n_searches"])
 
 
 _MODAL_APP_NAME = "openjarvis-toolorchestra-sandbox"
@@ -674,13 +688,25 @@ def _default_pool(
                 "concise extraction, formatting, arithmetic on given data."
             ),
         })
+    if ep == "openai":
+        search_type = "openai-web-search"
+        search_model = cloud_model
+        search_desc = "OpenAI hosted web search on the configured frontier model."
+    elif ep == "gemini":
+        search_type = "gemini-web-search"
+        search_model = cloud_model
+        search_desc = "Gemini Google Search grounding on the configured frontier model."
+    else:
+        search_type = "anthropic-web-search"
+        search_model = _DEFAULT_WEB_SEARCH_MODEL
+        search_desc = "Anthropic server-side web_search."
     pool.append({
         "id": len(pool),
         "name": "web-search",
-        "type": "anthropic-web-search",
-        "model": "claude-haiku-4-5",
+        "type": search_type,
+        "model": search_model,
         "description": (
-            "Anthropic server-side web_search. Use for facts that need a lookup "
+            f"{search_desc} Use for facts that need a lookup "
             "(recent events, rare names/dates, niche sources). Returns a digest."
         ),
     })
@@ -717,8 +743,13 @@ def _default_pool(
 #   `modal-python`   — One-shot Python exec in a fresh Modal Sandbox (the
 #                      paper's "Python sandbox" inside `enhance_reasoning`).
 _TOOLORCH_VALID_TYPES = (
-    "vllm", "openai", "anthropic", "anthropic-web-search", "gemini",
-    "tavily-search", "openrouter", "modal-python",
+    "vllm", "openai", "anthropic", "anthropic-web-search",
+    "openai-web-search", "gemini", "gemini-web-search", "tavily-search",
+    "openrouter", "modal-python",
+)
+_TOOLORCH_SEARCH_TYPES = (
+    "anthropic-web-search", "openai-web-search", "gemini-web-search",
+    "tavily-search",
 )
 
 # Default model used when an `anthropic-web-search` entry omits `model`.
@@ -739,10 +770,12 @@ def _resolve_worker_pool(
     the override is absent.
 
     Each user-supplied entry must be a dict with keys ``id``, ``name``,
-    ``type``, and (for non-search types) ``model``. ``type`` must be one
-    of ``vllm`` / ``openai`` / ``anthropic`` / ``anthropic-web-search``.
-    ``anthropic-web-search`` entries may omit ``model`` — it defaults to
-    ``claude-haiku-4-5``.
+    ``type``, and (for non-search types) ``model``. Search worker types are
+    ``anthropic-web-search``, ``openai-web-search``, ``gemini-web-search``,
+    and ``tavily-search``. ``anthropic-web-search`` entries may omit
+    ``model`` — it defaults to ``claude-haiku-4-5``. OpenAI and Gemini
+    search workers default to the configured cloud model. Tavily does not
+    require a model.
 
     Substitution: ``model = "$local"`` (or ``"<local>"``) resolves to
     ``local_model``; ``model = "$cloud"`` / ``"<cloud>"`` to ``cloud_model``.
@@ -804,14 +837,24 @@ def _resolve_worker_pool(
         elif isinstance(model, str) and model in ("$cloud", "<cloud>"):
             model = cloud_model
             entry["model"] = model
-        if wtype == "anthropic-web-search":
+        if wtype in _TOOLORCH_SEARCH_TYPES:
             if model in (None, ""):
-                model = _DEFAULT_WEB_SEARCH_MODEL
+                if wtype == "anthropic-web-search":
+                    model = _DEFAULT_WEB_SEARCH_MODEL
+                elif wtype in ("openai-web-search", "gemini-web-search"):
+                    model = cloud_model
+                else:
+                    model = wtype
                 entry["model"] = model
             elif not isinstance(model, str):
                 raise ValueError(
                     f"Invalid worker_pool entry [{wid}]: 'model' must be a string when set"
                 )
+            if wtype in ("openai-web-search", "gemini-web-search") and model not in PRICES:
+                raise ValueError(
+                    f"Invalid worker_pool entry [{wid}]: model {model!r} "
+                    f"is not in PRICES (known: {sorted(PRICES)})"
+                )
             # Search workers don't satisfy the "needs a solver" requirement.
         else:
             if not isinstance(model, str) or not model:
@@ -843,7 +886,7 @@ def _resolve_worker_pool(
     if not has_non_search:
         raise ValueError(
             "Invalid worker_pool entry [-]: worker_pool must contain at least "
-            "one non-search worker (vllm / openai / anthropic)"
+            "one non-search worker (vllm / openai / anthropic / gemini)"
         )
     return resolved
 
@@ -910,13 +953,31 @@ def _call_worker(
         )
         extra = n_searches * WEB_SEARCH_COST_PER_CALL
         return text, p, c, False, extra, n_searches
+    if wtype == "openai-web-search":
+        eff_temp = 1.0 if is_gpt5_family(worker["model"]) else temp
+        text, p, c, n_searches, _ = LocalCloudAgent._call_openai_agent(
+            worker["model"],
+            user=prompt,
+            max_tokens=max(max_tok, 16384) if is_gpt5_family(worker["model"]) else max_tok,
+            temperature=eff_temp,
+        )
+        extra = n_searches * OPENAI_WEB_SEARCH_COST_PER_CALL
+        return text, p, c, False, extra, n_searches
+    if wtype == "gemini-web-search":
+        text, p, c, n_searches, _ = LocalCloudAgent._call_gemini_agent(
+            worker["model"],
+            user=prompt,
+            max_tokens=max_tok,
+            temperature=temp,
+        )
+        extra = n_searches * GEMINI_SEARCH_COST_PER_CALL
+        return text, p, c, False, extra, n_searches
     if wtype == "tavily-search":
-        # Tavily costs are flat per call; charge `WEB_SEARCH_COST_PER_CALL`
-        # for parity with the Anthropic web-search worker. One call = one
-        # "n_search" for accounting.
         max_results = int(cfg.get("tavily_max_results", 5))
-        text, p, c = _call_tavily_search(str(prompt), max_results=max_results)
-        return text, p, c, False, WEB_SEARCH_COST_PER_CALL, 1
+        text, p, c, extra, n_searches = _call_tavily_search(
+            str(prompt), max_results=max_results,
+        )
+        return text, p, c, False, extra, n_searches
     if wtype == "openrouter":
         text, p, c = LocalCloudAgent._call_openrouter(
             worker["model"],
@@ -951,7 +1012,7 @@ def _swe_call_worker(
     caller can surface ``tool_calls`` per row. Fallbacks to one-shot
     workers return 0 bash turns (no agent loop ran)."""
     wtype = worker.get("type", "openai")
-    if wtype == "anthropic-web-search":
+    if wtype in _TOOLORCH_SEARCH_TYPES:
         # Search workers stay one-shot.
         text, p, c, is_local, extra, n_searches = _call_worker(worker, prompt, cfg)
         return text, p, c, is_local, extra, n_searches, 0
@@ -1197,7 +1258,8 @@ class ToolOrchestraAgent(LocalCloudAgent):
                 # Search workers are excluded — they answer fact-lookup
                 # questions, not synthesis.
                 non_search = [
-                    w for w in workers if w.get("type") != "anthropic-web-search"
+                    w for w in workers
+                    if w.get("type") not in _TOOLORCH_SEARCH_TYPES
                 ] or workers
                 worker = max(
                     non_search,

src/openjarvis/agents/manager.py

@@ -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:

src/openjarvis/agents/morning_digest.py

@@ -13,6 +13,7 @@ from typing import Any, List, Optional
 
 from openjarvis.agents._stubs import AgentContext, AgentResult, ToolUsingAgent
 from openjarvis.agents.digest_store import DigestArtifact, DigestStore
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.registry import AgentRegistry
 from openjarvis.core.types import Message, Role, ToolCall
 
@@ -21,7 +22,7 @@ def _load_persona(persona_name: str) -> str:
     """Load a persona prompt file by name."""
     search_paths = [
         Path("configs/openjarvis/prompts/personas") / f"{persona_name}.md",
-        Path.home() / ".openjarvis" / "prompts" / "personas" / f"{persona_name}.md",
+        get_config_dir() / "prompts" / "personas" / f"{persona_name}.md",
     ]
     for p in search_paths:
         if p.exists():
@@ -202,7 +203,7 @@ class MorningDigestAgent(ToolUsingAgent):
         tts_text = re.sub(r"\*{1,2}([^*]+)\*{1,2}", r"\1", tts_text)
         tts_text = tts_text.strip()
 
-        output_dir = str(Path.home() / ".openjarvis" / "digests")
+        output_dir = str(get_config_dir() / "digests")
         tts_call = ToolCall(
             id="digest-tts-1",
             name="text_to_speech",

src/openjarvis/agents/proactive_agent.py

@@ -43,6 +43,7 @@ from typing import Any, Dict, List, Optional, Set
 
 from openjarvis.agents._stubs import AgentContext, AgentResult, ToolUsingAgent
 from openjarvis.core.config import load_config
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.registry import AgentRegistry
 from openjarvis.core.types import Message, Role, ToolCall
 from openjarvis.tools.approval_store import (
@@ -342,8 +343,8 @@ class ProactiveAgent(ToolUsingAgent):
         return self._approval_store
 
     def _build_system_prompt(self) -> str:
-        user_md = _load_md_file(Path.home() / ".openjarvis" / "USER.md")
-        memory_md = _load_md_file(Path.home() / ".openjarvis" / "MEMORY.md")
+        user_md = _load_md_file(get_config_dir() / "USER.md")
+        memory_md = _load_md_file(get_config_dir() / "MEMORY.md")
         now = datetime.now()
         context_block = ""
         if user_md or memory_md:

src/openjarvis/agents/prompt_loader.py

@@ -15,16 +15,17 @@ from __future__ import annotations
 
 import json
 import logging
-import os
 from pathlib import Path
 from typing import Any
 
+from openjarvis.core.paths import get_config_dir
+
 logger = logging.getLogger(__name__)
 
 
 def _openjarvis_home() -> Path:
-    """Resolve $OPENJARVIS_HOME, defaulting to ~/.openjarvis."""
-    return Path(os.environ.get("OPENJARVIS_HOME", "~/.openjarvis")).expanduser()
+    """Resolve the OpenJarvis root, honoring OPENJARVIS_HOME / XDG_DATA_HOME."""
+    return get_config_dir()
 
 
 def load_system_prompt_override(agent_name: str) -> str | None:

src/openjarvis/channels/imessage_daemon.py

@@ -18,11 +18,13 @@ import time
 from pathlib import Path
 from typing import Any, Dict, List
 
+from openjarvis.core.paths import get_config_dir
+
 logger = logging.getLogger(__name__)
 
 _DEFAULT_DB_PATH = str(Path.home() / "Library" / "Messages" / "chat.db")
 _POLL_INTERVAL = 5
-_PID_FILE = str(Path.home() / ".openjarvis" / "imessage-agent.pid")
+_PID_FILE = str(get_config_dir() / "imessage-agent.pid")
 
 
 def poll_new_messages(

src/openjarvis/channels/slack_daemon.py

@@ -14,9 +14,11 @@ import sys
 from pathlib import Path
 from typing import Any
 
+from openjarvis.core.paths import get_config_dir
+
 logger = logging.getLogger(__name__)
 
-_PID_FILE = str(Path.home() / ".openjarvis" / "slack-daemon.pid")
+_PID_FILE = str(get_config_dir() / "slack-daemon.pid")
 
 
 def _to_slack_fmt(text: str) -> str:

src/openjarvis/channels/whatsapp_baileys.py

@@ -22,6 +22,7 @@ from openjarvis.channels._stubs import (
     ChannelStatus,
 )
 from openjarvis.core.events import EventBus, EventType
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.registry import ChannelRegistry
 
 logger = logging.getLogger(__name__)
@@ -38,7 +39,7 @@ if not _BRIDGE_SRC.exists():
     )
 
 # Default runtime directory (npm install + auth state).
-_DEFAULT_RUNTIME_DIR = Path.home() / ".openjarvis" / "whatsapp_baileys_bridge"
+_DEFAULT_RUNTIME_DIR = get_config_dir() / "whatsapp_baileys_bridge"
 
 
 @ChannelRegistry.register("whatsapp_baileys")

src/openjarvis/cli/_screen.py

@@ -0,0 +1,79 @@
+"""Screen capture for vision input (``jarvis ask --screen``).
+
+Captures the primary monitor to a temporary PNG so it can be handed to a
+vision-capable model. On Windows this uses the built-in .NET
+``System.Drawing`` stack (no third-party dependency). Other platforms fall
+back to ``mss`` or ``Pillow`` if installed.
+"""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import sys
+import tempfile
+
+# PowerShell: capture the PRIMARY monitor (more legible for a vision model
+# than a downscaled multi-monitor grab). {path} is filled in with forward
+# slashes, which .NET accepts on Windows and which avoids backslash escaping.
+_PS_CAPTURE = """
+Add-Type -AssemblyName System.Windows.Forms, System.Drawing
+$b = [System.Windows.Forms.Screen]::PrimaryScreen.Bounds
+$bmp = New-Object System.Drawing.Bitmap($b.Width, $b.Height)
+$g = [System.Drawing.Graphics]::FromImage($bmp)
+$g.CopyFromScreen($b.X, $b.Y, 0, 0, $bmp.Size)
+$bmp.Save("{path}", [System.Drawing.Imaging.ImageFormat]::Png)
+$g.Dispose(); $bmp.Dispose()
+"""
+
+
+def capture_screen_to_temp() -> str:
+    """Capture the screen to a temp PNG and return its absolute path.
+
+    Raises ``RuntimeError`` with actionable guidance if capture fails or the
+    platform has no available backend.
+    """
+    fd, path = tempfile.mkstemp(prefix="jarvis_screen_", suffix=".png")
+    os.close(fd)
+
+    if sys.platform.startswith("win"):
+        script = _PS_CAPTURE.replace("{path}", path.replace("\\", "/"))
+        proc = subprocess.run(
+            ["powershell", "-NoProfile", "-NonInteractive", "-Command", script],
+            capture_output=True,
+            text=True,
+            timeout=30,
+        )
+        if (
+            proc.returncode != 0
+            or not os.path.exists(path)
+            or not os.path.getsize(path)
+        ):
+            raise RuntimeError(
+                "screen capture failed: "
+                + (proc.stderr.strip() or "empty image written")
+            )
+        return path
+
+    # Non-Windows: optional backends.
+    try:
+        import mss  # type: ignore
+
+        with mss.mss() as sct:
+            sct.shot(mon=-1, output=path)
+        return path
+    except ImportError:
+        pass
+    try:
+        from PIL import ImageGrab  # type: ignore
+
+        ImageGrab.grab().save(path)
+        return path
+    except Exception as exc:  # noqa: BLE001
+        raise RuntimeError(
+            "screen capture on this platform needs 'mss' or 'Pillow' "
+            "(try: pip install mss)"
+        ) from exc
+
+
+__all__ = ["capture_screen_to_temp"]

src/openjarvis/cli/_version_check.py

@@ -9,9 +9,11 @@ import sys
 import time
 from pathlib import Path
 
+from openjarvis.core.paths import get_config_dir
+
 logger = logging.getLogger(__name__)
 
-_CACHE_PATH = Path("~/.openjarvis/version-check.json").expanduser()
+_CACHE_PATH = get_config_dir() / "version-check.json"
 _CACHE_TTL = 86400  # 24 hours
 _PYPI_API = "https://pypi.org/pypi/openjarvis/json"
 
@@ -21,7 +23,7 @@ def _config_path() -> Path:
     override = os.environ.get("OPENJARVIS_CONFIG")
     if override:
         return Path(override).expanduser()
-    return Path("~/.openjarvis/config.toml").expanduser()
+    return get_config_dir() / "config.toml"
 
 
 # Commands that surface the "new version available" nudge. We deliberately

src/openjarvis/cli/agent_cmd.py

@@ -12,15 +12,12 @@ from rich.table import Table
 
 def _get_manager():
     """Get or create the AgentManager singleton."""
-    from pathlib import Path
-
     from openjarvis.agents.manager import AgentManager
     from openjarvis.core.config import load_config
+    from openjarvis.core.paths import get_config_dir
 
     config = load_config()
-    db_path = config.agent_manager.db_path or str(
-        Path("~/.openjarvis/agents.db").expanduser()
-    )
+    db_path = config.agent_manager.db_path or str(get_config_dir() / "agents.db")
     return AgentManager(db_path=db_path)
 
 
@@ -273,6 +270,7 @@ def search(agent_id: str, query: str, limit: int) -> None:
     console = Console(stderr=True)
     try:
         from openjarvis.core.config import load_config
+        from openjarvis.core.paths import get_config_dir
         from openjarvis.traces.store import TraceStore
 
         config = load_config()
@@ -281,7 +279,7 @@ def search(agent_id: str, query: str, limit: int) -> None:
         if not agent:
             console.print(f"[red]Agent not found: {agent_id}[/red]")
             return
-        store = TraceStore(config.traces.db_path or "~/.openjarvis/traces.db")
+        store = TraceStore(config.traces.db_path or str(get_config_dir() / "traces.db"))
         results = store.search(query, agent=agent["name"], limit=limit)
         if not results:
             console.print("[dim]No results.[/dim]")
@@ -545,8 +543,7 @@ def run_agent(agent_id):
     updated = manager.get_agent(agent_id)
     runs = updated.get("total_runs", 0)
     console.print(
-        f"[green]✓[/green] Tick complete. "
-        f"Status: {updated['status']}, runs: {runs}"
+        f"[green]✓[/green] Tick complete. Status: {updated['status']}, runs: {runs}"
     )
 
     # Print the agent's actual output. summary_memory holds the latest tick's
@@ -662,6 +659,7 @@ def trace(agent_id, run_number, limit):
     import datetime
 
     from openjarvis.core.config import load_config
+    from openjarvis.core.paths import get_config_dir
     from openjarvis.traces.store import TraceStore
 
     manager = _get_manager()
@@ -671,7 +669,7 @@ def trace(agent_id, run_number, limit):
         raise SystemExit(1)
 
     config = load_config()
-    store = TraceStore(config.traces.db_path or "~/.openjarvis/traces.db")
+    store = TraceStore(config.traces.db_path or str(get_config_dir() / "traces.db"))
     traces = store.list_traces(agent=agent_id, limit=limit)
 
     if not traces:
@@ -841,8 +839,8 @@ def ask(agent_id, message, auto_approve):
     if auto_approve:
         executor._confirm_callback = lambda _prompt: True
     else:
-        executor._confirm_callback = (
-            lambda prompt: click.confirm(f"\n{prompt}", default=False)
+        executor._confirm_callback = lambda prompt: click.confirm(
+            f"\n{prompt}", default=False
         )
     # Run the tick with a live trace rather than blocking in silence — the
     # message we just queued is consumed as this tick's input, so the user

src/openjarvis/cli/ask.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import base64
 import json as json_mod
 import logging
 import sys
@@ -619,6 +620,21 @@ def _print_profile(
         "(default: ~/.openjarvis/knowledge.db)."
     ),
 )
+@click.option(
+    "-i",
+    "--image",
+    "image_paths",
+    multiple=True,
+    type=click.Path(exists=True, dir_okay=False),
+    help="Image file for a vision model (e.g. gemma3). Repeatable.",
+)
+@click.option(
+    "-S",
+    "--screen",
+    "capture_screen",
+    is_flag=True,
+    help="Capture the current screen and send it to the vision model.",
+)
 @click.option(
     "--persona",
     "persona_name",
@@ -645,6 +661,8 @@ def ask(
     research_mode: bool,
     knowledge_db: str | None,
     persona_name: str | None,
+    image_paths: tuple[str, ...] = (),
+    capture_screen: bool = False,
 ) -> None:
     """Ask Jarvis a question."""
     quiet = (ctx.obj or {}).get("quiet", False) or output_json
@@ -652,6 +670,27 @@ def ask(
     console = Console(stderr=True)
     query_text = " ".join(query)
 
+    # Vision: collect base64 images from --image files and/or --screen.
+    image_b64: list[str] = []
+    for _img_path in image_paths:
+        try:
+            with open(_img_path, "rb") as _fh:
+                image_b64.append(base64.b64encode(_fh.read()).decode("ascii"))
+        except OSError as exc:
+            console.print(f"[red]Could not read image {_img_path}: {exc}[/red]")
+            sys.exit(1)
+    if capture_screen:
+        try:
+            from openjarvis.cli._screen import capture_screen_to_temp
+
+            _shot = capture_screen_to_temp()
+            with open(_shot, "rb") as _fh:
+                image_b64.append(base64.b64encode(_fh.read()).decode("ascii"))
+            logger.debug("Captured screen to %s", _shot)
+        except Exception as exc:  # noqa: BLE001
+            console.print(f"[red]Screen capture failed:[/red] {exc}")
+            sys.exit(1)
+
     wall_start = time.monotonic() if enable_profile else None
 
     # Load config
@@ -671,11 +710,26 @@ def ask(
     # Without this fallback, `[agent].default_system_prompt` and the
     # SOUL.md / MEMORY.md / USER.md persona system are silently bypassed for
     # the most common command (`jarvis ask "..."`).
+    agent_explicitly_set = agent_name is not None
     if agent_name is None:
         configured_default = (config.agent.default_agent or "").strip()
         if configured_default:
             agent_name = configured_default
 
+    # Vision flows only through direct-to-engine mode. If an image/screenshot
+    # was supplied without an explicit --agent, route to direct mode so the
+    # picture reaches the model; if an agent was explicitly requested, say
+    # plainly that the image is being skipped rather than dropping it silently.
+    if image_b64:
+        if not agent_explicitly_set:
+            agent_name = ""
+        else:
+            console.print(
+                "[yellow]Note:[/yellow] --image/--screen only works in direct "
+                "mode; the image is ignored with --agent set. Re-run with "
+                '`--agent ""` to use vision.'
+            )
+
     # Track whether the user explicitly set --max-tokens
     user_set_max_tokens = max_tokens is not None
 
@@ -871,6 +925,27 @@ def ask(
         return
 
     # Direct-to-engine mode (no agent)
+    # Privacy guard: a screenshot/image is sensitive, and OpenJarvis is
+    # local-first. If the active engine isn't local, warn before the image
+    # leaves the machine rather than silently uploading it to a third party.
+    _LOCAL_ENGINES = {
+        "ollama",
+        "llamacpp",
+        "vllm",
+        "sglang",
+        "exo",
+        "nexa",
+        "uzu",
+        "apple_fm",
+        "gemma_cpp",
+    }
+    if image_b64 and engine_name not in _LOCAL_ENGINES:
+        console.print(
+            f"[yellow]Privacy warning:[/yellow] sending {len(image_b64)} "
+            f"image(s) to a non-local engine ('{engine_name}'). The image will "
+            "leave this machine. Use a local engine (e.g. ollama) to keep "
+            "vision on-device."
+        )
     messages = [Message(role=Role.USER, content=query_text)]
 
     # Memory-augmented context injection
@@ -897,6 +972,15 @@ def ask(
         except Exception as exc:
             logger.debug("Failed to inject memory context: %s", exc)
 
+    # Vision: attach images to the final user message *after* any context
+    # injection (which may rebuild the list). messages_to_dicts() forwards
+    # the "images" field to Ollama's /api/chat.
+    if image_b64:
+        for _m in reversed(messages):
+            if _m.role == Role.USER:
+                _m.images = image_b64
+                break
+
     # Generate (InstrumentedEngine handles telemetry + energy recording)
     try:
         with console.status("[bold green]Generating...[/bold green]"):

src/openjarvis/cli/config_cmd.py

@@ -276,6 +276,36 @@ def hardware() -> None:
 config.add_command(show_group, "show")
 
 
+@config.command("path")
+def show_path() -> None:
+    """Print the resolved OpenJarvis directories (home, config, cache).
+
+    All OpenJarvis state lives under a single root, resolved in priority
+    order: ``$OPENJARVIS_HOME`` > ``$XDG_DATA_HOME/openjarvis`` >
+    ``~/.openjarvis``. Use this to confirm where your data is stored after
+    setting an override.
+    """
+    from openjarvis.core.paths import get_cache_dir, get_config_dir, get_config_path
+
+    console = Console(stderr=True)
+    home = get_config_dir()
+    override = (
+        "OPENJARVIS_HOME"
+        if os.environ.get("OPENJARVIS_HOME")
+        else "XDG_DATA_HOME"
+        if os.environ.get("XDG_DATA_HOME")
+        else "default (~/.openjarvis)"
+    )
+    table = Table(show_header=True, header_style="bold")
+    table.add_column("Directory")
+    table.add_column("Path", style="cyan")
+    table.add_row("Home (root)", str(home))
+    table.add_row("Config file", str(get_config_path()))
+    table.add_row("Cache", str(get_cache_dir()))
+    console.print(table)
+    console.print(f"[dim]Resolved via: {override}[/dim]")
+
+
 def _probe_engine_host(url: str, console: Console) -> None:
     """Probe an engine host URL and print reachability status."""
     try:

src/openjarvis/cli/doctor_cmd.py

@@ -223,6 +223,47 @@ def _check_optional_deps() -> List[CheckResult]:
     return results
 
 
+def _check_speech_backend() -> CheckResult:
+    """Check whether the configured speech backend can load."""
+    try:
+        from openjarvis.speech._discovery import get_speech_backend
+
+        config = _get_config()
+        backend = get_speech_backend(config)
+        if backend is None:
+            return CheckResult(
+                "Speech backend",
+                "warn",
+                "Not configured",
+                details="Install desktop dependencies with `uv sync --extra desktop`.",
+            )
+
+        if backend.health():
+            return CheckResult(
+                "Speech backend",
+                "ok",
+                f"{backend.backend_id} ready",
+            )
+
+        details = None
+        last_error = getattr(backend, "last_error", None)
+        if callable(last_error):
+            details = last_error()
+        return CheckResult(
+            "Speech backend",
+            "warn",
+            f"{backend.backend_id} unavailable",
+            details=details
+            or "Install desktop dependencies with `uv sync --extra desktop`.",
+        )
+    except Exception as exc:
+        return CheckResult(
+            "Speech backend",
+            "warn",
+            f"Could not check: {exc}",
+        )
+
+
 def _check_security_profile() -> CheckResult:
     """Check if a security profile is configured."""
     try:
@@ -306,6 +347,7 @@ def _run_all_checks() -> List[CheckResult]:
     checks.extend(_check_models())
     checks.append(_check_default_model())
     checks.extend(_check_optional_deps())
+    checks.append(_check_speech_backend())
     checks.append(_check_nodejs())
     checks.append(_check_security_profile())
     return checks
@@ -354,7 +396,9 @@ def doctor(as_json: bool) -> None:
 
     # Background tasks section
     from openjarvis.cli._bg_state import get_status
+    from openjarvis.core.paths import get_config_dir
 
+    scripts_dir = get_config_dir() / ".scripts"
     console.print("[bold]Background tasks[/bold]")
     bg = get_status()
     bg_failed = False
@@ -364,8 +408,8 @@ def doctor(as_json: bool) -> None:
     elif bg.rust_extension == "failed":
         console.print(f"  [red]✗[/red] Rust extension: failed — {bg.rust_error[:80]}")
         console.print(
-            "    retry: ~/.openjarvis/.scripts/install-rust.sh && "
-            "~/.openjarvis/.scripts/build-extension.sh"
+            f"    retry: {scripts_dir}/install-rust.sh && "
+            f"{scripts_dir}/build-extension.sh"
         )
         bg_failed = True
     else:
@@ -380,7 +424,7 @@ def doctor(as_json: bool) -> None:
             console.print(f"  [green]✓[/green] {model_id}: ready")
         elif state == "failed":
             console.print(f"  [red]✗[/red] {model_id}: failed")
-            console.print(f"    retry: ~/.openjarvis/.scripts/pull-model.sh {model_id}")
+            console.print(f"    retry: {scripts_dir}/pull-model.sh {model_id}")
             bg_failed = True
         else:
             console.print(f"  [yellow]…[/yellow] {model_id}: downloading")

src/openjarvis/cli/log_config.py

@@ -7,6 +7,7 @@ from logging.handlers import RotatingFileHandler
 from pathlib import Path
 from typing import Optional, Union
 
+from openjarvis.core.paths import get_config_dir
 from openjarvis.security.credential_stripper import CredentialStripper
 
 _stripper = CredentialStripper()
@@ -68,7 +69,7 @@ def setup_logging(
         if log_file is None:
             from openjarvis.security.file_utils import secure_mkdir
 
-            log_dir = Path.home() / ".openjarvis"
+            log_dir = get_config_dir()
             secure_mkdir(log_dir)
             log_file = log_dir / "cli.log"
         file_handler = RotatingFileHandler(

src/openjarvis/cli/scan_cmd.py

@@ -11,6 +11,8 @@ from typing import Callable, List
 
 import click
 
+from openjarvis.core.paths import get_config_dir
+
 # Engine ports that should only be listening on localhost.
 _ENGINE_PORTS = {11434, 8080, 8000, 30000, 1234, 52415, 18181}
 
@@ -132,7 +134,7 @@ class PrivacyScanner:
     def check_icloud_sync(self) -> ScanResult:
         """Check whether ~/.openjarvis is inside iCloud Drive sync scope."""
         try:
-            config_path = Path("~/.openjarvis").expanduser().resolve()
+            config_path = get_config_dir().resolve()
             icloud_path = Path("~/Library/Mobile Documents/").expanduser().resolve()
             if str(config_path).startswith(str(icloud_path)):
                 return ScanResult(

src/openjarvis/cli/serve.py

@@ -11,6 +11,7 @@ from rich.console import Console
 from openjarvis.cli._banner import print_banner
 from openjarvis.core.config import load_config
 from openjarvis.core.events import EventBus
+from openjarvis.core.paths import get_config_dir
 from openjarvis.engine import (
     discover_engines,
     discover_models,
@@ -496,13 +497,9 @@ def serve(
     agent_manager = None
     if config.agent_manager.enabled:
         try:
-            from pathlib import Path
-
             from openjarvis.agents.manager import AgentManager
 
-            am_db = config.agent_manager.db_path or str(
-                Path("~/.openjarvis/agents.db").expanduser()
-            )
+            am_db = config.agent_manager.db_path or str(get_config_dir() / "agents.db")
             # The server owns the scheduler and is the authoritative tick
             # runner — on boot it holds no locks, so it (and only it) sweeps
             # any zombie running→idle left by a previous crash.
@@ -607,9 +604,7 @@ def serve(
         try:
             import tomllib
 
-            _cfg_path = str(
-                __import__("pathlib").Path.home() / ".openjarvis" / "config.toml"
-            )
+            _cfg_path = str(get_config_dir() / "config.toml")
             with open(_cfg_path, "rb") as _f:
                 _raw = tomllib.load(_f)
             api_key = _raw.get("server", {}).get("auth", {}).get("api_key", "")

src/openjarvis/cli/skill_cmd.py

@@ -11,6 +11,7 @@ from rich.table import Table
 
 from openjarvis.core.config import load_config
 from openjarvis.core.events import EventBus
+from openjarvis.core.paths import get_config_dir
 from openjarvis.skills.manager import SkillManager
 
 
@@ -28,12 +29,12 @@ def _get_trace_store():
 
 def _get_discovered_dir() -> Path:
     """Return the directory where discovered skill manifests are written."""
-    return Path("~/.openjarvis/skills/discovered/").expanduser()
+    return get_config_dir() / "skills" / "discovered"
 
 
 def _get_overlay_dir() -> Path:
     """Return the directory where optimization overlays are stored."""
-    return Path("~/.openjarvis/learning/skills/").expanduser()
+    return get_config_dir() / "learning" / "skills"
 
 
 def _get_skill_paths() -> List[Path]:
@@ -41,7 +42,7 @@ def _get_skill_paths() -> List[Path]:
     workspace = Path("./skills")
     if workspace.exists():
         paths.append(workspace)
-    user_dir = Path("~/.openjarvis/skills/").expanduser()
+    user_dir = get_config_dir() / "skills"
     paths.append(user_dir)
     return paths
 
@@ -174,8 +175,10 @@ def _get_resolver(source: str, url: str = ""):
         from openjarvis.skills.sources.github import GitHubResolver
 
         cache = _Path(
-            "~/.openjarvis/skill-cache/github/" + url.rstrip("/").rsplit("/", 1)[-1]
-        ).expanduser()
+            str(get_config_dir() / "skill-cache" / "github")
+            + "/"
+            + url.rstrip("/").rsplit("/", 1)[-1]
+        )
         return GitHubResolver(cache_root=cache, repo_url=url)
     raise click.BadParameter(f"Unknown source: {source!r}")
 

src/openjarvis/connectors/gcalendar.py

@@ -20,7 +20,6 @@ from openjarvis.connectors.oauth import (
     delete_tokens,
     load_tokens,
     resolve_google_credentials,
-    run_oauth_flow,
     save_tokens,
 )
 from openjarvis.core.config import DEFAULT_CONFIG_DIR
@@ -290,12 +289,18 @@ class GCalendarConnector(BaseConnector):
         """Handle the OAuth callback.
 
         If *code* looks like a ``client_id:client_secret`` pair (containing
-        ``.apps.googleusercontent.com``), store the credentials and trigger
-        the full browser-based OAuth flow.  Otherwise treat it as a raw
-        token / auth code.
+        ``.apps.googleusercontent.com``), persist the client credentials only.
+        The browser consent + code→token exchange is owned by the in-process
+        server flow (``/v1/connectors/{id}/oauth/start`` → ``/oauth/callback``),
+        which writes the real ``access_token`` to every Google credential file.
+
+        The previous daemon-thread browser flow (its own ``localhost:8789``
+        callback server) failed silently in the bundled desktop context and is
+        intentionally removed here (issue #512).
+
+        Any other *code* is treated as a raw token / auth code.
         """
         code = code.strip()
-        # If user pastes client_id:client_secret, store and run OAuth flow
         if ":" in code and ".apps.googleusercontent.com" in code:
             client_id, client_secret = code.split(":", 1)
             save_tokens(
@@ -305,20 +310,6 @@ class GCalendarConnector(BaseConnector):
                     "client_secret": client_secret.strip(),
                 },
             )
-            import threading
-
-            def _run() -> None:
-                try:
-                    run_oauth_flow(
-                        client_id=client_id.strip(),
-                        client_secret=client_secret.strip(),
-                        scopes=GOOGLE_ALL_SCOPES,
-                        credentials_path=self._credentials_path,
-                    )
-                except Exception:  # noqa: BLE001
-                    pass
-
-            threading.Thread(target=_run, daemon=True).start()
         else:
             # Raw token or auth code
             save_tokens(self._credentials_path, {"token": code})

src/openjarvis/connectors/gcontacts.py

@@ -20,7 +20,6 @@ from openjarvis.connectors.oauth import (
     delete_tokens,
     load_tokens,
     resolve_google_credentials,
-    run_oauth_flow,
     save_tokens,
 )
 from openjarvis.core.config import DEFAULT_CONFIG_DIR
@@ -195,12 +194,18 @@ class GContactsConnector(BaseConnector):
         """Handle the OAuth callback.
 
         If *code* looks like a ``client_id:client_secret`` pair (containing
-        ``.apps.googleusercontent.com``), store the credentials and trigger
-        the full browser-based OAuth flow.  Otherwise treat it as a raw
-        token / auth code.
+        ``.apps.googleusercontent.com``), persist the client credentials only.
+        The browser consent + code→token exchange is owned by the in-process
+        server flow (``/v1/connectors/{id}/oauth/start`` → ``/oauth/callback``),
+        which writes the real ``access_token`` to every Google credential file.
+
+        The previous daemon-thread browser flow (its own ``localhost:8789``
+        callback server) failed silently in the bundled desktop context and is
+        intentionally removed here (issue #512).
+
+        Any other *code* is treated as a raw token / auth code.
         """
         code = code.strip()
-        # If user pastes client_id:client_secret, store and run OAuth flow
         if ":" in code and ".apps.googleusercontent.com" in code:
             client_id, client_secret = code.split(":", 1)
             save_tokens(
@@ -210,20 +215,6 @@ class GContactsConnector(BaseConnector):
                     "client_secret": client_secret.strip(),
                 },
             )
-            import threading
-
-            def _run() -> None:
-                try:
-                    run_oauth_flow(
-                        client_id=client_id.strip(),
-                        client_secret=client_secret.strip(),
-                        scopes=GOOGLE_ALL_SCOPES,
-                        credentials_path=self._credentials_path,
-                    )
-                except Exception:  # noqa: BLE001
-                    pass
-
-            threading.Thread(target=_run, daemon=True).start()
         else:
             # Raw token or auth code
             save_tokens(self._credentials_path, {"token": code})

src/openjarvis/connectors/gdrive.py

@@ -20,7 +20,6 @@ from openjarvis.connectors.oauth import (
     delete_tokens,
     load_tokens,
     resolve_google_credentials,
-    run_oauth_flow,
     save_tokens,
 )
 from openjarvis.core.config import DEFAULT_CONFIG_DIR
@@ -178,15 +177,25 @@ class GDriveConnector(BaseConnector):
         """Handle the OAuth callback.
 
         If *code* looks like a ``client_id:client_secret`` pair (containing
-        ``.apps.googleusercontent.com``), store the credentials and trigger
-        the full browser-based OAuth flow.  Otherwise treat it as a raw
-        token / auth code.
+        ``.apps.googleusercontent.com``), persist the client credentials only.
+        The actual browser consent + code→token exchange is owned by the
+        in-process server flow (``/v1/connectors/{id}/oauth/start`` →
+        ``/oauth/callback``), which writes the real ``access_token`` to every
+        Google credential file.
+
+        Previously this spawned a daemon thread that popped a browser and ran
+        its own ``localhost:8789`` callback server; that thread failed silently
+        in the bundled desktop context, so the connector never gained an access
+        token and never appeared in Data Sources (issue #512). The background
+        flow is intentionally removed here.
+
+        Any other *code* is treated as a raw token / auth code.
         """
         code = code.strip()
-        # If user pastes client_id:client_secret, store and run OAuth flow
+        # A pasted client_id:client_secret pair is the app registration, not a
+        # completed credential — persist it and let the server flow finish auth.
         if ":" in code and ".apps.googleusercontent.com" in code:
             client_id, client_secret = code.split(":", 1)
-            # Save credentials immediately
             save_tokens(
                 self._credentials_path,
                 {
@@ -194,21 +203,6 @@ class GDriveConnector(BaseConnector):
                     "client_secret": client_secret.strip(),
                 },
             )
-            # Run OAuth flow in background thread to avoid blocking
-            import threading
-
-            def _run() -> None:
-                try:
-                    run_oauth_flow(
-                        client_id=client_id.strip(),
-                        client_secret=client_secret.strip(),
-                        scopes=GOOGLE_ALL_SCOPES,
-                        credentials_path=self._credentials_path,
-                    )
-                except Exception:  # noqa: BLE001
-                    pass
-
-            threading.Thread(target=_run, daemon=True).start()
         else:
             # Raw token or auth code
             save_tokens(self._credentials_path, {"token": code})

src/openjarvis/core/config.py

@@ -16,6 +16,14 @@ from dataclasses import dataclass, field
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Dict, List, Optional
 
+from openjarvis.core.paths import (
+    ConfigurationError,
+    get_cache_dir,
+    get_config_dir,
+    get_config_path,
+    get_data_dir,
+)
+
 if TYPE_CHECKING:
     # Only used by type-checkers (mypy/pyright) for the ``JarvisConfig.mining``
     # field annotation. The runtime import is deferred inside
@@ -33,15 +41,24 @@ except ModuleNotFoundError:
 # Hardware dataclasses
 # ---------------------------------------------------------------------------
 
-DEFAULT_CONFIG_DIR = Path.home() / ".openjarvis"
-DEFAULT_CONFIG_PATH = DEFAULT_CONFIG_DIR / "config.toml"
+# Legacy names, kept for the ~45 modules that import them. They are resolved
+# once at import via the env-aware resolver in ``openjarvis.core.paths`` (the
+# install-script model: ``OPENJARVIS_HOME`` / ``XDG_DATA_HOME`` are set before
+# the process starts). They are real module attributes — not computed lazily —
+# so existing tests can ``monkeypatch.setattr`` them and so dataclass-instance
+# defaults stay consistent. Code that must react to a mid-process env change
+# (or wants the override regardless of import order) should call
+# ``get_config_dir()`` / ``get_config_path()`` directly; the dataclass field
+# defaults below already do this via ``default_factory``.
+DEFAULT_CONFIG_DIR = get_config_dir()
+DEFAULT_CONFIG_PATH = get_config_path()
 
 
 def _ensure_config_dir() -> Path:
     """Ensure the config directory exists with restrictive permissions."""
     from openjarvis.security.file_utils import secure_mkdir
 
-    return secure_mkdir(DEFAULT_CONFIG_DIR)
+    return secure_mkdir(get_config_dir())
 
 
 @dataclass(slots=True)
@@ -742,7 +759,9 @@ class SkillsLearningConfig:
     optimizer: str = "dspy"  # "dspy" or "gepa"
     min_traces_per_skill: int = 20
     optimization_interval_seconds: int = 86400
-    overlay_dir: str = "~/.openjarvis/learning/skills/"
+    overlay_dir: str = field(
+        default_factory=lambda: str(get_config_dir() / "learning" / "skills")
+    )
 
 
 @dataclass(slots=True)
@@ -894,7 +913,7 @@ class StorageConfig:
     """Storage (memory) backend settings."""
 
     default_backend: str = "sqlite"
-    db_path: str = str(DEFAULT_CONFIG_DIR / "memory.db")
+    db_path: str = field(default_factory=lambda: str(get_config_dir() / "memory.db"))
     context_top_k: int = 5
     context_min_score: float = 0.0
     context_max_tokens: int = 2048
@@ -946,8 +965,10 @@ class AgentConfig:
     system_prompt_path: str = ""  # path to system prompt file (.txt, .md)
     context_from_memory: bool = True  # inject relevant memory context into prompts
     default_system_prompt: str = (
-        "You are a helpful AI assistant running locally on the user's own "
-        "hardware through OpenJarvis. You are not a cloud service. Respond "
+        "You are OpenJarvis, a helpful AI assistant running locally on the "
+        "user's own hardware. You are not a cloud service, and you are not "
+        "Claude, ChatGPT, Gemini, or any other branded assistant. If asked "
+        "who or what you are, identify yourself as OpenJarvis. Respond "
         "helpfully, concisely, and accurately."
     )
 
@@ -996,7 +1017,7 @@ class TelemetryConfig:
     """Telemetry persistence settings."""
 
     enabled: bool = True
-    db_path: str = str(DEFAULT_CONFIG_DIR / "telemetry.db")
+    db_path: str = field(default_factory=lambda: str(get_config_dir() / "telemetry.db"))
     gpu_metrics: bool = False
     gpu_poll_interval_ms: int = 50
     energy_vendor: str = ""  # auto-detect or force "nvidia"/"amd"/"apple"/"cpu_rapl"
@@ -1021,7 +1042,7 @@ class AnalyticsConfig:
     enabled: bool = True
     host: str = "https://34.231.106.201.sslip.io"
     key: str = "phc_ysKu72QaxzYNmDpHFcesD2ZZAe68zkdWJEKoYYkc5e3n"
-    anon_id_path: str = str(DEFAULT_CONFIG_DIR / "anon_id")
+    anon_id_path: str = field(default_factory=lambda: str(get_config_dir() / "anon_id"))
     flush_interval_seconds: int = 30
     flush_at_size: int = 100
 
@@ -1031,7 +1052,7 @@ class TracesConfig:
     """Trace system settings."""
 
     enabled: bool = True
-    db_path: str = str(DEFAULT_CONFIG_DIR / "traces.db")
+    db_path: str = field(default_factory=lambda: str(get_config_dir() / "traces.db"))
 
 
 @dataclass(slots=True)
@@ -1233,7 +1254,9 @@ class SecurityConfig:
     mode: str = "redact"  # "redact" | "warn" | "block"
     secret_scanner: bool = True
     pii_scanner: bool = True
-    audit_log_path: str = str(DEFAULT_CONFIG_DIR / "audit.db")
+    audit_log_path: str = field(
+        default_factory=lambda: str(get_config_dir() / "audit.db")
+    )
     enforce_tool_confirmation: bool = True
     merkle_audit: bool = True
     signing_key_path: str = ""
@@ -1244,7 +1267,9 @@ class SecurityConfig:
     local_engine_bypass: bool = False
     local_tool_bypass: bool = False
     profile: str = ""
-    vault_key_path: str = str(DEFAULT_CONFIG_DIR / ".vault_key")
+    vault_key_path: str = field(
+        default_factory=lambda: str(get_config_dir() / ".vault_key")
+    )
     capabilities: CapabilitiesConfig = field(default_factory=CapabilitiesConfig)
 
 
@@ -1365,7 +1390,7 @@ class SessionConfig:
     enabled: bool = False
     max_age_hours: float = 24.0
     consolidation_threshold: int = 100
-    db_path: str = str(DEFAULT_CONFIG_DIR / "sessions.db")
+    db_path: str = field(default_factory=lambda: str(get_config_dir() / "sessions.db"))
 
 
 @dataclass(slots=True)
@@ -1383,7 +1408,9 @@ class OperatorsConfig:
     """Operator lifecycle settings."""
 
     enabled: bool = False
-    manifests_dir: str = "~/.openjarvis/operators"
+    manifests_dir: str = field(
+        default_factory=lambda: str(get_config_dir() / "operators")
+    )
     auto_activate: str = ""  # Comma-separated operator IDs
 
 
@@ -1409,7 +1436,7 @@ class OptimizeConfig:
     benchmark: str = ""
     max_samples: int = 50
     judge_model: str = "gpt-5-mini-2025-08-07"
-    db_path: str = str(DEFAULT_CONFIG_DIR / "optimize.db")
+    db_path: str = field(default_factory=lambda: str(get_config_dir() / "optimize.db"))
 
 
 @dataclass(slots=True)
@@ -1417,18 +1444,20 @@ class AgentManagerConfig:
     """Persistent agent manager settings."""
 
     enabled: bool = True
-    db_path: str = str(DEFAULT_CONFIG_DIR / "agents.db")
+    db_path: str = field(default_factory=lambda: str(get_config_dir() / "agents.db"))
 
 
 @dataclass(slots=True)
 class MemoryFilesConfig:
     """Persistent memory-file paths and nudge settings."""
 
-    soul_path: str = "~/.openjarvis/SOUL.md"
-    memory_path: str = "~/.openjarvis/MEMORY.md"
-    user_path: str = "~/.openjarvis/USER.md"
+    soul_path: str = field(default_factory=lambda: str(get_config_dir() / "SOUL.md"))
+    memory_path: str = field(
+        default_factory=lambda: str(get_config_dir() / "MEMORY.md")
+    )
+    user_path: str = field(default_factory=lambda: str(get_config_dir() / "USER.md"))
     nudge_interval: int = 10
-    persona_name: str = ""  # named persona dir under ~/.openjarvis/personas/<name>/
+    persona_name: str = ""  # named persona dir under <config-dir>/personas/<name>/
 
 
 @dataclass(slots=True)
@@ -1467,13 +1496,15 @@ class SkillsConfig:
     """Configuration for agent-authored procedural skills."""
 
     enabled: bool = True
-    skills_dir: str = "~/.openjarvis/skills/"
+    skills_dir: str = field(default_factory=lambda: str(get_config_dir() / "skills"))
     active: str = "*"
     auto_discover: bool = True
     auto_sync: bool = False
     nudge_interval: int = 15
     index_repo: str = "https://github.com/openjarvis/skill-index.git"
-    index_dir: str = "~/.openjarvis/skill-index/"
+    index_dir: str = field(
+        default_factory=lambda: str(get_config_dir() / "skill-index")
+    )
     max_depth: int = 5
     sandbox_dangerous: bool = True
     sources: List[SkillSourceConfig] = field(default_factory=list)
@@ -1779,7 +1810,7 @@ def load_config(path: Optional[Path] = None) -> JarvisConfig:
     elif os.environ.get("OPENJARVIS_CONFIG"):
         config_path = Path(os.environ["OPENJARVIS_CONFIG"]).expanduser().resolve()
     else:
-        config_path = DEFAULT_CONFIG_PATH
+        config_path = get_config_path()
     if config_path.exists():
         with open(config_path, "rb") as fh:
             data = tomllib.load(fh)
@@ -2117,9 +2148,14 @@ __all__ = [
     "BrowserConfig",
     "CapabilitiesConfig",
     "ChannelConfig",
+    "ConfigurationError",
     "DEFAULT_CONFIG_DIR",
     "DEFAULT_CONFIG_PATH",
     "DiscordChannelConfig",
+    "get_cache_dir",
+    "get_config_dir",
+    "get_config_path",
+    "get_data_dir",
     "EmailChannelConfig",
     "EngineConfig",
     "FeishuChannelConfig",

src/openjarvis/core/credentials.py

@@ -10,13 +10,20 @@ import os
 import threading
 from pathlib import Path
 
+from openjarvis.core.paths import get_config_dir
+
 try:
     import tomllib
 except ModuleNotFoundError:
     import tomli as tomllib  # type: ignore[no-redef]
 
 _LOCK = threading.Lock()
-_DEFAULT_PATH = Path.home() / ".openjarvis" / "credentials.toml"
+
+
+def _default_path() -> Path:
+    """Resolve the credentials file under the OpenJarvis root (env-aware)."""
+    return get_config_dir() / "credentials.toml"
+
 
 TOOL_CREDENTIALS: dict[str, list[str]] = {
     "web_search": ["TAVILY_API_KEY"],
@@ -53,7 +60,7 @@ TOOL_CREDENTIALS: dict[str, list[str]] = {
 
 def load_credentials(path: Path | None = None) -> dict[str, dict[str, str]]:
     """Load credentials from TOML file."""
-    p = Path(path) if path else _DEFAULT_PATH
+    p = Path(path) if path else _default_path()
     if not p.exists():
         return {}
     with open(p, "rb") as f:
@@ -75,7 +82,7 @@ def save_credential(
     if not stripped:
         raise ValueError("Credential value must not be empty")
 
-    p = Path(path) if path else _DEFAULT_PATH
+    p = Path(path) if path else _default_path()
     with _LOCK:
         creds = load_credentials(path=p)
         if tool_name not in creds:

src/openjarvis/core/paths.py

@@ -0,0 +1,122 @@
+"""Central, env-aware resolution of OpenJarvis' home directory.
+
+OpenJarvis keeps all of its runtime state (config, databases, caches, logs,
+credentials, skills, recipes, …) under a single root so it never clutters the
+user's home directory beyond one directory. That root is resolved here, with
+the following precedence (highest first):
+
+1. ``$OPENJARVIS_HOME`` — explicit override (also honored by the shell
+   installer, see ``scripts/install/install.sh``).
+2. ``$XDG_DATA_HOME/openjarvis`` — when ``$XDG_DATA_HOME`` is set, follow the
+   XDG Base Directory spec by nesting a single ``openjarvis`` directory under
+   it. We deliberately use ONE directory rather than splitting across XDG
+   config/data/cache so the install tree stays self-contained and relocatable.
+3. ``~/.openjarvis`` — the historical default. With no env vars set, the
+   resolved path is exactly this, so existing installs are untouched.
+
+``config.py`` re-exports :func:`get_config_dir` results through the legacy
+``DEFAULT_CONFIG_DIR``/``DEFAULT_CONFIG_PATH`` names (computed dynamically) so
+the ~45 modules that import those names keep working while honoring the
+override. Modules that previously hardcoded ``Path.home() / ".openjarvis"``
+should call :func:`get_config_dir` (or :func:`get_data_dir` /
+:func:`get_cache_dir`) instead.
+
+Defense in depth: the resolved root must never live inside the OpenJarvis
+source tree (a misconfigured ``$OPENJARVIS_HOME`` pointing at the repo would
+otherwise scatter runtime artifacts into the working tree). This mirrors the
+guard in ``learning/spec_search/storage/paths.py`` and fails loudly per
+REVIEW.md's no-silent-failure discipline.
+"""
+
+from __future__ import annotations
+
+import os
+from pathlib import Path
+
+_DEFAULT_DIR_NAME = ".openjarvis"
+_XDG_SUBDIR_NAME = "openjarvis"
+
+
+class ConfigurationError(RuntimeError):
+    """Raised when the resolved home directory would violate isolation guarantees."""
+
+
+def _find_source_root() -> Path | None:
+    """Walk upward from this module to find the OpenJarvis source root.
+
+    Returns the directory containing the OpenJarvis ``pyproject.toml`` (the one
+    whose ``name = "openjarvis"``), or ``None`` when running from an installed
+    wheel rather than a source checkout.
+    """
+    here = Path(__file__).resolve()
+    for candidate in (here, *here.parents):
+        py = candidate / "pyproject.toml"
+        if py.exists():
+            try:
+                content = py.read_text(encoding="utf-8")
+            except OSError:
+                continue
+            if 'name = "openjarvis"' in content.lower():
+                return candidate
+    return None
+
+
+def _reject_source_tree(path: Path) -> Path:
+    """Raise if ``path`` resolves inside the OpenJarvis source tree."""
+    source_root = _find_source_root()
+    if source_root is not None:
+        try:
+            path.relative_to(source_root)
+        except ValueError:
+            pass  # Good — not inside the source tree.
+        else:
+            raise ConfigurationError(
+                f"OpenJarvis home ({path}) is inside the source tree "
+                f"({source_root}). OpenJarvis refuses to write runtime state "
+                "inside its own repo. Set OPENJARVIS_HOME (or XDG_DATA_HOME) "
+                "to a directory outside the repo (default: ~/.openjarvis)."
+            )
+    return path
+
+
+def get_config_dir() -> Path:
+    """Resolve OpenJarvis' single root directory, honoring env overrides.
+
+    Precedence: ``$OPENJARVIS_HOME`` > ``$XDG_DATA_HOME/openjarvis`` >
+    ``~/.openjarvis``. The result is always absolute and is rejected if it
+    falls inside the OpenJarvis source tree.
+    """
+    env_home = os.environ.get("OPENJARVIS_HOME")
+    if env_home:
+        resolved = Path(env_home).expanduser().resolve()
+        return _reject_source_tree(resolved)
+
+    xdg_data = os.environ.get("XDG_DATA_HOME")
+    if xdg_data:
+        resolved = (Path(xdg_data).expanduser() / _XDG_SUBDIR_NAME).resolve()
+        return _reject_source_tree(resolved)
+
+    return (Path.home() / _DEFAULT_DIR_NAME).resolve()
+
+
+def get_config_path() -> Path:
+    """Resolve the path to ``config.toml`` under the OpenJarvis root."""
+    return get_config_dir() / "config.toml"
+
+
+def get_data_dir() -> Path:
+    """Resolve the directory for persistent data (databases, blobs, …).
+
+    Consolidated under the single root; identical to :func:`get_config_dir`.
+    Provided as a distinct name so call sites read intentionally.
+    """
+    return get_config_dir()
+
+
+def get_cache_dir() -> Path:
+    """Resolve the directory for regenerable caches (eval datasets, etc.).
+
+    Lives at ``<root>/cache`` so caches stay inside the single OpenJarvis
+    directory instead of scattering across ``~/.cache``.
+    """
+    return get_config_dir() / "cache"

src/openjarvis/core/types.py

@@ -68,6 +68,10 @@ class Message:
     tool_calls: Optional[List[ToolCall]] = None
     tool_call_id: Optional[str] = None
     metadata: Dict[str, Any] = field(default_factory=dict)
+    # Base64-encoded image data for vision-capable models (e.g. gemma3,
+    # qwen2.5-vl). Forwarded to Ollama's /api/chat "images" field; None or
+    # empty for text-only messages (the common case).
+    images: Optional[List[str]] = None
 
 
 @dataclass(slots=True)

src/openjarvis/engine/_base.py

@@ -34,6 +34,10 @@ def messages_to_dicts(messages: Sequence[Message]) -> List[Dict[str, Any]]:
             ]
         if m.tool_call_id:
             d["tool_call_id"] = m.tool_call_id
+        # Vision: forward base64 images to the engine. Ollama's /api/chat
+        # accepts an "images" array on a message; text messages skip this.
+        if getattr(m, "images", None):
+            d["images"] = list(m.images)
         out.append(d)
     return out
 

src/openjarvis/engine/cloud.py

@@ -1,4 +1,7 @@
-"""Cloud inference engine — OpenAI, Anthropic, Google, and MiniMax API backends."""
+"""Cloud inference engine.
+
+OpenAI, Anthropic, Google, MiniMax, and DeepSeek API backends.
+"""
 
 from __future__ import annotations
 
@@ -48,6 +51,8 @@ PRICING: Dict[str, tuple[float, float]] = {
     "MiniMax-M2.7-highspeed": (0.60, 2.40),
     "MiniMax-M2.5": (0.30, 1.20),
     "MiniMax-M2.5-highspeed": (0.60, 2.40),
+    "deepseek-v4-flash": (0.27, 1.10),
+    "deepseek-v4-pro": (0.55, 2.19),
 }
 
 # Well-known model IDs per provider
@@ -83,6 +88,10 @@ _MINIMAX_MODELS = [
     "MiniMax-M2.5",
     "MiniMax-M2.5-highspeed",
 ]
+_DEEPSEEK_MODELS = [
+    "deepseek-v4-flash",
+    "deepseek-v4-pro",
+]
 
 # OpenRouter models — prefixed with "openrouter/" so they can be identified
 _OPENROUTER_POPULAR = [
@@ -111,6 +120,10 @@ def _is_minimax_model(model: str) -> bool:
     return model.lower().startswith("minimax")
 
 
+def _is_deepseek_model(model: str) -> bool:
+    return model.lower().startswith("deepseek")
+
+
 def _is_openrouter_model(model: str) -> bool:
     return model.startswith("openrouter/")
 
@@ -127,6 +140,35 @@ def _is_google_model(model: str) -> bool:
     return "gemini" in model.lower() and not _is_openrouter_model(model)
 
 
+# Positive prefix predicate for genuine OpenAI models. Kept in sync with
+# ``server/cloud_router.py:_OPENAI_PREFIXES`` so local-vs-cloud classification
+# agrees across the codebase. Used by ``_client_for_model``/``can_serve`` so the
+# cloud engine never claims it can serve an unrecognized (e.g. local Ollama)
+# model name just because an OpenAI key happens to be present (see #335).
+_OPENAI_PREFIXES = ("gpt-", "chatgpt-", "o1", "o3", "o4")
+
+
+def _is_openai_model(model: str) -> bool:
+    """True only for genuine OpenAI models (gpt-*, chatgpt-*, o1/o3/o4 series).
+
+    Defined positively so that an unrecognized model name (a local Ollama model
+    like ``qwen3.5:0.8b``, or a typo) is NOT treated as an OpenAI model. This is
+    the routing surface ``can_serve`` relies on; ``generate``/``stream`` keep
+    their OpenAI fall-through so an explicitly-requested unknown cloud model
+    still errors loudly at call time.
+
+    Caveat: a user may repoint the OpenAI client at an OpenAI-compatible server
+    (vLLM/LM Studio) via ``OPENAI_BASE_URL`` and legitimately serve non-gpt
+    names. That path is undocumented/untested in this engine; if it is added,
+    this predicate (or ``_client_for_model``) should treat a configured custom
+    base_url as "serves anything".
+    """
+    m = model.lower()
+    if m in (name.lower() for name in _OPENAI_MODELS):
+        return True
+    return m.startswith(_OPENAI_PREFIXES)
+
+
 def _is_openai_reasoning_model(model: str) -> bool:
     """Check if model is an OpenAI reasoning model that restricts temperature."""
     m = model.lower()
@@ -269,7 +311,7 @@ def _convert_tools_to_google(
 
 @EngineRegistry.register("cloud")
 class CloudEngine(InferenceEngine):
-    """Cloud inference via OpenAI, Anthropic, Google, and MiniMax SDKs."""
+    """Cloud inference via OpenAI, Anthropic, Google, MiniMax, and DeepSeek SDKs."""
 
     engine_id = "cloud"
     is_cloud = True
@@ -280,6 +322,7 @@ class CloudEngine(InferenceEngine):
         self._google_client: Any = None
         self._openrouter_client: Any = None
         self._minimax_client: Any = None
+        self._deepseek_client: Any = None
         self._codex_client: Any = None
         # Gemini thought_signatures: tool_call_id -> signature bytes
         self._thought_sigs: Dict[str, bytes] = {}
@@ -332,6 +375,17 @@ class CloudEngine(InferenceEngine):
                 )
             except ImportError:
                 pass
+        deepseek_key = os.environ.get("DEEPSEEK_API_KEY")
+        if deepseek_key:
+            try:
+                import openai
+
+                self._deepseek_client = openai.OpenAI(
+                    base_url="https://api.deepseek.com/v1",
+                    api_key=deepseek_key,
+                )
+            except ImportError:
+                pass
         # Codex — uses the OpenAI Responses API.
         # Supports both standard API keys (api.openai.com) and ChatGPT
         # OAuth tokens (chatgpt.com) via OPENAI_CODEX_BASE_URL override.
@@ -985,6 +1039,56 @@ class CloudEngine(InferenceEngine):
             ]
         return result
 
+    def _generate_deepseek(
+        self,
+        messages: Sequence[Message],
+        *,
+        model: str,
+        temperature: float,
+        max_tokens: int,
+        **kwargs: Any,
+    ) -> Dict[str, Any]:
+        if self._deepseek_client is None:
+            raise EngineConnectionError(
+                "DeepSeek client not available — set DEEPSEEK_API_KEY"
+            )
+        kwargs.pop("response_format", None)
+        create_kwargs: Dict[str, Any] = {
+            "model": model,
+            "messages": messages_to_dicts(messages),
+            "max_tokens": max_tokens,
+            "temperature": temperature,
+        }
+        t0 = time.monotonic()
+        resp = self._deepseek_client.chat.completions.create(**create_kwargs)
+        elapsed = time.monotonic() - t0
+        choice = resp.choices[0]
+        usage = resp.usage
+        prompt_tokens = usage.prompt_tokens if usage else 0
+        completion_tokens = usage.completion_tokens if usage else 0
+        result: Dict[str, Any] = {
+            "content": choice.message.content or "",
+            "usage": {
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "total_tokens": (usage.total_tokens if usage else 0),
+            },
+            "model": resp.model,
+            "finish_reason": choice.finish_reason or "stop",
+            "cost_usd": estimate_cost(model, prompt_tokens, completion_tokens),
+            "ttft": elapsed,
+        }
+        if hasattr(choice.message, "tool_calls") and choice.message.tool_calls:
+            result["tool_calls"] = [
+                {
+                    "id": tc.id,
+                    "name": tc.function.name,
+                    "arguments": tc.function.arguments,
+                }
+                for tc in choice.message.tool_calls
+            ]
+        return result
+
     def generate(
         self,
         messages: Sequence[Message],
@@ -1006,6 +1110,8 @@ class CloudEngine(InferenceEngine):
             return self._generate_openrouter(messages, **kw)
         if _is_minimax_model(model):
             return self._generate_minimax(messages, **kw)
+        if _is_deepseek_model(model):
+            return self._generate_deepseek(messages, **kw)
         if _is_anthropic_model(model):
             return self._generate_anthropic(messages, **kw)
         if _is_google_model(model):
@@ -1036,6 +1142,9 @@ class CloudEngine(InferenceEngine):
         elif _is_minimax_model(model):
             async for token in self._stream_minimax(messages, **kw):
                 yield token
+        elif _is_deepseek_model(model):
+            async for token in self._stream_deepseek(messages, **kw):
+                yield token
         elif _is_anthropic_model(model):
             async for token in self._stream_anthropic(messages, **kw):
                 yield token
@@ -1254,6 +1363,30 @@ class CloudEngine(InferenceEngine):
             if delta and delta.content:
                 yield delta.content
 
+    async def _stream_deepseek(
+        self,
+        messages: Sequence[Message],
+        *,
+        model: str,
+        temperature: float,
+        max_tokens: int,
+        **kwargs: Any,
+    ) -> AsyncIterator[str]:
+        if self._deepseek_client is None:
+            raise EngineConnectionError("DeepSeek client not available")
+        create_kwargs: Dict[str, Any] = {
+            "model": model,
+            "messages": messages_to_dicts(messages),
+            "max_tokens": max_tokens,
+            "temperature": temperature,
+            "stream": True,
+        }
+        resp = self._deepseek_client.chat.completions.create(**create_kwargs)
+        for chunk in resp:
+            delta = chunk.choices[0].delta if chunk.choices else None
+            if delta and delta.content:
+                yield delta.content
+
     # -- stream_full: rich streaming with tool_calls support ----------------
 
     async def _stream_full_openai(
@@ -1307,6 +1440,18 @@ class CloudEngine(InferenceEngine):
                 "stream": True,
                 **kwargs,
             }
+        elif _is_deepseek_model(model):
+            client = self._deepseek_client
+            if client is None:
+                raise EngineConnectionError("DeepSeek client not available")
+            create_kwargs = {
+                "model": model,
+                "messages": messages_to_dicts(messages),
+                "max_tokens": max_tokens,
+                "temperature": temperature,
+                "stream": True,
+                **kwargs,
+            }
         else:
             client = self._openai_client
             if client is None:
@@ -1473,24 +1618,40 @@ class CloudEngine(InferenceEngine):
             models.extend(_OPENROUTER_POPULAR)
         if self._minimax_client is not None:
             models.extend(_MINIMAX_MODELS)
+        if self._deepseek_client is not None:
+            models.extend(_DEEPSEEK_MODELS)
         if self._codex_client is not None:
             models.extend(_CODEX_MODELS)
         return models
 
     def _client_for_model(self, model: str) -> Any:
         """Return the provider client ``generate``/``stream`` will dispatch to
-        for *model* (mirrors the routing in those methods)."""
+        for *model*, or ``None`` for a model this engine cannot route.
+
+        Mirrors the routing in ``generate``/``stream``, but is intentionally
+        *stricter* on the OpenAI fall-through: only genuine OpenAI models map to
+        the OpenAI client. Unrecognized names (e.g. a local Ollama model like
+        ``qwen3.5:0.8b``) return ``None`` so ``can_serve`` declines them and the
+        cloud engine is not mis-selected as a fallback when the local engine is
+        transiently down and any (even dummy) ``OPENAI_API_KEY`` is set (#335).
+        ``generate``/``stream`` keep their OpenAI fall-through, so an
+        explicitly-requested unknown cloud model still fails loudly at call time.
+        """
         if _is_codex_model(model):
             return self._codex_client
         if _is_openrouter_model(model):
             return self._openrouter_client
         if _is_minimax_model(model):
             return self._minimax_client
+        if _is_deepseek_model(model):
+            return self._deepseek_client
         if _is_anthropic_model(model):
             return self._anthropic_client
         if _is_google_model(model):
             return self._google_client
-        return self._openai_client
+        if _is_openai_model(model):
+            return self._openai_client
+        return None
 
     def can_serve(self, model: str) -> bool:
         """Return ``True`` only if the provider client for *model* exists.
@@ -1512,6 +1673,7 @@ class CloudEngine(InferenceEngine):
             or self._google_client is not None
             or self._openrouter_client is not None
             or self._minimax_client is not None
+            or self._deepseek_client is not None
             or self._codex_client is not None
         )
 

src/openjarvis/engine/ollama.py

@@ -22,6 +22,62 @@ from openjarvis.engine._stubs import StreamChunk
 
 logger = logging.getLogger(__name__)
 
+# Qwen3 treats ``/think`` and ``/no_think`` as soft-switch control tokens that
+# toggle reasoning mode. Small models (e.g. qwen3:14b) fed a multi-line prompt
+# sometimes emit one of these as the sole tool argument, e.g.
+# ``{"command": "/no_think"}`` instead of the real command. Ollama parses that
+# into a fully-formed tool_call via the model's chat template, so we have to
+# drop it on our side before the agent executes garbage.
+_QWEN_CONTROL_TOKENS = frozenset({"/think", "/no_think"})
+
+
+def _is_control_token_only_args(raw_args: Any) -> bool:
+    """Return True if tool-call arguments contain nothing but a Qwen3 token.
+
+    ``raw_args`` may be a dict (Ollama's native shape) or a JSON / bare string.
+    A call is considered degenerate only when it carries at least one control
+    token and no other usable content, so legitimate calls such as
+    ``{"command": "date"}`` or ``{"command": "echo /no_think"}`` are kept.
+    """
+    parsed: Any = raw_args
+    if isinstance(raw_args, str):
+        try:
+            parsed = json.loads(raw_args)
+        except (json.JSONDecodeError, TypeError):
+            parsed = raw_args
+
+    if isinstance(parsed, str):
+        return parsed.strip().lower() in _QWEN_CONTROL_TOKENS
+
+    if not isinstance(parsed, dict) or not parsed:
+        return False
+
+    saw_token = False
+    for value in parsed.values():
+        if not isinstance(value, str):
+            return False  # a non-string value is real content
+        stripped = value.strip()
+        if not stripped:
+            continue
+        if stripped.lower() in _QWEN_CONTROL_TOKENS:
+            saw_token = True
+        else:
+            return False  # real string content
+    return saw_token
+
+
+def _default_num_ctx() -> int:
+    """Default context window (tokens). Override with ``JARVIS_NUM_CTX``.
+
+    Raised above Ollama's 4k default so an image (which costs many tokens)
+    plus a real conversation fit. 16k is comfortable for small models on a
+    typical consumer GPU.
+    """
+    try:
+        return int(os.environ.get("JARVIS_NUM_CTX", "16384"))
+    except ValueError:
+        return 16384
+
 
 @EngineRegistry.register("ollama")
 class OllamaEngine(InferenceEngine):
@@ -73,7 +129,7 @@ class OllamaEngine(InferenceEngine):
             "options": {
                 "temperature": temperature,
                 "num_predict": max_tokens,
-                "num_ctx": kwargs.get("num_ctx", 8192),
+                "num_ctx": kwargs.get("num_ctx", _default_num_ctx()),
             },
         }
         # Disable extended thinking by default (Qwen3.5 etc.).
@@ -155,14 +211,19 @@ class OllamaEngine(InferenceEngine):
         if raw_tool_calls:
             tool_calls = []
             for i, tc in enumerate(raw_tool_calls):
-                raw_args = tc.get("function", {}).get(
-                    "arguments",
-                    "{}",
-                )
+                fn = tc.get("function", {})
+                raw_args = fn.get("arguments", "{}")
+                if _is_control_token_only_args(raw_args):
+                    logger.warning(
+                        "Dropping Qwen3 control-token tool call %s(%r)",
+                        fn.get("name", ""),
+                        raw_args,
+                    )
+                    continue
                 tool_calls.append(
                     {
                         "id": tc.get("id", f"call_{i}"),
-                        "name": tc.get("function", {}).get("name", ""),
+                        "name": fn.get("name", ""),
                         "arguments": (
                             json.dumps(raw_args)
                             if isinstance(raw_args, dict)
@@ -170,7 +231,8 @@ class OllamaEngine(InferenceEngine):
                         ),
                     }
                 )
-            result["tool_calls"] = tool_calls
+            if tool_calls:
+                result["tool_calls"] = tool_calls
         return result
 
     async def stream(
@@ -189,7 +251,7 @@ class OllamaEngine(InferenceEngine):
             "options": {
                 "temperature": temperature,
                 "num_predict": max_tokens,
-                "num_ctx": kwargs.get("num_ctx", 8192),
+                "num_ctx": kwargs.get("num_ctx", _default_num_ctx()),
             },
         }
         # Mirror generate()'s default: disable extended thinking unless the
@@ -268,7 +330,7 @@ class OllamaEngine(InferenceEngine):
             "options": {
                 "temperature": temperature,
                 "num_predict": max_tokens,
-                "num_ctx": kwargs.get("num_ctx", 8192),
+                "num_ctx": kwargs.get("num_ctx", _default_num_ctx()),
             },
         }
         if "think" not in kwargs:
@@ -327,14 +389,22 @@ class OllamaEngine(InferenceEngine):
                         # OpenAI-delta fragment shape that agent_manager_routes
                         # expects in _merge_tool_call_fragments.
                         fragments: List[Dict[str, Any]] = []
-                        for i, tc in enumerate(raw_tool_calls):
+                        for tc in raw_tool_calls:
                             fn = tc.get("function", {}) or {}
                             raw_args = fn.get("arguments", "{}")
+                            if _is_control_token_only_args(raw_args):
+                                logger.warning(
+                                    "Dropping Qwen3 control-token tool call %s(%r)",
+                                    fn.get("name", ""),
+                                    raw_args,
+                                )
+                                continue
                             args_str = (
                                 json.dumps(raw_args)
                                 if isinstance(raw_args, dict)
                                 else str(raw_args)
                             )
+                            i = len(fragments)
                             fragments.append(
                                 {
                                     "index": i,
@@ -346,8 +416,9 @@ class OllamaEngine(InferenceEngine):
                                     },
                                 }
                             )
-                        yield StreamChunk(tool_calls=fragments)
-                        finish_reason = "tool_calls"
+                        if fragments:
+                            yield StreamChunk(tool_calls=fragments)
+                            finish_reason = "tool_calls"
 
                     if chunk.get("done", False):
                         reported_prompt = chunk.get("prompt_eval_count", 0)

src/openjarvis/evals/datasets/gaia.py

@@ -11,11 +11,12 @@ import shutil
 from pathlib import Path
 from typing import Iterable, List, MutableMapping, Optional, Sequence
 
+from openjarvis.core.paths import get_cache_dir
 from openjarvis.evals.core.dataset import DatasetProvider
 from openjarvis.evals.core.splits import apply_split
 from openjarvis.evals.core.types import EvalRecord
 
-_DEFAULT_CACHE_DIR = Path.home() / ".cache" / "gaia_benchmark"
+_DEFAULT_CACHE_DIR = get_cache_dir() / "gaia_benchmark"
 
 _DEFAULT_INPUT_PROMPT = """Please answer the question below. You should:
 

src/openjarvis/evals/datasets/liveresearch.py

@@ -17,6 +17,7 @@ import subprocess
 from pathlib import Path
 from typing import Any, Dict, Iterable, List, Optional
 
+from openjarvis.core.paths import get_cache_dir
 from openjarvis.evals.core.dataset import DatasetProvider
 from openjarvis.evals.core.splits import apply_split
 from openjarvis.evals.core.types import EvalRecord
@@ -24,7 +25,7 @@ from openjarvis.evals.core.types import EvalRecord
 LOGGER = logging.getLogger(__name__)
 
 LIVERESEARCH_REPO = "https://github.com/Ayanami0730/deep_research_bench.git"
-CACHE_DIR = Path.home() / ".cache" / "liveresearch_bench"
+CACHE_DIR = get_cache_dir() / "liveresearch_bench"
 
 
 def _load_jsonl(path: Path) -> List[Dict[str, Any]]:

src/openjarvis/evals/datasets/loghub.py

@@ -13,6 +13,7 @@ import re
 from pathlib import Path
 from typing import Any, Dict, Iterable, List, Optional
 
+from openjarvis.core.paths import get_cache_dir
 from openjarvis.evals.core.dataset import DatasetProvider
 from openjarvis.evals.core.types import EvalRecord
 
@@ -64,9 +65,7 @@ class LogHubDataset(DatasetProvider):
                 f"Choose from: {list(_DATASETS.keys())}"
             )
         self._subset = subset
-        self._cache_dir = (
-            Path(cache_dir) if cache_dir else Path.home() / ".cache" / "loghub"
-        )
+        self._cache_dir = Path(cache_dir) if cache_dir else get_cache_dir() / "loghub"
         self._records: List[EvalRecord] = []
 
     def load(

src/openjarvis/evals/datasets/paperarena.py

@@ -15,6 +15,7 @@ import random
 from pathlib import Path
 from typing import Any, Dict, Iterable, List, Optional
 
+from openjarvis.core.paths import get_cache_dir
 from openjarvis.evals.core.dataset import DatasetProvider
 from openjarvis.evals.core.types import EvalRecord
 
@@ -41,7 +42,7 @@ class PaperArenaDataset(DatasetProvider):
         cache_dir: Optional[str] = None,
     ) -> None:
         self._cache_dir = (
-            Path(cache_dir) if cache_dir else Path.home() / ".cache" / "paperarena"
+            Path(cache_dir) if cache_dir else get_cache_dir() / "paperarena"
         )
         self._records: List[EvalRecord] = []
 

src/openjarvis/evals/datasets/pinchbench.py

@@ -18,6 +18,7 @@ from typing import Any, Dict, Iterable, List, Optional
 
 import yaml
 
+from openjarvis.core.paths import get_cache_dir
 from openjarvis.evals.core.dataset import DatasetProvider
 from openjarvis.evals.core.splits import apply_split
 from openjarvis.evals.core.types import EvalRecord
@@ -25,7 +26,7 @@ from openjarvis.evals.core.types import EvalRecord
 LOGGER = logging.getLogger(__name__)
 
 PINCHBENCH_REPO = "https://github.com/pinchbench/skill.git"
-CACHE_DIR = Path.home() / ".cache" / "pinchbench"
+CACHE_DIR = get_cache_dir() / "pinchbench"
 
 
 def _parse_task_markdown(content: str, filename: str = "") -> Dict[str, Any]:

src/openjarvis/evals/datasets/taubench.py

@@ -12,9 +12,9 @@ import logging
 import os
 import subprocess
 import sys
-from pathlib import Path
 from typing import Iterable, List, Optional
 
+from openjarvis.core.paths import get_cache_dir
 from openjarvis.evals.core.dataset import DatasetProvider
 from openjarvis.evals.core.splits import apply_split
 from openjarvis.evals.core.types import EvalRecord
@@ -22,7 +22,7 @@ from openjarvis.evals.core.types import EvalRecord
 LOGGER = logging.getLogger(__name__)
 
 TAU2_REPO = "https://github.com/sierra-research/tau2-bench.git"
-CACHE_DIR = Path.home() / ".cache" / "tau2-bench"
+CACHE_DIR = get_cache_dir() / "tau2-bench"
 
 DOMAINS = ("airline", "retail", "telecom")
 

src/openjarvis/evals/datasets/webchorearena.py

@@ -21,6 +21,7 @@ import subprocess
 from pathlib import Path
 from typing import Any, Dict, Iterable, List, Optional
 
+from openjarvis.core.paths import get_cache_dir
 from openjarvis.evals.core.dataset import DatasetProvider
 from openjarvis.evals.core.types import EvalRecord
 
@@ -56,7 +57,7 @@ class WebChoreArenaDataset(DatasetProvider):
     ) -> None:
         self._subset = subset  # "all", "small", or a site name
         self._cache_dir = (
-            Path(cache_dir) if cache_dir else Path.home() / ".cache" / "webchorearena"
+            Path(cache_dir) if cache_dir else get_cache_dir() / "webchorearena"
         )
         self._headless = headless
         self._records: List[EvalRecord] = []

src/openjarvis/evals/scorers/swebench_harness.py

@@ -49,6 +49,7 @@ import tempfile
 from pathlib import Path
 from typing import Any, Dict, Optional, Tuple
 
+from openjarvis.core.paths import get_cache_dir
 from openjarvis.evals.core.scorer import Scorer
 from openjarvis.evals.core.types import EvalRecord
 
@@ -93,9 +94,7 @@ def _run_subprocess_hard_timeout(
     )
     try:
         stdout, stderr = proc.communicate(timeout=timeout_s)
-        return subprocess.CompletedProcess(
-            cmd, proc.returncode, stdout, stderr
-        )
+        return subprocess.CompletedProcess(cmd, proc.returncode, stdout, stderr)
     except subprocess.TimeoutExpired:
         # Kill the whole group, not just the direct child — Modal harness
         # subprocesses fork workers that would otherwise keep pipes open.
@@ -115,9 +114,7 @@ def _run_subprocess_hard_timeout(
             stdout, stderr = proc.communicate(timeout=10)
         except subprocess.TimeoutExpired:
             stdout, stderr = "", ""
-        raise subprocess.TimeoutExpired(
-            cmd, timeout_s, output=stdout, stderr=stderr
-        )
+        raise subprocess.TimeoutExpired(cmd, timeout_s, output=stdout, stderr=stderr)
 
 
 # ---------- Patch tracking ----------
@@ -208,11 +205,11 @@ def _patch_modal_sandbox_source() -> None:
         # Upstream changed the line — bail rather than apply blindly.
         return
     replacement = (
-        '        # ' + _CGROUP_SOURCE_SENTINEL + '\n'
-        '        try:\n'
+        "        # " + _CGROUP_SOURCE_SENTINEL + "\n"
+        "        try:\n"
         '            self.write_file("/sys/fs/cgroup/cpu/cpu.shares", "2048")\n'
-        '        except FileNotFoundError:\n'
-        '            pass  # cgroup v2 Modal sandbox — path missing is fine\n'
+        "        except FileNotFoundError:\n"
+        "            pass  # cgroup v2 Modal sandbox — path missing is fine\n"
     )
     new_src = src.replace(needle + "\n", replacement, 1)
     try:
@@ -324,22 +321,23 @@ def extract_patch(text: str) -> Optional[str]:
 
 # ---------- Harness invocation ----------
 
+
 def _harness_cache_dir() -> Path:
     """Where the swebench subprocess writes its report JSON + logs/ tree.
 
-    Defaults to ``$OPENJARVIS_HOME/.swebench-cache`` if set, otherwise to a
-    process-shared tempdir. Pin both so we don't pollute the project root.
+    Consolidated under the env-aware OpenJarvis cache root
+    (``<openjarvis-home>/cache/swebench``) so it never pollutes the project
+    root or scatters across ``$HOME``. Honors ``OPENJARVIS_HOME`` /
+    ``XDG_DATA_HOME`` via :func:`openjarvis.core.paths.get_cache_dir`.
     """
-    home = os.environ.get("OPENJARVIS_HOME")
-    if home:
-        cache = Path(home) / ".swebench-cache"
-    else:
-        cache = Path(tempfile.gettempdir()) / "openjarvis-swebench-cache"
+    cache = get_cache_dir() / "swebench"
     cache.mkdir(parents=True, exist_ok=True)
     return cache
 
 
-def _find_report(cache: Path, instance_id: str, run_id: str) -> Optional[Dict[str, Any]]:
+def _find_report(
+    cache: Path, instance_id: str, run_id: str
+) -> Optional[Dict[str, Any]]:
     """Find the harness's report JSON for one instance.
 
     swebench writes ``<model_name_or_path>.<run_id>.json`` inside the
@@ -427,26 +425,40 @@ def _run_harness(
     with tempfile.TemporaryDirectory() as tmp:
         tmp_path = Path(tmp)
         preds_path = tmp_path / "predictions.jsonl"
-        preds_path.write_text(json.dumps({
-            "instance_id": instance_id,
-            "model_name_or_path": "openjarvis-harness",
-            "model_patch": patch,
-        }) + "\n")
+        preds_path.write_text(
+            json.dumps(
+                {
+                    "instance_id": instance_id,
+                    "model_name_or_path": "openjarvis-harness",
+                    "model_patch": patch,
+                }
+            )
+            + "\n"
+        )
 
         cmd = [
-            sys.executable, "-m", "swebench.harness.run_evaluation",
-            "--predictions_path", str(preds_path),
-            "--max_workers", "1",
-            "--run_id", run_id,
-            "--dataset_name", "SWE-bench/SWE-bench_Verified",
-            "--instance_ids", instance_id,
+            sys.executable,
+            "-m",
+            "swebench.harness.run_evaluation",
+            "--predictions_path",
+            str(preds_path),
+            "--max_workers",
+            "1",
+            "--run_id",
+            run_id,
+            "--dataset_name",
+            "SWE-bench/SWE-bench_Verified",
+            "--instance_ids",
+            instance_id,
         ]
         if backend == "modal":
             cmd += ["--modal", "true"]
 
         try:
             proc = _run_subprocess_hard_timeout(
-                cmd, timeout_s=timeout_s, cwd=str(cache),
+                cmd,
+                timeout_s=timeout_s,
+                cwd=str(cache),
             )
         except subprocess.TimeoutExpired as exc:
             # The harness subprocess (and its Modal grandchildren) exceeded
@@ -497,6 +509,7 @@ def _run_harness(
 
 # ---------- Scorer ----------
 
+
 class SWEBenchHarnessScorer(Scorer):
     """SWE-bench Verified scorer that runs the official harness.
 
@@ -516,7 +529,7 @@ class SWEBenchHarnessScorer(Scorer):
         timeout_s: int = 1800,
         cell_name: Optional[str] = None,
         judge_backend: object = None,  # noqa: ARG002 — CLI factory compat
-        judge_model: str = "",         # noqa: ARG002 — CLI factory compat
+        judge_model: str = "",  # noqa: ARG002 — CLI factory compat
     ) -> None:
         self._timeout_s = int(timeout_s)
         # ``cell_name`` namespaces the ``run_id`` so concurrent cells scoring
@@ -538,16 +551,15 @@ class SWEBenchHarnessScorer(Scorer):
         if patch is None:
             return False, {"reason": "no_patch_extracted"}
 
-        instance_id = (
-            record.metadata.get("instance_id")
-            or record.record_id
-            or ""
-        )
+        instance_id = record.metadata.get("instance_id") or record.record_id or ""
         if not instance_id:
             return False, {"reason": "missing_instance_id"}
 
         result = _run_harness(
-            instance_id, patch, self._timeout_s, cell_name=self._cell_name,
+            instance_id,
+            patch,
+            self._timeout_s,
+            cell_name=self._cell_name,
         )
         details = dict(result.get("details", {}))
         details["patch"] = patch

src/openjarvis/evals/skill_benchmark.py

@@ -16,6 +16,8 @@ from datetime import datetime, timezone
 from pathlib import Path
 from typing import Any, Dict, List, Optional
 
+from openjarvis.core.paths import get_config_dir
+
 LOGGER = logging.getLogger(__name__)
 
 CONDITIONS = (
@@ -52,14 +54,12 @@ class SkillBenchmarkConfig:
     seeds: List[int] = field(default_factory=lambda: [42, 43, 44])
     max_samples: Optional[int] = None
     output_dir: Path = field(default_factory=lambda: Path("docs/superpowers/results/"))
-    skills_dir: Path = field(
-        default_factory=lambda: Path("~/.openjarvis/skills/").expanduser()
-    )
+    skills_dir: Path = field(default_factory=lambda: get_config_dir() / "skills")
     overlay_dir_dspy: Path = field(
-        default_factory=lambda: Path("~/.openjarvis/learning/skills-dspy/").expanduser()
+        default_factory=lambda: get_config_dir() / "learning" / "skills-dspy"
     )
     overlay_dir_gepa: Path = field(
-        default_factory=lambda: Path("~/.openjarvis/learning/skills-gepa/").expanduser()
+        default_factory=lambda: get_config_dir() / "learning" / "skills-gepa"
     )
 
 

src/openjarvis/learning/agents/ace_optimizer.py

@@ -30,6 +30,7 @@ from pathlib import Path
 from typing import Any, Dict, List
 
 from openjarvis.core.config import ACEOptimizerConfig
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.registry import LearningRegistry
 from openjarvis.learning._stubs import AgentLearningPolicy
 
@@ -48,7 +49,7 @@ except ImportError:
 
 
 def _default_save_dir(task_name: str) -> Path:
-    return Path.home() / ".openjarvis" / "learning" / "ace" / task_name
+    return get_config_dir() / "learning" / "ace" / task_name
 
 
 class _TraceDataProcessor:
@@ -87,9 +88,7 @@ class _TraceDataProcessor:
         if not predictions:
             return 0.0
         n_correct = sum(
-            1
-            for p, g in zip(predictions, ground_truths)
-            if cls.answer_is_correct(p, g)
+            1 for p, g in zip(predictions, ground_truths) if cls.answer_is_correct(p, g)
         )
         return n_correct / len(predictions)
 
@@ -149,8 +148,7 @@ class ACEAgentOptimizer:
             return {
                 "status": "skipped",
                 "reason": (
-                    f"only {len(traces)} traces, "
-                    f"min_traces={self.config.min_traces}"
+                    f"only {len(traces)} traces, min_traces={self.config.min_traces}"
                 ),
             }
 
@@ -158,8 +156,7 @@ class ACEAgentOptimizer:
             return {
                 "status": "error",
                 "reason": (
-                    "ace not installed (pip install "
-                    "'openjarvis[learning-ace]')"
+                    "ace not installed (pip install 'openjarvis[learning-ace]')"
                 ),
             }
 

src/openjarvis/learning/agents/skill_optimizer.py

@@ -19,6 +19,7 @@ from datetime import datetime, timezone
 from pathlib import Path
 from typing import Any, Dict, List, Optional
 
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.types import Trace, TraceStep
 from openjarvis.skills.manager import SkillManager
 from openjarvis.skills.overlay import SkillOverlay, write_overlay
@@ -99,7 +100,7 @@ class SkillOptimizer:
                 pass
             if overlay_dir is None:
                 overlay_dir = Path(
-                    "~/.openjarvis/learning/skills/"
+                    str(get_config_dir() / "learning" / "skills")
                 ).expanduser()
         overlay_dir = Path(overlay_dir).expanduser()
 

src/openjarvis/learning/spec_search/storage/paths.py

@@ -11,14 +11,21 @@ writing artifacts into the working tree.
 
 from __future__ import annotations
 
-import os
 from pathlib import Path
 
+from openjarvis.core.paths import ConfigurationError, get_config_dir
 from openjarvis.security.file_utils import secure_mkdir
 
-
-class ConfigurationError(RuntimeError):
-    """Raised when path configuration would violate isolation guarantees."""
+# ``ConfigurationError`` is re-exported from ``openjarvis.core.paths`` (it used
+# to be defined here). Spec search now resolves the home dir through the unified
+# core resolver, which raises the same exception type on a source-tree path, so
+# we alias rather than redefine to keep ``except ConfigurationError`` callers and
+# existing tests working.
+__all__ = [
+    "ConfigurationError",
+    "ensure_spec_search_dirs",
+    "resolve_spec_search_root",
+]
 
 
 def _find_source_root() -> Path | None:
@@ -42,11 +49,12 @@ def _find_source_root() -> Path | None:
 
 
 def _resolve_openjarvis_home() -> Path:
-    """Resolve the OPENJARVIS_HOME directory (env var or default)."""
-    env = os.environ.get("OPENJARVIS_HOME")
-    if env:
-        return Path(env).expanduser().resolve()
-    return (Path.home() / ".openjarvis").resolve()
+    """Resolve the OpenJarvis home directory via the unified core resolver.
+
+    Delegates to ``get_config_dir`` so spec-search honors the same env-aware
+    resolution (OPENJARVIS_HOME and XDG) as the rest of the framework.
+    """
+    return get_config_dir()
 
 
 def resolve_spec_search_root() -> Path:

src/openjarvis/mining/_constants.py

@@ -9,7 +9,7 @@ section 7.3 for the rev-bump workflow.
 
 from __future__ import annotations
 
-from pathlib import Path
+from openjarvis.core.paths import get_config_dir
 
 PEARL_REPO = "https://github.com/pearl-research-labs/pearl.git"
 # TODO at implementation time: replace with the specific commit/tag verified
@@ -47,9 +47,9 @@ DEFAULT_PEARLD_RPC_URL = "http://localhost:44107"
 MIN_FREE_DISK_GB = 200
 
 # Runtime sidecar location (single-session assumption — see spec §8.8).
-RUNTIME_DIR = Path.home() / ".openjarvis" / "runtime"
+RUNTIME_DIR = get_config_dir() / "runtime"
 SIDECAR_PATH = RUNTIME_DIR / "mining.json"
 SIDECAR_LOCK_PATH = RUNTIME_DIR / "mining.lock"
 
 # Pearl source cache for build-from-pin path (see spec §7.2).
-PEARL_CACHE_DIR = Path.home() / ".openjarvis" / "cache" / "pearl"
+PEARL_CACHE_DIR = get_config_dir() / "cache" / "pearl"

src/openjarvis/mining/cpu_pearl.py

@@ -17,6 +17,7 @@ import urllib.request
 from pathlib import Path
 
 from openjarvis.core.config import HardwareInfo
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.registry import MinerRegistry
 
 from . import _install
@@ -47,7 +48,7 @@ def _sidecar_path() -> Path:
 
 def _log_dir() -> Path:
     """Return the logs directory. Override in tests."""
-    return Path.home() / ".openjarvis" / "logs" / "mining"
+    return get_config_dir() / "logs" / "mining"
 
 
 def _parse_gateway_metrics(text: str, *, provider_id: str) -> MiningStats:

src/openjarvis/prompt/builder.py

@@ -5,6 +5,7 @@ from pathlib import Path
 from typing import List, Literal, Optional, Tuple
 
 from openjarvis.core.config import MemoryFilesConfig, SystemPromptConfig
+from openjarvis.core.paths import get_config_dir
 
 PromptCacheSegment = Literal["frozen_prefix", "dynamic_suffix"]
 
@@ -279,7 +280,7 @@ class SystemPromptBuilder:
                 f"Invalid persona name {name!r}: must be a simple "
                 "identifier (no path separators or '..')."
             )
-        base = Path.home() / ".openjarvis" / "personas" / name
+        base = get_config_dir() / "personas" / name
         return MemoryFilesConfig(
             soul_path=str(base / "SOUL.md"),
             memory_path=str(base / "MEMORY.md"),

src/openjarvis/recipes/loader.py

@@ -19,13 +19,14 @@ try:
 except ModuleNotFoundError:
     import tomli as tomllib  # type: ignore[no-redef]
 
+from openjarvis.core.paths import get_config_dir
 
 # Built-in recipes directory (package data)
 _PROJECT_RECIPES_DIR = Path(__file__).resolve().parent / "data"
 _PROJECT_OPERATORS_DIR = _PROJECT_RECIPES_DIR / "operators"
 # User-level directories
-_USER_RECIPES_DIR = Path.home() / ".openjarvis" / "recipes"
-_USER_OPERATORS_DIR = Path.home() / ".openjarvis" / "operators"
+_USER_RECIPES_DIR = get_config_dir() / "recipes"
+_USER_OPERATORS_DIR = get_config_dir() / "operators"
 
 
 @dataclass(slots=True)

src/openjarvis/security/guardrails.py

@@ -192,6 +192,7 @@ class GuardrailsEngine(InferenceEngine):
                             tool_calls=msg.tool_calls,
                             tool_call_id=msg.tool_call_id,
                             metadata=msg.metadata,
+                            images=msg.images,
                         )
             messages = processed
 

src/openjarvis/server/agent_manager_routes.py

@@ -2095,10 +2095,13 @@ def create_agent_manager_router(
             raise HTTPException(status_code=404, detail="Agent not found")
         try:
             from openjarvis.core.config import load_config
+            from openjarvis.core.paths import get_config_dir
             from openjarvis.traces.store import TraceStore
 
             config = load_config()
-            store = TraceStore(config.traces.db_path or "~/.openjarvis/traces.db")
+            store = TraceStore(
+                config.traces.db_path or str(get_config_dir() / "traces.db")
+            )
             traces = store.list_traces(agent=agent_id, limit=limit)
             return {
                 "traces": [
@@ -2120,10 +2123,13 @@ def create_agent_manager_router(
     def get_trace(agent_id: str, trace_id: str):
         try:
             from openjarvis.core.config import load_config
+            from openjarvis.core.paths import get_config_dir
             from openjarvis.traces.store import TraceStore
 
             config = load_config()
-            store = TraceStore(config.traces.db_path or "~/.openjarvis/traces.db")
+            store = TraceStore(
+                config.traces.db_path or str(get_config_dir() / "traces.db")
+            )
             trace = store.get(trace_id)
             if trace is None:
                 raise HTTPException(status_code=404, detail="Trace not found")

src/openjarvis/server/api_routes.py

@@ -893,7 +893,15 @@ async def transcribe_speech(request: Request):
     filename = getattr(audio_file, "filename", "audio.wav")
     ext = filename.rsplit(".", 1)[-1] if "." in filename else "wav"
 
-    result = backend.transcribe(audio_bytes, format=ext, language=language or None)
+    try:
+        result = backend.transcribe(audio_bytes, format=ext, language=language or None)
+    except Exception as exc:
+        logger.exception("Speech transcription failed")
+        raise HTTPException(
+            status_code=500,
+            detail=f"Speech transcription failed: {exc}",
+        ) from exc
+
     return {
         "text": result.text,
         "language": result.language,
@@ -908,9 +916,23 @@ async def speech_health(request: Request):
     backend = getattr(request.app.state, "speech_backend", None)
     if backend is None:
         return {"available": False, "reason": "No speech backend configured"}
+    try:
+        available = backend.health()
+        reason = None
+    except Exception as exc:
+        logger.exception("Speech health check failed")
+        available = False
+        reason = str(exc)
+
+    if not available and reason is None:
+        last_error = getattr(backend, "last_error", None)
+        if callable(last_error):
+            reason = last_error()
+
     return {
-        "available": backend.health(),
+        "available": available,
         "backend": backend.backend_id,
+        **({"reason": reason} if reason else {}),
     }
 
 

src/openjarvis/server/cloud_router.py

@@ -10,18 +10,18 @@ from __future__ import annotations
 import json
 import os
 from collections.abc import AsyncIterator
-from pathlib import Path
 from typing import Any, Sequence
 
 import httpx
 
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.types import Message
 
 # ---------------------------------------------------------------------------
 # Key / provider detection
 # ---------------------------------------------------------------------------
 
-_CLOUD_ENV_FILE = Path.home() / ".openjarvis" / "cloud-keys.env"
+_CLOUD_ENV_FILE = get_config_dir() / "cloud-keys.env"
 
 _OPENAI_PREFIXES = ("gpt-", "o1-", "o3-", "o4-", "chatgpt-")
 _ANTHROPIC_PREFIXES = ("claude-",)

src/openjarvis/server/connectors_router.py

@@ -3,7 +3,23 @@
 from __future__ import annotations
 
 import logging
-from typing import Any, Dict, Optional
+from typing import TYPE_CHECKING, Any, Dict, Optional
+
+# ``Request`` must be importable at *module* scope so that FastAPI can resolve
+# the stringized ``request: Request`` annotations on the OAuth endpoints below.
+# Because this module uses ``from __future__ import annotations``, every
+# annotation is a string that FastAPI evaluates against the module globals; a
+# ``Request`` imported only inside ``create_connectors_router()`` is invisible
+# there, which makes FastAPI mistake ``request`` for a required *query* param
+# (HTTP 422 on /oauth/start) or inject ``None`` (AttributeError on
+# /oauth/callback). Keep this import at top level. See issue #512.
+if TYPE_CHECKING:
+    from starlette.requests import Request
+else:
+    try:
+        from starlette.requests import Request
+    except ImportError:  # starlette ships with fastapi; absent only without it
+        Request = Any  # type: ignore[assignment,misc]
 
 logger = logging.getLogger(__name__)
 
@@ -78,7 +94,7 @@ def create_connectors_router():
     this package.
     """
     try:
-        from fastapi import APIRouter, HTTPException, Request
+        from fastapi import APIRouter, HTTPException
     except ImportError as exc:
         raise ImportError(
             "fastapi and pydantic are required for the connectors router"
@@ -125,6 +141,69 @@ def create_connectors_router():
             "chunks": chunks,
         }
 
+    def _maybe_oauth_client_pair(
+        connector_id: str, req: ConnectRequest
+    ) -> Optional[Dict[str, Any]]:
+        """Handle a pasted ``client_id:client_secret`` for an OAuth connector.
+
+        Returns an ``oauth_required`` directive (and persists the client
+        credentials to every credential file for the provider) when *req*
+        carries a Google ``client_id:client_secret`` pair, so the caller can
+        return early instead of triggering the silent background OAuth flow.
+        Returns ``None`` when there is no such pair (the caller then falls
+        through to the normal ``handle_callback`` / token path).
+
+        Raises ``HTTPException(400)`` when the pair is present but malformed or
+        the connector has no OAuth provider — per the silent-failure discipline
+        in REVIEW.md, a bad credential surfaces an actionable error rather than
+        a perpetual ``pending`` state.
+        """
+        from openjarvis.connectors.oauth import (
+            get_provider_for_connector,
+            save_client_credentials,
+        )
+
+        raw = (req.code or req.token or "").strip()
+        # Only the client-registration pair routes through the server flow.
+        # A raw access token (no ".apps.googleusercontent.com") is handled by
+        # the connector's handle_callback unchanged.
+        if ".apps.googleusercontent.com" not in raw or ":" not in raw:
+            return None
+
+        provider = get_provider_for_connector(connector_id)
+        if provider is None:
+            raise HTTPException(
+                status_code=400,
+                detail=f"No OAuth provider configured for '{connector_id}'",
+            )
+
+        client_id, client_secret = raw.split(":", 1)
+        client_id = client_id.strip()
+        client_secret = client_secret.strip()
+        if not client_id or not client_secret:
+            raise HTTPException(
+                status_code=400,
+                detail=(
+                    "Malformed credentials — expected 'CLIENT_ID:CLIENT_SECRET'. "
+                    f"Create an OAuth client at: {provider.setup_url}"
+                ),
+            )
+
+        save_client_credentials(provider, client_id, client_secret)
+        # Cached instances may have resolved a stale credentials path before
+        # these client creds existed; drop them so /oauth/callback rebuilds
+        # them against the freshly written files.
+        for cid in provider.connector_ids:
+            _instances.pop(cid, None)
+
+        return {
+            "connector_id": connector_id,
+            "connected": False,
+            "status": "oauth_required",
+            "oauth_start": f"/v1/connectors/{connector_id}/oauth/start",
+            "sync_status": None,
+        }
+
     # ------------------------------------------------------------------
     # Background-sync state tracking. Defined here (before the endpoints)
     # so that POST /connect can fire-and-forget into the same machinery
@@ -317,6 +396,19 @@ def create_connectors_router():
                     instance._connected = Path(req.path).is_dir()
 
             elif auth_type == "oauth":
+                # A pasted ``client_id:client_secret`` pair is NOT a completed
+                # OAuth credential — it is the app registration. Persist it and
+                # hand the UI a directive to run the in-process browser consent
+                # flow (/oauth/start → /oauth/callback), which is the only path
+                # that actually exchanges a code for an access_token. Previously
+                # this routed into the connector's handle_callback, which spawned
+                # a daemon thread that popped a browser + ran its own
+                # localhost:8789 callback server; that thread fails silently in
+                # the bundled desktop context, so the connector never became
+                # connected and never appeared in Data Sources (issue #512).
+                directive = _maybe_oauth_client_pair(connector_id, req)
+                if directive is not None:
+                    return directive
                 if req.code:
                     instance.handle_callback(req.code)
                 elif req.token:
@@ -433,9 +525,9 @@ def create_connectors_router():
     @router.get("/{connector_id}/oauth/callback")
     async def oauth_callback(
         connector_id: str,
+        request: Request,
         code: str = "",
         error: str = "",
-        request: Request = None,
     ):
         """Handle OAuth callback from the provider."""
         from fastapi.responses import HTMLResponse

src/openjarvis/server/routes.py

@@ -9,6 +9,7 @@ from typing import Any
 from fastapi import APIRouter, HTTPException, Request
 from fastapi.responses import StreamingResponse
 
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.types import Message, Role
 from openjarvis.server.models import (
     ChatCompletionChunk,
@@ -43,6 +44,51 @@ def _to_messages(chat_messages) -> list[Message]:
     return messages
 
 
+def _ensure_identity_prompt(messages: list[Message], app_config) -> list[Message]:
+    """Prepend OpenJarvis's identity system prompt when the client omits one.
+
+    The desktop UI's chat backend posts only user/assistant turns to
+    ``/v1/chat/completions`` (see ``frontend/.../Chat/InputArea.tsx``), so
+    nothing grounds the model's identity. Without a system prompt the model
+    answers from its training identity (e.g. "I'm Claude", "I am Qwen"),
+    which is what #540 reported. The CLI paths inject this via
+    ``SystemPromptBuilder`` / ``BaseAgent``; the engine-direct server paths
+    did not. This mirrors the agent fallback in ``agents/_stubs.py``.
+
+    If any message already carries a system role, the caller has supplied
+    their own grounding and we leave the list untouched (no double-prompting).
+
+    Resolution of the identity text: ``app_config.agent.default_system_prompt``
+    when a config is wired onto ``app.state``; otherwise fall back to
+    ``load_config()``. Config resolution is wrapped so a broken/missing
+    config degrades to "no injection" rather than crashing the endpoint, but
+    the failure is logged (per REVIEW.md — never silently swallow).
+    """
+    if any(m.role == Role.SYSTEM for m in messages):
+        return messages
+
+    prompt = ""
+    try:
+        if app_config is not None:
+            prompt = app_config.agent.default_system_prompt or ""
+        else:
+            from openjarvis.core.config import load_config
+
+            prompt = load_config().agent.default_system_prompt or ""
+    except Exception:
+        logging.getLogger("openjarvis.server").debug(
+            "Identity system prompt resolution failed; "
+            "serving request without identity grounding",
+            exc_info=True,
+        )
+        return messages
+
+    if not prompt:
+        return messages
+
+    return [Message(role=Role.SYSTEM, content=prompt), *messages]
+
+
 @router.post("/v1/chat/completions")
 async def chat_completions(request_body: ChatCompletionRequest, request: Request):
     """Handle chat completion requests (streaming and non-streaming)."""
@@ -149,7 +195,7 @@ async def chat_completions(request_body: ChatCompletionRequest, request: Request
         # from the engine for true real-time output.
         if request_body.tools:
             return await _handle_stream_tools(
-                engine, model, request_body, complexity_info
+                engine, model, request_body, complexity_info, app_config=config
             )
         return await _handle_stream(
             engine,
@@ -157,6 +203,7 @@ async def chat_completions(request_body: ChatCompletionRequest, request: Request
             request_body,
             complexity_info,
             trace_store=getattr(request.app.state, "trace_store", None),
+            app_config=config,
         )
 
     # Non-streaming: use agent if available, otherwise direct engine call.
@@ -192,6 +239,7 @@ async def chat_completions(request_body: ChatCompletionRequest, request: Request
         request_body,
         bus=bus,
         complexity_info=complexity_info,
+        app_config=config,
     )
 
 
@@ -201,9 +249,11 @@ def _handle_direct(
     req: ChatCompletionRequest,
     bus=None,
     complexity_info=None,
+    app_config=None,
 ) -> ChatCompletionResponse:
     """Direct engine call without agent."""
     messages = _to_messages(req.messages)
+    messages = _ensure_identity_prompt(messages, app_config)
     kwargs: dict[str, Any] = {}
     if req.tools:
         kwargs["tools"] = req.tools
@@ -380,6 +430,8 @@ async def _handle_stream_tools(
     model: str,
     req: ChatCompletionRequest,
     complexity_info=None,
+    *,
+    app_config=None,
 ):
     """Stream a raw OpenAI-compat function-calling response via SSE.
 
@@ -397,6 +449,7 @@ async def _handle_stream_tools(
     from openjarvis.server.cloud_router import is_cloud_model
 
     messages = _to_messages(req.messages)
+    messages = _ensure_identity_prompt(messages, app_config)
     chunk_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
     use_cloud = is_cloud_model(model)
 
@@ -491,6 +544,7 @@ async def _handle_stream(
     complexity_info=None,
     *,
     trace_store=None,
+    app_config=None,
 ):
     """Stream response using SSE format.
 
@@ -509,6 +563,7 @@ async def _handle_stream(
     )
 
     messages = _to_messages(req.messages)
+    messages = _ensure_identity_prompt(messages, app_config)
     chunk_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
 
     # Last user message — recorded as the trace query.
@@ -767,10 +822,9 @@ async def reload_cloud_engine(request: Request):
     key so that cloud models become available without a full app restart.
     """
     import os
-    from pathlib import Path
 
     # Re-read ~/.openjarvis/cloud-keys.env and update the running process env.
-    keys_path = Path.home() / ".openjarvis" / "cloud-keys.env"
+    keys_path = get_config_dir() / "cloud-keys.env"
     if keys_path.exists():
         for raw_line in keys_path.read_text().splitlines():
             line = raw_line.strip()

src/openjarvis/server/session_store.py

@@ -8,6 +8,8 @@ import sqlite3
 from pathlib import Path
 from typing import Any, Dict, List, Optional
 
+from openjarvis.core.paths import get_config_dir
+
 logger = logging.getLogger(__name__)
 
 _MAX_HISTORY_TURNS = 20
@@ -22,7 +24,7 @@ class SessionStore:
 
     def __init__(self, db_path: str = "") -> None:
         if not db_path:
-            db_path = str(Path.home() / ".openjarvis" / "sessions.db")
+            db_path = str(get_config_dir() / "sessions.db")
         from openjarvis.security.file_utils import secure_create
 
         secure_create(Path(db_path))

src/openjarvis/skills/importer.py

@@ -23,6 +23,7 @@ from typing import List
 
 import yaml
 
+from openjarvis.core.paths import get_config_dir
 from openjarvis.skills.parser import SkillParser
 from openjarvis.skills.sources.base import ResolvedSkill
 from openjarvis.skills.tool_translator import ToolTranslator
@@ -56,7 +57,7 @@ class SkillImporter:
         self._parser = parser
         self._translator = tool_translator
         if target_root is None:
-            target_root = Path("~/.openjarvis/skills/").expanduser()
+            target_root = get_config_dir() / "skills"
         self._target_root = Path(target_root)
 
     def import_skill(

src/openjarvis/skills/manager.py

@@ -7,6 +7,7 @@ from pathlib import Path
 from typing import Any, Dict, List, Optional
 
 from openjarvis.core.events import EventBus
+from openjarvis.core.paths import get_config_dir
 from openjarvis.skills.dependency import validate_dependencies
 from openjarvis.skills.executor import SkillExecutor, SkillResult
 from openjarvis.skills.loader import discover_skills
@@ -54,7 +55,7 @@ class SkillManager:
             except Exception:
                 pass
             if overlay_dir is None:
-                overlay_dir = Path("~/.openjarvis/learning/skills/").expanduser()
+                overlay_dir = get_config_dir() / "learning" / "skills"
         self._overlay_dir = Path(overlay_dir).expanduser()
 
     # ------------------------------------------------------------------
@@ -262,7 +263,7 @@ class SkillManager:
         discovered = discovery.analyze_traces(traces)
 
         if output_dir is None:
-            output_dir = Path("~/.openjarvis/skills/discovered/").expanduser()
+            output_dir = get_config_dir() / "skills" / "discovered"
         output_dir = Path(output_dir).expanduser()
         output_dir.mkdir(parents=True, exist_ok=True)
 
@@ -375,7 +376,7 @@ class SkillManager:
         manifest's ``name`` field equals ``name``.
         """
         if roots is None:
-            roots = [Path("~/.openjarvis/skills/").expanduser(), Path("./skills")]
+            roots = [get_config_dir() / "skills", Path("./skills")]
 
         matches: List[Path] = []
         for root in roots:

src/openjarvis/skills/sources/hermes.py

@@ -14,6 +14,7 @@ from typing import List
 
 import yaml
 
+from openjarvis.core.paths import get_config_dir
 from openjarvis.skills.sources.base import ResolvedSkill, SourceResolver
 
 LOGGER = logging.getLogger(__name__)
@@ -28,7 +29,7 @@ class HermesResolver(SourceResolver):
 
     def __init__(self, cache_root: Path | None = None) -> None:
         if cache_root is None:
-            cache_root = Path("~/.openjarvis/skill-cache/hermes/").expanduser()
+            cache_root = get_config_dir() / "skill-cache" / "hermes"
         self._cache_root = Path(cache_root)
 
     def cache_dir(self) -> Path:

src/openjarvis/skills/sources/openclaw.py

@@ -15,6 +15,7 @@ from typing import List
 
 import yaml
 
+from openjarvis.core.paths import get_config_dir
 from openjarvis.skills.sources.base import ResolvedSkill, SourceResolver
 
 LOGGER = logging.getLogger(__name__)
@@ -29,7 +30,7 @@ class OpenClawResolver(SourceResolver):
 
     def __init__(self, cache_root: Path | None = None) -> None:
         if cache_root is None:
-            cache_root = Path("~/.openjarvis/skill-cache/openclaw/").expanduser()
+            cache_root = get_config_dir() / "skill-cache" / "openclaw"
         self._cache_root = Path(cache_root)
 
     def cache_dir(self) -> Path:

src/openjarvis/speech/faster_whisper.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import logging
 import tempfile
 from typing import List, Optional
 
@@ -13,6 +14,13 @@ try:
 except ImportError:
     WhisperModel = None  # type: ignore[assignment, misc]
 
+try:
+    import ctranslate2
+except ImportError:
+    ctranslate2 = None  # type: ignore[assignment]
+
+logger = logging.getLogger(__name__)
+
 
 @SpeechRegistry.register("faster-whisper")
 class FasterWhisperBackend(SpeechBackend):
@@ -30,20 +38,60 @@ class FasterWhisperBackend(SpeechBackend):
         self._device = device
         self._compute_type = compute_type
         self._model: Optional[WhisperModel] = None
+        self._last_error: Optional[str] = None
+
+    def _resolve_compute_type(self) -> str:
+        """Pick a CTranslate2 compute type supported by the configured device."""
+        if ctranslate2 is None:
+            return self._compute_type
+
+        try:
+            supported = set(ctranslate2.get_supported_compute_types(self._device))
+        except Exception as exc:
+            logger.debug(
+                "Could not inspect CTranslate2 compute types for %s: %s",
+                self._device,
+                exc,
+            )
+            return self._compute_type
+
+        if self._compute_type in supported:
+            return self._compute_type
+
+        preferences = (
+            ("int8", "float32", "int8_float32", "int16")
+            if self._compute_type == "float16"
+            else ("float32", "int8", "int8_float32", "int16")
+        )
+        fallback = next((value for value in preferences if value in supported), None)
+        if fallback is None:
+            return self._compute_type
+
+        logger.warning(
+            "CTranslate2 compute_type=%r is not supported on device=%r; "
+            "using %r instead",
+            self._compute_type,
+            self._device,
+            fallback,
+        )
+        return fallback
 
     def _ensure_model(self) -> WhisperModel:
         """Lazy-load the Whisper model on first use."""
         if self._model is None:
             if WhisperModel is None:
-                raise ImportError(
+                self._last_error = (
                     "faster-whisper is not installed. "
-                    "Install with: uv sync --extra speech"
+                    "Install with: uv sync --extra desktop"
                 )
+                raise ImportError(self._last_error)
+            compute_type = self._resolve_compute_type()
             self._model = WhisperModel(
                 self._model_size,
                 device=self._device,
-                compute_type=self._compute_type,
+                compute_type=compute_type,
             )
+        self._last_error = None
         return self._model
 
     def transcribe(
@@ -54,20 +102,24 @@ class FasterWhisperBackend(SpeechBackend):
         language: Optional[str] = None,
     ) -> TranscriptionResult:
         """Transcribe audio bytes using Faster-Whisper."""
-        model = self._ensure_model()
+        try:
+            model = self._ensure_model()
 
-        # Write audio to a temp file (faster-whisper needs a file path)
-        suffix = f".{format}" if not format.startswith(".") else format
-        with tempfile.NamedTemporaryFile(suffix=suffix, delete=True) as tmp:
-            tmp.write(audio)
-            tmp.flush()
+            # Write audio to a temp file (faster-whisper needs a file path)
+            suffix = f".{format}" if not format.startswith(".") else format
+            with tempfile.NamedTemporaryFile(suffix=suffix, delete=True) as tmp:
+                tmp.write(audio)
+                tmp.flush()
 
-            kwargs = {}
-            if language:
-                kwargs["language"] = language
+                kwargs = {}
+                if language:
+                    kwargs["language"] = language
 
-            segments_iter, info = model.transcribe(tmp.name, **kwargs)
-            segments_list = list(segments_iter)
+                segments_iter, info = model.transcribe(tmp.name, **kwargs)
+                segments_list = list(segments_iter)
+        except Exception as exc:
+            self._last_error = str(exc)
+            raise
 
         # Build result
         text = "".join(seg.text for seg in segments_list).strip()
@@ -81,6 +133,7 @@ class FasterWhisperBackend(SpeechBackend):
             for seg in segments_list
         ]
 
+        self._last_error = None
         return TranscriptionResult(
             text=text,
             language=getattr(info, "language", None),
@@ -91,9 +144,17 @@ class FasterWhisperBackend(SpeechBackend):
 
     def health(self) -> bool:
         """Check if model is loaded or loadable."""
-        if self._model is not None:
+        try:
+            self._ensure_model()
             return True
-        return WhisperModel is not None
+        except Exception as exc:
+            self._last_error = str(exc)
+            logger.debug("Faster-Whisper health check failed: %s", exc)
+            return False
+
+    def last_error(self) -> Optional[str]:
+        """Return the last model load or transcription error, if any."""
+        return self._last_error
 
     def supported_formats(self) -> List[str]:
         """Supported audio formats (same as ffmpeg/Whisper)."""

src/openjarvis/system/builder.py

@@ -7,6 +7,7 @@ from typing import Any, List, Optional
 
 from openjarvis.core.config import JarvisConfig, load_config
 from openjarvis.core.events import EventBus, get_event_bus
+from openjarvis.core.paths import get_config_dir
 from openjarvis.engine._stubs import InferenceEngine
 from openjarvis.system.core import JarvisSystem
 from openjarvis.tools._stubs import BaseTool, ToolExecutor
@@ -232,12 +233,10 @@ class SystemBuilder:
         agent_manager = None
         if config.agent_manager.enabled:
             try:
-                from pathlib import Path
-
                 from openjarvis.agents.manager import AgentManager
 
                 am_db = config.agent_manager.db_path or str(
-                    Path("~/.openjarvis/agents.db").expanduser()
+                    get_config_dir() / "agents.db"
                 )
                 agent_manager = AgentManager(db_path=am_db)
             except Exception as exc:

src/openjarvis/templates/agent_templates.py

@@ -11,6 +11,8 @@ try:
 except ModuleNotFoundError:
     import tomli as tomllib  # type: ignore[no-redef]
 
+from openjarvis.core.paths import get_config_dir
+
 
 @dataclass(slots=True)
 class AgentTemplate:
@@ -72,7 +74,7 @@ def _builtin_templates_dir() -> Path:
 
 def _user_templates_dir() -> Path:
     """Return the path to user-defined templates (~/.openjarvis/templates/agents/)."""
-    return Path.home() / ".openjarvis" / "templates" / "agents"
+    return get_config_dir() / "templates" / "agents"
 
 
 def discover_templates(

src/openjarvis/tools/approval_store.py

@@ -19,6 +19,8 @@ from datetime import datetime, timedelta, timezone
 from pathlib import Path
 from typing import Any, Dict, List, Optional
 
+from openjarvis.core.paths import get_config_dir
+
 # ---------------------------------------------------------------------------
 # Decision constants
 # ---------------------------------------------------------------------------
@@ -147,7 +149,7 @@ class ApprovalStore:
 
     def __init__(self, db_path: str = "") -> None:
         if not db_path:
-            db_path = str(Path.home() / ".openjarvis" / "approvals.db")
+            db_path = str(get_config_dir() / "approvals.db")
         self._db_path = db_path
         Path(db_path).parent.mkdir(parents=True, exist_ok=True)
         self._conn = sqlite3.connect(db_path, check_same_thread=False)

src/openjarvis/tools/description_loader.py

@@ -16,10 +16,10 @@ The TOML file format (written by the applier) is::
 from __future__ import annotations
 
 import logging
-import os
-from pathlib import Path
 from typing import Dict, Optional
 
+from openjarvis.core.paths import get_config_dir
+
 logger = logging.getLogger(__name__)
 
 _cache: Optional[Dict[str, str]] = None
@@ -27,7 +27,7 @@ _cache: Optional[Dict[str, str]] = None
 
 def _load_overrides() -> Dict[str, str]:
     """Parse descriptions.toml and return {tool_name: description}."""
-    home = Path(os.environ.get("OPENJARVIS_HOME", "~/.openjarvis")).expanduser()
+    home = get_config_dir()
     desc_path = home / "tools" / "descriptions.toml"
     if not desc_path.exists():
         return {}

src/openjarvis/tools/memory_manage.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 from pathlib import Path
 from typing import Any
 
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.registry import ToolRegistry
 from openjarvis.core.types import ToolResult
 from openjarvis.tools._stubs import BaseTool, ToolSpec
@@ -14,7 +15,9 @@ from openjarvis.tools._stubs import BaseTool, ToolSpec
 class MemoryManageTool(BaseTool):
     """Manage persistent agent memory (MEMORY.md)."""
 
-    def __init__(self, memory_path: Path | str = "~/.openjarvis/MEMORY.md") -> None:
+    def __init__(self, memory_path: Path | str | None = None) -> None:
+        if memory_path is None:
+            memory_path = get_config_dir() / "MEMORY.md"
         self._memory_path = Path(memory_path).expanduser()
 
     @property

src/openjarvis/tools/skill_manage.py

@@ -5,6 +5,7 @@ from __future__ import annotations
 from pathlib import Path
 from typing import Any, List
 
+from openjarvis.core.paths import get_config_dir
 from openjarvis.core.registry import ToolRegistry
 from openjarvis.core.types import ToolResult
 from openjarvis.tools._stubs import BaseTool, ToolSpec
@@ -14,7 +15,9 @@ from openjarvis.tools._stubs import BaseTool, ToolSpec
 class SkillManageTool(BaseTool):
     """Manage agent-authored procedural skills."""
 
-    def __init__(self, skills_dir: Path | str = "~/.openjarvis/skills/") -> None:
+    def __init__(self, skills_dir: Path | str | None = None) -> None:
+        if skills_dir is None:
+            skills_dir = get_config_dir() / "skills"
         self._skills_dir = Path(skills_dir).expanduser()
 
     @property