release: v1.0.1 (#357)

2026-08-14 08:52:06 +00:00 · 2026-05-18 20:19:37 -07:00
parent 5b847cd081
commit fea8d3e872
24 changed files with 1572 additions and 23 deletions
@@ -0,0 +1,105 @@
+# Desktop auto-update
+
+The OpenJarvis desktop app ships with [Tauri's updater
+plugin](https://v2.tauri.app/plugin/updater/), which checks for new
+versions on launch and every 30 minutes. When a newer signed build is
+available, the app prompts the user to download and install it.
+
+## How it works
+
+```
+on launch / every 30 min
+        │
+        ▼
+GET https://github.com/open-jarvis/OpenJarvis/releases/download/desktop-latest/latest.json
+        │
+        ▼
+Parse manifest: { "version": "X.Y.Z", "platforms": { ... } }
+        │
+        ▼
+If manifest.version > installed_version:
+   download signed .dmg / .deb / .msi from manifest.platforms[target].url
+   verify against the minisign pubkey baked into the app
+   prompt user to install
+```
+
+The frontend code lives in
+[`frontend/src/components/Desktop/UpdateChecker.tsx`](../frontend/src/components/Desktop/UpdateChecker.tsx);
+the Tauri wiring is in
+[`frontend/src-tauri/tauri.conf.json`](../frontend/src-tauri/tauri.conf.json)
+under `plugins.updater`.
+
+## How releases reach the update endpoint
+
+The `Desktop Build & Release` GitHub Action
+([`.github/workflows/desktop.yml`](../.github/workflows/desktop.yml))
+publishes signed binaries plus a `latest.json` manifest to the
+`desktop-latest` GitHub release on every push to `main`. The
+`tauri-action` step with `includeUpdaterJson: true` generates the
+manifest automatically.
+
+Two release streams exist:
+
+- **`desktop-latest`** (rolling pre-release): updated on every push to
+  `main`. This is the channel the desktop app currently polls. Users
+  on this channel get the most recent build the CI produced.
+- **`desktop-vX.Y.Z`** (tagged stable): created when someone pushes a
+  `desktop-v*` git tag. Has the same artifacts but is marked as a
+  proper release rather than a pre-release.
+
+The current updater endpoint points at the rolling `desktop-latest`
+stream so that bug fixes (especially security and telemetry-policy
+changes) reach users without waiting for a manual stable tag. A future
+release may introduce a stable channel that points at
+`desktop-v*` tags only.
+
+## Signing
+
+Binaries are signed by `tauri-action` using the minisign key pair
+referenced via these GitHub Actions secrets:
+
+| Secret | Purpose |
+|---|---|
+| `TAURI_SIGNING_PRIVATE_KEY` | Private key (PEM-formatted minisign) |
+| `TAURI_SIGNING_PRIVATE_KEY_PASSWORD` | Passphrase for the private key |
+
+The matching public key is baked into the app at
+`tauri.conf.json:plugins.updater.pubkey`. If you ever need to rotate
+the key, replace the public key in the JSON file *and* update both
+secrets atomically — mismatched keys cause every update download to
+fail signature verification with no recovery path other than a manual
+reinstall.
+
+## Disabling the updater locally
+
+For frontend development, set `VITE_OPENJARVIS_NO_UPDATER=1` in your
+shell before running `npm run tauri dev`. Vite injects any
+`VITE_`-prefixed env var into `import.meta.env`, and the
+`UpdateChecker.tsx` component honors it to skip the 30-minute poll.
+
+```bash
+export VITE_OPENJARVIS_NO_UPDATER=1
+npm run tauri dev
+```
+
+This is purely a dev escape hatch — it has no effect on production
+builds (where `import.meta.env.VITE_OPENJARVIS_NO_UPDATER` will be
+`undefined` unless you explicitly set it at build time).
+
+## Verifying a release manually
+
+```bash
+# Download the latest manifest and confirm it parses cleanly
+curl -fsSL https://github.com/open-jarvis/OpenJarvis/releases/download/desktop-latest/latest.json | jq .
+
+# Fields:
+#   version       — semver string, must match the tag (without leading "v")
+#   notes         — release notes string
+#   pub_date      — RFC3339 timestamp
+#   platforms     — map keyed by "<target>-<arch>" e.g. "darwin-aarch64"
+#                   each entry has { signature: "...", url: "..." }
+```
+
+A 404 on the manifest URL means the most recent desktop CI run
+didn't complete or didn't have signing secrets — check the
+`Desktop Build & Release` workflow logs.
@@ -0,0 +1,135 @@
+# ACE optimizer (Agentic Context Engineering)
+
+OpenJarvis supports [ACE](https://github.com/ace-agent/ace) as a third
+optimizer alongside DSPy and GEPA. Where DSPy bootstraps few-shot
+examples and GEPA evolves prompts via reflective mutation, **ACE
+evolves a textual *playbook*** — annotated natural-language strategies
+the agent reads at inference time. The playbook is updated by a
+Generator / Reflector / Curator triad of LLM calls.
+
+## When to pick ACE
+
+| Task shape | DSPy | GEPA | ACE |
+|---|---|---|---|
+| Single-turn QA with crisp metric | strong | strong | weaker |
+| Long-running agent that should accumulate guidance | weak | medium | strong |
+| Open-domain where strategies matter more than templates | weak | medium | strong |
+| When you want to *read* what the optimizer learned | medium | medium | strong |
+
+ACE's headline artifact is `final_playbook.txt` — a human-readable
+file like:
+
+```
+## STRATEGIES & INSIGHTS
+[str-00001] helpful=5 harmful=0 :: When the user asks for unit
+                                   conversion, prefer the exact
+                                   rational form before rounding.
+[str-00002] helpful=3 harmful=1 :: Cite a primary source before
+                                   stating a date claim.
+```
+
+If reading those strategies feels like the form of "what learning
+should produce" for your task, ACE is the right choice.
+
+## Setup
+
+ACE is **not on PyPI** as of OpenJarvis v1.0.1, and the upstream
+repository is structured as a research codebase (multiple top-level
+directories) rather than a Python package. There's no `learning-ace`
+extra for that reason. Install ACE manually instead:
+
+```bash
+# 1. Clone ACE somewhere outside your OpenJarvis checkout
+git clone https://github.com/ace-agent/ace.git ~/code/ace
+cd ~/code/ace
+curl -LsSf https://astral.sh/uv/install.sh | sh   # if you don't have uv
+uv sync
+
+# 2. Make ACE's src/ importable from your OpenJarvis venv
+echo "$HOME/code/ace/src" > \
+  "$(python -c 'import site; print(site.getsitepackages()[0])')/ace.pth"
+
+# 3. Set the API key for whichever provider ACE will call
+cp ~/code/ace/.env.example ~/code/ace/.env
+# Edit ~/code/ace/.env to set API_KEY for your chosen provider.
+
+# 4. Verify the import resolves from OpenJarvis's venv
+python -c "from openjarvis.learning.agents.ace_optimizer import HAS_ACE; print(HAS_ACE)"
+# True
+```
+
+If `HAS_ACE` prints `False`, the `.pth` file isn't being picked up —
+verify the path matches `site.getsitepackages()[0]` for the same
+Python interpreter you're using to run OpenJarvis.
+
+## Configuration
+
+ACE is configured under `[learning.agent.ace]` in your OpenJarvis
+config TOML:
+
+```toml
+[learning.agent]
+policy = "ace"
+
+[learning.agent.ace]
+# ACE's three roles. Empty = inherit from the intelligence primitive's
+# default cloud model.
+generator_model = "claude-opus-4-7"
+reflector_model = "claude-opus-4-7"
+curator_model = "claude-sonnet-4-6"
+
+api_provider = "openai"     # sambanova | together | openai | commonstack
+
+num_epochs = 1
+max_num_rounds = 3
+playbook_token_budget = 80000
+max_tokens = 4096
+
+task_name = "openjarvis"
+save_dir = ""               # default: ~/.openjarvis/learning/ace/<task>/
+
+min_traces = 20
+```
+
+## Running
+
+Once configured, the same orchestrator that runs DSPy / GEPA also runs
+ACE — pick it via the `policy` field above. To force a one-shot run:
+
+```bash
+jarvis optimize agent --policy ace
+```
+
+ACE writes intermediate state and the final playbook to `save_dir`.
+The OpenJarvis runtime will pick up the playbook on next agent start
+(via the same sidecar overlay mechanism the Skills System uses).
+
+## Trace adapter behavior
+
+OpenJarvis traces are adapted into ACE's `train_samples` /
+`val_samples` / `test_samples` format via a 70 / 15 / 15 split
+(order-preserving for reproducibility). Each trace becomes a
+`{question: trace.query, ground_truth_answer: trace.result}` sample.
+Traces with empty `query` or `result` are dropped before splitting.
+
+The `DataProcessor` ACE expects is built from `_TraceDataProcessor`
+in `src/openjarvis/learning/agents/ace_optimizer.py` — it does a
+case-insensitive substring match for `answer_is_correct` and averages
+that for aggregate accuracy. If you're optimizing for a domain where
+substring matching is the wrong correctness signal (math problems,
+code, structured outputs), subclass `_TraceDataProcessor` and pass it
+through your own callsite to `ACEAgentOptimizer.optimize()`.
+
+## Limitations in v1.0.1
+
+- **No automatic install.** Document above is the only path.
+- **The trace adapter uses substring correctness.** Override for
+  domain-specific scoring.
+- **Single provider per run.** ACE assigns the same `api_provider` to
+  all three roles. To mix providers, run ACE outside OpenJarvis and
+  hand-deliver the resulting playbook into `save_dir`.
+
+These will get revisited once ACE publishes a PyPI package or stable
+provider interface — track
+[ace-agent/ace#issues](https://github.com/ace-agent/ace/issues) for
+upstream changes that would let us tighten the wrapper.
@@ -83,6 +83,33 @@ dropped. Tests covering the patterns: [`tests/analytics/test_redaction.py`](../t
 - **Never** sold, shared with advertisers, or used for anything other
  than improving OpenJarvis.

+## Opting out
+
+Three independent ways to disable analytics — any one is sufficient:
+
+1. **Set an env var** (no config file edit needed):
+   ```bash
+   export DO_NOT_TRACK=1            # W3C convention, honored by other tools too
+   # or
+   export OPENJARVIS_NO_ANALYTICS=1 # project-specific, leaves other DNT-aware tools unaffected
+   ```
+   Both are checked at runtime; any truthy value (`1`, `true`, `yes`,
+   `on`) disables analytics for that process. Truthy = anything other
+   than empty, `0`, `false`, `no`, `off`.
+
+2. **Edit `~/.openjarvis/config.toml`**:
+   ```toml
+   [analytics]
+   enabled = false
+   ```
+
+3. **Delete the anon ID** (`rm ~/.openjarvis/anon_id`) — events for
+   the prior identity are orphaned, but a new identity will be
+   created on the next run. Combine with #1 or #2 to fully stop.
+
+Env-var opt-out takes precedence over the config file, so setting
+`DO_NOT_TRACK=1` overrides `enabled = true` in the config.
+
 ## Retention

 - Default retention: **365 days**, then events are deleted by PostHog