v0.2.1

Added

• cli — add the batch command — one backend launch, many tasks

Fixed

• cli/batch — align catch bindings and remove underscore params in tests
• json — never repair end-of-stream truncation — give up instead
• json — repair an unclosed array at the object close, string-aware

Internal

• hooks — allowlist dl.google.com for chrome-builtin provisioning
• ci — neutral sample text — the verify prompt tripped the publish-doctrine scanner
• ci — make the verify job the lean consumer reference shape
• ci — cache only the model components + activation state, not profile litter
• ci — add the on-device model cache fill + offline-verify workflow

v0.2.0

Added

• cli — wire the dep-update reasoning family into the CLI
• llama-server — accept portless *.localhost loopback URLs
• odai — constrained decoding + generate-verify guardrails
• odai — best-of-N self-consistency + grounded determinism for decision tasks
• odai — code-first decision split — model extracts, code decides

Fixed

• release — baseline the three unprovenanced odai npm versions
• soak — drop the unpublishable bare stuie exclude
• cli — drop the phantom type parameter from parseJsonInput
• chrome-builtin — call clone and destroy through the session object

Internal

• fleet — resync -stable aliases + cascade and lockfile reconcile
• deps — align nock to the catalog pin + absorb sdk 4.1.3

v0.1.0

Added

• odai CLI bin with single-shot subcommands — summarize, commit-msg, triage, patch, plus a backends availability probe. Parsed JSON on stdout, diagnostics on stderr, a hard per-prompt budget via --timeout/ODAI_TIMEOUT_MS, and a stable exit-code contract: 0 success, 1 model failure, 2 usage error, 69 no backend available — the clean-skip signal for CI steps, printed with exact provisioning instructions.
• summarizeText, suggestCommitMessage, and triageAlerts tasks alongside the existing patch, dedupe, and lockfile tasks, exported from both entry points.
• Backend registry with createOdaiModel and selectBackend: backends simulator, gemini-nano-headless, llama-server, and apple-fm (declared, unavailable); selection precedence is explicit option, then ODAI_BACKEND env var, then availability probe order.
• llama-server backend adapter for any OpenAI-compatible /v1/chat/completions endpoint: live GET /health availability probe, ODAI_LLAMA_URL (default http://127.0.0.1:8080) and ODAI_LLAMA_MODEL config, SSE streaming, and prefill emulation through the existing JSON repair path.
• gemini-nano-headless backend bridge: in Node it launches real Google Chrome with --headless=new via playwright-core (optional peer dependency) and page-proxies the LanguageModel global — create, prompt, streaming, clone, destroy — through page.evaluate. System-Chrome mode clones the machine's already-downloaded Nano component into a odai-owned profile with copy-on-write (zero weights download, the live Chrome profile is never written; ODAI_CHROME overrides the executable). CI mode downloads the component once into a cacheable profile when ODAI_NANO_ALLOW_DOWNLOAD=1; ODAI_NANO_USER_DATA_DIR pins the profile for actions/cache. Chromium builds cannot run Nano — only real Chrome works.
• bench --backend=<name> scores any registry backend through the seam; real-model runs put Gemini Nano at 28.6% stable across reruns, 42.9% at best, on the 7-scenario battery.
• Bench reports now include per-scenario prompt latency in milliseconds, recorded by runEval and printed by formatReport.
• apple-fm backend adapter for Apple Foundation Models on macOS 26+: a Swift stdio shim compiled from embedded source with xcrun swiftc on first use and cached in node_modules/.cache/odai/, honest availability reporting the framework's own reason — deviceNotEligible, appleIntelligenceNotEnabled, modelNotReady — and a ODAI_APPLE_FM_SHIM override for prebuilt shims and tests. Single-shot prompting through the existing JSON repair path; streaming yields one chunk in v1.
• Initial fleet onboarding scaffolding.

Changed

• Renamed the package from @socketsecurity/gemini-nano to @socketsecurity/odai; the gnh evaluation harness is now bench — the ./gnh export and pnpm run gnh script are now ./bench and pnpm run bench.

Fixed

• apple-fm — move the shim cache out of node_modules to the repo-root store

v0.0.1

Release 0.0.1.