Skip to content

AI Chat v1.2.0

Latest
Latest

Choose a tag to compare

View all tags
@robertdevore robertdevore released this 07 Sep 10:44
v1.2.0
18e8667

AI Chat v1.2.0 adds durable execution recovery, bounded agent workflows, and documentation-aware chat.

  • Add encrypted execution journals, durable action receipts, retention tombstones, and explicit review/reconciliation/resume for HTTP and persistent native Codex sessions.

  • Drain active model work on shutdown, recover interrupted turns, and replay disconnected streams through bounded SSE buffers and durable cursors.

  • Budget complete provider context, including reasoning, tool schemas, arguments and receipts, using explicit or dated model limits.

  • Preserve user-authored chat constraints and decisions outside transcript compaction.

  • Add static page extraction with Markdown negotiation and shared URL/DNS policy, plus bounded alternate search backends with provenance.

  • Enforce browser network containment on macOS and Linux; unsupported or unavailable containment fails closed.

  • Route agents to relevant AI Chat manuals and discover deferred tools within request-scoped authorization.

  • Add configured RAG documentation lookup, source links, query timeouts, saved code-example language preferences and an optional composer language picker.

  • Unify auxiliary request and benchmark-queue cancellation; detect occupied ports before claiming storage.

  • Patch the qs dependency and harden local writes, external reads, tool dispatch, and Ollama documentation request limits.

  • Add real-family evaluation and reliability metrics. The eight-hour soak is not complete: its partial 2h14m run recorded 90 attempts, 70 correct results, 16 EAGAIN metric failures, and no recorded SSE overflow. Its termination cause is unknown; model-quality failures and unavailable monetary costs remain documented.

  • Add repeatable --model benchmark lanes with safe provider-profile resolution,
    change the no-flag default to Benchmarks 082626, and preserve saved
    pane-profile runs with explicit selection metadata and exact-lane resumption.

  • Add a managed loopback Hermes / Nous Portal profile seeded with the current zero-cost model catalog.

  • Add a separate managed xAI Grok OAuth profile for X subscription-backed model access, with Hermes used only as the local OAuth bridge.

  • Add explicit per-automation runtime tool selection and reject provider-generated textual tool-call envelopes as terminal protocol errors instead of saving them as successful answers.

  • Prevent tool-free benchmark runs from inheriting saved interactive runtime
    tool presets, keeping --tool-preset none deterministic and bounded.

  • Startup now repairs a missing better-sqlite3 native binding as well as an ABI-mismatched binding, so a clean install or pruned native build can recover automatically on the supported Node runtime.

Upgrade

Back up the database and matching encryption configuration before upgrading. Use Node 22.17.0 and run npm ci. Browser execution requires Playwright Chromium plus macOS Seatbelt or Linux Bubblewrap with permitted namespaces. Restart to load the fixed system prompt. See upgrade instructions.

Verification

  • Local suite: 387 passed, one Linux-only skip.

  • Authenticated offline-fixture smoke and browser-availability smoke passed against isolated local storage.

  • Production dependency audit: zero vulnerabilities.

  • The full eight-hour soak remains incomplete; this is not a deployed production-readiness certification. No production instance was restarted for this release.

  • Release commit passed Linux CI run 34112865139 and artifact guard 34112865149.

Assets 2
Loading