2.35.10

• f8922ed: Make /claude, /codex and /agent with --model formal-ai reliable end to end:

  • The Telegram bot keeps answering every chat while a command prepares its task (for example while a Formal AI image is pulled). Slow command work (the launch commands, /models, /auth, /log, /stop, /terminal_watch, /limits, /version) runs outside the update loop, logs are written asynchronously, gh calls have a timeout (HIVE_MIND_GITHUB_EXEC_TIMEOUT_MS), and lint rules keep blocking calls out of the bot process.
  • A run that changed nothing finishes with ⚠️ "Finished without changes" and exit code 5 instead of 🎉 (the queue and Docker monitor do not count it as a failure; HIVE_MIND_SOLVE_NO_CHANGES_EXIT_CODE=0 keeps exit 0), keeps a draft PR in draft, and reports a session that looped on one command. An echoed tool output is no longer posted as the working-session summary.
  • Telegram alias session recovery executes the real solve command, keeps resume hints, and reports, cleans up and retries snapshot recoveries that fail to start, on the issue when there is no pull request.
  • Docker sessions are only reported as finished after their container was observed (or the launch failed), and an unknown exit code is shown as "status unknown" instead of success. start-screen sessions record their exit status.
  • Formal AI images are resolved local-first, an accepted fallback image is reused instead of re-pulled, a timed-out pull keeps running detached instead of leaving a containerd lease, and sidecar leases survive while their session or its launcher is alive.
  • Codex runs against Formal AI keep a pinned CODEX_HOME and config across restarts and classify Formal AI failures.
  • Formal AI has the same two spellings on every tool (Codex no longer accepts openai/formal-ai, which ran against OpenAI), and hive validates it like solve.
  • A submitted PR review counts as progress, unreadable comments are not "no comment", and a Codex usage limit reported in a stream error is detected.
  • Recoverable agent tool errors no longer abort the run.
  • The DinD image clears stale dockerd/containerd pid files before starting.
  • Formal AI draft runs report their checks on the draft head, keep tool state between iterations, and can be called from other repositories.
  • npm run lint fails on any warning.
  • Pinned versions: @link-assistant/agent 0.26.15, start-command 0.37.8, Node.js 26.12.0, formal-ai 0.353.2, @dotenvx/dotenvx 2.35.0.

Related Pull Request: #3225

---