AgentChef 1.7.10 - readable team-run output

1.7.10 - 2026-10-16

• Every team-run command printed [object Object] without --json: output() stringified the result object. The plain form is now a short human summary: ok / not ok, the run line (id, status, repo, base), one line per worker (task, role, status, verdict), supervisor liveness and heartbeat age, scope warnings, checks, and any remaining fields. --json is unchanged. Regression test runs status without --json against a run record.

Full history: CHANGELOG.md. Update: npm run chef -- --update --apply.

AgentChef 1.7.9 - integrate --recheck --covers

1.7.9 - 2026-10-15

• team-run integrate --recheck --covers TASK-x records that the conductor completed a failed worker's scope in the integration branch (seen live in R20261015-4368: a root_cause_debugger worker only diagnosed and produced no changes, so the run could never reach verification). Accepted only for a failed, timed-out or rejected worker, only when the repair commits since the recorded integration touch that worker's write scope, and refused when they touch anything outside the plan's write scopes. The integration record and history carry coveredByConductor; independent verification and the user's merge approval are unchanged.
• The team-conductor skill tells the conductor to ask implementing roles to change code, not only diagnose, and how to recover when one does not.

Full history: CHANGELOG.md. Update: npm run chef -- --update --apply.

AgentChef 1.7.8 - integrate --recheck and team-run regression tests

1.7.8 - 2026-10-14

• team-run integrate --recheck reruns the run's checks on the integration branch after the conductor commits a repair of a failed integration. It refuses a passing integration, one with merge conflicts, or a worktree with uncommitted changes; it rescans the added diff and records integration_rechecked. Before this, a repaired run could not reach verification.
• Regression tests from team run R20261014-d760 (two Codex workers, conductor repair, independent verification: pass): supervisor-level memory guard (one worker always starts under a huge reserve and memory_hold is recorded; AGENTCHEF_TEAM_MEMORY_GUARD=off suppresses it) and usage-limit parser edge cases (ISO without zone, past ISO plus relative hint, far-future ISO, latest reset wins, missing finishedAt).

Full history: CHANGELOG.md. Update: npm run chef -- --update --apply.

AgentChef 1.7.7 - resume --wait and memory guard

1.7.7 - 2026-10-14

• team-run resume --wait sleeps until the usage limit resets, then resumes. The reset time is read from the paused workers' limit messages (clock times, relative hours/minutes, ISO timestamps; the latest wins, plus a 2-minute margin), falling back to 30 minutes and capped at 12 hours. status shows it as resumeAt for paused runs. New pure module scripts/lib/usage-limit-reset.mjs.
• Memory guard: the supervisor holds queued workers while free RAM is short (about 1.5 GB per worker plus a 2 GB reserve). One worker always starts when none is running, so a low-memory machine slows the run instead of deadlocking it. The run history records memory_hold once per hold. AGENTCHEF_TEAM_MIN_FREE_MB overrides the reserve and AGENTCHEF_TEAM_MEMORY_GUARD=off disables the guard. New pure module scripts/lib/memory-guard.mjs.
• Fix (from 1.7.6): resuming a paused run could fail with "The run changed while resuming" when its exiting supervisor wrote a last heartbeat; the heartbeat comparison now applies only to crash recovery.

Full history: CHANGELOG.md. Update: npm run chef -- --update --apply.

AgentChef 1.7.6 - crash recovery for team runs

1.7.6 - 2026-10-14

• team-run resume recovers a run left running after a sleep, crash or restart. When the supervisor process is gone and its heartbeat is older than 60 seconds, orphaned workers (queued, spawning, running or paused) return to the queue and continue their saved Codex thread in the same worktree; a worker without a saved thread starts fresh there. A live worker process blocks recovery, the plan hash is still checked, and the run history records recovered. Worker transitions spawning -> queued and running -> queued exist only for this path.

Full history: CHANGELOG.md. Update: npm run chef -- --update --apply.

AgentChef 1.7.5 - deterministic fan-out rule

v1.7.5 - 2026-10-14

AgentChef 1.7.5 makes the team conductor's split decision predictable: one worker per independent write scope, parallel workers only for disjoint and independent scopes, dependent work in a later run, never more than four.

What Changed

• Fan-out rule. Written into the team-conductor skill and the skills guide, so the same request splits the same way every time.

This page follows the release users should install now. Older engineering history remains available in CHANGELOG.md and CHANGELOG-0.5.md, so the public release guide stays useful instead of becoming an ever-growing archive.

AgentChef 1.7.4 - verifier receives host results

v1.7.4 - 2026-10-14

AgentChef 1.7.4 makes independent verification of team runs reliable. The read-only verifier cannot create temp directories in its sandbox, so it now receives the host's integration results with the diff and judges from those instead of trying to re-run test suites.

What Changed

• Verifier prompt. Host integration results (pass or fail per check, a short sanitized tail for failures) are part of the verification brief, with an explicit rule not to re-run writing suites.

This page follows the release users should install now. Older engineering history remains available in CHANGELOG.md and CHANGELOG-0.5.md, so the public release guide stays useful instead of becoming an ever-growing archive.

AgentChef 1.7.3 - resume after a usage limit

v1.7.3 - 2026-10-14

AgentChef 1.7.3 lets a team run continue after a Codex usage limit. When the limit resets, resume picks up the paused workers on their saved Codex threads in the same worktrees, with the plan hash checked so the approved plan cannot change.

What Changed

• Resume. team-run resume --run <runId> (or beyin takim resume --run <runId>) requeues only the workers a usage limit paused and continues each on its saved thread.
• Docs. The conductor skill and the operator guides explain when to resume and when to collect instead.

This page follows the release users should install now. Older engineering history remains available in CHANGELOG.md and CHANGELOG-0.5.md, so the public release guide stays useful instead of becoming an ever-growing archive.

AgentChef 1.7.2 - team mode hardening

v1.7.2 - 2026-10-14

AgentChef 1.7.2 closes the findings of an independent security audit and code review of team mode. Workers no longer see the conductor's secrets, their handoff text cannot inject approval lines, a plan cannot change after you approve it, and collection can no longer be bypassed or wedged.

What Changed

• Worker isolation. Secret-named environment variables stay with the conductor; handoff text is sanitized and capped.
• Plan integrity. start refuses briefs that changed after plan; brief files must live in the plan directory.
• Collection and supervision. Worker commits and dirty files are both scope- and secret-checked; a stuck abort, a board rejection or a re-run no longer wedges a run.
• Board. The review gate reads only the leading Verdict and Mutation check; Mutation check needs a reason or paths; glob conflicts and report paths are stricter.

This page follows the release users should install now. Older engineering history remains available in CHANGELOG.md and CHANGELOG-0.5.md, so the public release guide stays useful instead of becoming an ever-growing archive.

AgentChef 1.3.4

AgentChef 1.3.4 uses a skill or a role when a request needs it, by rule and by
a one-line routing hint, instead of by guesswork. Update with
npm run chef -- --update --apply, then trust hooks/routing-hint.json in
Codex's /hooks and start a new Claude Code session.

What Changed

• Routing catalog. Each of the 19 profiles (new: code-review) names its
  verifier, whether that verifier is required after file changes (security,
  release, MCP, frontend, data), and the skill to load first; explicit-only
  skills are suggested, not loaded. At most two agents start per task without
  you naming them. Turkish requests route: the scorer folds the dotless ı.
• Routing hint. On every prompt the plugin scores your request against the
  catalog in memory and, on a high-confidence match, adds one line of catalog
  identifiers for the model. It keeps no prompt text, never blocks a prompt,
  and can be turned off with AGENTCHEF_ROUTING_HINT=off. See
  security model and PRIVACY.
• Roles. Every role has one trigger-style description ("Use proactively
  when ...") and preloads the skill its work needs. npm run chef -- --routing --task "<request>" shows the matched profile and the hint line.