变更记录一览。

Features

• derive required capabilities at the CLAIM gate
• @developerz-ai review command on a PR comment enqueues a review
• escalation-delivery health vocabulary + unroutable-state alert
• provider-reconciliation cron for terminal-state keys

Bug Fixes

• alert when a published release has no ledger row
• stuck-ledger predicate reuses the claim's claimability gates

Features

• ensemble sampling — N parallel review samples with anchor-union merge
• grounding mounts + fan-out driver for the reviewer bench

Bug Fixes

• grade the autoroll's refusals, bound its read, audit its writes

Bug Fixes

• a box that cannot self-update says so, instead of clobbering Bun

Features

• the fleet updates itself when CI publishes a release

Bug Fixes

• a pr-outcome run proves its PR exists, or fails
• a dirty shared checkout recovers instead of wedging the box

Bug Fixes

• unquarantine donation-flow; oss-tier-gate needs an api CI has not got

Bug Fixes

• run every e2e lane, not just the ones that need nothing

Bug Fixes

• a non-integer stored generation destroyed the ring and reported success

Bug Fixes

• the join merge lost a line, kept stale credentials, and stranded secrets
• announce the stuck-ledger skip — an unregistered alarm was silent

Features

• fan the reviewer out over file-groups on a large diff

Bug Fixes

• untrack.dz-babysit — a tracked file the CI-fix lane deletes every run
• babysit merge gate refuses an all-tolerated check set
• make ready-for-agent a consequence of handoff delivery, and add a retract verb
• classify transport failures at the verify seam

Bug Fixes

• the join script merges the box env instead of replacing it

Bug Fixes

• the box env writers keep what they do not own

Features

• name the requester on every fleet PR body, not just plan slices

Bug Fixes

• answer 405 on GET /v1/mcp so streamable-HTTP clients can connect

Bug Fixes

• order the Agent fleet table by activity, live work first

Bug Fixes

• stop claiming a checkout and CodeGraph index the run may not have

Features

• let the bench driver measure kimi-k3 faithfully (temperature + reasoning suppression)

Bug Fixes

• fence the PR description into the reviewer prompt
• capability-escalation sweep is now placement-scope-aware
• drop self-refuting findings at the anti-noise seam
• report() could not read the manifest pin writes

Features

• add policy-validate CLI with --json (ok + errors)

Bug Fixes

• recognise Kimi's quota-exhaustion 403 as credits-exhausted

Features

• add the reviewer-bench model driver and the prompt A/B variant

Bug Fixes

• suppress reasoning on the review brain, not just the classifier

Bug Fixes

• a review that dies must say so on the PR
• CI state is never the maintainer bot's comment surface
• a CI completion is an observation, not a human directive
• give the z.ai tier column a real ladder

Bug Fixes

• pin OpenAI to /chat/completions like its ten siblings
• rewrite dotted tool names in HISTORY, not just the tool list
• fence review work away from boxes that cannot unwrap a sealed credential

Bug Fixes

• verify what inference does, and fix the 3 wires that lied

Bug Fixes

• MiniMax M-series could not complete a single structured call

Bug Fixes

• a valid key is refused, and the card never says why

Features

• POST /v1/plans, the scout entry's REST door
• reviewer-bench part 3 — matching, metrics, miss list, report
• a start control on /fleet/plans, dispatched-then-claimed
• copy-as-call, command palette, boundary fixes, mobile+theme+i18n

Bug Fixes

• register ASSIGNER_CLAIM_CONCURRENCY — refuse a bad value, never clamp it
• fence the Stripe usage drain to a metered-tier allowlist
• a review no model ran must say so, and may not approve

Features

• MCP agent ergonomics — the catalog tax and the pagination ceili…
• a watch path that does not eat the whole read budget