变更记录一览。

Bug Fixes

• every follow-up on a repo with secrets ran against a tree rebased off the PR head

Features

• the reviewer publishes a check run, so a running review is visible

Features

• wire the resource pass into the fleet-alerts tick

Features

• backlog, no-work, and idle alert tick
• config-drift alert + daily fleet digest

Features

• model/API failure classification alerts
• resource alerts + sustained-condition state

Features

• stuck task alert emissions — warden observation + reaper repair

Features

• task started/succeeded/failed alerts on the fleet.alert lane

Features

• the fleet.alert Discord lane + box offline/recovery alerts

Features

• add per-task hold writer for capability-escalation sweep
• delete DZ_TEAM_FANOUT and the unrun branch it forked
• a repo is not AI-first until something proved it

Features

• a prompt ceiling nobody measured is not a ceiling

Bug Fixes

• remove dead GateRefusedError escalation branch
• add missing escalation categories to OPERATIONAL enum

Features

• the signature proved the source, never the tenant

Bug Fixes

• an editor that exited before reading its brief killed the runner

Features

• the review disclosure is a footer, not a header
• invitations, and a tool surface measured before it was redesigned

Bug Fixes

• a Go template miss looked exactly like a missing image
• a cross-repo plan stamped one repo over its whole DAG

Features

• the devops role, retrievable memory, and why nothing is moving

Bug Fixes

• the digest resolver refused six images it could not explain

Bug Fixes

• six releases shipped with no attestation because a 250-byte file could not be uploaded

Features

• a deferred-to reviewer that cannot review left the PR unreviewed
• the bot's PR never said how the work reached the fleet
• three mechanics that notice, plus the groomer bug that retired its own work

Bug Fixes

• a snapshot lost what no migration dropped, and the next generate re-created it

Features

• a scout plan could never run at concurrency above one
• one tool call per round trip was the agent's speed ceiling
• a hand-maintained exclusion list put the box's own scratch in customer PRs

Performance

• the PR lifecycle asked GitHub what changed 3,754 times a day to learn about 381 changes

Features

• the raw task — one object, one admission door

Bug Fixes

• we published a "no egress" claim into customer repos that was untrue of the config we shipped
• a sub-agent's output existed only in memory, so a failed parse was a lost run

Bug Fixes

• the turn cap was a work budget, so it failed work

Bug Fixes

• a signed attestation named zero images while six of them shipped

Bug Fixes

• the re-entrant checkout lock leaked its slot on out-of-order dispose

Bug Fixes

• a report upload decided whether the code was good
• one cancelled follow-up parked a finished PR forever

Bug Fixes

• a babysit run queued behind itself and held the box for an hour

Features

• a refusing box reports its own refusal
• land box boot-refusal reports as account-anchored audit rows
• ship the repo's knowledge digest on the assignment, not the template
• code.recall — the coding lane can read a KB page, not just its index
• the area gate unlocks on proof that two file sets are disjoint
• the built-in harness is the default; spawning a CLI is the opt-in
• declare each PR's own files, so disjoint reviews stop queueing behind each other
• the agent finally reads the house rules it has been writing for the maintainer

Bug Fixes

• attribute a boot refusal to the box, not the release
• acknowledge DZ_TEAM_FANOUT as deliberate soak flag
• a failed dependency no longer deadlocks its dependents forever
• the KB gardener could not keep up with its own inflow
• monitor mode says out loud that it will never be offered work
• widen the boot() contract to match what monitor uses it for
• the KB digest stops naming tools the reader does not have
• a heartbeat that reports silence is no longer counted as progress
• refuse the assignment rather than spawn a coding CLI at the runner's cwd
• ui-debugger reached only repos that already had UI tests
• every box reported itself idle — derive the current task, drop the column
• a merged PR never retired its lane tasks, so every merge leaked a queue row
• an 8k output ceiling threw away seven minutes of finished review work
• the PR projection drifts from GitHub and nothing repaired it
• code.read handed the model 60,000 replacement characters for a PNG
• a dead inference connection is no longer indistinguishable from a thinking model
• six defects found by pointing the QA harness at a live customer account
• the stall guard was timing the model, and killing runs for thinking
• tenancy, credentials, and the boxes nothing could read
• the coding lane staged the harness's own scratch into customer commits

Bug Fixes

• wrap unguarded audit appends so a failed append doesn't drop the pass

Bug Fixes

• two deliveries of one error minted two tasks, two boxes, two PRs

Bug Fixes

• a scout plan opted out of the same-area gate while a box held the area