v4.495.2Bug Fixes one lane digest per account, a priority email budget, per-pass hold writers, babysit trees survive provider failures a router key's catalogue states its models' effort levels; an inherited default row walks its stored ladder
v4.495.1Bug Fixes an arbiter halt on an unaudited write fails the run, and a review spec names one side lane
v4.495.0Features the arbiter pass rules final on a decline the re-check held twice every key bypass, hold and diagnostic honours the per-key agent switch
v4.494.0Features every BYOK key carries an agent-switch deny-list, read by one rule in two spellings the arbiter's readers - the ruling marker, held re-check count, and a re-check that honours a ruling medium is the platform default reasoning effort; a probe records the level it proved the GitLab MR hint mints the head's platform review under the auto lane's gates a key captured without a priority ranks last in its pool every ranked key pick carries its agent, and the ranker skips a key that switches it off every metric card explains itself on hover, in plain English Bug Fixes the Stuck PRs read model stops advertising rescues no path mints the review mint's supersede, collision and revive are forge-fenced; the gated mint takes a forge code.edit lands a quote with re-typed typography and drifted indentation, and names the first differing line the BYOK verify probe honours a router throttle's dated reset instead of re-probing every 2 minutes exit 79 means "every check I ran passed, CI grades the rest" in every check-reading lane; a red save keeps the gate's report a box model call waits out a provider outage (up to 120 s) instead of failing after one 2 s retry the per-key gate admits and paces streams too; a stream holds its slot until it ends, errors, is cancelled or aborted a review no key can serve concludes gray and skipped, not queued forever the home tour's screen cannot render flipped, on any width a stacked-PR dependent whose links back-fill misses leaves a durable row a reopen cut off by a declined close or a lock keeps its cap slot accept reasoning effort on the router's OpenCode Go flash models Performance the fleet page answers from memory while its stall report refreshes behind it the fleet stall cache stays answerable for 5 minutes, sized to the per-pod gap between polls code.read pages 500 lines by default; a whole-file write needs a whole read; read volume is logged box loops clear stale tool results in 20k-token batches below the compaction threshold the repo memory index rides in the user turn, so the review system prompt is byte-identical across runs; per-call cache counts are recorded
v4.493.0Bug Fixes show Fleet AI account identity and priority the follow-up land refuses a change it already pushed, compared as the tree its commit would carry
v4.492.0Features a GitLab review rides the outbox: put, claim, post, status and disconnect retire Bug Fixes expose model and priority in Fleet key responses
v4.491.2Bug Fixes order account AI keys deterministically share key titles and show local quota reset times
v4.491.0Features the review outbox names its forge and poster, and its path-keyed reads pin the forge Bug Fixes a repair answer with the wrong supersedes ids is re-asked once, not escalated
v4.490.1Bug Fixes a repair's deploy edge onto an existing leg is checked, the judge sees failure kinds; coverage skips.d.<ext>.ts declarations distinguish full-result progress from repeated calls a plan repair ask is a factual report, and a dropped ask no longer blocks its goal
v4.488.0Features the review forge's GitLab arm for the read side Bug Fixes issue readiness runs the grader's own fence screen, so a piped grep -q is flagged at filing live-sha passes an app the release did not rebuild when no change since its sha reaches it
v4.487.5Bug Fixes the follow-up land's push is leased on the run-start sha, so a rewound PR head is refused
v4.487.0Features digest-pin image bases, report-only secret and image scans, retention docs that match DOCR Bug Fixes re-ask a reasoning-required refusal that arrives in-stream (dz-runner 0.1.44) price a run off the key it named, and page the keys that cannot run
v4.484.1Bug Fixes reviews on MiniMax work - re-ask a no-thinking refusal, read the router pool verdict (dz-runner 0.1.43) Performance index four live hot reads - scout drain, default-branch push, Fleet Live pick, admin health
v4.484.0Bug Fixes a large test-heavy PR is reviewed with its tests, most-changed first (dz-runner 0.1.40)
v4.480.1Bug Fixes an agent mode switch keeps keyboard focus through its own write, on the card and the panel row
v4.480.0Features stable KB and skills indexes ahead of policy in the prompt prefix; DeepSeek direct reasoning passback