变更记录一览。

Bug Fixes

• concern reviewers share one prompt prefix through the diff, so the provider caches it

Bug Fixes

• a screen with no usable key waits for the key instead of failing after three tries

Bug Fixes

• a cooled key's recovery probe asks a live model, not its walled own model

Features

• VMs running developerz.ai give the agent passwordless sudo that works inside runs

Features

• separate cross-border consent for mainland-China users

Bug Fixes

• the commit-hook probe reads the filesystem, so a repaired checkout is not refused until restart
• the narrowed gate runs workspace tests through the root runner, so fleet boxes load the test preload

Features

• an error's issue is told which release shipped its fix

Bug Fixes

• a dated per-model wall hops to the key's next live model mid-run, and one park audits once

Bug Fixes

• onboarding writes.mcp.json in biome's format, and its own CI-fix task is not held behind onboarding

Bug Fixes

• a box drops a service it no longer reaches, and the platform's REDIS_URL reaches the gate

Features

• the /fleet Agents panel shows a drained box as Drained

Features

• rescue our own reviewer's holds with one findings round or one fresh re-review
• undrain a box from /fleet, and show a drained box as Drained

Bug Fixes

• no OpenAI client leaves its base to the ambient OPENAI_BASE_URL
• a graph-shaped typecheck reserves its measured peak, and a run over its reservation says so
• one sub's dated quota parks its model, not the whole router key
• a re-join retires its proven predecessor, /join installs a matured build, a box can self-revoke, every revoke records why
• a box drain survives the runner sweep; router calls carry a per-conversation x-session-id
• the test-count corpus caches per blob, so a new HEAD parses only what it changed
• owed terminals survive a restart, the commit-hook probe reads isolated installs, a silent box is re-pushed, runner logs expire
• a provider-refused output ceiling is persisted on the key's model row
• a scout files a live-state question as unobservable, so a checkup survey settles instead of running for hours

Features

• re-file a held fix round past a task-owned hold, rescue an unrepeatable round, re-mint a killed rescue once
• refuse manual tier/seats on Stripe-billed accounts
• the admin org page shows tier and seats read-only on a Stripe-billed account
• /pricing says what your bill is made of, each card says its facts once, and a seat is four numbers
• /security, /enterprise and /how-it-works link each claim to the page that answers it

Bug Fixes

• LinkGrid drops the browser bullets, indent and full-ink bodies on every page that uses it
• a late session boot never rewrites an existing PR or issue row

Features

• a deployed fix is confirmed on its error's issue after N quiet hours

Bug Fixes

• read roll-drift target from ArgoCD desired images

Bug Fixes

• a block of only dead editors ends with the editors' own failure kind, and the silent-failure alert stops blaming the model

Features

• an error that returns after its fix files generation N+1 and reopens the original issue

Bug Fixes

• the test pool sizes to the box's free CPU, not its core count

Features

• a View queue button on each agent card lists the PRs and tasks that agent holds
• open gitee connects, release n+1 of the Gitee read path
• cache-prefix stability, loop recovery and prompt fixes across the in-cluster loop, the box harness and fan-out
• a language setting that overrides the browser language
• any coding park of an AI-managed PR idle 12h+ goes to the Escalation agent, and an unplaceable round reaches a human
• the Escalation queue lists upcoming hand-overs with a countdown
• detect audit kinds defined but with zero durable rows
• a Sentry error delivery stores its exception class, crash-site frame and group id as data
• PATCH /v1/runners/:id renames a box, with the rename and its runner.renamed row in one transaction
• audit-silent-kinds lists the audit kinds the code writes that have no row since the archive frontier, with their writers
• a nullable column can be a keyset sort key, nulls last in both directions, and the Fleet board sorts by last seen again
• a person reopening an issue files a new generation of its ask
• one disclosed GitHub issue per error group, capped at 5 per repo per day
• Stuck PRs lists every AI-managed PR idle 12h+ or open 72h+ with its park, blocker and rescue state
• onboarding classifies a stale dz/onboarding collision as its own reason
• onboarding PR body lists names the gate never proved it needs
• forge-qualified stored ids, expand then contract
• hold a Sentry error's tasks until its public issue is filed, then link it for Closes #n

Bug Fixes

• stale-lane closes are off unless the maintainer writes the window
• the agent queue's ISO timestamps are validated as ISO
• the warm-first gate opens on the first provider stream event, not the first content
• a rescue round no box ever started is not a spent rescue, and the box's decode refusal names the field
• an undecodable assignment is held for a runner release that can read it, and a live rescue round is not a spent rescue
• a large-diff review reads the whole diff, keeps finished work across a retry, and grades a key wall as the key's
• ops one-off to sweep disabled-repo pending-task orphans
• the onboarding lane's installation and repository rows land on the account's durable chain, not stdout
• an operator can re-run the backlog import for an already-enabled repo, and a re-run reaches past the asks already on file
• the typecheck hook scopes an edit in a clone to that clone, never a whole-graph tsc over the main checkout
• classify bun install --frozen-lockfile mismatch on dep-bot PRs as a tree fault, not an env fault
• an onboarding audit-append failure alerts under onboarding.run, not the dispatch lane
• a failing integration-test teardown fails the run instead of only logging
• typecheck hook resolves root from edited file via git
• relabel login footer to "Request access" → /waitlist
• a closed set the box decodes cannot be widened and written in one diff, and every prod psql read is bounded by a statement timeout
• a review on a subscription key runs its full wave past a busy plan, a refusal shrinks only the reserve, and the lead planner is no longer told an omitted file goes unreviewed
• post-edit hooks start in the project dir and lint a clone edit against that clone, never trusting the edited tree
• a box no longer sees the mothership's mux inbox on its fleet.task subscription
• a survey starved of a model slot ages past newer coding work instead of waiting forever
• biome post-edit hook resolves root from edited file or exits 0 with one line
• pin grace to runner stagger default
• a setup that calls sudo is graded tree with the fix, not unknown
• the schema-correction re-ask records its prompt audit, and constant system prompts are a checked exemption, paying four owed prompt-audit sinks
• stop 320px horizontal scroll on every code-window editorial split
• the scheduled-run alarm action loads again, Lighthouse starts marketing from its own directory, and action expressions may only read contexts an action can see
• drop the requested-page-count clause from the KB import system prompt
• a failed janitor act's error row carries both prompt audits
• the GitHub sign-in's state cookie is age-checked on open
• session replay drops the model's requested escalation channel
• the wip keep is the set match, not the count
• a failed setup's tail is the last error line, not the last N chars
• a FAIL → PASS transition on the same sha posts BOTH verdict comments
• a held fix round names its blocker; delete the writer-less held overlay
• schedules fire again, cross-account repo disable is fenced, gh capture never parses a cut page
• the pre-land gate narrows to the round's workspaces, and a task-cap OOM no longer burns the charged ladder
• a transferred repo cancels and audits the old owner's tasks in-tx; review gh is a required seam; pre-land scope includes autofix output
• pin the FAIL→PASS re-run transition
• KeyCard's provider lookups go through a Map, not a Record
• name the per-test scan cost on the two whole-tree corpus guards
• verify a self-hosted router key against its own gateway
• GPT reasoning effort over the Anthropic-wire router
• sessions, the first-time-contributor guard and the issue answer door read project membership by repo permission
• the merge floor, ask-before-acting, ask stamps and the boot screen grade a definite member project by repo permission

Features

• thread lead-planner call identity + prompt sections

Bug Fixes

• project create is advertised only when the App holds administration: write
• hold the review gate's own config for a human, and say only what is true in it

Features

• the public repo page shows the agent-authored numbers, also for consent-only repos
• one rescue round on the escalation model for a stuck AI-managed PR
• a GitLab source can live on jihulab.com as well as gitlab.com
• the PR lane judges comments, declines with a reason, and closes PRs stalled on their author
• an AI-managed PR's red check is taken over in a maintainer's order, and the PR says why it is held

Bug Fixes

• re-arm mark is no longer written for the run's own trigger and is cleared/renewed for the full run
• an answered ask is dropped from the hourly summary; recovery counts the bundles it skips
• the DPA names the real subprocessors and the privacy page states the retention the code enforces
• rotating a webhook source's signing secret keeps a GitLab read failure it cannot fix
• the stalled-PR sweep does not backdate a conflict, counts only a failed check as the author's, and states a true way back
• a stale lockfile on the pinned bun is a verdict on the tree, not a box fault
• a box can no longer forge the reply to a request sent to another box
• 401 test sees operator-credential cache from a prior file
• a blocked PR's hold sentence names its failing and missing checks before the approval
• a task whose PR merged after its run ended is settled to merged

Features

• the issue lane replies only when it moves the issue, closes what waits on a silent reporter, and can reopen and lock

Features

• add BottomSheet primitive + primitives demo
• every landed webhook-source health transition writes one audit row

Bug Fixes

• the escalations page names the three locked escalation categories, tied to the code by a guard

Bug Fixes

• a refused GitLab read marks its webhook source unhealthy until a read succeeds

Bug Fixes

• published contact addresses come from the configured inbox, not a bouncing literal

Bug Fixes

• pin survey-filing repo cap to RAW_TASK_REPOS_MAX

Bug Fixes

• Notifications page reads settings once

Features

• read a gitlab.com project's merge requests into the audit trail

Features

• own and dependency-bot PRs on auto-merge repos are AI-managed end to end