一本记录变更的账。

v4.174.2

Bug Fixes

• the merge question can be answered (#4390)
• a dead Devin citation, and two quota cells that ignore a free allowance (#4410)
• the frozen model is compared by identity, through one named table (#4412)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.174.1

Bug Fixes

• qualify five competitor claims stated as universals, and date the notes to a month (#4408)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.174.0

Features

• the run card shows the size class and the harness pairing (#4387)
• a plan records who started it, and the list can filter on it (#4349)
• register zh, and gate every locale instead of a hardcoded pair (#4407)

Bug Fixes

• an arm records the set of models it was served, and one that varied is refused (#4373)
• /proof says what the public record says, and every refusal it can see has a sentence (#4377)
• a setup pass parks the box's own dirt at task start, and a tree refusal retries (#4379)
• the transcript is a field, not a promise keyed on a record existing (#4380)
• the verdict panel reads the arm, not the pair (#4383)
• the served model carries its denominator (#4385)
• a BYOK key comes back by itself, or says why, and a blind review stops blaming the wrong thing (#4388)
• a locale that quotes CNY renders CNY, and the page says what it bills (#4350)
• one punctuation convention in zh.json, and a guard that keeps it (#4402)

Docs

• name both platform credentials, not just the OpenRouter key (#4351)
• the verifier tier is default-on since #4262, with its measured cost (#4363)
• enumerate every refusal the contract makes, and stop claiming enforcement for two fields that have none (#4371)

Tests

• hoist corpus-guard scans out of the timed test bodies (#4359)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.173.0

Features

• backfill the test-merge probes #3702 stored as merge commits (#4338)

Bug Fixes

• the reconcile cron stops probing revoked keys against the provider (#4352)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.172.2

Bug Fixes

• bin/check wins over DZ_TEAM_VERIFY_COMMAND (#4348)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.172.1

Bug Fixes

• write the lifecycle email.sent audit durably, anchored on the account (#4317)

Docs

• claude-code page — all four handoff.mode values, and convention support (#4326)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.172.0

Features

• abort a CI dispatch whose job is re-placed mid-run (#4315)

Bug Fixes

• a withheld auto_merge ask finally renders the informational template (#4325)
• retarget four onboarding file:line citations after the routes -> deps move (#4311)
• deferTriageJob rethrows DelayedError even when moveToDelayed faults (#4343)
• the session cap honours purchased seats, with the tier number as its floor (#4345)

Docs

• seats_checkout_configured mirrors BOTH 503 exits — say so at the wire field (#4323)

Other

• survey-pump: key sweep by UTC date so daily cron re-fires (#4093) (#4337)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.171.6

Bug Fixes

• repair 114 unparseable hold_reason rows and constrain the column to its vocabulary (#4281)
• the dz-env-vars ref guard tracks /* */ block comments across lines (#4324)
• a capacity-capped agent.run job waits for a slot instead of being dropped (#4329)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.171.5

Bug Fixes

• a parked ledger drains when the account's key is already usable, not only on a key transition (#4322)

Docs

• GLM-5.3-Flash placement — routing matrix over every LLM call site (#4318)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.171.4

Bug Fixes

• a key coming back re-drives all of an account's parked work, not the first 100 (#4320)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.171.3

Bug Fixes

• a key the provider still honours comes back by itself, and a keyless scout says so (#4319)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.171.2

Bug Fixes

• kill code.bash on silence and size the check budget absolutely (#4251)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.171.1

Bug Fixes

• mirror both 503 exits in seats_checkout_configured (#4314)
• give the disk-floor refusal its own UNCHARGED failure kind (#4309) (#4313)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.171.0

Features

• zh.json part 3 — proof and the long tail, 495 keys, still unregistered (#4308)

Bug Fixes

• gate gardener box-maintenance on ownership-verified repo row (#4287)
• gate the Buy seats card on Stripe price provisioning (#4052)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.170.3

Bug Fixes

• render ASK escalations with mergeAskWithheld: true as informational (#4299)
• take merge_commit_sha only from a merged payload, and alert when a merge produces no verdict (#4283)

Docs

• define target_class, the one run-card output the method named without defining (#4307)
• document both check runs on the checks:write row (#4303)

Refactors

• move the onboarding funnel use case from routes/ to deps/ (#4306)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.170.2

Bug Fixes

• reject comment-only refs in dz-env-vars Read-at guard (#4305)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.170.1

Bug Fixes

• a gate-refused merge discharges its expectation, and a transient refusal never stamps merge_decision='ask' (#4291)
• the harness is what this comparison varies, and an unpriced arm publishes no money delta (#4304)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.170.0

Features

• zh.json part 2 — the product pages, 703 keys, still unregistered (#4300)

Docs

• name both review check runs and the gate owner fence (#4302)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.169.5

Bug Fixes

• a DISMISSED review clears its author's prior verdict (#4218)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.169.4

Bug Fixes

• hold this repo's two generated artifacts out of its own review (#4296)
• zh.json catches up with three days of en copy (#4298)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.169.3

Bug Fixes

• /proof states the protocol is public and its rule selected nothing, and publishes all seven limits (#4295)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.169.2

Bug Fixes

• re-point 18 drifted file:line refs in dz-env-vars.md — main CI red since #4158 (#4292)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.169.1

Bug Fixes

• clear hold_reason on redispatchTask revive (#4285)

Docs

• name falsifying condition for deleted DependencyFailed hold (#4279)

Guard

• make comment-citations the sole source-docblock citation guard (#4181)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.169.0

Features

• audit boot_release claims reset, CI requeue and purge (#4083)

Bug Fixes

• include repo-less account-anchored sessions in weekly digest events (#4187)
• delete stalled_replay rows with lapsed 7-day leases during drain (#4189)
• add worker-side STDOUT_AUDIT_SITES registry with drift guard (#4150)
• redact the account-audit belt line before it hits stdout (#3954)

Tests

• pin that a plan stays open while a group is still unmerged (#4047)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.168.0

Features

• fold issues.closed/reopened onto issues.state on dispatch (#4184)

Docs

• make handoff.mode fleet the primary Claude Code handoff path (#4190)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.167.0

Features

• refuse cluster tooling in live fences at authoring (#4161)

Tests

• guard dz-env-vars.md file:line refs against drift (#4158)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.166.0

Features

• the repo-survey pump — daily cron so the survey writer keeps producing asks (#4282)

Bug Fixes

• carry no_credential/kek_unwrap_failed floor as blocked on context asks (#4270)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.165.9

Bug Fixes

• an issue-projection reconcile lane, mirroring the PR pair (#4280)

Other

• Widen MODEL_READ_CAP to noun-first caps and register the newly visible ones (#4275)

Tests

• permanent positive control for the post-deploy fence PASS path (#3383) (#4276)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.165.8

Bug Fixes

• fold scout_context into task result so return_context scouts answer (#4274)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.165.7

Bug Fixes

• pay the limit debt to zero — prompts sized from the model, no count refuses an answer (#4273)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.165.6

Bug Fixes

• derive page budget from the model window (#4271)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.165.5

Bug Fixes

• MCP answers the truth — blocked scouts, expiring blockers, a ready key, a fully reviewed PR (#4266)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.165.4

Bug Fixes

• grade pull-swapped boxes superseded, not confirm_deadline (#4043)

Api

• add sentry_environment to runner enroll and creds-reissue replies (#4044)

Docs

• state both listRepoBoard deltas in listAccountBoard docblock (#4045)

Tests

• serve hashed-css fixture from scratch dir via injected publicDir (#4265)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.165.3

Bug Fixes

• parked PRs escalate, human changes-requested gets a fix round, false Majors are judged (#4262)

Docs

• plan 101 site, docs and legal truth sync (#4261)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.165.2

Bug Fixes

• no agent bound cuts work on a huge repo — limits from the model, detected windows, compact-and-retry (#4258)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.165.1

Bug Fixes

• 1957-2035 AI economy section as a labelled timeline over a full-width projection (#4257)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.165.0

Features

• film v2 narration without the repeated "we believe" (#4255)

Chores

• push-scoped typecheck, locale copy-lint hook, fix regenerates artifacts, stable OG builds (#4253)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.164.0

Features

• home film, 1957-2035 AI economy section, page loops, canvas pricing, how-it-works and proof (#4252)

Docs

• document claude-opus-4-8's family-row resolution in output-ceiling.ts (#4248)

Guard

• widen agent-budgets corpus to model-feeding workspaces and all limit classes (#4247)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.163.0

Features

• film and page-loop media assets for the canvas reskin (#4249)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.162.0

Features

• fence reassignTask's UPDATE with an optional leaseEpoch CAS (#4245)

Bug Fixes

• no count ends team-lane work — waves, autofix, guard repair and manifests stop on progress (#4246)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.161.1

Bug Fixes

• a scout hears a steer and converges; long runs push; message answers the truth (#4243)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.161.0

Features

• Field and Tabs wrappers, stacked dialog trap, canvas onboarding (#4242)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.160.1

Bug Fixes

• drop "or ours" from BYOK comparison copy (#4049)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.160.0

Features

• canvas dashboard shell, live metric strip, one DataTable (#4237)

Bug Fixes

• plan DAGs survive no-work, redrive, approve and a lost merge webhook (#4238)

Docs

• correct stale docblocks claiming merge path reads PR projection (#4221)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.159.1

Bug Fixes

• honor per-reviewer DISMISSED in reviewGateFrom (#4224)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.159.0

Features

• redesign the site to the canvas, with photoreal Kubeez video backdrops (#4233)

Docs

• fix stale ogImageUrl fallback comment in blog schema (#4220)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.158.0

Features

• require --allow-drop-review-count to drop approving reviews to 0 (#4229)
• name VENOMTECH S.A.S. as the Colombian representative on /contact (#4230)

Bug Fixes

• ready_signals models current review state, not history (#4231)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.157.1

Performance

• skip browser e2e when no app is reached, fix empty release base, narrow issue readiness (#4227)

Docs

• record plan 101's 2026-09-16 slices, and stop finding.ts promising a truth gate it does not run (#4225)
• settle pricing — buyer decks at $1,000 per AI developer, pair at $2,000 (#4217)
• correct ReviewState docblock to name the real review-gate source (#4219)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.157.0

Features

• shared primitives, one breakpoint map, delete /world (#4223)

Bug Fixes

• keep a baseline-red task off the box that refused it, and stop it re-pending forever (#4215)

Chores

• dz-runner 0.0.11 (#4214)

Docs

• correct misleading blog-OG comment in og-images.ts (#4216)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.156.2

Bug Fixes

• name why a babysit follow-up landed nothing, and retry a raced push once (#4212)
• record the served model on review-lane run_provenance (#4213)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.156.1

Bug Fixes

• send a per-run OpenRouter session id on review calls, and record an unreported cache read as absent (#4208)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.156.0

Features

• reskin imagery — v2 art, OG cards, decks, Kubeez model seam (#4207)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.155.1

Bug Fixes

• re-mint the installation token once on a gh 401, and refresh it before every review write (#4203)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.155.0

Features

• brand foundation — new palette, Chivo + Geist Mono, cube mark (#4198)

Bug Fixes

• post one live APPROVE per PR, and show why a verdict was downgraded (#4201)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.154.5

Bug Fixes

• release review and scout at-head trees on every exit, and reclaim merged .wt trees (#4196)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.154.4

Bug Fixes

• keyless pass no longer advances the incremental cursor (#4079)

Other

• Stamp git_auth on GhCliError when gh refuses the credential (#4129)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.154.3

Bug Fixes

• cover tools ceilings in prompt-budget anti-vacuity bloat check (#4191)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.154.2

Bug Fixes

• fanout writes escalation delivery_state (#4188)

Tests

• pin stalled_replay retention verdict (claimed_until < now() - 7 d) (#4186)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.154.1

Bug Fixes

• use NotificationKind.Escalation instead of hard-coded string in stuckLedgerNotification (#4185)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.154.0

Features

• align BYOK SQL spend fence with domain isUsable (active, earmark) (#4183)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.153.9

Bug Fixes

• make ui-debugger refuse with a reason, and guard that .mcp.json is runnable (#4182)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.153.8

Bug Fixes

• stamp failure codes where the throw already had a kind, and pin the two ladders that block the rest (#4179)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.153.7

Bug Fixes

• correct retention.md's own R2 lane and derive the 72h CI-log window (#4177)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.153.6

Bug Fixes

• stop the stalled-replay drain deleting rows it cannot re-drive (#4175)

Docs

• enterprise reskin — one design system across marketing, dashboard and generated art (#4176)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.153.5

Bug Fixes

• babysit escalations reach a human, once per condition (#4169)

Tests

• pin the pr-lane cancel split, and correct a docblock that promised a cancel the door never sends (#4166)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.153.4

Docs

• repoint wiring docblocks at apps/agents-manager, drop dead tierToCostTier re-export (#4156)

Refactors

• replace hand-rolled error ternaries with domain errMessage (#4159)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.153.3

Bug Fixes

• align dev-seed policy blobs with policy/domain vocabularies (#4154)

Docs

• record the review-lane arc, and the verification gate it found (#4160)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.153.2

Bug Fixes

• remove duplicate systemPromptLayers member from AgentRunLoader (#4157)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.153.1

Bug Fixes

• add --conditions browser to test script (#4155)

Chores

• dz-runner 0.0.10 (#4153)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.153.0

Features

• the review lane's spend, memory, worktrees and instruments (#4152)

Docs

• re-anchor stale file:line refs in dz-env-vars.md (#4149)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.17

Bug Fixes

• stop Spanish nav overflowing at 992-1146px viewports (#4147)
• refuse a fence that reports FAILURE when its assertion SUCCEEDS, and pin the mutable refs (#4148)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.16

Bug Fixes

• stop pretending this repo prunes the registry, and assert the repo that does (#4146)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.15

Bug Fixes

• retry a gateway-cut structured response instead of losing the sample (#4145)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.14

Bug Fixes

• grade this repo on the bun it declares, and stop boxes installing an untagged one (#4144)

Docs

• close plan 101's slice-E rows and discharge the last untracked high (#4142)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.13

Performance

• bound the last full-history checkout, and stop the lanes waiting on a 5s health probe (#4140)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.12

Bug Fixes

• plan 101 slice E - the drift and the crosscut (#4136)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.11

Bug Fixes

• hygiene reaps aged dz- scratch under /tmp, behind a prefix licence (#4117)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.10

Bug Fixes

• plan 101 slice D - the surfaces (#4125)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.9

Bug Fixes

• delete remaining duplicate worker.on('failed') listeners (#4123)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.8

Bug Fixes

• plan 101 slice C - the machine that runs the work (#4119)

Docs

• align published rate-limit numbers with the code constants (#4113)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.7

Bug Fixes

• close the acquire-then-throw window, reject backslash prefixes, and widen the mkdtemp ratchet to JavaScript (#4108)

Docs

• failure-reporting header says shared line is additive, not a replacement (#4114)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.6

Bug Fixes

• plan 101 slice B - the seams the model and the caller reach (#4111)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.5

Bug Fixes

• refuse an unowned repoId on the repo prompt layer on both doors (#4112)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.4

Bug Fixes

• grade exit>128 without killedBy as killed, not red (#4110)

Chores

• drop unused bullmq dependency (#4107)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.3

Bug Fixes

• scratch dirs are a scoped resource, and a ratchet keeps raw mkdtemp out (#4106)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.2

Bug Fixes

• plan 101 slice A - retention, fences, test counts, and three babysit/review fixes (#4101)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.1

Bug Fixes

• strip ANSI where the gate tail is PERSISTED, and give a biome diagnostic one keeper (#4104)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.152.0

Features

• the scout files the handoffs its survey already authors (#4097)

Bug Fixes

• stale sweep writes the local issues row for its own close (#4091) (#4095)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.151.3

Bug Fixes

• demote decoration-only ✖ markers in gate diagnostic tiering (#4096)

Chores

• CLAUDE.md caps are 800 lines a page, and .claude is covered at all (#4087)

Db

• remove unused releaseRunnerClaimsOnBoot DbClient wrapper (#4086)

Docs

• note scripts/** and evals/** blind spot (#4084)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.151.2

Refactors

• convert runCiDispatch to an options object (#4082)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.151.1

Bug Fixes

• five ways a box stops working while still reporting healthy (#4075)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.151.0

Features

• add boot-release AuditEventKind to dispatch vocabulary (#4080)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.150.0

Features

• the survey inlet's reader admits scout handoffs as raw tasks (#4058)

Docs

• correct bare-specifier blind spot in test-code-boundary (#4072)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.149.0

Features

• POST /v1/tasks/:id/redispatch re-drives a terminal coding task (#4064)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.148.1

Bug Fixes

• an epic whose last group lands transitions to done, by the system (#4031)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.148.0

Features

• zh.json part 1 — the funnel, 416 keys, unregistered (#4028)
• epic <id> prints stuck stage + last error; epics --stuck prints the funnel (#4030)

Bug Fixes

• the notification row becomes the idempotency ledger, so a redelivered job stops re-sending (#4038)
• registered env vars read through the registry, and a fail-open preflight gate (#4042)
• the session leash bounds silence, not elapsed time (#4050)
• a lost entity lock ends the session, and the leash guard loses its blind spots (#4056)
• code the two transport throws that still land as runner_error (#4029)
• re-plan a conversation objective whose task exhausts environmentally; the structural terminal reason wins (#4014)
• plan-stuck's unfenced readers move to test-support (no production caller), funnel isolates a bad status, cmdEpics comment corrected (#4060)
• the CI lane admits before it owns, and a boot release purges the box subject (#4061)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.147.1

Bug Fixes

• one bounded drain, four callers (#4012)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.147.0

Features

• the catalog carries the CNY figures /zh will quote (#4024)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.146.1

Bug Fixes

• a mainland Zhipu or MiniMax key verifies against its own deployment (#4013)
• lift bun's 5s default on the fan-out its (#4019)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.146.0

Features

• a repo-local red-capable review check for owned repos (#3790) (#4017)

Bug Fixes

• route the lane guard through the pipelineConfig seam, and stop it red-flagging a legal posture (#4011)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.145.0

Features

• OG cards are generated and served per locale (#4003)

Docs

• correct lifecycle combination count to five (#4005)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.144.0

Features

• the repo's own autofix runs once when every gate failure is mechanically fixable (#3957)

Bug Fixes

• cross-arm served_model equality in benchmark admission (#3946)
• correct instrument-check figures in delivery-benchmark and run-record (#3947)
• provider bodies, belt audit lines and run events leave the platform unredacted (#3951)
• a kept biome block carries the summary that settles it (#3963)
• credential-bearing egress follows announced endpoints, redirects and rebinding hosts (#3964)
• a read-only caller gets a vault-write link, runner pre-auth routes are unthrottled, release bodies skip disclosure (#3980)
• pin all 42 third-party actions to commit shas, with dependabot and a guard (#3988)
• the footer offers every other locale, not the first that differs (#3989)
• the sans stacks name a CJK family (#3992)
• sign the release manifest, and verify it before reading which binary to install (#3995)
• rotate the fleet signing key over the credential roll, and make security.md true (#4000)
• a grounding pass survives a 2xx whose body could not be read (#4002)

Refactors

• one table for a lane's job options, so producers and lane declarations stop disagreeing (#3943)
• lane declarations read laneRetryOf instead of restating the table (#3949)

Tests

• the locale e2e lanes read SUPPORTED_LOCALES, not a typed /es (#3998)

🤖 Posted by developerz.ai — the maintainer agent, not a human.

v4.143.4

Bug Fixes

• close the enqueue-audit gap on retry-after-throw (#3942)

v4.143.3

Bug Fixes

• one merge gate, two callers, so gh.prs.merge no longer merges past ten refusals the babysit egress enforces (#3923)

Babysit

• forbid PR.md as a verdict channel on no-change rounds (#3849) (#3922)

v4.143.2

Bug Fixes

• surface policy_valid from repo detail into RepoPolicyTab (#3920)

v4.143.1

Bug Fixes

• the stored policy is the parser's output, so the free-OSS auto-merge opt-in gate has never fired (#3916)

Docs

• whole-project bug sweep, 187 findings across sixteen area files (#3906)
• the protocol is pre-registered in public, and the selection rule it carries selects nothing (#3909)
• the product and service portfolio in three languages (#3910)

Tests

• the blocker-fix leak check reads the body it was handed, not a clone of a spent one (#3913)

v4.143.0

Features

• the run record names the model that ANSWERED, and refuses one that contradicts what it declared (#3905)

Docs

• fix LlmPoolBudget docblock count (#3902)

v4.142.1

Bug Fixes

• a merged pull request records WHEN it merged, on both writers that settle it (#3901)

v4.142.0

Features

• onboarding runs an agent that proves the gate, asks for access once, and scores what it wrote (#3764)
• validate the babysit fix loop end to end (#2205) (#3734)
• pace a cooling router pool at the claim gate, not by re-claiming into it (#3783)
• persist deployment_status.description, so verified-vs-settled is one query (#3793)
• the benchmark page renders the absence of a result as carefully as it would render one (#3795)
• a recipe for splitting a dedicated box, and ask — the door a composed plan goes through (#3768)
• one ask, several repos, one ordered plan (#3780)
• the debugger, the reviewer and the scout become visible (#3812)
• a run provenance row says who funded it, not only who answered (#3833)
• the pre-registration is a state the page can render, and a run that drifts from it is refused (#3852)
• one arm recorded is a run in progress, not an artifact the page refuses (#3857)
• /pricing prices the payroll line the buyer actually has, not only other AI tools (#3862)
• the review footer names the box it ran on and prices OpenRouter models (#3880)
• a fleet PR body names the model, the box and the estimated cost of the run (#3882)
• a babysit fix round replies to and resolves the review threads it closed (#3889)
• merge waits on our own review threads, read by GraphQL, with a per-thread ladder (#3890)

Bug Fixes

• a fresh box enrolls correctly, and the fleet stops reading specs it no longer has (#3765)
• a repo with no commits onboards, and the onboarding wait stops being silent (#3771)
• a trailing GIT_SHA RUN per image, so a cached rebuild advances the config created (#3763)
• classify infra/docker/** as deployed surface, so a Dockerfile can merge (#3781)
• a stringified tool argument is repaired at both validation seams, and a rejected call is audited (#3778)
• drain the 400 the reasoning degrade discards, and correct its docblock (#3788)
• walk the rejection cause chain to the zod issues, making the #3745 repair real (#3791)
• the X_OK case asks the repo tree, not os.tmpdir() (#3804)
• a customer's own key may serve any model it names, and a spent key stops pinning the account (#3802)
• the proof hero goes back to one column, and a test stops the redesign returning a third time (#3807)
• deep review resolves the repo's bin/check and grades exits by the verify partition (#3423) (#3803)
• one-impl stripComments reads string/template/regex context (#3538) (#3785)
• a terminal task must not leave the box holding its claim (#3813)
• the review-samples cases read the env the code reads, not an empty one (#3817)
• pace the rate-limit wall per account, type the janitor audit trail, close the claim race (#3827)
• the box runs the toolchain the repo pins, and says so when it cannot (#3822)
• run the repo's bin/setup before every team coding session (#3840)
• the team gate runs without the run's delivered repo secrets (CI parity) (#3846)
• box_baseline_red gets its own counter and alarm, so excluding it stops looking like fixing it (#3844)
• an account is keyed by its numeric GitHub id, never its mutable login (#3851)
• the triage judge runs on the reasoning-suppressed classifier model (#3860)
• the account-identity guard stops grading paths git is told to ignore (#3861)
• every island renders the page's locale, so the Spanish catalog reaches a Spanish reader (#3863)
• the reviewer is told the file-size rule this repo enforces, not one nothing gates (#3864)
• one money convention per page, derived and authored alike (#3865)
• /proof stops scrolling sideways at 320px, in both locales (#3867)
• restore .github/workflows/** and .claude/** to the human-merge lane (#3810)
• join deploy verdicts on the merge commit, not the branch head (#3821)
• a git CONNECT failure is transport, not auth, so a blip stops being terminal (#3869)
• a team-lane PR without PR.md gets a body composed from its manifest and gate (#3874)
• a review says each thing once — no repeated fold, no prose for zero-finding lenses, capped summaries (#3877)
• act once the PR has settled, in one combined fix round (#3887)
• the grounding pass holds the stall window over an in-flight request and counts every model step (#3891)
• a team run with a phantom declared path gets one repair round instead of losing the run (#3892)
• a biome-only red gate keeps the diagnostic header that names the failing file (#3895)

Performance

• prune excluded subtrees during the walk, and start the glob walk at the pattern's prefix (#3815)

Docs

• straighten the arc 101 trail after slice 04 shipped (#3774)
• re-cut arc 101 slices 05, 06 and 07 against the code and prod (#3775)
• docs-keeper names the pin map and the compression rule, and the plan's live byte-baseline instructions retire (#3784)
• record slice 07 as merged-not-shipped, and say why a cluster-tooling fence grades nothing (#3829)
• a fence in an issue comment is never read, and grades as if none was written (#3836)

Investigate

• root-cause #3849 model_produced_no_work (no harness change) (#3871)

Other

• docs(bench),fix(marketing): the run card is frozen, and /proof stops refusing the registration it exists to accept (#3897)

Tests

• the narrow-viewport guard covers 390 and 768 too (#3868)

v4.141.0

Features

• fail a PR whose repo-wide test-declaration count drops (#3757) (#3760)

v4.140.2

Bug Fixes

• honour a refused output ceiling where the failing calls actually are (#3762)

v4.140.1

Bug Fixes

• recover a refused output ceiling on both generateText harness lanes (#3758)
• root-anchor the untracked read (--full-name + doubles) (#3751)
• relax deployed-surface YELLOW for proven comments-only diffs (#3724) (#3753)

v4.140.0

Features

• stamp every cancel path's reason into tasks.cancel_reason (#3339) (#3705)
• invite-only access, apply to allowed to signed in with seats (#3727)
• Alibaba as a provider, box add count under one enrolment lock, named org-budget refusal (#3754)
• the baseline probe row carries the duration-gap witness (#3698 criterion 4) (#3746)

Bug Fixes

• one tree expansion in e2e-typechecked, not one per case (#3721)
• exclude box_baseline_red from task failure rate (#3700) (#3716)
• the bench's harvest and run artifacts land where git keeps them (#3284) (#3704)
• the router's pool verdict outranks a member's spent wallet (#3730)
• a FAILED task keeps its box-condition errors (#3731)
• glob and grep see dot-directories, not credential dotfiles (#3729)
• the baseline-red probe runs the diff's scope, not the stashed tree's (#3717)
• re-ask once at the ceiling a provider names, instead of dying on it (#3747)
• close the credential gap dot:true opened (#3749)
• the baseline probe refuses a scope it cannot reproduce (#3750)
• refuse only the scripts/ paths the stash actually flips (#3759)

Docs

• BYO-cloud on-demand fleet — connect AWS/Hetzner, spawn from a golden image, idle-stop, wake on demand (#3725)
• output-ceiling's docblock states the SDK's 4096 substitution, not 'no ceiling' (#3719)

Tests

• pin build job's empty-apps skip gate (#3314) (#3718)
• the #3722 conjunct's tests, in the files the code says they live in (#3752)

v4.139.2

Bug Fixes

• refresh the 1h installation token on auth_expired, and keep the worktree (#3712)

v4.139.1

Bug Fixes

• ensure-deps installs unconditionally — a present bin never proved a fresh tree (#3713)

Docs

• investor-ready finish arc — eight sequential hive PRs, gap register, runbook (#3715)

v4.139.0

Features

• review-coverage eligibility read + close the three silent drop points (#3663)

Bug Fixes

• resolvePrev counts a success anywhere in a Deployment's status history (#3708)

v4.138.7

Bug Fixes

• a cancelled task leaves the failure-rate population (#3710)
• source the glm-5.3 and MiniMax-M3 ceilings, and ratchet the table (#3709)

Refactors

• a route states its own facts, REST and MCP share one use case, and two audit divergences close (#3703)

v4.138.6

Bug Fixes

• give fix_ci a legal ending for a verified-green no-op (#3690) (#3706)

v4.138.5

Bug Fixes

• one corpus sweep per whole-tree guard, hoisted out of the per-test ceiling (#3701)

v4.138.4

Bug Fixes

• grade each app against the release that built it, not one global sha (#3688)
• verify-live must not inherit a skip, so a roll is always settled (#3693)

Refactors

• an api module declares its own deps, createApp mounts the catalog, and each pod keeps one account audit store (#3685)

v4.138.3

Bug Fixes

• dz-enroll reads the committed policy's content, so a contradictory repo refuses instead of certifying (#3684)

v4.138.2

Refactors

• a closed vocabulary and a dashboard route are each stated once (#3681)

v4.138.1

Refactors

• the db barrel, the worker lane list and agent tool names register by convention, and the feature-tax ledger records its first lowered ceiling (#3680)

v4.138.0

Features

• the recipes reach merged, a bad argument teaches the fix, admins get a platform row, and every turn carries 16 KB less schema (#3678)

v4.137.0

Features

• every agent opens as one elite-engineer pilot, and the prompt cascade stops contradicting itself (#3675)

v4.136.3

Refactors

• one seam answers "can this tree accept a checkout?", and the three lanes that guessed stop guessing (#3644) (#3665)

v4.136.2

Performance

• the dev box stops OOM-ing, the gate loses its 90 s tail, main stops re-testing proven trees, and releases rebuild only what changed (#3674)

Docs

• nine claims the code contradicts, corrected against the code that disproves them (#3670)
• the nested corpus shrinks 87 percent and the briefs stop giving orders the house forbids (#3673)

v4.136.1

Performance

• the Playwright pole halves, the local gate stops waiting on a single-threaded tail, and a guard that read five sixths of the tree reads all of it (#3669)

v4.136.0

Features

• a managed repo can decline a PR, bump a manifest, sweep a quiet issue and see Dependabot's only CI signal (#3668)

v4.135.0

Features

• the wizard asks for a server and a chat client, and the fact ladder stops lying about who finished (#3667)

v4.134.0

Features

• the recipes reach a chat client as a tool, one call answers is-it-done, and a retried dispatch cannot run twice (#3652)

v4.133.0

Features

• dz enroll --apply performs exactly the diff the dry run printed, and refuses to narrow required checks (#3657)
• dz enroll reads a real repo — the live Reads client behind the existing port (#3653) (#3664)

Bug Fixes

• removeOnFail age-only + fake aligned with real BullMQ (#3556 follow-up) (#3660)
• the merge gate reads a reviewer's findings, not only its check status (#3647) (#3659)
• removeOnFail age-only, guard getState, align the fake with real BullMQ (#3654)

Chores

• re-seed the size budget after #3556 crossed it, so lint stops failing on every branch (#3656)

v4.132.3

Bug Fixes

• dz_migrate tombstoning — terminal jobs free the id + per-repo dedup discrimination (#3551) (#3556)

v4.132.2

Bug Fixes

• the collision probe catches file/directory wedges both ways, and a filename is a literal rather than a pathspec (#3640) (#3646)
• own-branch dirt PARKS in the team lane — carrying it is what destroyed the work (#3651)

v4.132.1

Bug Fixes

• a repo refused unknown_repo stops minting work (#3565)

Chores

• the one legal, satisfiable merge posture for a policy repo (#3648)

v4.132.0

Features

• one grant expresses N free AI devs, and every dispatch sink honours the seat cap (#3639)

v4.131.13

Bug Fixes

• a dead run's untracked files are parked instead of wedging the checkout the dirty judge called clean (#3640) (#3643)

v4.131.12

Bug Fixes

• the commit-hook probe asks whether THIS checkout declares the hook, instead of asserting our own lefthook against every repo (#3637) (#3641)

v4.131.11

Bug Fixes

• a blocked onboarding pass settles as the failure it is, instead of the status that promises a landed PR (#3489) (#3638)

v4.131.10

Bug Fixes

• the quiescence tests assert sample ORDER instead of racing the scheduler that produces it (#3628)
• a bounded classifier failure names the provider that refused, instead of defaulting to unknown (#3616) (#3627)

v4.131.9

Performance

• the typecheck gate stops paying a 334-module import four times to answer one question (#3626)

v4.131.8

Refactors

• the truncate arithmetic this repo already got right once reaches the thirteen lanes that redid it (#3624)

Tests

• six registers stop grading a reason by its length, and two dead citations fall out of the change (#3625)

v4.131.7

Bug Fixes

• the git step that refused reaches the timeline, so checkout_failed stops being two faults wearing one word (#3531) (#3622)

v4.131.6

Bug Fixes

• the styles.css gate gets a budget sized for the compiler it spawns, not for a unit test (#3621)

v4.131.5

Bug Fixes

• one issue gets ONE plan in flight, so a re-triaged issue stops minting a fifth copy of the same work (#2703) (#3620)

Tests

• a Playwright retry becomes a verdict, and every CI lane must have a local equivalent (#3619)

v4.131.4

Refactors

• askExcerpt stops being two byte-identical copies, and a shape ratchet pins it (#3618)

v4.131.3

Bug Fixes

• the saturating leg of the gate runs alone, because a 25% false-red rate is not a scheduling win (#3601)
• the both-locales rule says what is enforced, and the app with 2,496 keys finally has the guard (#3608)
• the changelog is generated where the token is, so prod stops telling users there are no releases (#3611)
• a terminal holder's branch stops holding the shared checkout, so a corpse cannot wedge a box (#3523) (#3612)
• a parked model stops selecting its own key, so the triage judge is not handed the one model that cannot answer (#3491) (#3615)

Ci

• the whole-tree lane runs through the sharded runner, in the lanes its files have locally (#3599)

Docs

• the CI lanes mapped with their measured cost, local equivalent and red-run signatures (#3600)
• the gate's memory peaks measured on a no-swap box, and the levers before an OOM (#3605)
• slice 18's .claude/ condition is retired on both halves, and the row does not move (#3607)
• plan-101 round 16, the first round where every move is work and none is a correction (#3613)

Refactors

• both bare repeatable registrations go through the shared helper, and the pinned set is deleted (#3603)
• account becomes a verb, and its residue map stops calling anything owed (#3604)

Tests

• the suite's real-clock waits measured, the release ladder faked through a knob, and a guard that keeps literal sleeps out (#3602)
• the meta-gate reads a generator's output as a BINDING, so the next ungated artifact fails whatever its gate would have been called (#3606)
• the agents' shared block gets the guard that was holding it up until now, which was nothing (#3609)
• the generated assets get gates that can fail, and the one that cannot is excluded with its mechanism (#3610)
• a reason is graded on what it names, not on how long it is (#3617)

v4.131.2

Bug Fixes

• every gate script reaches node_modules/.bin through one seam that needs no node (#3598)

Ci

• config/test.yml reaches all 30 workspaces in CI, and a guard executes the job to prove it (#3596)

Tests

• the size ratchet learns to fail below, and 124 banked units are the proof it could not (#3597)

v4.131.1

Docs

• plan-101 round 12, where slice 01 loses complete over a scoring convention it wrote down itself (#3592)
• two pages get the how-to their debt register deferred to them by name (#3593)

Refactors

• three bodies split along their real seams, and packages/mcp/src has no function over a repo cap (#3594)

Tests

• a debt register names who pays it, and the tenth gate stops being green over nothing (#3595)

v4.131.0

Features

• eleven more call sites parse the wire, and seventeen scroll ports become reachable by keyboard (#3591)

Performance

• the shard cap is obeyed by the branch that exists to obey it, and the bundler stops writing where tsc emits (#3590)

v4.130.0

Features

• a fourth seam face nobody had named, where every optional action was stated twice one arm apart (#3588)

Docs

• the how-to bullet is scored on its shape, which is what its own baseline measured (#3589)

v4.129.0

Features

• the task detail body becomes a resource, the enabled-repo fence collapses nine to one, and problem+json hits zero (#3587)

Bug Fixes

• a scroll port contains its own absolutely positioned descendants, and three waived routes stop scrolling (#3585)

Refactors

• a bare jobAuditStore stops compiling, and the last unbounded module-scope store is bounded (#3586)

v4.128.2

Refactors

• callWithRetry is renamed for what it does, and the deferral that kept it rested on a false premise (#3584)

v4.128.1

Performance

• the dependency bootstrap stops loading twelve workspaces to read one integer (#3583)

v4.128.0

Features

• a keyed single-flight kit replaces seven hand-rolled busy stores, and the try/finally bullet is retired (#3582)

v4.127.0

Features

• gap A closes — a verb's schema is assembled where the wiring is known, and an unavailable arm is dropped (#3581)

Docs

• the last dep debt is paid, and the .claude bullet is shown to rest on duplication the tree does not have (#3580)

v4.126.0

Features

• task result reads what a run delivered — the report with its source, the raw tasks it filed, the blocker it declared (#3576)

Docs

• the corpus hands back 11,291 bytes, a drifted action inventory is deleted, and the ceiling is lowered to bank it (#3578)
• plan-101 round 11 re-measures every slice, and the file's own headline was stale by two rounds (#3579)

Tests

• a comment that cites a path must resolve, and seven stranded citations are fixed (#3575)

v4.125.1

Bug Fixes

• the findings budget stops telling the model its surplus is discarded, and a contract test holds the sentence to the code (#3569)
• a report run is briefed for REPORT.md, its final turn is the fallback, and delivering nothing is not completed (#3574)

Refactors

• a fifth fence class that is checked, and the checker learns to see a bare ownerId (#3568)
• the review brain splits four ways, and its two latch lifetimes become a signature (#3570)

Tests

• the overlay shells are swept open, and the btn burn-down stops counting its own guards (#3571)
• the feature tax gets a live seam instrument, because both of its old ones read frozen history (#3573)

v4.125.0

Features

• a ratchet holds the harness leash-free for main agents and subagents, and its last work budget is deleted (#3559)

Bug Fixes

• the devops lane prepares, probes and grounds inside the tree the session runs in, not before any lane has one (#3562)

Chores

• stop emitting declaration and source maps, and guard that they stay off (#3554)
• every e2e directory is reached by a typecheck project, and a guard derives that set from the typecheck script (#3561)

Refactors

• one seedAccount for 27 integration suites, and a ratchet that counts definitions rather than lines (#3560)

Tests

• all ten settle arms assert they bank the ledger, and the zero-token residual is attributed out of this slice (#3557)
• the repeatable detector matches by argument list, and the hand-attached census is partitioned by declaration home (#3558)
• the id-mint concern enumerates its nine survivors, so a move and a new mint fail differently (#3566)
• the SSE heartbeat is proved on a real socket, and the idleTimeout pairing is ratcheted (#3567)

v4.124.2

Refactors

• the army fence classification becomes a guard's input, and one more verb reaches the seam (#3550)
• the gh port leaves the ci-fix feature, and three excerpt copies become one (#3552)

v4.124.1

Docs

• plan-101 round 10 re-measures every slice, and the headline falls 65 to 64 (#3547)
• onboard-a-repo runbook generated from dz-enroll state machine (#3548)

Refactors

• the job audit store leaves the blog consumer, and its 21 real call sites become a guard's input (#3549)

v4.124.0

Features

• dz-enroll state machine, dry-run (#3395) (#3546)

v4.123.2

Bug Fixes

• ground the scout on an at-head worktree, not the box's shared checkout (#3541)

v4.123.1

Refactors

• the mock plane splits per family under a 300-line ceiling, and 14 dark components get covered (#3544)

v4.123.0

Features

• a runtime-total REST/MCP body parity gate, and the digest detail body becomes one resource (#3543)

v4.122.1

Refactors

• the last two audit ladders collapse onto the seam, and a guard stops the next one (#3540)
• two platform-floor cascades become one, and four prompt layers stop being fleet-only (#3542)

v4.122.0

Features

• parameterise the autonomy report by repo list and target issue (#3536)

Docs

• nine workspace pages state their own deps, and two debt lists empty out (#3539)

v4.121.1

Refactors

• the audit envelope is spelled once, and nine kit rows are refusals with their arguments checked (#3537)

v4.121.0

Features

• plan_start scouts 1..n repos, and plan context reads their merged answer (#3359) (#3452)
• the merged scout fan-out answer gets its REST door (#3451) (#3454)
• the review lead may decline a file, and a fence decides (#3461)
• every janitor pass reports the cohort it considered (#3463)
• the janitor lane reads the prompt cascade (#3471)
• the agent-tuner gets its second signal, and its skills claim is corrected (#3474)
• the MCP server serves recipes, so a chat client learns a journey instead of re-deriving one (#3482)
• coding work on an un-onboarded repo mints the pass that releases it (#3483)
• the onboarding wizard stops calling an account ready that cannot run anything (#3488)
• code.edit locates old_string through a match ladder, so an indentation-drifted quote lands instead of burning a round trip (#3505)
• the compaction summarizer reads a clipped transcript and updates the prior note, instead of re-reading every tool result whole (#3509)
• compaction gains a prune stage that clears stale tool-result bodies, so bloat inside the verbatim tail has an answer (#3510)
• a stall reason carries the fact its kind cannot say (#2929) (#3515)
• name a partial roll instead of listing it (#3528)
• renderAutonomousPolicy proposes the autonomous posture (#3532)

Bug Fixes

• the ensemble census says degradation is all-or-nothing, and stops calling its own share a lower bound (#3464)
• the ensemble census reports Q1 per day, so an incident cannot read as a rate (#3465)
• a related-context prefetch that loses its caller sources stops reporting complete (#3466)
• the kb-gardener's model calls get one corrective re-ask (#3468)
• the kb-gardener's page branches read the page catalog, and run (#3469)
• four defects found by earlier rounds and unfixable from where they were found (#3479)
• a leaked mock fixture took main red, and every stateful family gets a reset (#3485)
• fan-out members asserted by identity, not by a random-id coin flip (#3486)
• a universal box hands the setup lane its git token, so the onboarding pass can clone (#3487)
• a screen is settled when its reads have STARTED and finished, not when two frames agree (#3490)
• a raw ask carries its permalink as data, not only as prose (#3453) (#3456)
• unlimited_spend becomes an issuable grant, and a guard holds the vocabulary to the domain (#3494)
• the grant flag rule stops being restated inline, so unlimited_spend is issuable and discoverable (#3496)
• the signed body names the event, so a relabelled header is refused (#2921) (#3497)
• the home axe sentinel stops pinning a reading of the wall clock (#3498)
• the fixture behind the age stops pinning a calendar instant, and a guard holds it there (#3500)
• the box write fence moves into the two statements that were holding it by call order (#3076) (#3504)
• an extension-less .test file is a test, so the reviewer stops grading four bash suites blind (#3070) (#3506)
• every timestamptz is precision 3, so a later stamp can no longer sort before the row that carries it (#3068) (#3508)
• slice 19's CI targets are measured, two are refused, and the refusal gets a guard (#3519)
• surface verify tail in baseline-red refusal messages (#3534)

Chores

• TypeScript 7 native compiler, and the guard that proves it is the one running (#3467)

Docs

• correct release-deployment header's first-status clause to match decideInitialStatus (#3457)
• janitor lifecycle deep dive — five stages graded against prod rows (#3458)
• the janitor MCP verb has two blockers, both measured (#3475)
• the onboarding and triage specs stop describing a world two PRs ago (#3493)
• a customer-onboarding runbook, and the unlimited_spend gap it has to warn about (#3499)
• the plan-101 ledger is re-derived from the tree, and four Done-when bullets that could never be ticked are retired (#3518)

Other

• docs(bench),feat(scripts): the delivery-benchmark method is frozen before any number exists, and its claim rules become refusals (#3263) (#3507)

Refactors

• the db barrel becomes subpaths, two lanes join the registry, and four guards stop being green over live defects (#3462)
• the lane registry grows a per-row audit anchor, and three blind guards start seeing (#3473)
• one audit store per process, and the page that fed a guard comes back (#3476)
• a capability is one declaration, and the fence that must answer before the body is read (#3477)
• a denial is an answer, and the viewport lane measures what a person meets (#3478)
• six god files become barrels, and two path-keyed registers follow their symbols (#3492)
• six more god files become barrels, and four suites split under an exact test-count invariant (#3501)
• the last runner size offenders, and five path-keyed registers instead of the one the pre-flight saw (#3503)
• team/fanout.ts becomes a barrel, and the register that only a workspace-relative grep can find (#3511)
• apps/runner has no production file over 800, five registers follow their symbols, and a lease window stops comparing two clocks (#3512)
• thirteen test suites split, and three guards learn that test-support is test code (#3516)
• the last two suites over the cap, and apps/runner has no file over 800 (#3517)
• every button class string has one writer, and a wrong variant is now a compile error (#3520)
• three verbs join the seam, and the eight that cannot say why in a type (#3521)
• forty by-id doors get a named fence, and one class is checked instead of believed (#3526)
• nineteen suites split, and a support module is production corpus to six guards (#3529)
• queues.ts becomes five planes, and the metric that would have called it progress is re-scoped (#3533)

Tests

• the connected Sentry lane, proved against real Postgres — ten deliveries, one raw task (#3362) (#3455)
• the stall remedy cannot reach a screen without someone choosing a locale seam (#2947) (#3513)
• the client SSE seam gets the gate the plan asked for, one rung up from where it asked (#3535)

v4.120.1

Refactors

• nine god files become barrels, four hashers become one, and two exemption lists learn to fail (#3450)

v4.120.0

Features

• a lane can declare its concurrency, and gets it (#3449)

v4.119.10

Refactors

• the last review document moves to its package, and three guards learn what test code is (#3448)

v4.119.9

Refactors

• the repo-owned tenant fence is spelled once, and by-id doors are enumerated (#3446)
• three SSE consumers collapse onto one loop, and both enumerate their outcomes (#3447)

v4.119.8

Refactors

• the Button rehearsal, one problem-body vocabulary, and the raw-task lane (#3445)

Tests

• the two remaining real-clock dial-count assertions move onto the virtual scheduler (#3337) (#3443)

v4.119.7

Refactors

• three more god files become barrels, three lanes declare their audit, and a false audit row is retired (#3442)

v4.119.6

Refactors

• one scope vocabulary and one path helper, with the typo class made a compile error (#3441)

v4.119.5

Bug Fixes

• a re-asked turn bills both attempts; a foreign bot's thread is not ours (#3440)

v4.119.4

Refactors

• plan 101 round 2 — five slices, and two silent holes closed (#3439)

v4.119.3

Refactors

• eleven more lanes onto defineLane, the dead webhook attach deleted, and two ratchets lowered (#3437)

Tests

• measured ceilings for the two apps/api tests that do real subprocess and WASM work (#3438)

v4.119.2

Bug Fixes

• an evals-only diff runs the evals corpus, not the whole-graph gate (#3435) (#3436)

Refactors

• plan 101 round 1 — wire Resources, db barrels, and the review test split (#3433)

Tests

• each rejection assertion is awaited inside the test that owns it (#3434)

v4.119.1

Refactors

• seven more lanes onto defineLane, and two false safety claims corrected (#3432)

v4.119.0

Features

• caller-file context for the review prefetch — blast-radius names read from the tree (#2551) (#3246)

v4.118.5

Chores

• attested review gate with autonomous classification parks (#3379)

Refactors

• a runtime-port channel for the lane seam, and the dead attach mirrors deleted (#3431)

v4.118.4

Bug Fixes

• narrow the tree-gap emit to the arms where the tree is the next row to read (#3416)

v4.118.3

Performance

• the review burst's top-up round is bounded by the same deadline (#3429)

v4.118.2

Performance

• the review wave slides inside a pass and barriers only at the ladder rung (#3427)

v4.118.1

Bug Fixes

• a frozen payload renders as frozen, and an accept path cannot leave the repo (#3424)

Performance

• a review sampling burst stops waiting for a straggler once a peer has answered (#3426)

v4.118.0

Features

• Ruby rung in detectStack and a machine-readable repository profile (#3420)
• a branch-protection client that reads the ruleset back (#3419)

Chores

• move repo config under .dz/ (#3413)

v4.117.0

Features

• report AA30 (agent-authored share) across a repo list in repo-autonomy (#3417)

Bug Fixes

• dual-read agent-pipeline config (#3421)

v4.116.0

Features

• install the toolchain manager and build baseline in the curl|sh join door (#3418)

v4.115.5

Bug Fixes

• tree-gap audit fires on unproven, not ladder arm (#3402) (#3415)

v4.115.4

Bug Fixes

• read putRes.content.sha in the blob-identity assertion, not the absent top-level sha (#3407) (#3412)
• the survey-unwired arm keeps its full attempt ladder (#3410)

Tests

• a fourth suite grades what the scout hands back (#3360) (#3411)

v4.115.3

Bug Fixes

• put the mode in the dz_migrate jobId so a dry-run cannot tombstone the real run (#3389)
• thread per-job installationId through the guard git.run seam (#3404)
• a truncation after an accepted answer surrendered the answer too (#3405)
• make the review-blind alarm answerable, and stop it missing an empty lane (#3408)
• the autonomy report refuses a pull-request-blind listing instead of printing an issues-only rate (#3364) (#3406)

v4.115.2

Bug Fixes

• split the scout grounding fault so an unproven tree is not a box fault (#3382)
• the account board reported itself saturated forever, serialising every repo (#3381)
• bound the pin-map total, and test the merge property that nothing tested (#3387)
• hand the cluster CA to the Kubernetes API TLS layer (#3378) (#3388)

Tests

• positive control for the post-deploy fence (#3386)

v4.115.1

Bug Fixes

• wire dz-config-migrate guard seams to live observations (#3367) (#3373)

v4.115.0

Features

• run each closing issue's live-test fence after the deploy and write the verdict to the issue (#3316)

Bug Fixes

• wire the triage judge into the fleet-alerts cron (#3340) (#3366)
• the CLAUDE.md byte ratchet pins per-file, closing the merge race (#3347)

v4.114.1

Bug Fixes

• negotiate a refused forced tool_choice instead of failing the call (#3330)

Chores

• the plan-chain refusal names today's number and the mechanism, and re-measuring it is one command (#3361) (#3369)

v4.114.0

Features

• the scout can hand its context back, and mints nothing when it does (#3358) (#3368)

v4.113.1

Bug Fixes

• drop engines.ladder — its pre-launch probe parks every mode-6 dispatch (#3344)

Tests

• drive reconnect-pacing test with fake timers (#3363)

v4.113.0

Features

• watch the stall report, so a queue held for a day is said out loud (#3350) (#3357)

v4.112.2

Bug Fixes

• a triage claim is counted, and an ask out of claims is settled instead of re-claimed forever (#3352) (#3356)

v4.112.1

Bug Fixes

• a box-maintenance run asks for no git credential, so the weekly fleet-upkeep lane stops failing 31 of 31 (#3351) (#3355)

v4.112.0

Features

• /healthz reports the build sha through one release reader (#3283)
• CFR7 beside A30 in the nightly report; progress.md counts verified merges (#3282)
• deploy only on green CI, record a production Deployment per sha, verify the roll live (#3302)

Bug Fixes

• deploy-proof records real deployment_status deliveries and drops the fictional fallback (#3285)
• the .dz/ policy home joins the human-merge lane it was given parity into (#3240)
• an unreadable policy file is not an absent one (#3279)
• no gate compared the migrations to schema.ts, and the rotation that prevents a delivery outage had zero tests (#3290)
• policy_dual_present never fires on a box hosting no tool service (#3253)
• ui-debugger tool-service pair moves under .dz/ui-debugger/ (#3239)
• every surface names the policy home the runner resolves (#3288)
• two lanes that destroy bytes on a schedule left no queryable trace (#3294)
• main is red because a rename could not reach a test that did not exist yet (#3297)
• the pipeline config is NOT green — both homes join the human-merge lane (#3296)
• a box that cannot run the commit hook now repairs itself instead of discarding the work (#3272)
• the qa.bug fingerprint, the digest period, and the box-warden roster each lied about dedupe (#3315)
• four scheduled sweeps that could never converge now drain or say why (#3317)
• cron lanes claimed a BullMQ retry they never registered, and one blind gather could silence every fleet alarm (#3318)
• the maintenance lanes' own reports now say what they swept, skipped and why (#3319)
• escalation nudges must name a performable action, not a label the repo lacks (#3320)
• the merge gate waits for an incumbent reviewer, bounded by the window it already measured (#3192)
• lift the "no unobserved claims" rule to the platform floor (#3322)
• degrade to reasoning-on when OpenRouter refuses suppression (#3313)
• append /v1 to z.ai Anthropic wire base URL (#3275) (#3328)
• record the model the provider SERVED, not just the one we asked for (#3324)
• a superseded sha is not a failed release (#3334)
• ceiling the whole-tree corpus-guard rescans at 30s, not bun's 5s default (#3327)
• phantom-edit guard names a gitignored write (#2988); wire-stamp scan budget (#3331)
• extract the triage-judge module and its alert kind (#3341)
• mask credential shapes in git output, not only URL userinfo (#3323)
• a provider reset stated without a year is dropped, and a rolled calendar day is fabricated (#3348) (#3353)
• the dev Dragonfly stops keeping every 10-minute snapshot forever (#3349) (#3354)
• the status-POST retry ladder holds the process alive, and a spawned-process guard proves it (#3312)

Chores

• commit the provider-failover ladder so a spent quota fails over (#3291)

Docs

• a coding task proves the dev env before its first edit — nothing runs bin/setup today (#3277)
• plan 101's ledger said not_started for fifteen slices that had already shipped (#3292)
• close what six merges actually closed, and record the stale-base incident that went with them (#3299)
• the lane table still taught the audit posture #3294 removed (#3300)
• the bug-sweep deliverable is a verified merge; autonomy-report names CFR7 (#3301)
• route a prod-fence FAIL on a healthy rollout, scope the /healthz guard to sha readers, prefer git revert (#3311)

Refactors

• the two slice-04 concerns that were invisible to the gate (#3293)
• review/core.ts was 3,058 lines with a 642-line function inside it (#3295)
• 22 lanes onto the defineLane seam, and three audit trails that were writing to stdout or throwing (#3308)

Tests

• team-lane checkout tests neutralise the ambient inference env (#3244) (#3346)

v4.111.9

Bug Fixes

• ensure_mise probes the principal the unit actually runs as, and says so when no mise is runnable (#3247)
• make a corpus freeze git discards impossible to write (#3249)
• the reasoning suppression #3225 shipped is refused by the endpoint it targets, so tesote reviewed nothing for seven hours (#3271)

v4.111.8

Bug Fixes

• a review that could not look approved the PR — 118 real GitHub approvals in 7d from runs with no codegraph (#3237)

v4.111.7

Bug Fixes

• the lane that deletes release bytes four times a day audited to stdout, so prod holds zero rows for every pass it ever ran (#3236)

v4.111.6

Bug Fixes

• a reasoning model reached through a gateway escapes the suppression gate, so a bounded review call spends its whole budget thinking (#3225)
• the at-head index syncs a throwaway copy and deletes it, so every review re-pays the checkout's whole drift (#3226)
• "no_demand" was a false statement — the lane refusal collapsed four causes and the capability axis had no count at all (#3228)
• demand counted rows no box could claim, so ten idle boxes chased 29 phantom tasks since 2026-08-28 (#3229)
• the arbiter reads the two new held axes, and the hold pass keys on the gates whose ONLY writer is this claim (#3232)

v4.111.5

Bug Fixes

• a cache keyed so nothing can ever read it, and a merge that cancels the only gate the commit behind it will ever get (#3223)

Performance

• a scripts edit typechecked the whole repo three times, and the fast path everyone reaches for is unsound (#3224)

v4.111.4

Bug Fixes

• the harness named the wrong registry, prescribed the command every brief bans, and prompted for the reads it tells agents to run (#3220)

Docs

• 73.5% of the root brain reached no reader, and the SRP chain 25 files point at answers a different question (#3222)

v4.111.3

Bug Fixes

• a review that paid for tokens and printed no cost line, MCP children that leaked on the polite path, and four more prompt seams than anyone had counted (#3218)
• ten screens whose data body was an error box, one socket per widget, and a client type wrong on every field it disagreed with the server about (#3219)

v4.111.2

Bug Fixes

• a by-id key read whose fence was a fourth-caller coin flip, and the id-brand split that would have made the plan's next step an outage (#3216)
• a kill path that cancelled its own SIGKILL, a drain that could not release a queued member, and three guards that had gone blind (#3217)

v4.111.1

Bug Fixes

• a repeatable sweep that pruned nothing for its whole life, and a lane that declares its audit instead of forgetting it (#3215)

v4.111.0

Features

• the last six gates, where the census was wrong in five of six, plus an agent roster that had been prescribing the whole-tree gate to subagents (#3213)

Refactors

• a capability declares its scope once, and the wire layer reaches the doc and the SPA (#3214)

v4.110.8

Refactors

• one wire shape per entity, and a fixture layer that graded nothing (#3212)

v4.110.7

Refactors

• one id minter behind the schema that defines the prefix, and a gate that had been scanning live code as blank (#3211)

v4.110.6

Refactors

• the last three gated concerns, where the census refused two of them (#3210)

v4.110.5

Refactors

• the rest of the logger sweep — the plain alias reaches ONE, and it is the canonical (#3208)

v4.110.4

Refactors

• one structured-log port, and 60 of the 81 aliases collapse into it (#3207)

v4.110.3

Refactors

• plan 101 slice 04 — four primitives get one implementation, and the ceilings that lock it in (#3206)

v4.110.2

Bug Fixes

• plan 101 slice 03 — every confirmed bug, red-first, and the ten slice claims that did not survive re-measurement (#3205)

v4.110.1

Bug Fixes

• a plain comment stopped clearing a CHANGES_REQUESTED, and the merge gate stopped grading 30 of a PR's reviews (#3203)

v4.110.0

Features

• the feature tax is seven numbers priced against real commits, the time budget its own measurement refuted is reverted, and a lane that timed out in silence now says why (#3202)

Docs

• the transpiled-statement denominator would score a third of the repo over 100% (#3200)

v4.109.0

Features

• a per-test time budget that fails in place, seeded at today's slowest deliberate test (#3199)

v4.108.2

Bug Fixes

• merging two lcovs was unsound, and the coverage ratchet runs in no CI job (#3196)

v4.108.1

Bug Fixes

• the repair runs BEFORE the baseline probe and the fix wave, and its tail is capped (#3197)

v4.108.0

Features

• a gate that refused to grade gets one bin/setup and one more gate (#3195)

v4.107.1

Bug Fixes

• the autonomy population comes from the repository listing, and a scripts-only CI run stops falling through to a bare bun test (#3121) (#3190)

Tests

• one live-DB fixture that owns the pool, and the CI ceiling it lets us delete (#3193)
• say "your database is behind" once, and stop two concurrent lanes colliding on hardcoded ids (#3194)

v4.107.0

Features

• the honest coverage number, per workspace, and a ratchet that cannot lower it (#3188)

v4.106.4

Bug Fixes

• name the OFF-SCHEMA non-answer too, not just the not-JSON one (#3183) (#3186)
• a test's wait budget stops being its own timeout, and a pinned clock stops making a ceiling vacuous (#3187)

v4.106.3

Bug Fixes

• scope the catalog-budget ledger guard to the NEWEST entry (#3185)

Chores

• the catalog budget's raise gets its ledger entry, and a silent one gets a guard (#3182)

Refactors

• one BullMQ queue fake, the ten files that hand-rolled their own, and the wall-clock flake that blocked it (#3184)

v4.106.2

Refactors

• one drizzle chain fake, and the five densest query suites moved onto it (#3180)

v4.106.1

Bug Fixes

• a test file stops handing the next one a pinned clock, a wrong Redis client and a DOM's streams (#3179)

v4.106.0

Features

• the structural gates, seeded at today's measured numbers, and the one bypass the compiler can take over (#3178)

v4.105.6

Bug Fixes

• a retry is eligible for another box, never guaranteed one (#3176)

v4.105.5

Bug Fixes

• a gate that could not run gets its own failure kind and the full attempt ladder (#3164) (#3174)
• the scout heals its own codegraph, the install ceilings stop killing slow boxes, and a lead overrun stops firing on a millisecond tick (#3175)

v4.105.4

Bug Fixes

• one grep hit stops evicting the whole context window, an ineffective compaction stops being invisible, and a truncated turn is re-asked instead of binned (#3173)

v4.105.3

Bug Fixes

• a ceiling-killed toolchain install is not a successful one, and a drain during it ends the run (#3171)

v4.105.2

Bug Fixes

• the per-claim git token is minted per claim and the box re-mints it, instead of sharing one 55-minute cache and pushing with a dead credential (#3169)

v4.105.1

Bug Fixes

• the coding lane installs the toolchain the checkout declares (#2950) (#3167)

Docs

• .dz/maintainer/maintainer.yml in docs, marketing, generators; .dz/README template (#3151)

v4.105.0

Features

• dz-config-migrate ready lane — re-run the guard on a draft move PR and mark it ready only when it still passes (#3153)

Bug Fixes

• a scripts/ path outside the test harness stops widening to the whole-graph gate (#2975) (#3165)

Docs

• measured plan to replace hand-wired boilerplate with one declarative seam per subsystem

v4.104.1

Bug Fixes

• a box adverts a service it can reach without a restart, and a report slice stops killing the work sequenced behind it (#3065, #2949) (#3143)

Docs

• record the 2026-09-03 review-lane arc in Progress (#3162)

v4.104.0

Features

• every box carries the latest codegraph and the review tree gets its index, the lead prefetches related code for every sub-reviewer, a per-repo reviewer prompt is read at head, and the PR states the review's run economics (#3160)

Bug Fixes

• a key out of quota parks as temporarily unavailable until the provider's reset, instead of being re-probed every two minutes or revoked (#3161)

v4.103.2

Bug Fixes

• a dead key is never re-dialled, four sub-reviewers in waves, a push never cancels a running review, and the janitor compacts the review KB (#3159)

v4.103.1

Bug Fixes

• no wall clock kills a run and no house number caps its output - liveness is the only kill, the vendors' published ceilings the only bound (#3155)

Docs

• record the 2026-09-03 re-audit of plans 101 and 102 and the follow-up PRs it produced (#3156)

v4.103.0

Features

• ratchet guard for the nested CLAUDE.md corpus (#2992) (#3125)
• dz-config-migrate consumer + ops verb + draft/ready flow (#3128)
• dual-read .dz/maintainer/maintainer.yml in the review adapter and setup; scaffolder proposes .dz/ (#3152)

Bug Fixes

• a retry inherits the prior attempt's terminal reason (#3131)
• a fully-cooled key pool is graded provider_rate_limit, and the box holds its pulls until the router's reset (#3050) (#3130)
• the devops lane names its own deliverables and the no-op case is no longer a fault (#3014)
• the fix wave recovers the paths real gate output prints, and a red base is probed before the first wave, not after the last (#3136)
• unblock main and make a CLAUDE.md ratchet trip attributable (#3137)
• the runner-release age window is re-derived at 29 releases/day, and the prune tick moves with it (#3134)
• a box that defers its swap behind a live run reports the deferral, and the rollout grades it pending instead of confirm_deadline (#3138)
• the DOCR retention cron excludes the two out-of-remit mail images and exits by what the prune did (#3105) (#3140)
• the reviewer sees whole files under a budget derived from the serving model's window, and its prompt says what a full file and a green typecheck each prove (#3100) (#3141)
• the hive reader attributes rows to the binary that wrote them, and says no_provenance when no lane stamped one (#3142)
• pin the ensemble validator's #2196 merge stamp in prose and tests (#3145)
• the native harness compacts at 70% of the model's window, the summary carries no house ceiling, and MiniMax gets its window rows (#2979, #3106) (#3146)
• every hive row is stamped with the binary that wrote it, at the one seam every lane passes through (#3148)

Refactors

• the box installer knows no language - ensure_ruby is deleted, a generic build baseline and mise take its place, and the setup lane writes the repo's toolchain as text (#3079) (#3147)

Tests

• pin the assignment schema's unknown-key strip on the bare contract (#3116) (#3132)

v4.102.2

Bug Fixes

• a JetStream grant whose effect rides the request body is not a permission (#3118) (#3129)

v4.102.1

Bug Fixes

• a verify failure on an already-red base is box_baseline_red and spends no attempt (#3123)

v4.102.0

Features

• the hive's rates get a reader, and the first thing it found was a dead producer (#3109)

Bug Fixes

• ensure-deps install-failed exits CHECK_EXIT_PRECONDITIONS_UNMET, not 1 (#3027) (#3112)
• purge the arc-101 retired flags, which no writer has ever removed (#3117)
• the autonomy population is derived twice, and a disagreement refuses (#3114) (#3120)

Docs

• close plan 102 with the deployment evidence, and say why the fix wave has 0 rows (#3119)

v4.101.0

Features

• record refuted review findings on the babysit envelope (#3107)

Tests

• derive skew-guard stand-in from parent schema (#3039) (#3113)

v4.100.4

Bug Fixes

• the scout refuses the plan call when its promised grounding never arrived (#3037) (#3104)

v4.100.3

Bug Fixes

• revert the harness --no-verify and fix the pre-push hook it was working around (#2980) (#3017)

v4.100.2

Bug Fixes

• a red gate buys one bounded fix wave instead of ending the session (#3099)
• a free-form BYOK model id degrades the review_sampled slots instead of deleting the event (#3101)

v4.100.1

Bug Fixes

• the runtime ceiling settles a quota-stalled run by its cause, not run_timeout (#2996) (#3055)

v4.100.0

Features

• the review lane becomes a hive, and the four passes that never ran now run (#3095)

Bug Fixes

• an empty survey fails the run instead of minting a decision_needed blocker (#3054)

v4.99.0

Features

• the debugger lane gets its first hive, and two defects found while building it (#3091)

v4.98.2

Bug Fixes

• the Ruby toolchain converge installs libffi-dev, so wurk's bundle can build (#3067) (#3088)

v4.98.1

Bug Fixes

• a box can only write for a task it may serve, and the cross-tenant pool is retired (#3077)

v4.98.0

Features

• one source for the hive rules, seated only where a model actually leads (#3087)

v4.97.0

Features

• sub-scouts get a channel a model can read, and the scout lead corrects a live leaf (#3083)

v4.96.0

Features

• the lead gets a mouth - rule 9's channel had a receiver, one human sender, and zero guidance rows in 60 days (#3080)

v4.95.2

Bug Fixes

• the converged Ruby toolchain can install gems as the runner user, and an Unknown stack with a test dir is not told to build a harness (#3067) (#3078)

v4.95.1

Bug Fixes

• the coding leaf gets a retry ladder, a coordinator can refuse an assignment, and the task attempt is a real number (#3073)
• ensure_ruby carries its own --no-systemd guard, like both siblings (#3067) (#3075)

v4.95.0

Features

• a box provisions the Ruby toolchain, so a Ruby repo's gate can run at all (#3067) (#3069)

Bug Fixes

• the ruby toolchain converge reaches an enrolled box (#3067) (#3072)

v4.94.2

Bug Fixes

• the instrumentation watermark answers for the repos being graded, not the whole table (#2937) (#3066)

v4.94.1

Bug Fixes

• the repo-service converge reaches an enrolled box (#2936) (#2944)

v4.94.0

Features

• one hive round contract, and run provenance that names the binary that wrote it (#3062)

Bug Fixes

• the excusal streak gets a durable row (the commit #2932 merged without) (#2935)

Docs

• the enterprise deck in English and Romanian
• explain how CLAUDE.md files are loaded in the runner (#3060)

v4.93.1

Bug Fixes

• the diff classifier parks the trees that define the merge gate (#2951) (#3052)

v4.93.0

Features

• every self-update hop the box takes lands as an audit row, a rollout survives a worker redeploy, and update refusals reach the log egress (#3029)

Bug Fixes

• review worktrees stop leaking into the shared checkout (#2995) (#3031)
• a path the harness owns is never a declared manifest file (#2981)
• one run ceiling, a ledger that resumes only where it is read, a post-edit check that degrades instead of disabling, a recorded blocked rung, and provider failures that stop reading as box faults (#3034)
• a reaped box stops running the task it lost, a non-advancing credential re-issue is refused, and every dropped fleet event lands a row (#3041)
• absence is a sentinel — an empty 2xx body is not "file missing" (#3042)
• the boot credential roll reports too, instead of being dropped (#3047)
• one tenant's rollout no longer starves every other tenant's (#3048)
• hygiene learns both halves of the at-head review layout (#2995) (#3049)
• fleet PRs close the governing issue the task's envelope names (#2271) (#3051)
• a request id no longer spends the review ensemble's only fallback dial, and the sibling is budgeted from its own ceiling (#3053)
• a babysit follow-up can answer with a reasoned no-change (#3056)

Docs

• close plan 101 with the after-probes and the prod rows that prove it (#3046)

v4.92.0

Features

• .dz/ migration — mechanical guard + config inventory (#3020)
• revoke reaches the box and halts it, one act decommissions a box, and re-enrolling with --replace retires the predecessor (#3025)

Bug Fixes

• an ambient ANTHROPIC_BASE_URL can no longer redial an Anthropic-direct review (#2790) (#3026)
• classify lefthook.yml as GREEN so hook edits stop parking as unknown-tier (#3028)

v4.91.1

Bug Fixes

• connectivity plane, transport half — ping before the reaper, heartbeat from the beat, mint before rotate, fleet status says unknown, every fleet alert carries its tenant (#3015)
• sweep the box-side leaks, one setsid spawn seam for every git/gh/docker child, a producer for adapterPgids, and RAII at every teardown (#3018)
• exit 75 / 126 / 127 from a gate are not a verdict on the edit (#2934) (#3021)
• a third-party coding-agent App can be allowlisted for review (#2994) (#3019)
• propagate a failing leg's real exit code out of the gate reporters (#3016)
• the box converges its own env file — retired keys purged, backups pruned, a root-owned file graded loudly, and the crash-egress DSN delivered from the platform (#3022)

v4.91.0

Features

• dual-read .dz/maintainer/maintainer.yml in policy-ingest and the push predicate (#2959) (#2999)
• box-probe.sh + infra fleet-health — the measurements that prove each runner lifecycle slice (#3010)
• a box can be decommissioned, the installers agree on one contract, and the env file is owned by the user the unit runs as (#3013)

Bug Fixes

• a red gate reports the failure, not the end of the pass stream (#2972) (#2993)
• give the #513 area gate the NON_EDITING_ORIGINS conjunct (#2982) (#2986)
• a drain releases the in-flight run instead of failing it, a fleet push waits for idle, and every exit path runs one teardown (#3011)

Docs

• runner lifecycle deep dive — plan 101 (2026-09-02)
• record run 1 (Claude Code) against PR #2911 (#3009)
• hive mind across the lanes — plan 102 (2026-09-02)

v4.90.7

Bug Fixes

• close five lanes with no producer (#2990)

v4.90.6

Bug Fixes

• re-check the deference per head (#2872) (#2989)

v4.90.5

Bug Fixes

• pass --no-verify on the harness's own git commit and push (#2977) (#2984)

v4.90.4

Bug Fixes

• audit the push-supersede review cancel (#2772) (#2978)

v4.90.3

Bug Fixes

• name the drain signal's abort reason (#2943)

v4.90.2

Bug Fixes

• the at-head tree fetches the reviewed commit, and says why when it cannot (#2957)

v4.90.1

Bug Fixes

• give every lane-fence purpose an approved-model source (#2955)

v4.90.0

Features

• post the nightly wurk autonomy report to #2770 (#2913)

v4.89.0

Features

• report time-to-first-bot-action per item (#2914)

Bug Fixes

• the committed merge posture refuses every PR (#2948)
• give the per-lane approved-model gate a committed home it can fire from (#2842)

Tests

• the lead-round tests wait for the window, not for the clock (#2939)
• the overran assertion settles the round on a microtask, not a timer (#2942)

v4.88.0

Features

• hive rules 6, 8 and 9 — the lead names the check, works, and steers (#2933)

v4.87.1

Bug Fixes

• a refusal id the platform never declared, and an excusal nobody counted (#2223) (#2932)

v4.87.0

Features

• hive rule 4 — seam_changes on the task envelope (#2927)

v4.86.0

Features

• no_service_capacity — a task nobody can claim now says why (#2928)
• the converge runs the Redis a repo's gate needs, so the advert has something true to report (#2930)

Bug Fixes

• a babysit follow-up may not edit the gate whose verdict triggered it (#2931)

v4.85.1

Bug Fixes

• dispatchHandlerDeps dropped two sinks on the floor, and nothing could see it (#2926)

v4.85.0

Features

• a repo can declare the services its gate needs, and a box that cannot reach them stops claiming its work (#2925)

Bug Fixes

• one tenant's boxes cannot withdraw a release from every tenant (#2223) (#2924)

v4.84.0

Features

• grade the survey's no_answer/prose rates per account (#2917)

v4.83.4

Bug Fixes

• a touch is a fact about a delivery, not an artifact of a session (#2904)

v4.83.3

Bug Fixes

• drop-noise reachable — dupe-of, escalate, lookup finding (#2916)

v4.83.2

Bug Fixes

• state taxonomy — give the failure, hold and plan-stall columns reachable producers (#2915)

Tests

• pin every escalation category to a reachable channel (#2897)

v4.83.1

Bug Fixes

• the gate interpreter resolves as the daemon principal, and a box that cannot run the gate refuses to boot (#2912)

v4.83.0

Features

• pin the autonomy window start with --since (#2910)

Chores

• opt this repo into autonomous classification parks (#2905)

Docs

• document release-check.ts purpose (#2911)

v4.82.0

Features

• compute A30 from our own rows, and refuse the window our rows cannot cover (#2896)
• a guard that can tell a reviewed diff from a summarised one (#2901)

Bug Fixes

• one seam for comparing a GitHub login, and the two live defects that existed without it (#2892)
• the MCP surface advertised what it could not serve (#2894)
• the scope picker reports its own state in the design system's fill (#2900)
• BYOK verify failures land in recoverable states, and a model check never kills a key (#2906)

Other

• Revert "feat(marketing): price the buyer's own headcount beside ours on /pricing" (#2874) (#2876)

v4.81.0

Features

• record who touched an item, and GitHub's own clock (#2869)

v4.80.0

Features

• price the buyer's own headcount beside ours on /pricing (#2873) (#2874)

v4.79.1

Bug Fixes

• two of the three review passes could never report what they spent (#2870)
• the blind rate says WHEN, so a healed collapse stops reading as a live fault (#2871)

v4.79.0

Features

• ask the reviewer for documentation accuracy, the largest measured miss class (#2868)

Bug Fixes

• delete the per-session USD hard stop (#2864)
• the review fan-out dispatches in one wave, and retries by narrowing (#2865)
• expiry fixtures are written against the clock, not against a date (#2867)

Docs

• record the token-naming and typecheck-gate work in the Progress log (#2836)

v4.78.0

Features

• bun run infra review — the review lane in one hop (#2862)

Chores

• remove maxBudgetUsd, a per-run dollar ceiling nothing ever read (#2863)

v4.77.2

Bug Fixes

• the inline cap folds findings instead of destroying them (#2860)
• a dispatch that seals no credential leaves a durable row (#2861)

v4.77.1

Bug Fixes

• stop declaring an output-token ceiling nobody measured (#2858)
• make "no turn ceiling" expressible, and delete the two that were invented (#2859)

v4.77.0

Features

• the stored shape and wire class for customer-owned SSH keys (#2854)

Docs

• correct 101-factory-of-code/status.yml slice-03 auto-merge note (#2856)

Tests

• pin truncation onto the backlog-import audit row (#2851)

v4.76.0

Features

• status MCP verb — one composed answer (#2843)

Docs

• one rulebook — AGENTS.md becomes a symlink to CLAUDE.md (#2850)

Tests

• make the fresh-checkout gate green on any git build and any shell env (#2849)

v4.75.0

Features

• alert when the AI reviewer goes blind (#2846) (#2847)

Docs

• the contract matrix and the origin crosswalk — agent-mode spec (#2840)

v4.74.1

Bug Fixes

• approve the model spellings dispatch actually seals (#2825) (#2841)

v4.74.0

Features

• readiness as one implementation (repo-scoped, verb-named, DB-proven) (#2837)

v4.73.0

Features

• a PAT can be named, and the scope picker is grouped (#2828, #2829) (#2831)

Bug Fixes

• the scratch-exclusion guard deletes only what it created (#2834 follow-up) (#2835)

Chores

• the typecheck program no longer roots gitignored scratch (#2833) (#2834)

v4.72.1

Bug Fixes

• budget the six deterministic failure kinds to one attempt (#2572) (#2832)

v4.72.0

Features

• report the authorship split beside A30, so the rate cannot be misread (#2827)
• project setup as the advertised greenfield path (#2830)

v4.71.5

Bug Fixes

• the team lane never read PR.md, so every PR it opens says the agent wrote nothing (#2822)

v4.71.4

Bug Fixes

• wire the byok_key_purpose port so the verb is reachable in prod (#2826)

Docs

• The architect role (cast row, role-architect.md, generator/no-scaffold contract, selector tests) (#2823)
• publish the deployments read scope (#2824)

Tests

• verify plans backfill migration on a throwaway Postgres schema (#2820)

v4.71.3

Bug Fixes

• decide PR.md by HEAD membership, and unstage a copy the agent staged (#2819)

v4.71.2

Bug Fixes

• the harness deleted a customer's tracked PR.md and wedged the box (#2817)

v4.71.1

Bug Fixes

• the coding lane staged PR.md on a well-onboarded repo (#2815)

Tests

• triage plan_groups writer against real Postgres (#2814)

v4.71.0

Features

• expedite priority band + plan-group release enumeration (#2813)

Tests

• pin the raw-task fold's remaining invariants (#2811)
• pin the no-scratch landing contract properly (#2812)

v4.70.2

Bug Fixes

• the babysit lane committed PR.md onto other people's branches (#2809)

v4.70.1

Bug Fixes

• five fail-open paths in the autonomy metric (#2770)

Db

• make parallel_safe inexpressible at the triage mint (#2808)

v4.70.0

Features

• A30/T50 — make "wurk is 100% AI-managed" a number a command computes (#2806)

Ci

• refuse a PR whose squash would erase an issue's audit trail (#2799)

Docs

• commit the OSS dogfood plan and correct two dead premises (#2805)

Tests

• pin the overlapping-files race and the lapsed-lease job key (#2802)

v4.69.1

Bug Fixes

• guard the janitor containment handler's audit appends (#2228) (#2793)

Tests

• the mid-conversation fix link is single-redeem and write-only (#2797)

v4.69.0

Features

• scout conversation — mailbox drain, turn loop, verdict tests (#2796)

Tests

• the scout eval doc 07 step 1 asked for (#2795)

v4.68.0

Features

• give the grounding tree the reviewed head SHA (#2687) (#2791)
• The setup pass decides; the renderers write (#2788)
• a per-task attempt ladder on the task detail page (#2794)

Chores

• split CLAUDE.md into per-section docs (#2792)

v4.67.0

Features

• make split-runner-claude-md target selectable (#2777)
• every fleet alert leaves a durable audit row, including its delivery outcome (#2576) (#2785)
• project task failure_kind + failure_reason (#2574) (#2782)
• the ranked failure-class table in the ops digest (#2693) (#2786)

Bug Fixes

• stop advertising a nullable column as a widget sort key, and repair the 7 boards (#2775)

v4.66.0

Features

• cancel a held plan — one stop, the whole plan (#2696) (#2773)

Bug Fixes

• stamp provider_quota_exhausted on a status-less planner failure (#2710) (#2783)

Docs

• the reliability bar gates customer admission (#2577) (#2779)

v4.65.1

Bug Fixes

• aitm purge fixtures — arrangement, not assertions (#2653) (#2780)
• task-failure-report excludes both D4 origins (#2690) (#2778)

v4.65.0

Features

• wire deploy-proof lane end-to-end, INERT until App holds deployments:read (#2767)

Docs

• Claude mobile in the MCP client list; clients substantiated today (#2776)

v4.64.0

Features

• make PromptRole.Architect reachable (#2683) (#2722)

v4.63.0

Features

• pin dz/onboarding to human-merge, surface the stall (#2765)

Bug Fixes

• the team coordinator's manifest rejection is terminal after three identical asks (#2763)

Docs

• split CLAUDE.md into per-section docs (#2755) (#2766)

v4.62.2

Bug Fixes

• refuse a widget whose sort or filter key its reader cannot honour (#2761)

Tests

• live-DB proof onboarding mints exactly one setup task (#2756)
• prove tenant isolation on a minted credential, not on a verb's query (#2764)

v4.62.1

Bug Fixes

• run the pre-commit path on native binaries so a box with no node can commit (#2758)

v4.62.0

Features

• the dispatch conversation contract and its error taxonomy (#2724)

v4.61.0

Features

• CI capability without env flag + enrollment flow (#2754)

v4.60.0

Features

• fleet-bearing boxes grant on free/oss_verified (#2753)

Docs

• Close § 6 in deferred.md — in-cluster compaction loop now measurable (#2752)

v4.59.8

Bug Fixes

• the fleet failure-rate alarm had no denominator, so it could only say 100% (#2751)

v4.59.7

Bug Fixes

• an unconfigured escalation email threw, and the throw did the damage (#2748)
• every Spanish body link walked the reader back into English (#2749)
• an operator cancel left its review check wedged for 45 minutes (#2750)

v4.59.6

Bug Fixes

• five surfaces claimed things they could not know (#2747)

v4.59.5

Bug Fixes

• two guards were green over the class they exist to catch (#2745)

v4.59.4

Bug Fixes

• donation create could give away a member's PERSONAL key and private box (#2742)
• a REFUSED KB write emitted an audit row for a page nobody stored (#2743)
• box releases hid every halted release behind an undisclosed cap (#2744)

v4.59.3

Bug Fixes

• a RetryError-wrapped refusal was filed as a residual provider failure (#2739)
• the model gate's rollback restored the very lane it protects against (#2740)
• a cancelled review left its check run queued forever (#2741)

v4.59.2

Bug Fixes

• two lanes minted a BullMQ jobId BullMQ refuses (#2738)

v4.59.1

Bug Fixes

• the coding lane could not explain its own failures (#2736)

v4.59.0

Features

• hold a plan for approval and a verb to release it (2680) (#2721)

Docs

• add /bug-sweep, an unattended QA-and-fix campaign command
• authorize /bug-sweep to land on main directly, scoped
• /bug-sweep always opens a PR, and may admin-merge it

v4.58.4

Bug Fixes

• an expired credential is not a wiring fault (#2733)

v4.58.3

Bug Fixes

• 59% of issue triage was a non-answer recorded as a success (#2730)
• a credential re-issue minted the expiry it was replacing (#2731)

v4.58.2

Bug Fixes

• a gate that could not run is not a diff that is red (#2729)

v4.58.1

Bug Fixes

• the factory wasted half its runs and could not say why (#2726)

v4.58.0

Features

• surface the conversation_ref rollup as a plan resource action (#2681) (#2725)

v4.57.0

Features

• give conversation_ref its first reader — a per-conversation rollup (#2681) (#2718)

Bug Fixes

• a babysit follow-up that landed nothing reported the task done (#2719)

v4.56.0

Features

• give the tool contract a real version — SERVER_INFO only, zero catalog cost (#2686) (#2715)

Bug Fixes

• a 429-paused BYOK key has no automatic way back — stamp a retry so the re-probe recovers it (#2678) (#2717)

v4.55.1

Bug Fixes

• a reply sent while the scout is thinking is dropped, not queued - no human has ever talked back (#2675) (#2712)
• dispatch.created records no actor - audit_events.user_id is null on every row (#2677) (#2713)

v4.55.0

Features

• delete the spawn lane, and refuse a box still configured for it (#2629)

v4.54.5

Bug Fixes

• the spawned scout surveyed the runner's own cwd, not the repo it was asked about (#2610)

v4.54.4

Bug Fixes

• renumber byok_key_models onto 0163, price the catalog rebase (#2597)

v4.54.3

Bug Fixes

• triage is flight control — one per org, and nobody rewrites its procedure (#2709)

v4.54.2

Bug Fixes

• one org-slug seam — per-PR worktree isolation has been standing aside since #1556 (#2706) (#2707)

v4.54.1

Bug Fixes

• an empty DZ_RUNNER_ORG must not defeat the account_id fallback (#2702) (#2704)

v4.54.0

Features

• alarm when an account's lane credentials fall below a floor (#2697)

v4.53.0

Features

• a sustained failure-rate alarm, replacing the per-attempt spam (#2692)

Bug Fixes

• a review that could not provision its grounding tree may not approve (#2688) (#2689)

Docs

• correct #2196's merge date and flag the contaminated baseline window (#2207) (#2695)

v4.52.5

Bug Fixes

• stamp failure codes on the [team] lane (#2685)

v4.52.4

Bug Fixes

• resolve the devops harvest and report collector per assignment (#2664) (#2670)
• close the two swallowed failures in the review lane (#2674)

v4.52.3

Bug Fixes

• the no-tools clause belongs at the end of the prompt, not the top (#2661) (#2665)

v4.52.2

Bug Fixes

• classify provider quota exhaustion as a CHARGED failure kind (#2669)

v4.52.1

Bug Fixes

• gh.prs.merge asks whether a review happened (#2099) (#2666)

v4.52.0

Features

• per-sample provider fallback for review ensembles (#2203) (#2559)

Docs

• the KNOWN DEAD alert section is false — name the live producer (#2663)

v4.51.0

Features

• CI checks, prior discussion and sibling tests reach the reviewer (#2204) (#2596)

Bug Fixes

• gate-4 reachability — read merge state off the PR-owning task (#2635) (#2658)
• the ensemble-validation verdict is four-way + Q1 caveat (#2660)

Tests

• guard the no_usable_keys pacer is wired into the claim (closes #2649) (#2656)
• derive the hold reason from its symbol — #2649's last criterion (#2662)

v4.50.1

Bug Fixes

• the events SQL and its reader now name one key (closes #2646) (#2651)

Review

• validate post-#2196 ensemble telemetry against the 64.4% baseline (#2207) (#2655)

v4.50.0

Features

• refuse an inference endpoint that answers 200 without being an API (#2630)

Bug Fixes

• bound the unbounded reapStaleRunner lane — stamp a kind, charge an attempt (#2647)

v4.49.0

Features

• a mandatory pass matrix — hunk coverage cannot say WHICH kind of review ran (#2642)

v4.48.0

Features

• an approved-model list per lane, refused loudly at dispatch (#2567) (#2640)

Bug Fixes

• key the box workspace on account_id when DZ_RUNNER_ORG is absent (#2569) (#2644)

v4.47.2

Bug Fixes

• stamp the team lane's no-checkout refusal too (#2643)

v4.47.1

Bug Fixes

• a quiet provider is a transport fault, not a model that said nothing (#2636)
• stamp machine error codes at every named failure seam (#2570) (#2639)

v4.47.0

Features

• category regression fixtures and recall (#2206) (#2595)
• per-pass model routing + per-pass cost observability (#2550) (#2637)

v4.46.0

Features

• back the fleet dispatch tool with a real ToolContext port (#2632)

Bug Fixes

• the none lane still records the #513 area hold (#2633)

Docs

• the drain docblock claimed an 11-minute ceiling and a free redelivery (#2628)

v4.45.1

Bug Fixes

• main is red on a fixture for a throw that no longer exists (#2631)

Docs

• one PR in flight at a time, and the incident that proves why (#2634)

v4.45.0

Features

• give the maintainer session a fleet dispatch tool (#2011) (#2626)
• ensemble-hardening-validation (#2207) (#2558)
• an approved-model list per lane, refused loudly at dispatch (#2627)

Bug Fixes

• a wall clock was the shadow sweep's CAS token, so a stale write won (#2624)

Docs

• two rules the fan-out has no field for, and evidence for the three unbuilt (#2623)

v4.44.0

Features

• a customer sends a PR url, and one command says what broke (#2621)
• alert on the three failures that reported success (#2620)

Bug Fixes

• babysit called a PR mergeable that GitHub refuses, then PUT it 19 times (#2618)
• a deleted janitor kept firing daily and box-warden ran two schedules at once (#2619)
• bound the whole session with the wall-clock leash, not just the model loop (#2622)

v4.43.2

Bug Fixes

• the coding harness ran exactly one turn, because omitting stopWhen means stepCountIs(1) (#2616)
• the tree contract's coding evidence read idempotent redelivery as a wrong tree (#2617)

Docs

• capture the prod-debugging path a fresh session re-derives from zero (#2615)

v4.43.1

Bug Fixes

• triage escalated 7 asks to a human and hid all 7 from them (#2614)

v4.43.0

Features

• give the plan DAG a dead-tail exit, and make a model that never reaches turn one visible (#2605)
• declare which tree each lane may read, and refuse the scout on an unproven one (#2613)

Bug Fixes

• the repo gate took a fleet box down, so scope it to the change (#2612)

v4.42.2

Bug Fixes

• the reviewer filed major findings about code its own fan-out withheld from it (#2611)

v4.42.1

Bug Fixes

• a migration whose journal when is not the newest is silently never applied (#2602)
• the team coding lane handed setsid a sentence, so every native box failed in 2ms (#2607)

Chores

• give the fleet this repo's own gate, not a generic default ending in bare bun test (#2608)

v4.42.0

Features

• the reviewer discloses a takeover, salvages a sample, and can finally read the repo's memory (#2601)

v4.41.0

Features

• the fleet self-heals, and a personal box can serve its owner across orgs (#2600)

v4.40.0

Features

• operator-initiated credential re-issue verb (#2547) (#2593)
• persist NATS credential expiry and alert on failure to refresh (#2543) (#2590)

Bug Fixes

• the coding lane failed 100% of runs on a coordinator that echoed the manifest schema (#2599)

v4.39.0

Features

• enumerate the deterministic failure kinds and budget them at today's values (#2565) (#2592)
• task-failure-report — the repeatable failure measurement (#2564) (#2591)

v4.38.3

Bug Fixes

• write-time tenant derivation + computed_at CAS on the shadow upsert (#2202) (#2587)

Chores

• stop casting blocking verdicts, so ci-only is reachable (#2585)

v4.38.2

Bug Fixes

• a box re-issues its NATS credential twice on boot (#2546) (#2584)

v4.38.1

Bug Fixes

• name the files the devops powers read, so the lane can produce anything (#2428) (#2583)

Docs

• credential runbook, and correct the never-refreshed claim (#2582)

v4.38.0

Features

• --unit-only converge (#2555) (#2581)

v4.37.0

Features

• shadow-match pipeline — per-PR OUR-vs-CodeRabbit comparison rows (#2202) (#2579)
• provider model enumeration, pre- and post-capture (#2556)

Bug Fixes

• address the #2579 review round — fence re-stamp, one-seam terminal statuses, narrowed catch, shutdown close (#2580)

v4.36.0

Features

• wire fleet runner through review pipeline (#2557)

v4.35.2

Bug Fixes

• let the runner write its own profile dir, and name a blocked write (#2539) (#2560)

v4.35.1

Bug Fixes

• discover repos that predate the App install, dormant (#2553)

v4.35.0

Features

• token expiry policy, last_used_at, OAuth grant provenance (#2541)

v4.34.0

Features

• connect panel + signed-out redirect + CORS + tests (#2538)

v4.33.0

Features

• waitlist wall, invite by email, pending list, durable audit (#2536)

v4.32.1

Bug Fixes

• embed failing box hostname in task_failed alerts (#2422) (#2537)

v4.32.0

Features

• invite landing, return-to leg, and verdict tests (#2535)

v4.31.0

Features

• origin-scoped LLM credentials + all-lane prose cascade (plan 101 slice 11, steps 3-4) (#2534)

v4.30.0

Features

• require editor tags, prove fan-out idempotency (#2533)

v4.29.3

Bug Fixes

• a drain mid-verify was misreported as a red gate (#2531)

Docs

• the user lifecycle arc — invite to org scale
• lineage, the worked example, and the collapsed plan levels
• a deferred register for the arc, and the multi-people axis
• triage as air traffic control — owner-scoped mutex, dupe-of, tools
• onboarding as an agent lane, and spend the structural doors now
• what buying more VMs and subscriptions actually buys
• the two capacity paths — top up a balance, or add a plan
• "how is developerz.ai doing" — status and capacity over MCP
• parameterized MCP tools — the catalog fix that is worth doing
• one core, two consumers — agents and humans call the same MCP
• record the two compaction regimes
• detect the context window instead of hardcoding it
• deduce price and context window from the provider
• record the conversion finding in the tracker, drop personal names
• restore status.yml evidence as folded block scalars (#2532)

v4.29.2

Bug Fixes

• four janitor lanes ran 2,634 times in a week and could not have acted once (#2529)
• one path adjudicator, a per-repo board that admits its cap, and 529 stalled slices counted as in flight (#2530)

v4.29.1

Bug Fixes

• verify the runner_error storm's secondary clusters against main, pin the team fetch's auth env (#2524)

v4.29.0

Features

• converge join-owed env keys on credential roll (#2527)
• triage decides and mints, and a schedule verdict that mints nothing is now unrepresentable (#2525)

v4.28.0

Features

• 1,052 escalations reached nobody, and the detector would have been an email (#2521)

Bug Fixes

• 529 plan groups were stalled and nothing could say so (#2523)
• resolve the ssh user from ~/.ssh/config, and name the override on auth failure (#2423) (#2522)

v4.27.6

Bug Fixes

• require the two fleet-name fields, and cover the hostname rung (#2519)

v4.27.5

Bug Fixes

• the Agents panel renamed every named box, so one fleet read as two (#2515)
• the raw-task lane had five producers and three postures, and the divergent one silently re-armed nothing (#2517)

v4.27.4

Bug Fixes

• three raw-task producers filed asks that nothing was ever told about (#2516)

v4.27.3

Bug Fixes

• a blocker answer filed an ask labelled as the wrong inlet, carrying no access, that no index could find (#2510)

Docs

• both docs pages said we publish no check run, and #2379 shipped one (#2513)
• the spec said shipped work was missing and missing work was shipped (#2514)

v4.27.2

Bug Fixes

• 73 triage settles left no audit row, and the guard that should have caught it named every kind but this one (#2507)
• a raw-task nudge got one attempt, so a single Redis blip stranded the ask (#2508)
• a webhook dedupe key was a prefix of untrusted text, so two distinct asks collapsed into one (#2509)

v4.27.1

Bug Fixes

• every babysit follow-up declared an empty file set, re-serialising the repo it ran on (#2504)
• a folded or glob-spelled plan group declared free concurrency it had not proven (#2505)
• parallel_safe could be granted by a config key and by the wire, with nothing checked either way (#2506)

v4.27.0

Features

• 2,421 unread files, and nothing could say which of three problems to fix (#2502)

v4.26.0

Features

• the reviewer reasoned about hunks and was told it read files in full (#2501)

v4.25.1

Bug Fixes

• the reviewer could not read the prices #2490 had just added (#2498)

v4.25.0

Features

• 151M tokens of real work displayed as free (#2496)

v4.24.1

Bug Fixes

• 99.99% of all recorded spend is a fallback ceiling wearing the costume of a price (#2490)

Chores

• CodeRabbit auto-review was off for a quota reason, which is not a reason (#2491)

v4.24.0

Features

• a 2,388-node run timeline with no filter, and a refusal that rendered as "no events yet" (#2488)

v4.23.0

Features

• a run belonging to another org was diagnosable only by raw psql (#2484)
• 100% of our error volume is on the box, and the box reported to nothing (#2485)

Bug Fixes

• round-over-round dedupe suppressed ONE finding in a month, and three silent decisions had no number (#2486)

v4.22.2

Bug Fixes

• setup mode minted no per-claim token, so its landing could never authenticate (#2480)

Refactors

• the git-auth env fold was a function in one lane and a convention in three (#2481)

v4.22.1

Bug Fixes

• a diff nobody fetched whole could post a real APPROVE, and the fan-out ignored the model's budget (#2478)

Docs

• the onboarding marker was backfilled by a migration, not by the escape hatch (#2479)

v4.22.0

Features

• freedom + power + quality > cost — the caps that cost coverage come out (#2475)

v4.21.4

Bug Fixes

• the team lane's clone gate could not see the token the claim had just minted (#2473)
• per-PR worktree isolation never worked once, because a bare clone has no remote-tracking refspec (#2474)

v4.21.3

Bug Fixes

• a full confirm page made every PR absence unprovable, so no coding run could open one (#2472)

v4.21.2

Bug Fixes

• the stacked-PR-links lane edits a customer's PR body and recorded nothing (#2471)

v4.21.1

Bug Fixes

• a 13-file manifest was thrown away because two prose blurbs ran twenty characters long (#2468)
• box logs has never returned a report — the runner-log presign port was never wired (#2470)

v4.21.0

Features

• the platform that watches your errors has never seen one of its own (#2463)

Bug Fixes

• the install page's .maintainer.yml silently turned ON what it claimed to limit (#2462)
• setup proposed refining the CLAUDE.md it had just written, and the e2e pinned that as correct (#2466)

Performance

• ratchet the catalog budget onto a measurement, and make the ceiling observably trip (#2465)

v4.20.2

Bug Fixes

• a coordinator that never answered was reported as three missing manifest fields (#2461)

v4.20.1

Bug Fixes

• onboarding a repo is what stopped its coding lane committing (#2458)
• a model that ignores forced tool_choice threw away a whole survey (#2459)

v4.20.0

Features

• setup writes a specialist roster and the editors that write those files never saw one (#2455)

Bug Fixes

• every task the platform ever assigned left no durable record (#2457)

v4.19.0

Features

• the role slot existed for months and the lane every report and babysit run spawns from passed no role (#2453)

Bug Fixes

• outcome_kind said pr for 495 scout tasks that never open one (#2451)

v4.18.1

Bug Fixes

• a model that answered ALMOST right lost the entire answer (#2450)

v4.18.0

Features

• a repo could not say how its own scout or reviewer should behave (#2449)

v4.17.0

Features

• the scout composed its prompt at import, so no repo could ever shape it (#2448)

v4.16.0

Features

• an out-of-credits account enqueued a review no box could finish (#2446)

v4.15.0

Features

• a raised blocker told nobody, so one was raised and none was ever answered (#2445)

Bug Fixes

• a skipped checkout was surveyed as if it were a tree, and the refresh verdict reached nobody (#2443)

Docs

• the scout paperwork described a scout that does not exist (#2447)

v4.14.2

Bug Fixes

• a worker-pinned box claimed scout sessions and ran them through the coding adapter (#2442)

v4.14.1

Bug Fixes

• the scout's plan call promised repo_ask and mounted only response, so 83% of plans were floor plans (#2440)
• the read-only tool set was restated in seven places and derived in none (#2441)

v4.14.0

Features

• make the onboarding pass leave an old repo actually AI-first (#2432)

Bug Fixes

• the scout reasoned on the cheapest tier because its route was declared and never returned (#2439)

v4.13.0

Features

• mount the blocker verb and give a missing credential a link that supplies it (#2427)
• a blocked scout raises a durable blocker, and it expires (#2431)

v4.12.5

Bug Fixes

• every conditions tick died at gatherSnapshot — unqualified runners.id in the idle subquery (#2420)

v4.12.4

Bug Fixes

• a null optional field threw away the whole review (#2418) (#2419)

v4.12.3

Bug Fixes

• the team lane's git had no credential on a pinned checkout (#2415) (#2416)
• the install stamped a funnel state that never happened (#2400) (#2417)

v4.12.2

Bug Fixes

• the box strings told an agent to fix a box that was correct (#2404) (#2413)
• the 401 with no credential named no way to get one (#2404) (#2414)

v4.12.1

Bug Fixes

• the funnel's only progress signal could be self-declared (#2404) (#2412)

v4.12.0

Features

• /join installs and starts the signed dz-runner service (#2404) (#2411)

v4.11.1

Bug Fixes

• the join lane installed a coding agent and never named it (#2410)

v4.11.0

Features

• an enrolled box was never told where the fleet is (#2408)

Bug Fixes

• box_unhealthy null-guard — an unassigned runner fed '' into the uuid column (#2409)

v4.10.0

Features

• an agent driving MCP could not find out what to do next, or why it was refused (#2406)

Bug Fixes

• the Select repos step contained no repos, and two more steps had no way out (#2407)

v4.9.7

Bug Fixes

• a scout task was held forever by a gate meant for coding work (#2405)

v4.9.6

Bug Fixes

• the published join lane never starts the runner, and the quickstart hands out a file it calls safe (#2403)

v4.9.5

Bug Fixes

• approve resolved a login to one row at random, so three people sat locked out for a month (#2401)
• the install funnel wrote its trail to stdout, so prod holds zero onboarding audit rows (#2402)

Docs

• a PDF engine, and the Spanish commercial deck built on it

v4.9.4

Bug Fixes

• wire REVIEW_GENERATION_BUDGET to its callers, where it was never routed (#2399)

v4.9.3

Bug Fixes

• a re-review no longer repeats a finding it already posted (#2398)

v4.9.2

Bug Fixes

• a failed sample now records WHY, instead of the model's echoed answer (#2397)
• a schema-default change never reached a repo whose config had not changed (#2396)

v4.9.1

Bug Fixes

• a finding is only ever cut at a markdown-safe offset (#2395)

v4.9.0

Features

• the review says WHICH model read the diff and what it cost on your key (#2394)

Bug Fixes

• a queued check run now has a deadline, so it cannot strand a merge (#2393)

v4.8.0

Features

• the reviewer files a real GitHub verdict, and dismisses its own stale block (#2391)

v4.7.4

Bug Fixes

• a truncated diff posted as a complete review, with full approve authority (#2390)

v4.7.3

Bug Fixes

• four filter-chain defects that deleted real findings before anyone could see them (#2389)

v4.7.2

Bug Fixes

• an incremental pass anchored findings against a diff wider than the PR, and GitHub refused the whole review (#2387)
• the reviewer was told to call tools the wire does not bind, and every obedient sample died (#2388)

v4.7.1

Bug Fixes

• every follow-up on a repo with secrets ran against a tree rebased off the PR head (#2386)

v4.7.0

Features

• the reviewer publishes a check run, so a running review is visible (#2379)

v4.6.0

Features

• wire the resource pass into the fleet-alerts tick (#2360 follow-up) (#2377)

v4.5.0

Features

• backlog, no-work, and idle alert tick (#2371)
• config-drift alert + daily fleet digest (#2374)

v4.4.0

Features

• model/API failure classification alerts (#2376)
• resource alerts + sustained-condition state (#2375)

v4.3.0

Features

• box_unhealthy + keys_exhausted alerts (#2372)
• stuck task alert emissions — warden observation + reaper repair (#2373)

v4.2.0

Features

• task started/succeeded/failed alerts on the fleet.alert lane (#2370)

v4.1.0

Features

• the fleet.alert Discord lane + box offline/recovery alerts (#2369)

Docs

• arc 101 status was nine slices behind the code (#2365)

v4.0.0

BREAKING CHANGES

• delete DZ_TEAM_FANOUT and the unrun branch it forked (#2344)

Features

• add per-task hold writer for capability-escalation sweep (#2148) (#2160)
• delete DZ_TEAM_FANOUT and the unrun branch it forked (#2344)
• a repo is not AI-first until something proved it (#2364)

v3.1.0

Features

• a prompt ceiling nobody measured is not a ceiling (#2354)

v3.0.1

Bug Fixes

• remove dead GateRefusedError escalation branch (#2208)
• add missing escalation categories to OPERATIONAL enum (#2114) (#2155)
• raise max_connections to 600 for integration tests (#2240)

Chores

• commit gate-1 (#2201) result files for the fixtures work (#2206) (#2265)

Tests

• add concurrent-append test for janitor audit chain (#2185) (#2188)

v3.0.0

BREAKING CHANGES

• the signature proved the source, never the tenant (#2340)

Features

• the signature proved the source, never the tenant (#2340)

Bug Fixes

• an editor that exited before reading its brief killed the runner (#2353)

v2.2.0

Features

• the review disclosure is a footer, not a header (#2350)
• invitations, and a tool surface measured before it was redesigned (#2351)
• the pr_open stage, so a pushed branch becomes a pull request (#2352)

Docs

• slice 02 was marked done off the PRs carrying its number, not its steps (#2349)

v2.1.1

Bug Fixes

• a Go template miss looked exactly like a missing image (#2347)
• a cross-repo plan stamped one repo over its whole DAG (#2348)

v2.1.0

Features

• the devops role, retrievable memory, and why nothing is moving (#2343)

Bug Fixes

• the digest resolver refused six images it could not explain (#2346)

v2.0.0

BREAKING CHANGES

• six releases shipped with no attestation because a 250-byte file could not be uploaded (#2342)

Bug Fixes

• six releases shipped with no attestation because a 250-byte file could not be uploaded (#2342)

v1.21.0

Features

• a deferred-to reviewer that cannot review left the PR unreviewed (#2332)
• the bot's PR never said how the work reached the fleet (#2335)
• three mechanics that notice, plus the groomer bug that retired its own work (#2336)

Bug Fixes

• a snapshot lost what no migration dropped, and the next generate re-created it (#2338)

Docs

• record what arc 101 actually shipped, and what is blocked (#2339)

v1.20.0

Features

• a scout plan could never run at concurrency above one (#2329)
• one tool call per round trip was the agent's speed ceiling (#2331)
• a hand-maintained exclusion list put the box's own scratch in customer PRs (#2333)

Docs

• the parallel_safe page still described the channel #2329 removed (#2330)

v1.19.1

Performance

• the PR lifecycle asked GitHub what changed 3,754 times a day to learn about 381 changes (#2328)

v1.19.0

Features

• the raw task — one object, one admission door (#2319)

v1.18.6

Bug Fixes

• we published a "no egress" claim into customer repos that was untrue of the config we shipped (#2316)
• a sub-agent's output existed only in memory, so a failed parse was a lost run (#2317)

Docs

• restate the vision as a role cast and plan the arc to the north star (#2304)
• work sources, the claim gesture, and the onboarding checklist (#2305)
• one home for harness scratch, and a harness that wakes the agent (#2306)
• deferred register, and CodeGraph's lifetime + the egress claim we publish (#2307)

v1.18.5

Bug Fixes

• the turn cap was a work budget, so it failed work (#2303)

v1.18.4

Bug Fixes

• a signed attestation named zero images while six of them shipped (#2302)

v1.18.3

Bug Fixes

• the re-entrant checkout lock leaked its slot on out-of-order dispose (#2301)

v1.18.2

Bug Fixes

• a report upload decided whether the code was good (#2300)
• one cancelled follow-up parked a finished PR forever (#2299)

v1.18.1

Bug Fixes

• a babysit run queued behind itself and held the box for an hour (#2298)

v1.18.0

Features

• a refusing box reports its own refusal (#1917) (#2184)
• land box boot-refusal reports as account-anchored audit rows (#1917) (#2190)
• ship the repo's knowledge digest on the assignment, not the template (#2273)
• code.recall — the coding lane can read a KB page, not just its index (#2279)
• the area gate unlocks on proof that two file sets are disjoint (#2278)
• kb_page_get — the coding lane can now actually fetch a page (#2282)
• the built-in harness is the default; spawning a CLI is the opt-in (#2284)
• declare each PR's own files, so disjoint reviews stop queueing behind each other (#2291)
• the agent finally reads the house rules it has been writing for the maintainer (#2292)

Bug Fixes

• attribute a boot refusal to the box, not the release (#1917) (#2176)
• acknowledge DZ_TEAM_FANOUT as deliberate soak flag (#2152)
• a failed dependency no longer deadlocks its dependents forever (#2272)
• the KB gardener could not keep up with its own inflow (#2274)
• monitor mode says out loud that it will never be offered work (#2275)
• widen the boot() contract to match what monitor uses it for (#2276)
• the KB digest stops naming tools the reader does not have (#2277)
• a heartbeat that reports silence is no longer counted as progress (#2280)
• refuse the assignment rather than spawn a coding CLI at the runner's cwd (#2281)
• ui-debugger reached only repos that already had UI tests (#2283)
• every box reported itself idle — derive the current task, drop the column (#2285)
• a merged PR never retired its lane tasks, so every merge leaked a queue row (#2286)
• an 8k output ceiling threw away seven minutes of finished review work (#2288)
• the PR projection drifts from GitHub and nothing repaired it (#2289)
• code.read handed the model 60,000 replacement characters for a PNG (#2290)
• a dead inference connection is no longer indistinguishable from a thinking model (#2293)
• six defects found by pointing the QA harness at a live customer account (#2294)
• the stall guard was timing the model, and killing runs for thinking (#2295)
• tenancy, credentials, and the boxes nothing could read (#2296)
• the coding lane staged the harness's own scratch into customer commits (#2297)

Chores

• un-yank releases condemned by autoroll cascade (#2170)

Docs

• north-star — what the platform is for, from homework to monorepo (#2287)

Other

• Security design: IdP re-authentication flow (#2209)

v1.17.3

Bug Fixes

• wrap unguarded audit appends so a failed append doesn't drop the pass (#2263)

v1.17.2

Bug Fixes

• two deliveries of one error minted two tasks, two boxes, two PRs (#2260)

v1.17.1

Bug Fixes

• a scout plan opted out of the same-area gate while a box held the area (#2261)

v1.17.0

Features

• the agent that did the work writes the pull request (#2258)

v1.16.5

Bug Fixes

• the supersede guard was declared, implemented, and never connected (#2259)
• the provider label named the wire, so every OpenRouter run said openai (#2256)

v1.16.4

Bug Fixes

• every commit the fleet ever pushed was attributed to nobody (#2257)

v1.16.3

Bug Fixes

• an abandoned attempt's edits rode into the next run's pull request (#2255)

v1.16.2

Bug Fixes

• the stall guard killed runs for using a timeout the tool granted (#2253)
• multi_bash declared no budget at all, and its timeout is per command (#2254)

v1.16.1

Bug Fixes

• the floor named the coding task after the scout that produced it (#2249)
• a reply_bot follow-up pushed a file it invented onto the branch (#2250)
• make matching report-outcome redelivery a real no-op (#1969 follow-up) (#2251)
• revert #2250 and fix what actually caused it — an impossible verb (#2252)

v1.16.0

Features

• top-up degraded ensemble bursts and floor approve authority on sample completion (#2083) (#2196)
• let .maintainer.yml map qualification labels to a repo's own names (#1555) (#2244)
• keep a PR's follow-ups on the box that holds its tree (#2242)

Bug Fixes

• a CI fix that never reached the PR it was minted for (#2236)
• a superseded CI run made the fleet fix green pull requests (#2243)
• a box that restarts parked its task for three hours (#2241)
• the harness was committing its own telemetry to customer branches (#2245)
• the follow-up's git fetch carried no credential at all (#2246)
• a follow-up that can never be re-minted now says so (#2247)
• a completed report task never had its outcome_kind corrected (#1969) (#2248)

v1.15.4

Bug Fixes

• a degraded review ensemble may not post a real APPROVE (#2083) (#2237)

v1.15.3

Bug Fixes

• add resolved state to prevent terminal declines (#2227)
• add resolved state to prevent terminal declines (#2227) (#2238)

Chores

• add task marker for #2210 (#2234)

v1.15.2

Bug Fixes

• write_creds converges ownership when creds already exist (#2215) (#2233)

v1.15.1

Bug Fixes

• the integration lane ran out of Postgres connections, not assertions (#2231)
• a native box could code and could not scout (#2230)
• every shipped PR was recorded as a failed run (#2229)

v1.15.0

Features

• a built-in coding harness, so the box stops shelling out (#2220)

Bug Fixes

• a queued review stalls every merge behind it (#2221)
• an embedding provider can hang the task-create request, twice (#2222)
• more boxes bought nothing, and a wedged lock took the pool (#2224)
• a scout on a BYOVM box plans against a repo it cannot read (#2219)

Docs

• the enrollment runbook did not survive a fresh box (#2218)

Tests

• add cross-pod concurrent-append test for (session_id, prev_hash) retry

v1.14.2

Bug Fixes

• classify config/ deliberately — dev/test GREEN, everything else parks (#2210) (#2212)
• boot-guard every login redirect against the route table (#2173) (#2217)
• compose the follow-up commit subject from the staged paths (#2192) (#2225)

Tests

• live-Postgres cross-pod race + non-chain never-retried pins (#2185) (#2216)

v1.14.1

Bug Fixes

• publish authorization_endpoint on the API origin (#2183)
• inject DZ_INSTALLATION_TOKEN into the coding lane's checkout env (#2211)

v1.14.0

Features

• consume nc.status() at the dial seam, alert on permission denials (#2193) (#2197)

Bug Fixes

• add 7-day expiry to NATS creds and clear natsCredsRef on revocation
• gate the completed line on the same pr the checklist reads (#2115) (#2199)

v1.13.1

Bug Fixes

• drain needs-area / needs-size and remember a maintainer-side decline (#2162) (#2194)

v1.13.0

Features

• tell an issue author what the pipeline cannot run, at filing time (#2123)

Bug Fixes

• OIDC_REDIRECT_URI points at /auth/sso/callback (#2173) (#2180)

v1.12.0

Features

• narrate the terminal failure as a bounded Reason line (#1966) (#2187)

Docs

• declare this repo's admissible owner handles (#2165)

v1.11.0

Features

• durable janitor audit chains on platform parent (#2116) (#2179)

Tests

• the failed-ledger-read case refuses something, instead of being starved (#2166)

v1.10.0

Features

• resolve BASE_REF from newest successful run (#2178)

v1.9.2

Bug Fixes

• add SESSION_COOKIE_DOMAIN to production.yml (#2182)

v1.9.1

Bug Fixes

• the origin's HTML pages were hardcoded light, and each held its own palette (#2175)

v1.9.0

Features

• add unlimited_spend GrantKind (#2163)

Bug Fixes

• defer completion when ledger read fails on completing round (#2164) (#2169)

v1.8.5

Bug Fixes

• derive pull ledger bound from channel window constant (#2167)

v1.8.4

Bug Fixes

• make box-capability probe fault visible in setup (#2159)
• queue depth cap allows one task for Free/OSS tiers (#2161)
• translate the AI-SDK base URL at one seam, and resolve the flat-rate default at read time (#2122)

Tests

• add .mcp.json integrity guard (#2158)

v1.8.3

Bug Fixes

• a fleet PR is rendered before its run can say which model wrote it, so it never names one (#2154)

v1.8.2

Bug Fixes

• refusing a cyclic tool-service config costs 2.2 seconds of blocked event loop (#2153)

v1.8.1

Bug Fixes

• the box names a provenance path it never created, so the harness's append ENOENTs into silence (#2150)

Tests

• the #513 board test asserts two stream chunks land in one tick, which is true only on a loaded box (#2151)

v1.8.0

Features

• a fleet PR names the model that actually drove it, and links to the run that shows how (#2146)

v1.7.0

Features

• add full escalation vocabulary to issue/pr loops (#2139)

v1.6.0

Features

• add observer_blind to box offline payloads (#2126) (#2137)

v1.5.0

Features

• advertise and sync box capability tokens on presence beats (#2136)

v1.4.1

Bug Fixes

• recover escalation category from audit trail (#2114) (#2135)

v1.4.0

Features

• reconcile diverged local base via parked ref (#2134)

Bug Fixes

• the rollout target moves forward, so a build published mid-pass is not stranded (#2124)

v1.3.0

Features

• stamp observer_blind on sweep offline alerts (#2133)

Tests

• pin .mcp.json as fail-closed UNKNOWN (#2132)

v1.2.0

Features

• add per-event email list-unsubscribe (#2131)

v1.1.0

Features

• add resolver_refused widget page refusal (#2130)

Docs

• delete stale 1847 live-test-fence-ref plan doc (#2129)

v1.0.0

BREAKING CHANGES

• delete the digest policy block, which was documented and unreadable (#2095)
• the spend cap meters real model spend, and has no platform default (#2097)

Features

• delete the digest policy block, which was documented and unreadable (#2095)
• give RunnerCapability.Ci a way to be granted, and prove every capability is reachable (#2096)
• a first-class router provider on the Anthropic wire (#1904) (#2102)
• setup mode drafts a per-repo specialist roster (#2121)

Bug Fixes

• a rate limit is not a permanent verdict on a BYOK key (#2094)
• the spend cap meters real model spend, and has no platform default (#2097)
• each failure gets the remedy that fits it, and the comment stops promising a retry the guard refuses (#2093)
• fold written-back memories into the boot prompt instead of discarding them (#2105)
• the install outcome carries its cause, so each failure gets one true sentence (#2101)
• give every capped code.* read a door to the rest (#2108)
• size the loop's ceilings for the world it runs in (#2110)

Chores

• run gsd on kimi k3 for the overnight sweep (#2106)

Docs

• correct the drift that would have someone re-add a leash (#2113)
• write down the end goal, and the test that decides what serves it (#2119)

v0.44.1

Bug Fixes

• a bearer credential sends the bearer alone, not an empty x-api-key beside it (#2090)
• normalize the status bun actually returns, and stop asserting numbers nobody measured (#2092)

v0.44.0

Features

• the router provider factory, built and tested standalone (#2078) (#2088)

Bug Fixes

• a fresh worktree bootstraps its own dependencies before the gate runs (#2082)
• stamp the collection into every keyset cursor (#2053) (#2072)
• the bootstrap survives a concurrent gate leg, and its own guards can fail (#2087)

Docs

• the scout-plan consumer is not the Dispatcher, and it is not hypothetical (#2089)

v0.43.0

Features

• give kb.background.max_pages a reader, and guard the next orphan (#2041) (#2073)

Bug Fixes

• the structured probe says which fault it means, instead of one sentinel for two (#2084)

v0.42.4

Bug Fixes

• the scout dispatch seam named the wrong mothership consumer (#2042)

v0.42.3

Bug Fixes

• an undecodable widget page cursor refuses, instead of serving page 1 (#2049) (#2070)

v0.42.2

Bug Fixes

• a halt condemns on the durable record, not on its own sample of one (#2080)

Chores

• run gsd on claude-sonnet-5 (#2079)

v0.42.1

Bug Fixes

• an empty route and a missing board render as themselves, not as a blank page and a generic error (#2074)

v0.42.0

Features

• a daily lane that levels every account's built-in dashboards (#1901) (#2047)

Bug Fixes

• key the replay gate on the signed body, not the unsigned delivery id (#2058)
• make the review lane and the fleet git_token verb write rows that survive (#2063)
• reach both ownership anchors, and delete the heartbeat route (#2068)
• stop re-probing a permanently broken BYOK key on the customer's own dime (#2061)
• classify z.ai credit exhaustion, and size compaction off the registry (#2069)
• give the seam audit row a target, so two opposite decisions stop looking identical (#2064)
• lease a concurrency slot per session, not per counter key (#2065)
• make the public docs say what the code actually does (#2066)
• refuse the input instead of silently doing the wrong thing (#2067)

v0.41.4

Bug Fixes

• deny-by-default env for the coding shell, and a failed run stops reporting success (#2043)
• a mail scanner opening an action link is no longer the approval (#2045)
• whole-tree guards stop being gated on a subset of the repo (#2048)
• namespace the SSO identity anchor per connection (#2046)

v0.41.3

Bug Fixes

• sweep 4 - frontend, support packages and the guards (#2039)

v0.41.2

Bug Fixes

• sweep 3 - the HTTP edge and the agent surface (#2035)

v0.41.1

Bug Fixes

• sweep 2 - orchestration (#2030)

v0.41.0

Features

• a scout that finds a blocker stops instead of dispatching the plan (#2031)

Bug Fixes

• drop runners.offline_since — a column with no reader and no writer (#2023) (#2028)

Tests

• drive the run-view and epic stop controls in a browser (#1981) (#2026)

v0.40.0

Features

• an Org Health built-in dashboard over the served read models (#1988) (#2025)

Bug Fixes

• fence the fleet git_token verb on the caller's held claim (#2021)
• declare agent.default devops and stop rendering a null deploy line (#1989) (#2011)

v0.39.0

Features

• add offlineSince column to runners table (#1929)

Bug Fixes

• pin the rollback-preserved binary to 0755 (#1922)

Performance

• Anthropic prompt-cache breakpoint on the system part (#1881)

v0.38.1

Bug Fixes

• sweep 1 - 22 defects in the core contract and data layer (#2020)

v0.38.0

Features

• sync a box's address from NATS, which already knows it (#2010)

Chores

• run gsd on kimi k3, not the kimi-for-coding alias (#2008)
• grant standing merge authorization (#2009)

v0.37.4

Bug Fixes

• the swap preflight stops claiming more than it proves (#1993)
• no box has recorded an address since liveness moved to NATS (#2004)

v0.37.3

Bug Fixes

• repo-kind sessions were unreapable, and the lock key was on the row all along (#1998)

Chores

• register the gsd engine on this repo (#1996)

Docs

• the email layer is our own SMTP, and a guard so it stays said that way (#1997)

Refactors

• retire every feature flag in the repo, each with the branch it guarded (#1999)

v0.37.2

Bug Fixes

• the attest job ENOENTs on every runner-only merge (#1994)

Refactors

• delete dead code, and wire the billing path that was dead (#1992)

v0.37.1

Refactors

• delete the deprecation channel, the empty shadow registry, and the ?limit alias (#1991)

v0.37.0

Features

• retention for every published artifact, and a 5-minute adoption cadence (#1984)

Bug Fixes

• a box that can never swap says so at boot, not on every release (#1985)

v0.36.0

Features

• surface the stop control on the fleet views (#1906) (#1980)

Tests

• attribute the sealed-fence assertions to the boxes the test created (#1976) (#1979)

v0.35.4

Bug Fixes

• fleet PRs now carry a Closes reference, so delivered issues close (#1822) (#1978)

v0.35.3

Refactors

• collapse the fleet event sink's three fences into one (#1951) (#1975)

v0.35.2

Bug Fixes

• a floor review no longer satisfies review presence on the bot lane (#1974)

v0.35.1

Bug Fixes

• backfill the dead-owner orphans #1913's rescue can never reach (#1955) (#1973)

v0.35.0

Features

• a devops agent template and its GitOps constraint (#1897) (#1970)

v0.34.1

Bug Fixes

• decide the account-wide key veto account-wide (#1953)

v0.34.0

Features

• shard the MCP write budget per conversation, under a token ceiling (#1895) (#1971)

Tests

• the dead-owner rescue's real wiring never exercised the Cancelled arm (#1968)

v0.33.2

Bug Fixes

• dedup DEAD_OWNER_STATUSES, add wired-path integration test for #1913's dead-owner rescue (#1959)

v0.33.1

Bug Fixes

• a malformed release manifest reads identically to absent and healthy (#1965)

v0.33.0

Features

• a daily digest of dead keys, quiet boxes and stuck work (#1900) (#1958)

Policy

• restore pr.require_review — review absence refuses a merge again (#1963)

Tests

• the env-seam guard scans the tree once, not three times (#1956) (#1960)
• make the stream tests load-immune — assert the scheduled delay, not the wall clock (#1962)

v0.32.0

Features

• review a fleet PR against its issue's acceptance criteria (#1912)

v0.31.2

Bug Fixes

• anchor the scout-plan ingest to the scout task, not to the payload (#1943) (#1957)
• reassign locks the row it reports the prior runner from (#1946) (#1948)

v0.31.1

Bug Fixes

• a fleet PR orphans when its owning run dies after opening it (#1913) (#1926)

v0.31.0

Features

• a stop control for a running task or plan (#1898) (#1944)

Bug Fixes

• a task held on no usable keys says so on the read path (#1899) (#1947)

v0.30.0

Features

• carry a conversation_ref from the ask to the PR body (#1942)
• duration-based box_offline alerts — a box dark 30min pages hourly (#1921) (#1924)

Bug Fixes

• findable rename, filled tiles, no highlight borders, stoplight status (#1954)

Docs

• the requireSecret seam is sanctioned, and its set is pinned (#1940) (#1952)

v0.29.0

Features

• the goal board's controls look like controls (#1949)

v0.28.0

Features

• filter the goal board, and rename a goal at the source (#1945)

v0.27.0

Features

• add 'dz' BYOK key pointing to router pool (#1933)

Bug Fixes

• flaky sessionStream tests under sharded full-suite run (#1934)

Refactors

• split update-check.ts per invariant — arming and the pull ledger move out (#1801) (#1931)

v0.26.0

Features

• the autonomous goal monitor (#1918)

Tests

• pin the #1894 reserved-env refusal to both write edges (#1941)

v0.25.5

Bug Fixes

• fence agent_templates.invocation.env against reserved names (#1928)

v0.25.4

Bug Fixes

• the anthropic provider ignores base_url (#1930)

Tests

• pin the restored binary's 0o755 mode bit in update-boot (#1882) (#1923)
• pin the ops_action lane as unreachable and record its trip condition (#1939)

v0.25.3

Bug Fixes

• carve the fleet's own PRs out of the bot-actor defer (#1911)

v0.25.2

Bug Fixes

• a scouted bug notifies the account that owns it (#1870) (#1880)
• rollout actions leave a durable record, not a log line (#1844) (#1883)
• a destructive verb records the token that acted (#1845) (#1884)

v0.25.1

Bug Fixes

• the review lane provisions its grounding tree instead of guessing (#1614) (#1879)

v0.25.0

Features

• bind a commit to the bytes it produced, and check it (#1612) (#1878)