Versionshinweise
Versionshinweise zum Produkt
Hier siehst du die Änderungen der veröffentlichten Agent-Studio-Versionen.
Die Einträge stammen beim Website-Build aus der CHANGELOG.md des Produkts. Die Originalnotizen stehen auf Englisch. Änderungen eines noch nicht veröffentlichten Stable-Checkouts erscheinen erst nach einem Produkt-Release.
v0.9.5 · 2026-10-04
Agent Studio v0.9.5
Nachtrag nach der Betriebswoche: 189 Commits seit 0.9.4 und 43 abgeschlossene Karten. Das Token-Ledger bepreist GPT-6 Sol wieder; der Stable-Katalog verwendet TokenEconomy 0.3.5. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Added: Agent Studio consumes TokenEconomy 0.3.5: registry entries, CLI floors, provider-rejection fallbacks and migration proposals for claude-opus-5-5, gpt-6-sol and gpt-6-luna; the ledger prices GPT-6 Sol and Luna runs (AGT-2892).
- Added: Installer: one executable for Windows and Linux that installs Agent Studio on Docker or as native services (AGT-2738).
- Added: Operations Server backchannel beside the Task Server (AGT-2907).
- Added: Decision cards: kind 'decision' with structured options, decide and reopen API and computed blocking (AGT-2927); badge with decider, project header count and card detail with options (AGT-2928).
- Added: Review concerns get one automatic fix round; every run records who triggered it and why (AGT-2887).
- Added: Escalation view: parked operator session with legible contrast and an on-demand excerpt of the run's model and prompt (AGT-2902).
- Added: Connector security for a remotely attached Studio: Credential Manager, Origin and CSRF checks (AGT-2984).
- Added: Flaky quarantine with expiry, and guard tests as the first gate step (AGT-2918).
- Added: Run budget per card: adaptive run timeout and continuation instead of 90-minute timeout escalations (AGT-2998).
- Added: Diagnosis contract: every gate and review failure is classified by experiment before it charges a card (AGT-2916).
- Added: Credential registry and effective-source monitoring (AGT-2971), deployment secret transport and WireGuard renewal (AGT-2977), installation upgrade and rollback ownership (AGT-2947), connectivity manifest with a single link owner (AGT-2945), one-box profile (AGT-2942).
- Added: Work is routed to capable host classes with bounded admission (AGT-2939).
- Added: Workspace usage projection and shared usage chips (AGT-2960, AGT-2961).
- Changed: Integrated-only human review: admission guard, accept confirmation, guarded archive, self-healing integration (AGT-2920).
- Changed: The card's Result summarises the task, not the last round (AGT-2880).
- Changed: Remote stop requests are durable and attempt scoped (AGT-2935); review settlement continuation is persisted before acknowledgement (AGT-2936).
- Changed: Operator workstations use the runner-host contract (AGT-2940).
- Changed: Gate economics: review reuse and impact selection shorten the full-suite merge gate (AGT-2997); the per-project delivery queue no longer waits behind agent rounds and one gate slot (AGT-2994).
- Changed: Deterministic bounce rail for the first recoverable conflict-skipped round (AGT-2919).
- Changed: Bounded task core served from the index (AGT-2953); persistent recovery overlays removed from task navigation (AGT-2959).
- Changed: Link migration: the running forward is adopted and AgentRunner-TunnelKeeper retired after the seven-day soak (AGT-2764); tunnel loss and fenced recovery rehearsed for coding and review (AGT-2937).
- Changed: Documentation drift audit: every page under docs/system and docs/operations verified against the current code (AGT-3004); Dossier review round 2 synthesised into AGT-W53 (AGT-2900).
- Fixed: Merge gate on Windows: cmd.exe no longer destroys the quoted dotnet test logger arguments (AGT-2912).
- Fixed: An oversized result upload (HTTP 413) no longer discards a finished remote delivery (AGT-2915).
- Fixed: Frontend unit suite is green on a de-DE Windows workstation (AGT-2980); the red side-sheet pin spec baseline on develop (AGT-3007).
- Fixed: The card gate renders the scenario Compose stack so overlay drift is caught before the promotion train (AGT-2981); the train's scenario smoke tolerates a loaded runner host (AGT-2982).
- Fixed: Review lane silent for 24 h: library-v1 plans no longer require toolchain capabilities the executor never registers (AGT-2987).
- Fixed: Test-suite audit: time bombs, clock dependencies, line-ending and platform assumptions removed from the test suites (AGT-3003).
- Fixed: Acceptance-rail correctness findings from the Quality Studio review (AGT-2990).
- Fixed: Security hygiene wave: vulnerable runtime dependencies, secret-scan baseline, bounded whole-file reads (AGT-2991).
- Fixed: Stable hotfixes from 4 October, merged back from main (88bee6805..c076d2da9): backend build and test gates run only on the remote Linux worker and fail closed without it; integration status reads each Git origin once per projection instead of once per card (180 to 2-13 Git spawns per index run); deployment smoke budgets the worker cleanup before run.completed; board review reads no longer depend on authority persistence.
v0.9.4 · 2026-09-27
Agent Studio v0.9.4
Fehlerkorrektur für den Produktions-Runner: Seine Abschlüsse wurden mit 0.9.3 wegen einer abweichenden Attempt-ID abgewiesen. Die Fortsetzungsbelege verwenden wieder die zum Lease gehörige Attempt-ID. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
v0.9.3 · 2026-09-27
Agent Studio v0.9.3
Zweites Durchsatz-Release des Betriebswochenendes: 63 Commits seit 0.9.2 durchliefen die Work-Package-Prüfungen auf dem Windows-Gate-Host. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Added: Agent Studio is Docker-deployable: one Compose file runs the complete product from published images, with backup, restore and a source-built path (AGT-2736).
- Added: Single-host Task Server rehearsal: control plane in Docker on the runner host, Runner credentials fenced from Studio routes, operations runbook and coordinated backup (AGT-2737).
- Added: Integration dead ends get an exit: a continuation extends the task with the conflict context instead of parking it (AGT-2909).
- Added: Bounded session continuation for mechanical rebase rounds with a durable per-generation token ledger (AGT-2921).
- Added: Per-SHA gate cache: the gate verdict is bound to the tested tree and an unchanged re-run is skipped (AGT-2923).
- Added: Claimable gate step: GateSubject, GateAttempt, lease and fenced report for the build and test gate (AGT-2925).
- Added: The Task Server steering boundary is enforced for continuation, stop and report paths (AGT-2933).
- Added: Quota fallback picks the comparable model from the Token Economy catalogue for every Claude call (AGT-2889).
- Added: Stable task-switch baseline correlated for the immediate task switching Dossier (AGT-2952); the Dossier is registered in the North star map (AGT-2958).
- Changed: Runs no longer stop with NeedsInput to request approvals that the pipeline grants after delivery (AGT-2931).
- Changed: Frontend unit tests run with a pinned en-US locale so the suite is deterministic on every workstation; the merge gate on a de-DE host is green again (operator commit 4da36b4a7).
- Changed: New Dossier on develop: credentials and logins as a managed process (AGT-2968).
- Fixed: Deployment regression scenario: the Compose overlay follows the bootstrap-generated secrets topology of AGT-2736 (injected scenario tokens, no published images), so the release contract test and the promotion train render the stack again.
- Fixed: Preparation cache: an empty block is no longer published; an incomplete block is rejected and rebuilt (AGT-2891).
- Fixed: Pickup no longer runs a card on claude-haiku-4-5 when the pinned Claude model is unknown to the installed CLI (AGT-2893).
- Fixed: Prompt enrichment no longer appends Agent Studio instruction blocks to cards of other projects (AGT-2908).
- Fixed: v1 plane: result finalization and lease release after the fenced completion carry the exact outbox authority instead of answering 409 (AGT-2967).
- Fixed: Interactive chat turns no longer wait behind the coding slots; chat turns have priority and heavy turns account against coding capacity (AGT-2969).
v0.9.2 · 2026-09-26
Agent Studio v0.9.2
Durchsatz-Release vom 25. September: 51 Commits seit 0.9.1 wurden über die Work-Package-Prüfungen integriert. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Added: Project Git view: resizable columns, a sweep bar inside its panel, a scrolling commit graph and commit details on the side (AGT-2886).
- Added: Pipeline aspect steps show an honest concerns status, direct report access, parsed verdict markers and readable tables (AGT-2888).
- Added: Oversized result artifacts no longer cost the delivery: code is delivered first, the artifact upload is bounded and a 413 is typed (AGT-2890).
- Added: Execution Hosts shows and alarms when a runner host release lags the Stable release (AGT-2826).
- Added: Deterministic test scoping: a diff-to-project mapping filters the gate suite without spending tokens (AGT-2924).
- Added: Runner: a provider marked limited says why, until when and how to clear it; an operator-terminated CLI is reported as such (AGT-2873).
- Changed: A card may only report delivered with proof: an integrated delivery or a named deliverable without code (AGT-2817).
- Changed: The six merged Dossiers are recorded as superseded and the rolling log as historical; they move to History (AGT-2922).
- Changed: New Dossiers on develop: Decision cards (AGT-2795), the Task Server as the bus between the engine and the daemons (AGT-2905), the deployment story from one box to many runners (AGT-2906), immediate task switching with a bounded core (AGT-2910), the header as a responsive usage cockpit (AGT-2913), the remote host integrates and the local Studio observes (AGT-2881), Dossier review round 2 for the active workbenches (AGT-2899).
- Fixed: Merge gate: a torn preparation cache (missing .nupkg in the cache run directory) is classified as a gate environment failure and retried, not as a code failure (AGT-2901).
- Fixed: LinkSupervisor: a dead client session that still holds the remote listener no longer makes the cleanup probe time out; the stale session is fenced and replaced (AGT-2911).
- Fixed: integration/retry accepts a passed delivery whose review attempt passed on the delivery SHA (AGT-2914).
- Fixed: Runner: post-completion result uploads on the v1 Task Server plane carry the exact outbox authority; without it the standalone Task Server answered 409 and the evidence artifact was lost (AGT-2890 follow-up, found by the deployment regression scenario in the 0.9.2 promotion gate).
v0.9.1 · 2026-09-25
Agent Studio v0.9.1
Fehlerkorrektur für Work-Package-Prüfungen auf Windows: Git Bash erhält die Quotierung der dotnet-test-Befehle, die zuvor durch cmd.exe verloren ging. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
v0.9.0 · 2026-09-25
Agent Studio v0.9.0
Betriebs- und Integrationsrelease mit 80 Commits seit 0.8.0. Es korrigiert unter anderem Integrationsstatus, Modell-Fallback, Gate-Wiederholung und die Berechnung gecachter Codex-Tokens. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Fixed: Auto Review wait classification now uses the restart-resume policy's integration ancestry decision, so a delivery merged by a later gate is reported as awaiting its completion transition instead of awaiting delivery integration (AGT-2860).
- Fixed: Lost-worker recovery now retains the latest diagnostic stderr and system lines across long worker logs, so earlier crash evidence is not silently dropped after each 2,000 scanned lines (AGT-2870).
- Fixed: Automatic timeout-salvage continuation now rolls back its prompt note and clean-context override when the card cannot be moved to Ready, so a refused continuation does not leave partial preparation on the task (AGT-2861).
- Fixed: Worker resource envelope: the task ceiling is 512 per core of fair share (minimum 1024) instead of 128. pids.max counts threads, and a worker with a .NET test host hit the 307-task ceiling of a 12-core, 5-slot host; the runtime aborted with PALSEHException and the run was lost.
- Fixed: Runner hosts no longer accumulate processes that outlive their run. Detached coding and review workers start with MSBUILDDISABLENODEREUSE=1 and DOTNETCLIUSEMSBUILDSERVER=0 and shut the build servers down on teardown, worker teardown kills whatever is left in the worker cgroup and reports the count on the worker-envelope line, and daemon start moves processes no worker owns any more out of the unit cgroup (killing only aged strays of finished attempts) so cgroup delegation succeeds on a host that has run before instead of logging applied=no and running uncapped (AGT-2868).
- Fixed: A result transfer that fails because the Task Server is restarting is now retried from the persisted runner slot instead of waiting for the next daemon restart. The Coding slot stays in phase finalizing with its retry bookkeeping and the delivery it already secured, the worktree and durable worker result are retained, the lease is not released, and the running daemon's poll loop re-drives the same idempotent startup-reconciliation step once /api/system/about answers again (15 s, 30 s, then every 60 s). Retries do not stop at the run timeout; outliving it only adds a journal line. The journal carries one line per retry with its reason and the final completion names the retry count (AGT-2869).
- Fixed: A remote run whose heartbeat is older than the lease it was renewing, or a remote-routed Progress card whose job-folder replay has stopped, is projected as remote-stale with the last runner event named, instead of continuing to read as remote-running. An operator and the acceptance rail can now tell a phantom run from a live one (AGT-2869).
- Added: Onboarding declares the kernel limits parallel test suites need: /etc/sysctl.d/90-agent-runner.conf sets fs.inotify.maxuserinstances to 1024 and fs.inotify.maxuserwatches to 1048576, applies them, and verifies both before the systemd phase (AGT-2868).
v0.8.0 · 2026-09-18
Agent Studio v0.8.0
Betriebsrelease: getrennte cgroups begrenzen Coding- und Review-Worker auf Linux-Hosts; Update-Vorbedingungen werden vor dem Neustart geprüft. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Added: Per-worker resource envelope on Linux runner hosts: each detached coding and review worker runs in its own cgroup under the role unit (Delegate=cpu pids, DelegateSubgroup=daemon) with cpu.max, cpu.weight and pids.max derived from RUNNERHOSTCODINGSLOTS and RUNNERHOSTREVIEWSLOTS; every run reports a worker-envelope line with CPU seconds and peak tasks, review admission clamps to the same budget, and onboarding writes the unit lines and peer slot counts (AGT-2866).
- Added: Update Service preflight evaluates the preconditions of every post-restart verification step against the running instance and refuses with a named, actionable error when one cannot pass (AGT-2865).
- Added: Review gate coverage: runner.Tests and task-server.Tests run as deterministic verify commands, so runner and Task Server races are caught in review instead of in the promotion gate (AGT-2864).
- Changed: The db-touch sentinel (POST /api/internal/probe) has its own gate, UpdateService:ProbeEnabled; unset, it is open exactly while .metadata/stable-approved-tag exists. DevTools:UpdateStableEnabled still opens it for compatibility but is no longer part of the contract (AGT-2865).
- Changed: CliProcessReaperTests runs in a non-parallel collection: it asserts the process-global reap total, which another collection reaping a live workspace could advance mid-test (AGT-2864).
- Fixed: Remote review: deleting a settled review slot record is final. A heartbeat renewal whose persistence was scheduled after the terminal delete used to write the record back, leaving an orphan slot that the next daemon generation adopted and reported a second time (AGT-2864).
- Fixed: setup.Tests: the join-token file expectation is normalised with Path.GetFullPath on every platform, so the Windows pre-develop gate no longer fails on a path separator (ef81146ae).
v0.7.0 · 2026-09-17
Agent Studio v0.7.0
Betriebsrelease mit Korrekturen am Update Service, einer Fortsetzungsrunde nach einem Laufzeitlimit und genauerer Zuordnung von Review-Versuchen zu Releases. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Added: Automatic continuation round after a run timeout with a salvage commit, bounded to one per delivery generation; escalations name the salvage ref (AGT-2861).
- Added: Review attempt provenance: release id and worker binary in the grade and in the adoption log line; drain mode for the review daemon restart (AGT-2863).
- Fixed: Update Service: the frontend probe tries localhost, 127.0.0.1 and ::1; a frontend-only verification failure after a successful backend restart ends the run as degraded with listener evidence instead of a rollback (AGT-2862).
- Added: Review worker release provenance: every review attempt records the agent-host release and resolved binary path of the detached worker that produced its verdict. The daemon logs worker-release / daemon-release on adoption and one notice per adopted attempt running a superseded build, the grade file's Immutable subject proof block names both releases, and the card carries a reviewgradedbysupersededrelease timeline entry and the release pair on its review projection (AGT-2863).
- Added: Opt-in review release drain: agent-runner-deploy --restart-review-drain drains Review before the unchanged promote step, and RUNNERREVIEWRELEASEDRAIN=1 makes an already restarted daemon hold claims until its adopted superseded attempts finish. Neither mode ever ends an adopted worker (AGT-2863).
v0.6.0 · 2026-09-17
Agent Studio v0.6.0
Betriebsrelease mit gezielten Testpaketen vor dem Merge, stabileren Update- und Review-Prüfungen und einer Annahmeansicht, die denselben Integrationsstand wie die Karte liest. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Added: Pre-develop gate work-package matrix: backend-only, frontend-only, mixed and docs-only diffs select the matching test commands; the selection is logged per gate run (AGT-2854).
- Added: Stale-branch sweep: all-namespace classification, report and operator-confirmed reclamation with a project panel (AGT-2794).
- Added: Activity view renders Claude toolprogress and agent events as events, with synced CLI docs and a runner-group DOM spec (AGT-2814).
- Added: Branch-defect attribution and integration-branch gate-health alerting; the three largest components split with lowered baselines; one lint chain (AGT-2819).
- Added: Automatic continuation after a task-server restart during the pre-develop gate: the un-gated merge is detected and re-verified (AGT-2849).
- Added: Temp and cache hygiene: test fixtures remove their temp roots, a suite-level guard keeps the temp root from growing, the review worker redirects and cleans its own temp residue, and the pilot preparation cache lives under a bounded, documented product-owned root (AGT-2858).
- Fixed: UpdateServiceRestartIdentityDrillTests no longer fails under gate load (AGT-2855); BoardConditionalReadTests, DuplicateTaskKeyTests and V1ReviewPlaneDiagnosticsEndpointTests share no state with their neighbours any more, including a real race in TaskJsonFile / TaskMutationService (AGT-2859).
- Fixed: The acceptance rail consults the card's integration record and the delivery ancestry instead of a separate "pending" computation, and no longer retries every 30 seconds (AGT-2856).
- Fixed: The review executor's TmpMountTornDown classifier ignores signature text inside test names and output once a test-run summary exists (AGT-2857, operator hotfix fcdb68b24).
- Fixed: A task-server restart no longer exhausts the auto-review post-processing deferrals for cards with a passed review; killed gates whose merge was pushed later complete on the next pass (AGT-2860).
- Fixed: The dead overview-pane/lifecycle-phase.util re-export shim is gone (AGT-2819).
v0.5.0 · 2026-09-17
Agent Studio v0.5.0
Betriebsrelease: Das Windows-Integrationsgate wiederholt einmal fehlgeschlagene Tests auf demselben Build. Board, Token-Anzeige, Remote-Ergebnis und Update Service zeigen bzw. verwenden präzisere Zustände. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Added: A build gate that never reaches a verdict is recorded as an interrupted run. The pre-develop gate writes a durable in-flight record before the merge, and startup recovery rolls the integration branch back to the exact pre-merge tip - the oldest one when several un-gated merges stacked up - or resumes the merge whose verdict for that exact SHA is already durable. Affected cards are integrated again without a new remote review. A branch whose rollback anchor is gone or already published escalates instead of being rewritten (AGT-2849).
- Added: Windows integration gate: one targeted re-run of exactly the failed tests on the same build, green re-run keeps the gate green and records the names as flaky under the review executor's classification (AGT-2853).
- Added: Review executor: baseline verify results are cached per repository, baseline SHA and command (AGT-2843).
- Added: A parked card says why it is parked (AGT-2816); a card that cannot be picked says so where it claims to be queued (AGT-2818); "delivered" requires an integrated delivery or a named deliverable without code (AGT-2817).
- Added: Token panel on the task Overview shows the reasoning level and separates the run hierarchy from the pipeline steps (AGT-2811); "Current" marks the live run, not the newest one (AGT-2810).
- Added: ETag/304 for task list and grouped reads with a trimmed payload (AGT-2703).
- Changed: integration.status=integrated now requires the delivery to be reachable from the pushed remote integration branch. A delivery that only a local branch or a worktree ref can see reports the new merged-locally status, naming the commits origin/<branch> cannot reach. Acceptance and the acceptance rail keep moving such a card, because the origin push is a separate, backstopped step (AGT-2849).
- Changed: The completion contract, the archive guard, and the Human Review accept action read the same merge question, so a delivery awaiting only its push is no longer refused as an unintegrated delivery and is no longer offered a second merge. The badge still names the missing push (AGT-2849).
- Fixed: The pre-develop merge gate honours the gate-run budget policy (AGT-2846).
- Fixed: Review command watchdogs (silence, no-CPU-progress) are carried into the detached review worker and no longer kill healthy backend test runs (AGT-2851).
- Fixed: Adaptive review parallelism advisor reads the remote attempt queue instead of the local post-processing queue (AGT-2848).
- Fixed: A degraded result summary no longer turns a delivered remote run into a lost run (AGT-2850).
- Fixed: Update Service: the runtime identity check reads the intended build manifest so an upgrade can pass (AGT-2847).
- Fixed: UpdateServiceRestartIdentityDrillTests isolate their update-service factory state (AGT-2852); GateFlakyRerunBehaviorTests run on Windows (operator fix 651d7267a).
v0.4.0 · 2026-09-16
Agent Studio v0.4.0
Betriebsrelease mit automatischer Wiederholung nach Review-Infrastrukturfehlern, Branch-Bereinigung und einem zuverlässiger vorbereiteten Windows-Integrationsgate. Karten zeigen ihren tatsächlichen Integrationsstand. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Added: A card held in a pickup lane now says so where it claims to be queued (AGT-2818). TaskInfo.pickupHold is a read-time projection built by the pure PickupHoldPolicy from the same facts the runner admission gate consults, and carries the mechanism (dependency gate, dispatch rejection, epic container, crash backoff, pickup policy), the specific reason, when the hold started, how long it has lasted, and the ways out. A releaseGate edge pointing at an archived, never released target is now classified unsatisfiable rather than "waiting", earns the same once-per-card runner warning a dependency cycle already got, and renders as "this gate can never open" instead of "waits for release". The durable remoteDispatchRejection is rendered on the board card as well as the detail header, with its code and instant. GET /api/pickup-holds and a boot sweep list every held card across all projects. Everything here is report-only: releasing a validation gate stays an operator decision, so the two resolutions are offered and never taken. See docs/concepts/pickup-hold-visibility.md.
- Added: Branch reclamation (AGT-2793) is now wired into the product path instead of sitting unreachable: BranchReclaimTriggerService fires after a successful delivery integration, after a card archives, and (via a new POST /api/git/branch-reclaim/promotion endpoint called from scripts/release/promote-develop-to-main.sh --project <name>) after develop is promoted to main. Every deletion is appended to a per-project reports/git-branch-reclaim.jsonl evidence file and, for the two per-task triggers, echoed onto that task's own timeline as a branchesreclaimed entry.
- Added: A bare-remote integration test seeds all six managed ref namespaces (task/, runner/, delivery/, agent-studio/results/, agent-studio/salvage/, agent-studio/quarantine/) and asserts exactly the expected refs are deleted while proof commits stay reachable from main; a replay test proves a task can be reissued after its results/ ref is deleted, and a wiring test suite proves each of the three triggers fires on its transition and stays silent when that transition fails or never lands.
- Fixed: The local merge gate no longer hands a verify command the Studio's own listener (AGT-2840). api.sh start runs the backend through dotnet run, whose launch profile exports ASPNETCOREURLS into the process environment even though the port comes from --urls; the gate passed that environment to dotnet test, and the five ConnectorProfileTests that boot WebApplicationFactory<Program> read the Studio's URL as the connector's own and failed with "The connector may listen only on http://[::1]:5031" - only on the machine that runs the Studio, sending three already-reviewed cards (AGT-2825, AGT-2826, AGT-2827) to Human Review. HostListenerEnvironment now names the keys that carry a listener (ASPNETCOREURLS, URLS, DOTNETURLS, the ASPNETCOREPORTS pair, and anything under Kestrel) and BuildTestGateRunner drops them from every verify child, the way the preparation gate already curates the prepare script's environment. The connector tests additionally boot with those keys set aside, so the class is hermetic against any launcher.
- Fixed: A delivery blocked by the commit candidate gate said so instead of escalating without a reason (AGT-2828). WEB-21 captured 12 screenshots, the gate warned binary-surprise on every one of them, nothing was committed, and the card landed in 5e-escalated with cause no-completion-signal and an empty parked reason. Now: the gate records what it held back as withheld-commit-candidates.json on the card (path plus finding codes per file); the no-completion-signal park carries its reason through the lane move, so the parked marker names the escalation category, the gate, and the count instead of being empty, and the card lists which files and why; evidence assets under a path the project declares in ProjectSettings.EvidenceAssetPaths (a recognized asset type at or below 5 MiB) commit without a manual step; and POST /api/tasks/{id}/git/withheld-candidates/commit is the operator action that commits the withheld set after review. Explicit review clears the warnings that withheld the files, never a block, so secret material stays refused.
- Fixed: Concurrent Remote Review workers shared one host-global .NET build server (AGT-2831). The attempt workspace fenced every writable path per attempt, but Roslyn's VBCSCompiler listens on /tmp/<pipename> and reusable MSBuild worker nodes on /tmp/MSBuild<pid>, neither of which honours TMPDIR, and both keep the working directory of the attempt that started them. At RUNNERMAXPARALLELISM=4 a later attempt connected to a server answering from a deleted tree and blocked, holding its review slot at near-zero CPU for the whole command budget. Every review command - candidate, baseline and dependency preparation - now runs with its own server-free build namespace (MSBUILDDISABLENODEREUSE=1, DOTNETCLIUSEMSBUILDSERVER=0, UseSharedCompilation=false, attempt-local MSBUILDDEBUGPATH), applied over the immutable plan so a plan can never re-enable a shared server.
- Fixed: A review verify command that blocks instead of working is now reaped on a bounded budget. RUNNERREVIEWNOCPUPROGRESSSECONDS (default 900, 0 disables) kills a command whose whole process tree fails to burn one percent of one core within the window and reports the attempt as ReviewInfra/NoCpuProgress, checked before baseline comparison so a hang is never graded as a product regression.
- Fixed: BranchRetentionAction.Namespace and .TaskKey were never populated on the actions GitBranchRetentionService.RunRepository/ReclaimForTask actually return (always null), so the evidence and reason-text those actions carry were silently empty for the periodic sweep and every per-task reclaim; both fields are now set on every returned action.
- Fixed: Integration of a local project no longer runs in the developer checkout (AGT-2832). Delivery merges happen in a Studio-owned worktree derived from the repository path and reset before each integration, so unrelated uncommitted edits in that checkout can no longer refuse an integration ("Integration working tree has uncommitted changes; refusing to merge", most recently QS-100) and a successful merge no longer switches the checkout's branch. The worktree stays detached, so the integration branch remains available to the person working in the project.
v0.3.0 · 2026-09-14
Agent Studio v0.3.0
Modellrouting benennt Modell und Reasoning-Stufe je Tier ausdrücklich. Der Update Service prüft Neustart und Laufzeitidentität; die Repository-Definition kann einen Release-Abschnitt tragen. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Added: Economy-mode model routing floor: sonnet-low (claude-sonnet-5/gpt-5.6-sol at low) is now the lowest tier economy mode may select for feature and bug cards, closing a gap where an unrecognized vendor catalogue's positional fallback could silently route a Claude CLI card to claude-haiku-4-5 with no thinking level (AGT-2793, AGT-2807). ModelRoutingPolicyRegistry.Recommend() now always returns an explicit thinking level.
- Added: POST /api/admin/maintenance/backfill-thinking-levels reports (default) or applies (apply=true) a policy-derived thinkingLevel for existing cards that carry a model with no level.
- Added: The model-level badge now renders a visible placeholder when a model has no thinking level, instead of omitting the level silently.
v0.2.0 · 2026-09-13
Agent Studio v0.2.0
Erstes getaggtes Stable-Release mit unveränderlichem Tag, Build-Manifest, gesperrter Abhängigkeitswiederherstellung und Prüfung der Laufzeitidentität. Darunter stehen die englischen Originaleinträge aus dem Produkt-Changelog.
- Added: Project-owned Stable release identity and restore rules, with lock-file and exact-pin-plus-registry-integrity validation, fixture coverage, and a legacy Stable first-release runbook.
- Added: Stage M1 repository execution definitions, shared preparation manifests, content-addressed npm, NuGet, and Playwright caches, Linux Runner stable checkouts with leased subject worktrees, proposal cards, and the project Settings > Execution surface.
- Added: Durable restart continuity for local and Remote runs, review aspects, and pipeline post-steps, including operator-visible bridge and loss events plus a Windows Studio release drill.
- Added: Initial open source community and release-hygiene baseline.
- Added: Dossier relevance-review metadata, managed review writes and history, configurable review-due policy, and review tags, tooltips, filtering, sorting, and viewer recording controls.
- Added: Local run teardown now owns and ends complete Windows and Linux process trees, retries worktree removal, renames persistent stale directories aside, and periodically sweeps orphan directories that lost their .git link. Worktree preparation failures are visible on cards and timelines, retry with bounded backoff, and park with a path-specific blocker after five attempts. Local run admission also warns when a configured project URL port already has a listener.
- Changed: Dossier list cards now use a wide clamped summary, anchored footer actions, copyable keys, and labelled lane-colour references with immediate complete tooltips.
- Fixed: Stable release manifest generation and candidate preflight no longer assume a product-wide backend/packages.lock.json; Agent Studio now declares and commits its NuGet lock, and promotion rejects missing or stale lock state.
- Fixed: Develop-to-main promotion now supplies a configurable annotated-tag identity independently of host Git configuration and records post-gate tag or atomic push failures in the durable promotion record.
- Fixed: Production builds no longer load the task-detail component graph with the initial task board bundle, restoring release-budget headroom without raising the configured limit.
- Fixed: Process launches across backend, retention, setup, and repository Node scripts now suppress Windows console windows, drain redirected CLI stderr, and carry repository-wide regression guards for future spawn sites.
- Fixed: Public-demo execution-route inventory now includes POST /api/v1/reviews/attempts/{attemptId}/reclaim in the Continue path.
- Fixed: Frontend release coverage is green again for the CliAdminPanelComponent smoke test, the CliModelsPanelComponent known-CLI grouping test, the ExecutionAssignmentCardComponent delivery-failure test, the WorkspaceOverlaysComponent smoke test, and the Dossier and Wiki GlobalSearchComponent navigation tests. The additionally exposed WorkbenchTabHostComponent catalogue-backfill test is isolated from persisted tab state as well.