AIWorkHub v0.12.67

Added

• reject_review(to="pending", allowed_writes_amendment=[...]) widens a rework card's allowed_writes add-only, so a scope gap found at review (a test that asserts the changed contract, a call site the card must wire) no longer makes every rework unwinnable. Paths go through task_create's own rules and are refused when they contain .., are absolute, match forbidden, or lead into a reserved root (.git, .aiworkhub/runtime), including through a glob. Each amendment is recorded on the card (NF-2026-01641).
• accept_review and reject_review take an inline lesson_candidate or a no_new_lesson_reason token, so the decision closes its learning out in the same call. Contradictory input is refused before the decision runs, and a failed close-out never undoes a decision that already landed (NF-2026-01672).
• Every published MCP tool parameter carries a description from its docstring Args entry or Field hint, tool descriptions carry no trailing whitespace, and no property publishes an empty examples list. The invariants are locked on the in-process, write-gated and stdlib tools/list listings.
• A failed validation that matches a measured sandbox signature ([WinError 1314] in the Windows AppContainer lane, a filename-less SemLock PermissionError under landlock) is stamped on its row, and the finalizer attaches a one-call replay plan for the manager. Each failure must prove the signature in its own pytest block, the closing summary must list every failure it counts, and a signature counts only on the backend it was measured on. Nothing is ever run on the host automatically (NF-2026-01671).
• Read-only operator recipes preflight_card_winnability, accept_readiness_bundle and blocked_queue_triage (NF-2026-01700).
• The dashboard task list, in the browser and in VS Code, has an Effort column with the reasoning effort each worker actually runs at, or — when none was recorded (NF-2026-01698).
• LLM clients get a manager tool surface. The extension writes AIWORKHUB_MCP_SURFACE=manager into every LLM client config it registers. That surface hides the aiworkhub_dashboard_* tools only the dashboard webview calls, except aiworkhub_dashboard_needfix_transition and aiworkhub_dashboard_snapshot, which have no manager counterpart, and sends each tool result once as text instead of text plus structuredContent. Measured on the stdlib backend: tools/list drops from 217 tools and 249,979 bytes to 175 tools and 209,085 bytes, and a large call from 11,838 to 6,181 bytes. A hidden tool answers with the ordinary unknown-tool error; the webview's own MCP child and an unset value keep the full surface (NF-2026-01676).
• A dependency-ready pending card whose bound route has an open quota or auth circuit is moved through reroute_launch_identity, with every gate of a manual reroute, to the best-ranked available worker those gates accept, and then launched by the normal path. It moves at most once per circuit episode and never on a transient, model_not_found, half-open or closed circuit; with no capable route it stays pending with one no_capable_route blocker (NF-2026-01674).
• recover_blocked_rework takes allowed_writes_amendment with the same add-only path rules and refusal codes as reject_review; it is refused together with validation_only_replay, which keeps the exact sealed scope (NF-2026-01701).

Changed

• Reasoning effort is tiered by difficulty and risk instead of always max. MAX is kept for a card the manager declares critical; bounded low/medium work runs at medium and standard work at high. A reviewer follows the card under review: MEDIUM by default, HIGH for the security lens, a high-risk target, or complex, security or architecture work. The Claude CLI and VS Code LM routes derive effort the same way (NF-2026-01697, NF-2026-01703).

Fixed

• A rate, quota or session-limit refusal is recorded as provider_account_limit, so the reviewer successor skips every model on the exhausted adapter. A wall-clock reset such as "resets 2:40am (Asia/Tbilisi)" resolves to its next UTC instant (NF-2026-01684).
• verify_audit_ledger fails closed on malformed ledger lines. A line that is bad JSON, is not an object, has a foreign schema_id or lacks its hmac counts in entries_malformed and returns audit_ledger_malformed_lines:<n>; such lines are no longer skipped silently (NF-2026-00500).
• validation_toolchain_authority_secret_unavailable now names its cause: the key loader's reason, the OSError type/errno/winerror, or key_absent (NF-2026-01082).
• VS Code LM failure diagnostics stay within their 2048-character budget. The oldest turns are dropped first, so the newest evidence survives and the JSON stays parseable (NF-2026-01309).
• The combined-tree delta lists both the source and the destination of a staged rename (NF-2026-00499).
• task_template_show states the default mandatory changed outputs of bugfix_with_regression, test_only and docs_change and how to narrow them (NF-2026-01059).
• The stale storage snapshot test proves the read is non-blocking by event ordering instead of a 0.1 s timing bound that failed under full-suite load (NF-2026-01362).
• A python -m aiworkhub.<module> recipe such as aiworkhub.operator.worktree_diff now runs under the server's own interpreter, with this package on PYTHONPATH. It no longer dies with ModuleNotFoundError in a repository that does not install AIWorkHub. Every other recipe keeps the repository's interpreter (NF-2026-01693).
• quality_reviewer_launch no longer hands back a reviewer the manager archived. The next launch for that lens starts a fresh reviewer, so a wrong high finding no longer forces a rework of the whole card (NF-2026-01692).
• A retried attempt carries the failed attempt's retained in-scope changes instead of starting again from HEAD; an out-of-scope change is recorded and never carried (NF-2026-01662).
• Validation-only replay rebuilds a predecessor worktree that retention collected from its sealed rework delta, instead of refusing with validation_only_replay_workspace_invalid and paying for a worker relaunch (NF-2026-01655).
• accept_review runs once per request: a second call joins the running accept, status queries see it in flight, and MCP progress keeps the client from losing the receipt (NF-2026-01626).
• The reviewer packet carries the card's whole objective and acceptance instead of cutting them at 2000 characters (NF-2026-01593).
• Reviewer routing honors model disables and moves a lens off an adapter that could not inspect the candidate (NF-2026-01620).
• The AppContainer site shim names asyncio and multiprocessing pipes in the \\.\pipe\LOCAL\ namespace, so an asyncio subprocess in the sandbox no longer fails with WinError 5 (NF-2026-01689), and its PermissionError fallback resolves a path that already carries the \\?\ prefix (NF-2026-01702).
• A sandbox drive letter (DefineDosDevice) left behind by a crashed launcher is leased and reclaimed instead of leaking until sign-out; reclaim waits when the ownership scan is incomplete (NF-2026-01695).
• Source Graph file mode drops primitive-annotation and stdlib-import edges before the bounded edge slice (NF-2026-01699) and never silently drops a requested path: a path with spaces stays one path and truncation is judged by characters. VS Code LM reviewer file-mode reads count as progress (NF-2026-01704).
• The tool-policy sync rollback restores a missing host file by a no-clobber rename when the AppContainer denies the hard link, and its test now runs in that lane instead of being skipped (NF-2026-01628).
• The stdio schema freeze pins cover the new optional reject_review parameters. The concurrent stdio smoke now fingerprints schema shape without descriptions, like its two sibling freeze scripts.
• reroute_launch_identity accepts any adapter in the target runner's identity tuple, such as vscode_lm for a GLM runner, and records the catalog adapter on the card; an adapter outside the tuple is still refused (NF-2026-01602).
• reroute_launch_identity onto a Claude CLI route compares the catalog model through the launch's own model normalizer, so a catalog alias such as sonnet no longer refuses the reroute and the quota-circuit auto reroute can pick a Claude worker (NF-2026-01709).
• The Effort column shows the effort a worker actually runs at: the isolated launcher writes reasoning_effort into the running process event as well as the launch receipt (NF-2026-01711).
• reroute_launch_identity returns a lifecycle receipt (card_sha256/card_bytes, optional include_card) instead of the full card, which was about 55 KB (NF-2026-01688).
• A mid-run OAuth expiry that the Claude CLI reports as a status-less authentication_failed stream takes the single bounded auth retry instead of failing the attempt (NF-2026-01685).

Verify downloaded artifacts against the attached SHA256SUMS.