@agentskit/harness 0.12.0 → 0.14.0

package/CHANGELOG.md

@@ -1,5 +1,100 @@
 # Changelog
 
+## [0.14.0] - 2026-09-19
+
+A queue that several machines can drain, a review whose strictness matches the risk, and four fixes that
+were sitting unpublished. Motivated by a real 24/7 loop that had been silently doing nothing: the config
+was invalid, the queue asked for the wrong assignee, and the diagnostic misnamed its own subject.
+
+### The queue stops being "my issues"
+
+- **`linear.queueOwnership`** (`person` | `unassigned`, default `person`, nothing changes unless you opt
+  in). Under `unassigned` the queue lists with `--assignee null` and the assignee becomes a **transient
+  claim**: written right after a dispatch succeeds, cleared when the item comes back. That is what lets
+  several machines share one priority-ordered queue without two of them picking the same issue.
+  The claim is non-fatal and gets its own `queue.claim-failed` event — what actually removes an issue from
+  the queue is the status transition, so a failed claim must not cost the transition or the comment.
+- **`linear.anyLabels`** — "at least one of these" (OR), because `requireLabels` is AND: listing two
+  layers there demands both on the same issue and matches **nothing**. A queue that returns zero is
+  indistinguishable from "no work to do", which is the worst failure mode this loop has.
+- `loop doctor` now says **which** queue it read. Under `unassigned` ownership it used to print
+  "for \<person\>" — the opposite of what it listed, and that is how an empty queue goes unnoticed.
+
+### Review strictness that matches the risk
+
+- **`reviewOverrides`** — stricter review for the slices that deserve it, keyed by label. First match
+  wins and only the named fields are replaced: an override that raises `votes` must not silently reset
+  the deadline or swap the CLI. The matched label travels to the deliver log, because a gate that costs
+  more without explaining itself reads as a bug.
+- The labels come from the **dispatch record**, not a fresh Linear read, so editing a label mid-flight
+  cannot change the gate a running item is judged by.
+
+### Telling the worker the truth about the base branch
+
+- **`knownFailures`** — suites already red on the base, declared with the tracking issue (mandatory: a
+  quarantine without an owner becomes permanent). The harness does **not** run `verifyCommand` — the
+  worker does, in its own worktree — so tolerating known breakage is information in the brief, not output
+  parsing. Without it, every item touching a broken package fails verification for someone else's defect.
+
+### Memory: recurrence instead of guesswork
+
+- Learning ids are content-derived, so a lesson that reappeared was silently deduplicated and a pattern
+  looked exactly like a one-off. **`sightings` now counts**, and `loop retro` offers the lessons that hit
+  `memory.recurrence.minSightings` with the promote command already filled in.
+- Promotion still requires a human (`HUMAN_APPROVAL_REQUIRED`, ADR-0019). Memory is read into every
+  worker brief: a wrong lesson promoted without a human is a wrong instruction on every future task.
+
+### One human approval covers the goal's own effects
+
+- **`tracking.authorization`** (`goal` | `separate`, default `goal`). Approving the verification result
+  now authorizes the declared external effect too, recording `authorization.recorded` at the same
+  instant. `separate` keeps the old two-gate behaviour.
+
+### Fixes that had never shipped
+
+- Phase age and worker age are different numbers: `ageMin` counts from dispatch, `phaseAgeMin` from the
+  event that started the phase. An item in review for 10 minutes used to show the dispatch age and looked
+  stuck for hours.
+- Stale delivery state is reset on redispatch, dead stage locks recover, queue alerts are ignored during
+  scheduled stages, and expected pre-PR delivery gaps stop being reported as problems.
+
+## [0.13.0] - 2026-09-14
+
+A full-codebase test-coverage sweep (every module in `src/kernel/`, `src/execution/`, `src/adapters/`, and most of
+`src/loop/` brought to 90%+ statements/branches) that surfaced eight real bugs along the way, each fixed in its own
+focused PR rather than folded silently into a test change.
+
+- **Per-issue state files are now written atomically**: `dispatch.json`, `delivery.json`, and `contract.json` were
+  each written with a plain `writeFileSync` straight to the final path. `tick` and `deliver` run as separate
+  scheduled processes against the same state directory, and `readDeliveryState` already treats malformed JSON as
+  "no state yet" rather than erroring — so a crash mid-write or a read racing a write could silently reset
+  `fixRounds`/`nudges`/`finalOutcome` instead of surfacing the corruption. Consolidated the three duplicated unsafe
+  writes into one shared `src/loop/fs-atomic.ts` (temp file + atomic rename).
+- **`events.ndjson` rotation is now serialized**: `appendLoopEvent`'s size-based rotation (`statSync` →
+  `renameSync` → `appendFileSync`) had no lock, so two scheduled processes rotating the same file near-
+  simultaneously could overwrite one process's archive or drop events. Only the rotation decision is now gated
+  behind a lock file; a busy lock skips rotation for that call rather than racing it, and a lock older than 5s is
+  treated as an abandoned crash artifact and cleared.
+- **PII scanner recognizes current-format secrets**: added `sk-proj-...` (current OpenAI project keys), fine-
+  grained GitHub PATs, Google API keys, Stripe live keys, PEM private-key blocks, and the AWS secret-access-key
+  half (previously only the `AKIA` access-key id was matched).
+- **`Ctrl-C` now actually stops `loop watch`**: its SIGINT handler only set `process.exitCode` without calling
+  `process.exit()`, so the long-running poll loop (and its `gh`/`orca` shell-outs) kept running in the background
+  after a cancelled watch.
+- **Raw throws reclassified as `HarnessError`**: `retro.ts`'s `--since` parsing and `doc-bridge.ts`'s index
+  freshness/readability checks threw plain `Error`, so these bad-input/bad-state failures fell through the CLI's
+  generic exit-1 catch-all instead of the classified exit-code path every other validation failure uses.
+- **Fixed two `execution/agent.ts` session-recorder bugs**: a dead `executing` Set that could never affect
+  control flow (removed), and a structurally-invalid-but-non-throwing runtime result that was being swallowed into
+  a generic, retryable `RUNTIME_ERROR` instead of surfacing as its own distinct failure.
+- **`findExecutable` now checks the execute bit**: a non-executable regular file sitting on `PATH` with a matching
+  name was reported as a runnable binary, only to fail with `EACCES` at actual spawn time.
+- **`loop debrief --issue X` no longer returns an empty report** for a normal, not-yet-dispatched issue: the
+  "always include an explicitly requested issue" fast path was shadowed by an unconditional second skip check
+  right after it.
+
+Everything else in this release is test-only: no other production behavior changed.
+
 ## [0.12.0] - 2026-09-13
 
 Closes gaps found reusing Orca instead of reinventing it. Orchestration mutations (`run-create`, `task-create`,

package/README.md

@@ -36,7 +36,7 @@ ak-harness doctor --json
 ak-harness plan approved --by human
 ak-harness start
 ak-verify run --json
-ak-verify approve <run-id> approved --by human --json
+ak-verify approve approved --by human --json
 ak-harness cancel <run-id> --by human --reason "Requirements changed"
 ak-harness benchmark --manifest benchmarks/harness-phase-0.json --json
 ```
@@ -90,7 +90,12 @@ that must match the `verification.completed` event before human approval. Human
 approvals, rejections, and tracking authorizations are then recorded as
 hash-chained `approval.recorded` or `authorization.recorded` events bound to
 that digest, source revision, and contract hash. The stable `run.json` remains
-the CLI projection and evidence index.
+the CLI projection and evidence index. The run ID and hashes are audit fields,
+not inputs a human needs to provide: `ak-verify approve approved` resolves the
+latest pending run. By default, a declared tracking target is covered by the
+same goal approval and records both lifecycle events. Set
+`tracking.authorization` to `"separate"` only when a project explicitly needs
+a second human decision.
 
 Use `ak-harness audit [run-id]` to reconcile a run projection with its verified
 events. `ak-harness status` performs the same reconciliation before reporting
@@ -157,7 +162,7 @@ contract is frozen:
 
 `runtime.kind` chooses the executor used by an integration: `process` is a bounded shell-free local child process; `docker` adds the Docker sandbox. The choice is frozen in the resolved contract and therefore changes its hash. Docker remains fail-closed when its daemon or image is unavailable.
 
-`autonomy: "yolo"` removes the generic final review only after every applicable check passes, tracking is disabled, and the frozen contract has no ambiguity. It never auto-approves a material decision, external tracking, or a tool rule that requires approval.
+`autonomy: "yolo"` removes the generic final review only after every applicable check passes, tracking is disabled, and the frozen contract has no ambiguity. It never auto-approves a material decision, external tracking, or a tool rule that requires approval. A human goal approval covers declared intermediate work and goal-scoped tracking; it does not waive evidence, freshness, or ambiguity gates.
 
 The phase executor applies the same rule to a declarative SDLC profile. A profile
 declares dependencies, inputs/outputs, gates, bounded retries, budgets, and an