@agentskit/harness 0.8.0 → 0.9.0

package/CHANGELOG.md

@@ -1,5 +1,15 @@
 # Changelog
 
+## [0.9.0] - 2026-09-12
+
+- **Failure-classification fix**: `classifyProviderFailure` now recognises real Claude/Codex/Grok usage-limit phrasing ("You've hit your session limit", "usage limit reached", "credit balance is too low", "spend limit reached", "temporarily limiting requests", "Overloaded") as `quota` instead of falling through to `other`, and `extractResetsAt` parses a relative (`resets in 3h`) or clock-time (`resets 10:40pm`) reset out of the message. A code-review exit is classified the same way, marking the reviewer's provider (not the review-CLI transport id) cooling down instead of retrying every tick. Root cause of a 2026-09-11/12 pilot bug: 19 unclassified `contract.failed` retries across 4 issues and 12 incomplete reviews over 7h with no cooldown ever recorded.
+- **Auto-pause after repeated failures**: `resilience.maxConsecutiveFailures` (default 3) pauses a single issue after that many consecutive `contract.failed`/`worker.dispatch-failed` events — one deduplicated Linear comment, the `resilience.pausedLabel` (default `loop:paused`), and it is skipped locally until the label is removed or `ak-harness loop resume <issue>` runs. `resilience.stagePauseAfterRuns` does the same for a `loop stage tick|deliver` run that throws repeatedly, via `loop resume --stage tick|deliver` and a new `loop paused` listing.
+- **Skills pinned into the worker brief**: `brief.skills` lists Markdown files (relative to `project.root`) embedded verbatim, sha256-digested, and truncated at `brief.maxSkillChars` in every worker brief's new "Skills (pinned)" section. A missing file fails the dispatch closed. The rendered brief is persisted to `<stateDir>/issues/<id>/brief.md`; `dispatch.json` records `briefDigest` and per-skill digests. A handoff never re-reads `brief.skills`, so editing a skill file after dispatch never affects an in-flight worker. `loop doctor` gained a `brief.skills` check.
+- **Worktree setup command**: `project.setup.command` (argv, no shell) runs once in a freshly created worktree before the worker terminal opens (e.g. `pnpm install --frozen-lockfile`), bounded by `project.setup.timeoutSec` (reserved out of the tick budget). `project.setup.required` (default true) fails the dispatch — same cleanup and consecutive-failure accounting as any other dispatch failure — on a non-zero exit or timeout; set to false to only warn.
+- **Reasoning effort per role**: `models.effort.<role>` (`low`\|`medium`\|`high`\|`xhigh`) is rendered via `providers.<id>.effortFlag` (e.g. codex `-c model_reasoning_effort={effort}`, grok `--reasoning-effort {effort}`) into `tui`/`headless`; a provider without `effortFlag` ignores it. `dispatch.json` and `loop retro`'s `dispatches.byProvider` record/group by `provider/model@effort`.
+- **GitHub label intake**: `github.intakeLabel` (default `loop:review`; `null` disables it) makes `deliver` review and comment on any open PR carrying the label, even one the loop never dispatched (tracked as `pr-<n>`, no Linear issue involved). Every nudge lands as a PR comment instead of a terminal send; `github.reviewOnly` is a fixed schema guarantee — a clean review always ends held ("merge is human") with the label removed, never auto-merged.
+- Fixed `scripts/verify-release-manifest.mjs`, which pinned the release version literal to `0.5.0` and would have failed every release since 0.6.0.
+
 ## [0.8.0] - 2026-09-12
 
 - Worker **handoff**: when a dispatched worker is idle (or its terminal is gone) and its provider is unavailable (usage/cooldown), deliver relaunches another builder on the **same** Orca worktree + branch with a continuation brief (`renderHandoffBrief`). Lease stays; `dispatch.json` updates terminal/provider/model; `delivery.handoffs[]` + `worker.handed-off` event. Caps via `delivery.handoff.maxHandoffs` (default 2).