progress-reviewer

Imported from xmanatee/kota (src/modules/autonomy/workflows/progress-reviewer/AGENTS.md). Install upstream with npx skills add xmanatee/kota --skill progress-reviewer. Copyright stays with the author.

Progress Reviewer Workflow

This workflow owns cross-run systemic operational and product learning.

• Collect structured evidence first, then let the reviewer assess it.
• Automatic reviews consume a pinned, coalesced evidence window when builder demand leaves capacity. Compare delivery, repair and review outcomes; source growth and elapsed time alone do not justify work. The agent decides evidence sufficiency for the proposed decision and can choose no action.
• Retain baseline/current outcome summaries with scoped raw references. Follow generated interventions through canonical task state, integrated changes and later outcomes that may disprove the original hypothesis. The default compact packet prioritizes the pinned comparison summary over individual raw references.
• Select current outcomes and prior interventions together, reserving manifest coverage across runs before diagnostic files. Canonical state details travel through redacted review projections, never agent access to operational stores.
• Publish cited architecture/guidance handoffs through the existing semantic publication transaction, preserving the shared improvement topic key. The receiving owner applies its own evidence and authority rules. Handoff evidence identity excludes admission revisions and citation order; scoped file content is pinned at collection. Carry that identity to the receiver independently of the shared topic and run-artifact provenance. While the topic's task is open or blocked, retain the handoff and its references as pending across task updates and unrelated reviews. A newer resolution or owner-question disposition cancels the topic's pending handoff; publication reconciles remaining handoffs against task completion before consuming them. Dispatcher invokes the same reconciliation on idle, so release does not require a new review. Both owners stage receipts with delivery in the runtime transaction.
• Consume each automatic semantic input revision once. Automatic requests use their own latest-only event slot, while the explicit request event is lossless, so owner/system requests cannot replace or be replaced by a superseded automatic revision.
• Reject consumed automatic revisions through the workflow's canonical triggerAdmission watermark before durable run admission. The watermark is one runtime state row; RunStateDatabase owns queued work and delivery-attempt keys own replay dedupe.
• Treat the complete current open queue, anchors, dependencies, durable issues, durable run state, and owner decisions as canonical state. Recent terminal task history can add context but must not stand in for queue truth.
• Bind runtime-authored evidence to its pre-agent digest. The reviewer receives a machine-enforced per-run agent-output/ directory while sibling evidence, step state, and run metadata remain runtime-owned inputs rather than agent-authored authority. Its project write scope is deny-all.
• Keep the reviewer evaluative: it may create normal follow-up tasks or owner questions, but it must not directly mutate product code.
• Every review artifact must state its scope, trigger kind, evidence window, included evidence, excluded evidence, and applied actions.
• Dedupe before creating tasks or owner questions so recurring reviews do not spam the queue. Publication records explicit source runs independently of the automatic watermark. Per-topic observations preserve newer dispositions while allowing out-of-order explicit requests about unrelated topics.
• Validate citations against the collected evidence inside the agent contract. Exhausted output correction retains a typed rejection and finishes with warnings, without actions or proposal publication. It consumes the rejected automatic revision through runtime compare-and-set so changed evidence can open a later window without replaying exhausted correction. Runtime and evidence integrity failures remain terminal.
• Reconcile generated tasks and owner questions through the shared proposal lifecycle when canonical state disproves their premise.
• Stage task changes and semantic publication evidence in the writer run. Owner-question reconciliation and the consumed semantic watermark publish in the original run's shared success finalizer, using compare-and-set against the latest runtime state row. A failed finalizer retains the original run for recovery without repeating agent work.
• A consumed review watermark does not cancel paired owner effects for an integrated task change. Publication reconciles against canonical task state and the retirement disposition recorded in archived tasks, so delayed effects complete without undoing a later disposition.