chore(skills): run /ship review loops as Workflow scripts - #610
Merged
Conversation
Encode the plan and code review loops as deterministic Workflow scripts instead of prose the coordinator follows by hand: - references/plan-review-loop.workflow.mjs — the whole 5-pass plan loop in one run: serialized chatgpt-review plan passes, schema-validated verdict (fail-closed to REVISE), parallel read-only verification of every finding, and a revise agent that folds accepted findings into the plan file in place (rejected ones become "## Review responses" rebuttals). Cap 5 is a loop bound. - references/code-review-pass.workflow.mjs — one PR pass per run: review (publishes the comment), parallel verification, then a fix agent that applies accepted findings with tests, loops the full local gate to green, and commits locally only. The coordinator still owns push, CI wait, re-invocation, and the gate; the 3-pass cap stays enforced by the chatgpt-review script. - references/review-loops.md — the invocation contract: args, return statuses and required coordinator actions, and the hard rules (coordinator-only, one review workflow at a time, question files carry the verdict protocol, findings never silently dropped). SKILL.md steps 2.2/2.3/3.5, the operating rules, per-issue-cycle.md's plan review section, and the footguns are rewired accordingly. One behavior change: the loop's revise agent (not the waiting worker) edits the plan file, so the worker re-reads the approved plan before implementing — the file, not the worker's draft, is the contract. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01E4bdx3RwYDjVdPdFCggKVQ
1 task
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Follow-up to #609: the
/shipreview loops now run as Workflow scripts instead of prose instructions, so the loop mechanics are enforced by code:plan-review-loop.workflow.mjs— the full plan loop in one run. Each pass: one serializedchatgpt-review plancall (with one resume-retry on incomplete status), schema-validated verdict parsing (missing/malformed ⇒ REVISE, fail-closed), one read-only verifier agent per finding in parallel, then a revise agent that folds accepted findings into the plan file in place (path = session identity) and records rejections as## Review responsesrebuttals. The 5-pass cap is afor-loop bound. Returnsapproved/needs_human(FULL STOP) /error.code-review-pass.workflow.mjs— exactly one PR pass per run: review (publishes the PR comment), parallel finding verification, and — when findings are accepted — a fix agent that applies them with tests, loops the full local gate to green, and commits locally only. The coordinator keeps push, CI-wait keyed on head SHA, re-invocation, and the merge gate; the 3-pass cap remains enforced by thechatgpt-reviewCLI. Returnscertified-pending-proofs/fixed-await-push/no-accepted-findings/fix-failed/needs_human/error.review-loops.md— the invocation contract: args, return-status tables with required coordinator actions, and the hard rules (coordinator launches review workflows one at a time — agent Chrome is a single session; question files must carry the verdict protocol; accepted/rejected findings are always returned and recorded).SKILL.md (operating rules, steps 2.2/2.3/3.5),
per-issue-cycle.md(plan-review worker part), footguns, and.wiki/Project-Skills.mdrewired to match. Both scripts pass a syntax check under the workflow runtime's async-wrapper semantics and avoid the forbidden runtime APIs (Date.now/Math.random/new Date).Behavior change: review revisions to a plan are made by the loop's revise agent, not the waiting worker — the worker re-reads the approved plan file before implementing, making the file (not worker memory) the contract.
Process/tooling only — no
src/changes, no CHANGELOG entry.🤖 Generated with Claude Code
https://claude.ai/code/session_01E4bdx3RwYDjVdPdFCggKVQ