Skip to content

SKILL.md

Latest commit

 

History

History
56 lines (38 loc) · 3.74 KB

SKILL.md

File metadata and controls

56 lines (38 loc) · 3.74 KB
Raw
name review-checkpoint
description Run blocker-only branch reviews through a native read-only reviewer subagent. Direct review requests are read-only; explicit fix loops and coordinating workflows may apply deterministic in-scope fixes.

Review Checkpoint

Use one native read-only reviewer subagent as the branch-diff gate. Review is read-only unless fix_loop is explicitly authorized or selected by a coordinating workflow.

Inputs

  • mode: review_only (direct default) or fix_loop (explicit fix/coordinator default).
  • max_iterations: default 3.
  • review_base: optional caller-provided base ref.
  • manifest_path: optional and only for Shipyard's integration review; child reviews keep isolated state in their handoffs.

Memory

Direct invocation owns $agent-memory; nested Workbench or Shipyard invocation defers it to the caller. Capture only accepted durable review rules or reusable root causes.

Findings and authorization

  • Classify findings as spec_blocker (acceptance or unrequested behavior), standards_blocker (repository rules), safety_blocker (data, secrets, security, or forbidden paths), test_blocker (missing or misleading validation), or non_actionable.
  • A blocker is actionable only when deterministic, in scope, and fixable without a product decision. Cosmetic, speculative, stale, broad, unclear, contradictory, or out-of-scope findings are non_actionable.
  • review_only never edits. fix_loop fixes only actionable findings and each iteration must reduce that set. Do not loop on non-actionable findings.
  • Stop when the budget is spent, a finding repeats after its targeted fix, or a later finding contradicts an accepted one. Record contradictions as non_actionable: contradictory semantics.

Review loop

  1. Require a clean worktree except local .context/progress.md and a named local branch. Resolve review_base, base_sha, and head_sha; the reviewed branch does not need an upstream.
  2. For the current (base_sha, head_sha), spawn exactly one native read-only reviewer subagent with fork_turns=none, choosing the least-cost sufficient reviewer role under repository instructions. Do not use an external review service or substitute a local self-review.
  3. Give the reviewer the absolute worktree, branch, base and head SHAs, reproducible diff refs, task requirements, repository instructions, and validation evidence. Require it to inspect the actual diff and relevant callers/tests, make no edits or state changes, and return only PASS or blocker findings with taxonomy, file/line evidence, impact, and the smallest credible fix.
  4. Validate each finding against scope and current HEAD; reclassify unsupported findings as non_actionable.
  5. The final gate is the latest completed subagent review with no later commit. If no reviewer subagent is available or it fails without a result, return BLOCKED with the tooling failure.

Append compact review history to .context/review-events.jsonl; do not create per-review folders.

Complete or fix

  • No actionable blocker -> PASS.
  • Actionable blockers in review_only -> BLOCKED with findings and no edits.
  • In fix_loop, apply the smallest fix, inspect the diff, run the smallest relevant check, commit inspected files, and return to Step 1 for the new HEAD.

When manifest_path is set, record pass, blocker, or tooling-failure events through:

python3 <shipyard_dir>/scripts/manifest.py --manifest <manifest_path> set-review --file <event_file>

A passing integration event must contain status:"PASS", current branch, base_sha, and exact reviewed head_sha; return PASS only after the manifest accepts it.

Output

Return PASS or BLOCKED with mode, reviewer role, reviewed base/head SHAs, checks, and unresolved actionable blockers.