Changeset Codereview

Provide a code review for the given local changeset.

Use this skill after .agents/skills/codereview/SKILL.md routes the task here. Read ../references/triggering-and-signal.md and ../references/review-depth-and-delegation.md first. When composing the final report, also read ../references/report-composition.md.

If the user explicitly asks for architecture review in addition to local changeset review, if an upstream prmr-codereview workflow forwards that requirement, or if the reviewed scope materially changes layering, ownership boundaries, dependency direction, compatibility/public surfaces, or other architecture-review dimensions, apply ../architecture-review/SKILL.md as a paired logical review against the same scope. Keep this skill as the main report owner and summarize the paired architecture result in Related review inputs. The paired review normally shares the same review subagent. Use ../architecture-review/SKILL.md by itself only when architecture is the sole requested review dimension.

Follow ../references/review-depth-and-delegation.md. The main agent owns guidance discovery and candidate validation. Each heavy changeset-review round defaults to one strong review subagent covering all applicable review dimensions.

To do this:

1. Determine the reviewed changeset.

   • Identify whether the target is a commit, branch diff, staged diff, working tree diff, patch, or explicit file set.
   • Choose the smallest reasonable diff or file scope that satisfies the request.
   • When the change alters public or observable behavior, derive an interface surface before narrowing the review: applicable CLI help/config, public examples, README/docs, runbooks, packaged Agent Skills, and their install/package resources. Inspect related artifacts for drift even when they have no diff; an explicit file list is not by itself an exclusion.
   • For re-review, default to the full current effective changeset unless the user explicitly asks for incremental-only validation.
   • Keep a prior-findings ledger when needed and classify each item as fixed, unresolved, or no longer applicable.
   • If the target is ambiguous, make a reasonable local choice and state it in the report metadata.

2. Decide whether architecture review is also applicable.

   • Add the architecture-review dimensions when explicitly requested or when the scope materially changes layering, ownership, dependency direction, compatibility/public surfaces, or other architecture boundaries.
   • Keep architecture candidates separate so they can be summarized as a paired review input without launching another subagent by default.

3. Gather and read the relevant repository guidance locally, including:

   • This repository's root AGENTS.md file, if it exists
   • Any directory-scoped AGENTS.md files that apply to modified files
   • Relevant .agents/instructions/, .agents/references/, or .agents/skills/ files referenced by those AGENTS.md files
   • For a durable documentation claim that depends on externally mutable state, such as a remote job outcome, storage state, or publication status, identify the authoritative source and verify it from fresh live/readback evidence or an inspectable retained snapshot. A local temporary summary alone does not substantiate the claim; if evidence is unavailable, report the claim as unverified rather than confirmed.

4. Summarize the reviewed changeset locally before issue discovery.

5. Launch one strong review subagent for the current round by default. Give it the full effective changeset, applicable guidance paths, prior findings, and any paired architecture context. Require one structured candidate list that covers:

   • scoped repository-guidance compliance
   • bugs, correctness, security, and significant performance risk
   • contracts, compatibility, and caller-facing behavior
   • architecture boundaries and minimality when Step 2 applies

   The reviewer must exhaust the scope for high-signal issues rather than stop after the first finding. Keep only candidates with a concrete location, impact, evidence, and confidence. Ignore style nits and speculative issues.

6. Validate every candidate locally against the current code, call sites, tests, and scoped guidance. Confirm that each cited guidance rule is in scope and actually violated. Keep the current reviewer responsible for verifying fixes to its own findings. After its ledger is resolved, follow the shared review-loop rule and use a fresh reviewer for the next issue-discovery round. Do not launch separate validator agents by default.

7. Filter and converge.

   • Remove unvalidated issues.
   • De-duplicate overlapping issues and assign final severity.
   • Compare the validated set with reviewer output and any prior-findings ledger before reporting so one pass contains all current findings.

8. Output a summary using REPORT_TEMPLATE.md.

   • Summarize paired logical review inputs in Related review inputs; do not duplicate their full findings in the main findings sections.
   • If architecture review was paired, include its required summary.
   • If no issues were found, use the exact text: No issues found. Checked for bugs and scoped guidance compliance.
   • If issues were found, include only validated, de-duplicated high-signal findings.