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 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, launch ../architecture-review/SKILL.md as a paired review against the same reviewed scope. Keep this skill as the main report owner and summarize the paired architecture result in Related review inputs. Use ../architecture-review/SKILL.md by itself only when architecture is the sole requested review dimension.

Agent assumptions (applies to all agents and subagents):

• All tools are functional and will work without error. Do not test tools or make exploratory calls. Make sure this is clear to every subagent that is launched.
• Only call a tool if it is required to complete the task. Every tool call should have a clear purpose.
• Choose subagents by capability tier rather than vendor-specific model names: use a lightweight reviewer for scope discovery or file lists, a general reviewer for balanced summaries or guidance checks, and the strongest available reviewer for bug finding or issue validation.
• Reviewer and validator subagents should try to exhaust the reviewed scope for high-signal issues. Do not stop after the first good finding or intentionally cap the response to only the top 1-2 issues.
• This workflow requires subagents. Do not silently collapse it into a single-agent review. If delegation is unavailable or not yet authorized, stop and ask for explicit delegation permission before proceeding.
• Keep the review fan-out bounded. Close completed discovery, reviewer, and validator agents as soon as their outputs have been incorporated instead of leaving earlier review waves idle while later phases run.
• The numbered reviewer counts in this workflow are mandatory minima. Do not silently launch fewer subagents, merge distinct reviewer roles into one agent, or skip the validation stage.

To do this, follow these steps precisely:

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.
   • If this is a re-review of an updated target, default to the full current effective changeset rather than only the incremental patch since the last review, unless the user explicitly asks for incremental-only validation.
   • If prior findings or user-specified concerns exist, keep a working ledger and classify each item against the current scope 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. If the user or upstream review context explicitly requested architecture review in addition to this changeset review, or if the reviewed scope materially changes layering, ownership boundaries, dependency direction, compatibility/public surfaces, or other architecture-review dimensions, launch a paired architecture-review subagent against the same reviewed scope before continuing. Keep its output as a separate paired-review input for the final report.

3. Launch a lightweight review subagent to return a list of file paths (not their contents) for all relevant repository guidance files including:

   • The root AGENTS.md file, if it exists
   • Any directory-scoped AGENTS.md files that apply to modified files
   • Any files under .agents/instructions/, .agents/references/, or .agents/skills/ referenced by those AGENTS.md files and relevant to the review scope Close this discovery agent once the applicable guidance file list has been captured.

4. Summarize the reviewed changeset locally before issue discovery. Do not keep a dedicated summary agent open for this step.

5. Launch 3 agents in parallel to independently review the changes. Each agent should return the list of issues, where each issue includes a description and the reason it was flagged. Each reviewer must return every high-signal issue it can find in scope, not only the single strongest issue.

   Agent 1: repository guidance compliance reviewer Audit changes for AGENTS.md / .agents guidance compliance. When evaluating guidance compliance for a file, only consider the guidance files that are in scope for that file, including applicable parent AGENTS.md files and the .agents instruction/reference/skill files they reference.

   Agent 2: deep bug reviewer Scan for obvious bugs. Focus only on the reviewed changeset. Flag only significant bugs; ignore nitpicks and likely false positives.

   Agent 3: deep bug reviewer Look for problems introduced by the reviewed changeset. This could be security issues, incorrect logic, or clear contract breakage.

   CRITICAL: We only want HIGH SIGNAL issues. Flag issues where:

   • The code will fail to compile or parse
   • The code will definitely produce wrong results regardless of inputs
   • Clear, unambiguous repository guidance violations where you can quote the exact rule being broken Close these reviewer agents once the candidate-issue set has been consolidated for validation.

6. Validate the issues in small batches. Keep at most 2 validator agents open at once: one general validator for repository guidance issues and one strongest-available validator for bugs and logic issues. The validator's job is to confirm that the issue is truly real with high confidence in the reviewed scope. For repository guidance issues, validate that the cited AGENTS.md / .agents rule is actually in scope and actually violated. Reuse or relaunch validators as needed, but close each validator as soon as its assigned validations have been incorporated.

7. Filter out any issues that were not validated. De-duplicate overlapping issues across all reviewers, then assign a final severity to each remaining issue. Before moving on, run a convergence pass against the reviewer outputs, validator outputs, and any prior-findings ledger. If a new validated issue surfaces during write-up, fold it into this round instead of leaving it for a later follow-up review.

8. Output a summary using REPORT_TEMPLATE.md.

   • If this workflow used any paired review skill, include a short Related review inputs summary for each paired skill. Keep those summaries concise and reference the paired report instead of copying its findings into the main findings sections.
   • 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.

Source: HorizonRobotics/RoboOrchardLab — distributed by TomeVault.