wimpysworld/nix-config/review-code

Claude Code and agent skill by wimpysworld/nix-config

Trust tier 2
  • community
LLM review pending
Source check pending

Checking that the SKILL.md file is still in the repository.

Install

npx vskill@latest install wimpysworld/nix-config/review-code

SKILL.md

Review Code

Run a substantive review of a change and write a findings report. The caller supplies the review target, the lens, the severity bar, and the report name. This skill supplies the method.

Keep GitHub access read-only. For PR metadata, diffs, comments, reviews, threads, and status, follow the global GitHub read rule: prefer dedicated gh reads, then gh-api-safe; otherwise use only documented, clearly read-only GitHub MCP operations. Never use a GitHub MCP mutation or a tool whose effect is unclear. Posting is a separate step.

Execution mode

Direct worker mode applies when the caller selects it or this skill runs inside a worker. Complete all assigned concerns directly, including security. Do not launch agents or execute generated launch wrappers. A complete-review worker owns the full review and final report. A review-lane worker completes only its assigned lane and returns directly to its parent.

In direct worker mode, read surrounding source, verify evidence, deduplicate findings, and pressure-test blocking findings yourself. Where practical, build and run relevant tests, restore repository state afterwards, and distinguish environmental failures from change-caused failures. Keep the caller's lens, severity bar, and coverage.

Coordinator mode applies only to a coordinator that does not select direct worker mode. It retains the fan-out, delegation packets, fallback files, retries, resumed workers, and selective independent verification below. Loading this skill grants no coordinator authority.

For a complete review in either mode, add a topic sweep. Search Linear for related issues and Slack for recent conversations on the same domain, not only what the change links. Use that context to understand prior contributions and avoid undoing them. Keep the sweep read-only, with no comments or posts.

Input Resolution

Resolve the target to a diff before anything else:

InputReviewed diff
PR URL or numberGitHub read path: diff plus description, metadata, comments, reviews, and status
Branch namemerge-base with the default branch, to the branch tip
The default branch itselforigin/main..main, the unpushed commits; if there are none, stop and ask what to review
Worktree pathstaged and unstaged changes against HEAD
Commit SHAthat commit alone
No argumentthe current worktree's changes

Record the full head commit SHA in the private report. It is a guard for post-code-review and never appears in a review comment.

Report Location

Use review-report-path for report storage. The calling command supplies <review-name>.md.

Process

For a complete review, follow every step below that applies to your mode. Review-lane workers perform only their assigned checks and return findings. The parent owns the final report for those lanes.

  1. Apply communication-rules before writing anything. Read it first unless its complete, current instructions are already in this context.
  2. Apply contribution-voice when wording findings. Read it first unless its complete, current instructions are already in this context. The report itself stays private, but draft-code-review lifts these findings into a comment posted under the user's name, so they must already read as the user wrote them.
  3. Resolve the input to a diff and gather context, per Input Resolution.
  4. Apply review-report-path to the resolved target before writing the report or, in coordinator mode, starting workers. Reuse supplied paths. A complete-review worker allocates one exclusive run only when none is supplied.
  5. In direct worker mode, inspect all assigned concerns yourself. In coordinator mode only, fan out to workers, per Fan-out. Name each worker's fallback findings file in its packet, <run-dir>/findings-<concern>-<worker-id>.md, so no two collide. Never reuse a fallback path.
  6. In coordinator mode only, re-request once from any worker that went idle without returning findings. The follow-up carries a one-line recap of its scope, the two or three questions that matter most named concretely, and an instruction to reply in text rather than write a file. A worker that fails twice is your own work to finish, to the same standard, not a gap in the report.
  7. Deduplicate overlapping findings before verification. Pressure-test every qualifying blocking finding, per Adversarial pressure-test.
  8. Synthesise one report at the derived path: resolved target, full reviewed head SHA, caller-supplied lens and severity bar, summary of the change, verification performed, deduplicated findings, and conclusion. Put Target, Reviewed SHA, Lens, and Severity bar fields before the Summary heading, so follow-up and posting commands can recover the review contract. In coordinator mode, worker replies and durable fallback files are the record. Read a findings file only as a convenience where one exists. In direct worker mode, use your own verified evidence. Drop duplicate findings in either mode. Every section except Findings is evidence for the user, never material for a comment, so mark none of it for reuse. Write each finding to the three-sentence budget below, because Findings is the only section draft-code-review reads.
  9. Deliver the conclusion and every finding the user must act on, in house style (the communication-rules skill). Report the path. The file keeps the full report.
Fan-out

This section applies only in coordinator mode. Delegate to a wide fan-out of workers, in parallel where possible. Divide the review by concern, or by area or file group when the diff is large: for example correctness and logic, security, and tests and behavioural regressions.

The coordinator alone plans and dispatches this review. Review-lane workers complete their assigned lane and return directly. They never launch agents or invoke orchestrating commands.

Keep each packet's attack list short, around three or four concrete targets. A long multi-target packet correlates with a worker stalling and returning nothing. Split the concern across two workers instead of lengthening one list.

Route the security concern to dibble workers. Donatello implements; Dibble is the security specialist.

Each worker's delegation packet must instruct it to:

  • Never launch another agent or invoke an orchestrating command. Complete the assigned review lane and return its findings directly to the caller.
  • Read the surrounding code in the working tree to understand the change in context, within its assigned concern or area.
  • Where practical, verify conclusions by building and running the relevant tests on the reviewed code (for example in a temporary worktree), restoring repo state afterwards. Distinguish environmental test failures (also failing on the base branch) from failures the change caused.
  • Apply the lens and severity bar the caller set. Do not widen them.
  • Apply contribution-voice to every finding. Read it first unless its complete, current instructions are already in the worker's context.
  • Return its findings in its final reply. The reply is the deliverable: a reply that does not contain the findings is a failed task, whatever else it did. Each finding carries file:line references, severity, and why it matters.
  • Reply "my area is clean" when it is. That is a complete, valid reply and still has to be sent.
  • Copy the findings to the file its packet names as a fallback, never as the primary channel. Where the Write tool is refused, a shell heredoc writes the file instead. Never retry or fight a refused write, and never let a blocked write stop the reply.
  • Keep GitHub access read-only. Never use GitHub MCP mutations.
Adversarial pressure-test

Accept clean reports as complete. Do not launch verifiers for reports with no actionable findings. In coordinator mode only, use an independent verifier for a concrete unresolved question or an explicit user request.

For each finding rated medium or higher that would justify blocking, adversarially verify its preconditions against deployment reality. In direct worker mode, perform this check yourself without delegation. In coordinator mode only, send one follow-up to the worker that raised it (resume its existing context). Does the threat or failure mode arise in the deployed configuration? Check the actual runtime context (what executes where, isolation, who can read what, what gets logged or persisted), not just the diff. Downgrade findings whose preconditions do not hold. Where the worker cannot be reached, pressure-test the finding yourself to the same standard, rather than letting it through or dropping it.

This step stops false positives reaching a human. Do not skip it and do not soften it.

Constraints
  • British English throughout. Lead with conclusions. No filler.
  • Every worker and the final report must keep feedback succinct and actionable. In coordinator mode only, name contribution-voice in each delegation packet and require its complete, current instructions in that worker's context.
  • A finding is three sentences at most: the defect, the proof, the fix. One file:line reference is the proof; a second instance of the same defect adds nothing. No headings inside a finding, no restating the diff back at the reader, and no paragraph explaining that the surrounding code is correct. A finding that runs to five paragraphs is over budget, whatever its severity.
  • The report is the only deliverable. Do not draft a review comment and do not state a verdict; draft-code-review owns that.

Security

  • Security scanPassed, 100 of 100Pattern checks on the skill's files.
  • Model reviewNot runA model review is scheduled for the next rescan.

Quality evaluation

  • Task evaluationNot runNo evaluation yet. Run one in Skill Studio.

Source

Popularity

7-day trend score
0

Not refreshed yet.

Works with

Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.

  • Claude Code
  • Cursor
  • GitHub Copilot
  • Windsurf
  • Codex
  • Gemini CLI
  • Cline
  • Amp
  • Roo Code
  • Goose
  • Continue
  • Kilo Code
  • Junie
  • Trae
  • OpenHands
  • Qwen Code
  • AI Agent
  • Tabnine Chat
  • Sourcegraph Cody
  • Amazon Q
  • Firebase Genkit
  • Sweep
  • Devin
  • Phind
  • Mentat
  • Claude Desktop
  • Aider
  • GPT Pilot
  • Plandex
  • bolt.new
  • v0
  • Replit Agent
  • Anse
  • TypingMind
  • BetterChatGPT
  • LibreChat
  • Warp AI
  • Pieces
  • Fig AI
Embed badge
vskill VERIFIED badge
[![vskill VERIFIED](https://verified-skill.com/api/v1/skills/wimpysworld/nix-config/review-code/badge)](https://verified-skill.com/skills/wimpysworld/nix-config/review-code)

Something wrong with this skill or its page? Tell us.