AGENT SKILL

code-review

Deep review of the working branch — code rules compliance, bugs, regressions, security, soundness. Use before opening a PR, or when a change needs scrutiny beyond /code-rules-check.

License 许可
Apache-2.0
所属仓库
LNSD/lorecraft
最近更新
2026-11-10
Stars
★ 0
Forks
1

01 前言 Front Matter

name
code-review
description
Deep review of the working branch — code rules compliance, bugs, regressions, security, soundness. Use before opening a PR, or when a change needs scrutiny beyond /code-rules-check.
allowed-tools
Bash(git diff *) Bash(git status *) Bash(git merge-base *) Bash(grep *)

03 SKILL.md 正文

Code Review

A thorough review of the current branch, run locally. It performs /code-rules-check at review depth — fanned out across groups of code rules — and adds what a compliance check cannot see: logic gaps, regressions, security, safety, and soundness.

The subject is the uncommitted work by default (git diff HEAD, git status). For a whole branch use git diff $(git merge-base HEAD main)...HEAD. Review only what the diff touches — an unchanged file that breaks a code rule is not this change's finding.

When to Use This Skill

  • Before opening a PR
  • After a large, risky, or long-running implementation
  • Reviewing someone else's branch locally
  • When /code-rules-check is clean but the change still warrants scrutiny

For the routine "does this follow the code rules?" pass after finishing a piece of work, use /code-rules-check alone — it is a fraction of the cost and is the gate in the development workflow.

Review Checklist

Please review this code change and provide feedback on:

1. Security Concerns

Review for security vulnerabilities:

  • Exposed secrets or credentials — a literal token, key or password, and anything that reaches a log line or an exception message
  • Input validation and sanitization at the boundaries: a document read from disk, its YAML frontmatter, a JSON Schema file, a path supplied by the caller
  • yaml.load on document frontmatter where yaml.safe_load is the only safe choice
  • A document or schema path joined from caller-supplied parts without containing it inside the corpus directory, so ../ escapes the tree being checked
  • A subprocess spawned with shell=True, or a command line assembled by string interpolation

2. Principles Violations

The principle-* code rules are the design principles, and the /code-rules catalog states each one in full — work from the catalog rather than a list here, which would cover a subset and go stale on the next edit.

Judge the change against all of them. Read a full principle code rule when you need to argue a finding: the examples and the Pragmatism Caveat are what separate a violation from a deliberate, documented exception, and a principle finding that ignores the caveat will be rejected.

3. Potential Bugs

Look for common programming errors such as:

  • Off-by-one errors, especially in line numbers and section boundaries
  • Incorrect conditionals
  • Use of the wrong variable when several of the same type are in scope
  • min vs max, first vs last, flipped ordering
  • Iterating a dict or set in an order-sensitive operation — findings reported in a different order on every run are a defect, not noise
  • A mutable default argument, or a dataclass field sharing one instance across objects

4. Unhandled Failure Paths

Identify the paths that raise where nothing locally proves they cannot:

  • Indexing or dict[key] on data that came from outside — parsed frontmatter, a schema file, a config file
  • int(), float() or a datetime parse on an untrusted string
  • A bare except, or an except Exception that swallows the failure and returns a default — a malformed document must produce a finding, never a silent pass
  • assert used to enforce a runtime invariant; -O removes it
  • A resource left unreleased when the failing path skips the cleanup

Note: this overlaps with docs/code/error-handling.md, docs/code/error-types.md and docs/code/error-boundaries.md. Verify compliance with the project's error handling standards.

5. Backwards Compatibility

Verify backwards compatibility is maintained:

  • A changed public dataclass field — a rename, a removed default, a narrowed type — breaks every caller that constructs it by keyword
  • A tightened frontmatter schema or structure file that documents already in a consuming repository cannot satisfy, turning a clean corpus into a wall of findings
  • A renamed check identifier, or a changed finding shape, that a caller parsing the output depends on
  • A renamed or removed public name in a package __init__.py

6. Code Rule Compliance

Run /code-rules-check, forcing its fan-out path regardless of diff size: one agent per group of code rules, spawned in a single message, each applying its code rules' ## Checklist items to the diff. That skill owns the procedure — which code rules govern a change, how groups are derived, and the report format. Do not restate the code rules here; they change when docs/code/ changes.

A finding in this dimension is a violation with a code rule behind it. Anything a reviewer notices that no code rule states belongs in the dimensions above and below, not here.

7. Testing

Evaluate test coverage and quality:

  • Reduced test coverage without justification
  • Tests that do not actually test the intended behavior
  • Tests with race conditions or non-deterministic behavior
  • A test that mocks its own subject, or patches module internals instead of exercising the public surface
  • A document-shaped fixture inlined as a string where a checked-in file under tests/ would let the test read the same thing an agent would
  • Changes to existing tests that weaken assertions
  • Changes to tests that are actually a symptom of breaking changes to user-visible behaviour
  • A new marker that is not declared in pyproject.toml; --strict-markers turns it into a collection error
  • A test that needs a network service or credentials: the suite is unit-only, so there is no tier for it and just test-unit would either reach it or skip it silently — the dependency has to go

8. Performance

Check for performance issues:

  • The same document read and parsed once per check instead of once per document
  • A regex compiled, or a JSON Schema loaded, inside a per-document loop rather than once up front
  • A quadratic scan over sections or findings where a single pass, or a lookup by key, would do
  • Reading and parsing a whole document when only its frontmatter is needed

9. Documentation

Ensure documentation is up-to-date:

  • README.md and docs/ reflect current behaviour, and the justfile recipes they quote still exist
  • AGENTS.md marks planned things as planned: a section that calls something "planned, not created" must be updated in the same change that creates it
  • A new or changed check is reflected in the spec it enforces, under docs/__meta__/ — the requirement a check applies and the prose stating it change together

10. Dead Code

Find dead code that is not caught by ruff:

  • A value assigned then overwritten before it is read
  • A module, class or function nothing imports, and that no registry discovers by name
  • breakpoint(), a debugging print(), or a commented-out block left behind
  • A TODO or FIXME added by this change without an owner or an issue

11. Inconsistencies

Look for inconsistencies between comments and code:

  • A docstring whose Args, Returns or Raises no longer matches the signature or the body
  • Misleading variable names or comments
  • Outdated comments after refactoring

12. Code Rule Validation

If the change touches docs/code/ or docs/__meta__/, check the code rules against the specs that govern them: the corpus spec docs/__meta__/code.md, narrowed by the code-<prefix>.md namespace spec for the code rule's prefix. No checker is wired up, so this is done by reading.

Notes

Focus on Actionable Feedback

  • Provide specific, actionable feedback on actual lines of code
  • Avoid general comments without code references
  • Reference specific file paths and line numbers
  • Suggest concrete improvements

Code Rule Compliance is Critical

Code rule violations should be treated seriously as they:

  • Reduce codebase consistency
  • Make maintenance harder
  • May introduce security vulnerabilities, in the code that parses documents from outside this repository above all
  • Conflict with established architectural decisions

Always run the code rule compliance review (section 6) as part of every code review.

Review Priority

Sections are ordered by priority — review from top to bottom:

  1. Security concerns (§1, highest priority)
  2. Principles violations (§2)
  3. Potential bugs and unhandled failure paths (§3–4)
  4. Backwards compatibility (§5)
  5. Code rule violations (§6)
  6. Testing (§7)
  7. Performance (§8)
  8. Documentation, dead code, and inconsistencies (§9–11)

Next Steps

After completing the code review:

  1. Provide clear, prioritized feedback
  2. Distinguish between blocking issues (bugs, security) and suggestions (style, performance)
  3. Reference specific patterns from docs/code/ when flagging violations
  4. Suggest using /code-format, /code-check, and /code-test to validate fixes

04 同仓库更多 Skills