Code Review Hardening

Code Review Hardening

Strengthen code against defects through systematic, type-aware review. Every reviewed line is understood or explicitly scoped-out. Findings are classified by severity and fixability. The agent applies direct fixes autonomously and re-reviews before producing a final report.

Gotchas

Passive review is incomplete: A static report is a human review. For an autonomous agent, after classifying findings the skill must enter the self-repair loop — apply fixes, re-review, iterate.

Wrong strategy for the type: A hotfix and a new feature require opposite postures. Applying a full-scope feat review to a hotfix causes unnecessary delay. Always determine the change type first.

Blocking on style preference: Only block on style if the project's style guide explicitly requires it. Personal style preferences are Nit: — never gates.

Skipping lines: Every human-written line must be understood or explicitly scoped-out with a note.

Mentoring vs. blocking: Label every educational comment Mentoring: so the author knows it is not a gate.

Over-engineering: Block implementations that solve future problems the author doesn't know they'll face. Solve today's problem well.

Quick Start

Input:  Files or directories to review, optional change type tag
        (feat | fix | hotfix | refactor | migrate | docs)
Output: Structured markdown review report with fix status

Workflow:
1. Determine change type (auto-detect or explicit)
2. Gather project context (style guide, conventions, docs)
3. Review per type strategy — read every line
4. Classify findings by severity AND fixability
5. Self-repair loop — apply direct fixes, re-review (max 2 passes)
6. Author review — remaining author-required items
7. Final verdict + escalation if needed

Activation

Load this skill when:

• User asks for a code review of any kind
• User mentions PR review, pull request review, code review, change review, or review this
• User asks to review specific files, a branch, or a diff
• User wants feedback on a CL (changelist)

Do NOT load this skill for:

• Simple one-liner fixes the user already validated
• Questions about code that don't involve reviewing a change
• Formatting-only tasks (convert JSON to YAML, etc.)

Determine Change Type

Detect the change type in priority order:

1. Explicit tag: User provides a type tag (feat, fix, hotfix, refactor, migrate, docs). Use directly.

2. Branch name scan: Check the branch name for conventional-commit prefixes:

   • feat/ → feat
   • fix/ → fix
   • hotfix/ → hotfix
   • refactor/ → refactor
   • chore/ → migrate
   • docs/ → docs

3. Commit message scan: Examine the last 5 commit messages for conventional-commit prefixes (feat:, fix:, hotfix:, refactor:, chore:, docs:). Use the most frequent prefix.

4. Heuristic guess: If no prefix found, infer from prompt language: "bug fix" / "fix this" → fix; "hot patch" / "production issue" → hotfix; "restructure" / "clean up" → refactor; "migrate" / "upgrade" → migrate; "update docs" / "add readme" → docs; otherwise → general-purpose review.

5. Fallback: If no signal and no confident guess, apply a balanced general-purpose review across all types.

Pre-flight

Before reviewing, gather project context:

1. Check for a project style guide (search for STYLEGUIDE.md, .styleguide, style-guide.md, or inline style references).
2. Check for existing review conventions or patterns (search for REVIEW_GUIDELINES.md, CONTRIBUTING.md, or established comment styles).
3. Check for relevant subsystem documentation related to the changed files.
4. Load references/change-type-strategies.md to get the review strategy for the detected type.

If no project conventions are found, proceed with general standards.

Review Workflow

Phase 1: Determine Type

Run the type detection chain above.

Phase 2: Load Type Strategy

Read references/change-type-strategies.md for the detected type.

Phase 3: Pre-flight

Gather project context: style guide, conventions, relevant docs.

Phase 4: Review per Type

Apply the focused checklist from references/review-checklist.md, adjusting depth per type. For every file: read every line, understand it or note why it was scoped out.

Phase 5: Classify Findings

Read references/finding-severity.md to classify each finding:

Step A — Severity:

• Blocking — must fix before approval
• Nit — optional polish, not a gate
• Mentoring — educational, not a gate

Step B — Fixability (new):

• direct-fix — fix is unambiguous, agent applies it now
• author-required — fix requires author judgment or is ambiguous
• informational — mentoring or good-notice finding

Phase 6: Self-Repair Loop

Apply direct fixes autonomously. Max 2 passes.

Pass 1:

1. For every direct-fix blocking finding: apply the fix.
2. Re-read modified files to verify fix didn't break anything.
3. Re-review modified files and their immediate callers for new blocking findings.
4. If new blocking findings appear: rollback the fix, mark original finding as author-required.
5. If no new blocking findings: mark original finding as fixed.
6. Collect remaining unfixed direct-fix items.

Pass 2:

• Repeat Pass 1 on remaining unfixed direct-fix items.
• If a Pass 2 fix fails (introduces new blocking finding): rollback, mark as author-required.

After 2 passes, any remaining direct-fix items are marked author-required and escalated via references/escalation.md.

Phase 7: Author Review

Remaining author-required findings are output in the final report with clear Decision Needed descriptions. The agent does not guess.

Phase 8: Final Verdict + Escalation

• If no blocking findings remain: Approve
• If author-required blocking findings remain: Request Changes
• If your scoped review is complete but other areas need other reviewers: Approval with Notes

Any unresolved author-required finding that blocks approval is escalated via references/escalation.md.

Type-Selected Priorities

For detailed per-type goals, posture, and block-vs-nit guidance, read references/change-type-strategies.md.

Severity Taxonomy

Blocking criteria (read references/finding-severity.md for the full decision tree):

• Code health degrades — overall maintainability, readability, or testability worsens
• Safety violation — security vulnerability, memory leak, unhandled error
• Style guide violation — required style point not followed
• Correctness bug — logic error, off-by-one, wrong assumption, missing edge case
• Missing or broken tests — test doesn't cover the scenario it claims

Nit criteria:

• Style preference not in the style guide
• Minor readability improvement
• Over-engineering suggestions for future needs
• Naming that is "good enough" but could be more descriptive

Mentoring criteria:

• Teaches a language feature, framework pattern, or design principle
• Shares knowledge that improves code health over time

Output Format

Produce this structured markdown report:

# Code Review — {change-type}

**Branch:** {branch-name-or"N/A"}
**Files reviewed:** {n} ({list or "see below"})
**Verdict:** Approve | Request Changes | Approval with Notes

## Summary
{2-3 sentence overall assessment of code health impact}

## Direct Fixes Applied
| File | Line | Finding | Fix Applied | Status |
| ---- | ---- | ------- | ----------- | ------ |
| src/foo.ts | 42 | Null check missing | Add `if (!user) return null;` | fixed |

## Blocking Findings (Author Required)
| File | Line | Finding | Decision Needed | Status |
| ---- | ---- | ------- | -------------- | ------ |
| src/bar.ts | 18 | Inverted condition | Confirm: `> 0` or `< 0`? | pending |

## Nit Findings
| File | Line | Finding | Suggestion | Status |
| ---- | ---- | ------- | ---------- | ------ |
| src/baz.ts | 55 | Non-descriptive name `x` | Rename to `recordCount` | fixed |

## Mentoring / Good Things
| File | Line | Observation |
| ---- | ---- | ----------- |
| src/qux.ts | 12 | Nice early return pattern |

## Escalation Record (if any)
{from references/escalation.md}

Fix Status values:

• fixed — auto-fixed by the agent in the self-repair loop
• pending — awaiting author decision
• escalated — passed to escalation path after max iterations

Verdict definitions:

• Approve — All blocking findings resolved. CL improves code health and is merge-ready.
• Request Changes — author-required blocking findings present. Author must address before approval.
• Approval with Notes — Your scoped review is complete. Other reviewers should address remaining findings. Nits left open by author preference are noted.

Reference Loading

Load reference files on demand, not upfront: