ywc-impl-review

name: ywc-impl-review description: >- (ywc) Use when implementation is complete and before creating a PR, and the user wants to validate code matches the spec, check implementation quality, or run a comprehensive review. Triggers: "구현 검증", "impl review", "implementation review", "사양 적합성", "코드 리뷰", "구현 리뷰", "PR 전 검증", "check my implementation", "実装レビュー". Do not use for active code generation, for spec-only review (use ywc-spec-validate), for product/business-level review (use ywc-product-review), for capturing or reading accumulated review preferences as durable learnings (use ywc-review-learnings), or for a security-only audit of auth / external-input / sensitive-data code (use ywc-security-audit).

ywc-impl-review

Announce at start: "I'm using the ywc-impl-review skill to run a five-axis (architecture / design / devex / security / QA) implementation review."

Implementation conformance review skill. Runs five parallel reviewers (Phase 1: 4 Sonnet + 1 Haiku) and escalates only ambiguous findings to a short Opus advisor pass (Phase 2). The four code-review aspects (architecture / design / devex / security) follow the gstack review-army pattern; QA stays as a separate axis because coverage analysis is mechanical and benefits from a lighter model. See Advisor Pattern for why this shape is used.

Rationalization Defense

When tempted to skip a step, check this table first:

Violating the letter of these rules is violating the spirit. Review without honest severity is theater.

Arguments

Advisor Pattern

This skill uses Pattern B (Two-Phase Review) from advisor-pattern.md. The rationale: review findings range from mechanical (hardcoded secret, missing null check, trivial OWASP match) to genuinely ambiguous (architectural judgment call, severity debate between two OWASP categories, spec-conformance question with more than one reasonable reading). Running every reviewer on Opus wastes frontier capacity on the mechanical cases; running every reviewer on Sonnet undersells the ambiguous ones. Phase 1 handles the mechanical cases at Sonnet/Haiku cost; Phase 2 escalates only the ambiguous ones to Opus with tightly bounded context.

Budget discipline (see advisor-pattern.md §6): default cap is 5 Opus calls per invocation, shared across all categories. Use fewer when possible. Never forward full files or full project context to the Phase 2 advisor — only the finding text, a bounded snippet, and the relevant spec excerpt.

Execution Steps

0. Load Review Learnings (skip if --skip-learnings) — Invoke ywc-review-learnings --mode read --target <changed files / --code path> to load the project's accumulated review preferences from docs/review-learnings.md. The result is a compact "Applicable Review Learnings" block: DO / DO-NOT entries become extra checks injected into the matching reviewer subagent (Step 3), and FALSE-POSITIVE entries tell that reviewer to stop raising a known non-issue. If the file is absent, proceed with an empty set — never block a review on missing learnings. This is the compounding mechanism that makes review quality rise per repository over time (see coderabbit-methodology.md §7).

1. Collect Project Context — Read CLAUDE.md, package.json to identify conventions, tech stack, and PR gate conditions. If docs/ubiquitous-language.md exists, read it — the Reviewer subagent must flag identifiers that match a "Synonyms to Avoid" entry instead of the canonical term. Per coderabbit-methodology.md §3, treat the spec / PR description as the statement of intent and trace each changed symbol to its callers/callees before judging it — a locally-fine change can still break a caller's contract.

2. Read Spec + Code — When --git-range is provided instead of --code: run git diff --name-only <range> to obtain the changed-file list and treat those files as the review target. Read the specification file and all target code files. This context stays with the parent; do not forward it wholesale to Phase 2.

3. Phase 1 — Parallel Executor Review — Use the Task tool to spawn five subagents in parallel. Pass model explicitly on each call so the executor layer stays at Sonnet or Haiku cost:

   • Architecture subagent (model: sonnet) — Module boundaries, layering, structural patterns, dependency direction, simplicity / over-abstraction, structural spec conformance. Reference: references/architecture-agent.md. When the diff touches DB schema or migrations, also apply the shared schema review checklist (../references/schema/core.md Part C); raise cascade ↔ API status (B2) and multi-tenant scope (B6) gaps as one-line cross-references to the Security subagent rather than duplicating them.
   • Design subagent (model: sonnet) — API/interface design, naming, signatures, error models, return shapes, public-surface discipline, contract spec conformance. Reference: references/design-agent.md.
   • Devex subagent (model: sonnet) — Readability, error messages, logging, documentation, debuggability, config UX. The operator-experience dimension. Reference: references/devex-agent.md.
   • Security subagent (model: sonnet) — OWASP Top 10 analysis. Reference: references/security-agent.md. When the Claude Code named-agent catalog at tools/claude-code/agents/ is installed, prefer subagent_type: ywc-security-engineer so the subagent carries the dedicated worker persona (tools/claude-code/agents/ywc-security-engineer.md).
   • QA subagent (model: haiku) — Test coverage gaps and missing test cases. Reference: references/qa-agent.md. Haiku is appropriate because coverage-gap detection is largely mechanical (file enumeration, assertion counting, branch enumeration) and does not typically require frontier reasoning.

   Inject the Step 0 learnings into each subagent prompt, filtered to that aspect's category: a DO/DO-NOT learning is an extra check; a FALSE-POSITIVE learning is an explicit instruction not to raise that finding here. Apply the --profile dial: in chill (default), a subagent suppresses its Style/Docs/Devex-polish nitpick tail (Low/Info) and surfaces only correctness/security/logic/runtime-risk findings; assertive emits the tail too (Critical/High/Medium are never suppressed). Verify before surfacing (the precision lever, coderabbit-methodology.md §2): every finding must cite primary evidence — exact file:line, the symbol traced to its definition/callers, or fresh command output — and a finding that cannot be substantiated by re-reading the code or running a scoped check is dropped, not surfaced with a hedge. Where the project ships linters/scanners (ruff, eslint, golangci-lint, shellcheck, semgrep, gitleaks, ast-grep), run them and feed their output to the relevant subagent as evidence to triage — not as the verdict (§4). Surgical-changes check (all aspects, incl. Tier-2 language reviewers): flag any changed hunk that does not trace to a spec line or PR-description intent as an out-of-scope finding — drive-by refactors, formatter/style churn, and edits outside the change's stated purpose. Every changed line should trace to the request.

   Each subagent must return two artifacts:

   • Confirmed findings — issues the executor is confident about (Phase 1 complete, no escalation needed).
   • Advisor candidates — findings the executor flags for Phase 2 review. Each candidate must include: the finding text, a bounded code snippet (≤100 lines), the relevant spec excerpt (if any), and a one-sentence reason the executor wants a second opinion.

   Each reference file (references/*-agent.md) explains what "confident" vs "needs advisor" means for that category. The four code-review aspects (architecture / design / devex / security) stay in their own lanes: an Architecture finding does not include naming polish (Design) or error-message wording (Devex). Cross-aspect concerns surface as one-line cross-references, never as duplicated findings.

   Language-specific reviewer dispatch (Tier 2) — when the changed-file list is dominated by a single language and the Claude Code named-agent catalog at tools/claude-code/agents/ is installed, replace the Design and Devex generic model: sonnet subagents with the matching language reviewer for sharper findings. Currently shipped:

   • TypeScript / TSX / Vue / Svelte → subagent_type: ywc-typescript-reviewer (covers type-system depth, async correctness, framework idioms, tsconfig strictness, ESM/CJS interop)
   • Python / .py / .pyi / pyproject.toml → subagent_type: ywc-python-reviewer (covers type-system depth, asyncio correctness, framework idioms — Django / FastAPI / Pydantic v2 / Flask, mypy strict mode, GIL implications, lifecycle patterns)
   • Go / .go / go.mod / go.sum → subagent_type: ywc-go-reviewer (covers goroutine lifecycle, channel patterns, interface design — accept interfaces return concrete, error wrapping %w / errors.Is / errors.As, pointer vs value semantics, generics post 1.18, defer + sync primitives)

   Tier 2 reviewers for other languages (Swift / Rust) are follow-up PRs; in the meantime continue with the generic Design / Devex subagent prompts for those files.

4. Aggregate and Select Phase 2 Candidates — Combine candidate lists from all five subagents:

   • Deduplicate findings that share {file}:{line} across categories.
   • Cap the total at --advisor-budget (default 5). If candidates exceed the cap, prioritize: Critical > High > Medium, and within the same severity prefer Security > Architecture > Design > Devex > QA. (Architecture / Design / Devex order reflects irreversibility — structural decisions are hardest to walk back, then contracts, then operator UX.)
   • Log the candidates that were dropped due to the cap in the final report so the user can see what was not escalated.

5. Phase 2 — Advisor Pass (skip entirely if --no-advisor) — For each surviving candidate, spawn a short Opus subagent via the Task tool with model: opus. When the candidate's source category is Architecture and the Claude Code named-agent catalog at tools/claude-code/agents/ is installed, prefer subagent_type: ywc-architect so the advisor carries the dedicated architectural-decision persona (tools/claude-code/agents/ywc-architect.md). When the candidate's ambiguous axis is performance (latency / throughput / memory / bundle-size / Web Vitals regression on an Architecture or Devex candidate), prefer subagent_type: ywc-performance-engineer (tools/claude-code/agents/ywc-performance-engineer.md) so the advisor carries the dedicated Performance persona with concrete remediation (specific query rewrite, missing-index DDL, dynamic-import split point, Web Vitals fix path, profiler invocation recommendation). This is a routing hint — Phase 2 subagent count and budget are unchanged. Other categories use the generic model: opus dispatch.

   • Context payload: only the candidate's finding text, the bounded snippet, the spec excerpt, and the category-specific severity rubric from the matching reference file. Do not forward the full spec, the full file, or the Phase 1 transcripts.
   • Expected output: a short verdict (≤200 words) containing: confirmed severity, a one-line rationale, and either "confirmed" or "adjusted" (with the adjustment if any). The 200-word cap is tighter than advisor-pattern.md §3's observed ceiling of "typically <500 words" — 200 is an operational choice for this skill. If a genuinely complex candidate needs more room, invoke the §6 override: exceed the cap and justify the overrun in the final report's Advisor Budget Report section.
   • Opus calls are sequential, not parallel — each is small and fast, and sequential execution keeps the budget enforcement simple and auditable.

6. Merge and Output Report — Combine Phase 1 confirmed findings with Phase 2 verdicts. Mark each finding in the final report with its provenance (Phase 1 vs Phase 2) so the user can see which decisions involved frontier judgment.

7. Capture Learnings (skip if --skip-learnings) — After the report, offer to promote durable lessons into docs/review-learnings.md via ywc-review-learnings --mode update --source review: confirmed findings that recur (same class across files or recent reviews) become DO/DO-NOT learnings caught earlier next time, and any finding the user dismisses as a false positive becomes a FALSE-POSITIVE learning with the dismissal reason. This is what makes the next review on this repository sharper — do not write learnings without the user-confirmation CHANGESET that ywc-review-learnings enforces.

Output Format

## Implementation Review Result: {spec} vs {code}

### Summary
- Phase 1 findings (Sonnet/Haiku executor): Architecture A, Design D, Devex V, Security M, QA K
- Phase 2 advisor calls (Opus): X of Y budget used
- Phase 2 adjustments: N confirmed as-is, M severity-adjusted

### Architecture
1. [severity] [P1|P2] {file}:{line} — Description
   (if P2) Advisor verdict: {one-line rationale}

### Design
1. [severity] [P1|P2] {file}:{line} — Description
   (if P2) Advisor verdict: {one-line rationale}

### Developer Experience (Devex)
1. [severity] [P1|P2] {file}:{line} — Description
   (if P2) Advisor verdict: {one-line rationale}

### Security
1. [severity] [P1|P2] {file}:{line} — Description
   (if P2) Advisor verdict: {one-line rationale}

### Testing (QA)
1. [severity] [P1|P2] — Description
   (if P2) Advisor verdict: {one-line rationale}

### Fix Priority
1. (Sorted by Critical first, with [P1|P2] markers preserved)

### Advisor Budget Report
- Used: X of Y calls
- Dropped (over budget): {list, if any}

### Completion Status
(One of: DONE | DONE_WITH_CONCERNS | BLOCKED | NEEDS_CONTEXT)

Reporting Symbols: Replace the bracketed [severity] placeholder above with the corresponding symbol from symbols.md. Severity vocabulary is shared across all ywc-* review skills — use the same symbol for the same tier.

Example finding line:

🚨 [P2] src/api/users.ts:42 — SQL injection: input concatenated into query.
   Advisor verdict: confirm Critical; remediation requires parameterization.

The [P1] / [P2] marker is preserved unchanged — it describes Phase 1 vs Phase 2 escalation history, while the severity symbol describes finding impact. Both pieces of information are needed; do not collapse them.

For the Confidence Gate score in the report header, use the band marker from symbols.md: ✅ for PROCEED, ⚠️ for REVIEW, ❌ for STOP.

Completion Status rules:

[P1] marks findings confirmed entirely by the Phase 1 executor. [P2] marks findings that went through the Phase 2 advisor. This distinction matters when the user calibrates trust in the output — Phase 2 items represent the decisions the executor deemed genuinely ambiguous.

HTML mode (--format html) — emits the same findings as a self-contained HTML report: severity color coding, tab navigation, and a Copy as Markdown button. Structure and conventions follow html-output.md. The Markdown surface is preserved inside the file, so downstream integration is unaffected.

Agent Prompt References

Read the corresponding reference file when spawning each subagent and include the relevant section in the subagent prompt. The reference files describe both the standard review dimensions and the "advisor candidate" criteria specific to each category:

• references/architecture-agent.md — Architecture's review dimensions (structural spec conformance, pattern consistency, module interface, simplicity, surgical changes) + advisor triggers
• references/design-agent.md — Design's review dimensions (contract spec conformance, naming, signatures, error model, return shapes, public-surface discipline) + advisor triggers
• references/devex-agent.md — Devex's review dimensions (readability, error messages, logging, documentation, debuggability, config UX) + advisor triggers
• references/security-agent.md — Security's OWASP Top 10 checklist and advisor-escalation triggers
• references/qa-agent.md — QA's coverage analysis and advisor-escalation triggers

If a reference file does not yet contain an "Advisor Candidate Criteria" section, fall back to the three-property test in advisor-pattern.md §5: objective trigger, irreversibility, ambiguity. A finding must satisfy all three to be a Phase 2 candidate.

CodeRabbit-derived methodology — the verbosity dial (--profile), the verify-before-surface precision gate, the intent/context phase, static-analysis-as-evidence, and the per-project learnings loop are specified in references/coderabbit-methodology.md. These are the techniques imported to raise review quality toward bot-reviewer parity, in a runtime-agnostic form (no CodeRabbit dependency). The learnings half pairs with ywc-review-learnings, which owns docs/review-learnings.md.

Recurring real-world defects catalog — every reviewer agent additionally consults references/recurring-defects.md, a derived-from-data catalog of the defect classes that production bot reviewers (CodeRabbit, Codex Review) flag most often: data-layer access-boundary (ownership / tenant isolation), data-integrity / NULL handling, error-swallowing, external-call resilience, validation / fail-fast, HTTP-status semantics, and test fidelity. Each agent's "High-frequency real-world checks" section points at its catalog slice. The catalog tells the reviewer where real bugs cluster; the per-aspect rubric still governs severity and escalation. This is the same surface the executors' --review flag targets, so these classes get caught before the PR opens — reducing bot-review round-trips after the PR is created.

Confidence Gate

This skill applies the Confidence Gate to the aggregated review output before emitting the final report. The gate sits between Phase 2 advisor consolidation and report emission.

In addition, when the report's gate band lands in PROCEED and findings include a DONE (or DONE_WITH_CONCERNS) completion claim against the implementation, the surface must follow ywc-verify-done: the verification block (command, output excerpt, exit code) appears before the status line, no should / probably / seems wording appears in the conclusion, and any "this finding's fix was verified" claim cites the fresh command output that proves it. A report that reads "all clear, looks good" without an evidence block is not a Confidence-Gate PROCEED — downgrade to REVIEW until the evidence is attached.

Required dimensions (must each score ≥ 70):

• Evidence quality — Every finding must cite a verified source: file path, line number, or test output. Findings phrased as "this might be wrong" without primary evidence reduce the score even if the finding itself is plausible.
• Root cause identified — A finding that names only the symptom ("test fails") without identifying the underlying cause ("test fixture is shared across cases") cannot be remediated correctly. The reviewer must reach root cause before recording the finding.

Band-to-status mapping for this skill:

The gate score must appear in the report header. Per-finding [P1] / [P2] markers remain unchanged — they describe escalation history, while the gate describes overall report confidence.

Integration

• upstream: ywc-sequential-executor or ywc-parallel-executor (auto-invoked via --review flag)
• downstream: PR creation
• pairs with: ywc-review-learnings — loaded in Step 0 (read) to apply accumulated review preferences and called in Step 7 (update --source review) to capture new ones; docs/review-learnings.md is the per-project memory that compounds review quality
• pattern source: references/advisor-pattern.md, references/coderabbit-methodology.md