AGENT SKILL

review

Code review for PRs or staged changes. >80% confidence threshold. Verdict: APPROVE, REQUEST CHANGES, or COMMENT. Triggers: review, pr, code review.

使用场景代码审查
License 许可
Apache-2.0
最近更新
2026-11-14
Stars
★ 1.6k
Forks
491

01 前言 Front Matter

name
review
kernel
{ "kind": "validator", "version": 1, "confirmation": "none", "side_effects": "none" }
description
Code review for PRs or staged changes. >80% confidence threshold. Verdict: APPROVE, REQUEST CHANGES, or COMMENT. Triggers: review, pr, code review.
allowed-tools
Read, Bash, Grep, Glob
user-invocable
true

02 资源文件 1

03 SKILL.md 正文

<skill id="review">

<purpose> Review code changes for quality, correctness, security. Only report issues with >80% confidence.

Load: skills/build/reference/testing.md, skills/tearitapart/reference/security.md Reference: _meta/research/ai-code-anti-patterns.md </purpose>

<context> ai_code_stats: buggier: 1.7x more issues than human code security: 40-62% contain vulnerabilities findings: 10.83 per AI PR vs 6.45 human

priority: check Big 5 first (what AI actually breaks) </context>

<on_start>

agentdb read-start

<identify_scope>

gh pr diff {number}        # For PRs
git diff --staged          # For staged
git diff HEAD~1            # For recent

</identify_scope>

<deterministic_lane> Machines first, then judgment. Run before reading a single diff hunk:

"${CLAUDE_PLUGIN_ROOT:-.}/scripts/deterministic-review.sh" <repo-dir> [base-ref]

Parallel gitleaks / semgrep(p/security-audit) / eslint / ruff / shellcheck / actionlint / zizmor / osv-scanner — whatever is installed; diff-scoped when base-ref given; exit 1 on HIGH (secrets, security SAST, high CVEs). Feed findings.tsv into the review as ground truth. Lanes marked SKIPPED are reported as NOT CHECKED, never as clean. </deterministic_lane> </on_start>

<confidence_threshold>

Confidence Category Report?
95%+ Definite bug YES
85-95% Likely issue YES
70-85% Possible issue MAYBE
<70% Style preference NO
</confidence_threshold>

<big5 name="BIG 5: AI-SPECIFIC CONCERNS"> Check these FIRST - what AI actually breaks:

<check id="1_input_validation">

  • Zod/Pydantic schema for every API endpoint?
  • Parameterized queries (no string concat)?
  • File uploads validated (size, type, extension)? detection: grep -r "req.body" | grep -v "parse|validate|z." </check>

<check id="2_edge_cases">

  • Null/undefined handling present?
  • Empty arrays handled (length check)?
  • Zero-length strings rejected?
  • Timeout handling for external calls? </check>

<check id="3_error_handling">

  • No empty catch blocks?
  • Errors logged with context?
  • User-facing messages generic? detection: grep -r "catch.*{}" (empty catch) </check>

<check id="4_duplication">

  • Same logic repeated in multiple places?
  • Should be extracted to shared utility? </check>

<check id="5_complexity">

  • Run the configured project verify command first. If complexity is not on that armed path, report the enforcement gap; a manual measurement is not a gate.
  • Any function above its .ccnrc budget? A number, not a guess.
  • If a baseline exists, does the diff say regressed=0?
  • For JS/TS, did stderr name eslint AST? Lizard's fallback does not see object-literal methods.
  • Nested ternaries > 2 levels? detection: ${CLAUDE_PLUGIN_ROOT}/scripts/complexity.sh <repo> <base-ref>; ${CLAUDE_PLUGIN_ROOT}/scripts/complexity.sh --diff <baseline.tsv> <repo> fix: /kernel:simplify on the listed functions </check> </big5>

<checklist> <section name="Logic & Correctness">

  • [ ] Edge cases handled
  • [ ] Error paths covered
  • [ ] Null checks present
  • [ ] Type safety </section>

<section name="Security">

  • [ ] Input validation (Zod schema)
  • [ ] No hardcoded secrets
  • [ ] SQL injection prevented (parameterized queries)
  • [ ] XSS prevented (context-aware output encoding / framework auto-escaping first; DOMPurify only where raw HTML is injected, e.g. dangerouslySetInnerHTML, v-html, innerHTML)
  • [ ] Auth tokens in httpOnly cookies </section>

<section name="Performance">

  • [ ] No N+1 queries
  • [ ] Appropriate caching </section> </checklist>

<refute_before_report> Before a finding enters the report, check it against reference/refutation-patterns.md (11 false-alarm families + 5 refutation moves). A finding matching a family needs evidence distinguishing it from the refuted precedent. Severity gates from the same file apply: falsification check before CRITICAL, consensus deflation, authority stripping on CVEs, and the judge's own additions get refuted too. Refuted candidates go in the report WITH their refutations; if the project keeps a refutation-corpus file, append new refuted false alarms to it in the same session. </refute_before_report>

<output_format> CODE REVIEW

Files: X changed Findings: Y (Z critical)

CRITICAL

[file:line] Issue (confidence: XX%) → Fix: suggestion

HIGH

[file:line] Issue (confidence: XX%) → Fix: suggestion

Summary: APPROVE | REQUEST CHANGES | COMMENT </output_format>

<ask_user> Use AskUserQuestion when: a finding is between 70-85% confidence (ambiguous) Ask: "Found {issue} at {file:line} (confidence {X}%). Intentional, or should I flag it?" Options: intentional — skip, flag it, investigate deeper </ask_user>

<verdict_rules> Review is finite because its purpose is AUTHORIZATION, not exhaustion. Approve once the change improves the health of the codebase, not once nothing else can be found. Nothing else can ever be found. (#204)

  • APPROVE: no finding clears the block bar. Open non-blocking comments do NOT prevent APPROVE — say the issue number out loud and approve. Forcing another round to re-check a nit costs more than the nit.
  • REQUEST CHANGES: at least one finding clears the block bar below.
  • COMMENT: findings worth saying, none blocking, and the author asked for a read rather than a decision.

<block_bar> A finding blocks only with ALL of:

  1. pasted output, not a described procedure
  2. a distance proof threshold cleared — distance sets how much evidence is needed, never whether a finding may block: d0 changed code fails · d1 violates a declared invariant · d2 pre-existing defect this change exposed, needs a user-visible consequence · d3+ needs outside assumptions, blocks only on an executed demonstration or a cited prior failure. Never auto-close d3; taste never clears it, a real outcome does.
  3. a named observable failure: predict what breaks and how we would see it
  4. violates the acceptance profile: tag the finding's dimension, and let the profile's blocks_at map decide the threshold for that dimension. A production-hardening finding on a demo is not a blocker however real it is, and a privacy finding on a demo IS one when the profile says so. The stage label never decides; the dimensions do.

Everything else is filed with its distance:N label under the quarantine milestone. Recurrence is signal, silence is a verdict.

Complexity clears the block bar only when changed code exceeds a declared budget, a regression diff names the increase, or the acceptance record requires the project gate and that armed path does not run it. A lizard result over JS/TS object-literal methods cannot support APPROVE or FAIL; it is NOT CHECKED until an AST-aware parser runs. </block_bar>

Read the acceptance record before reviewing: claims, declared invariants, and tradeoffs already accepted. If the commit carries one, it is FROZEN and adjudication will refuse a FAIL. Reopening takes new_failing_input, changed_dependency, missed_requirement, disproven_assumption, profile_changed, or owner_promotion. A new reviewer is not a reopen event. Never ask a reviewer whether criticism is complete; it cannot answer and will always say no.

An acceptance-profile review never ships ALONE. Pair it with one outcome-shaped discovery pass (DISCOVERY AXIS: outcome) that never receives the profile, invariants, or ACCEPT WHEN list. Measured (EXP-RTP1 variant M, #206): on a real defect the contract never names, the bounded reviewer caught it 0/20 and returned PASS 18/20, while the open reviewer caught it 18/19. The contract terminates review; alone it certifies what it was never told to see. </verdict_rules>

<on_complete> Emit findings as a kernel.verdict/v1 document and let the acceptance function decide. The reviewer never grades its own completeness.

python3 scripts/adjudicate.py findings.json --text --strict   # 0 PASS · 1 FAIL · 2 INVALID
agentdb write-end '{"command":"review","verdict":"X","critical":N,"high":N,"big5_violations":N,"quarantined":N}'

cannot_falsify is mandatory and non-empty. State what this review structurally could not see (no real device, conformance only, no live data, a suite that did not finish). Silence about coverage reads as coverage. </on_complete>

</skill>

04 同仓库更多 Skills

hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-review-with-annotations

Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.

★ 1.6k
hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-find-and-synthesize-context

Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.

★ 1.6k
hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-create-or-update-docs

Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.

★ 1.6k
hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-create-or-manage-records

Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.

★ 1.6k
hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-create-or-manage-interactive-html

Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.

★ 1.6k
hashgraph-online
hashgraph-online/awesome-codex-plugins

worktable-collaborate-threads

Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.

★ 1.6k