Autonomous code review pipeline with asymmetric dual-engine dispatch. Codex runs ARP's correctness + adversarial framings (2 dispatches); Gemini runs its own /ce:review compound-engineering pipeline (1 dispatch). Confidence-weighted consensus, bounded auto-fix loop, loop-thrash kill switch, safe-off defaults.
This plugin requires configuration values that are prompted when the plugin is enabled. Sensitive values are stored in your system keychain.
dryRunIf true, run review stages and print findings + proposed fixes without applying any Edit, commit, or PR comment.
${user_config.dryRun}autoCommitIf true, autonomously commit fixes at the end of the pipeline. Default false for safety — user should review before committing.
${user_config.autoCommit}failOnErrorIf true, abort pipeline when a stage can't PASS within maxIterations instead of escalating.
${user_config.failOnError}Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
geminiModelv5.3.0 pin: model is hardcoded to gemini-3-flash-preview in the dispatch wrapper regardless of this userConfig. Changing this value has no effect. Rationale: v5.2 cascade (Flash → Flash-Lite → operator-override Pro) shipped with two flaws — (a) Flash-Lite fallback doc admits '[] empty findings under load' (degraded quality), (b) Pro tier 429'd 28× in 2 min on v5.2 release day (unreliable). v5.3 ships Flash-only with API key Tier 1 auth + parallel persona spawn. Re-adding Pro requires a fork of this plugin.
${user_config.geminiModel}defaultEngineEngine selection when none specified on CLI. Values: both, gemini, codex.
${user_config.defaultEngine}maxIterationsMax auto-fix retry loops before escalating. Range: 1-10. Unlimited is intentionally not supported (cost safety).
${user_config.maxIterations}postPrCommentIf true, auto-post executive summary to the GitHub PR via 'gh pr comment'. Default false — review output locally first before making public.
${user_config.postPrComment}Autonomous dual-engine code review pipeline for Claude Code. Asymmetric dispatch — Codex runs ARP's dual-framing (correctness + adversarial) while Gemini CLI runs its own /ce:review compound-engineering pipeline. 3 parallel perspectives per iteration. Dedups findings by confidence, auto-fixes inline, escalates unresolvable findings. Verification delegated to your CI.
/ce:review. Codex gets ARP's framing discipline; Gemini uses its own multi-persona compound-engineering pipeline (P0-P3 tiering) because it already has one. No redundancy.file:line:severity:normalize(issue):sha1(fix_code[:200]). Multi-source agreement boosts confidence by +0.15 per extra source. Findings below 0.60 confidence are dropped.fix_code inline, re-runs until PASS or maxIterations (1-10, default 3). Unlimited loops are intentionally not supported.autoCommit and postPrComment default to false. --dry-run previews findings without editing. Dependency precheck fails fast if Codex / Gemini CLI missing..arp_session_log.json records Codex↔Gemini agreement rate. Tune dual-engine cost/value over time.flowchart TD
Start([/arp PR]) --> Pre
subgraph Stage0["Stage 0: Pre-flight"]
Pre[deps · flock · gh auth · PR resolve<br/>+ PR conversation context fetch]
end
subgraph Stage1["Stage 1: Review (dispatch → merge → fix → loop)"]
Disp{3 parallel dispatches}
Disp --> C1["Codex × Correctness<br/>Agent: codex-rescue<br/>read-only contract"]
Disp --> C2["Codex × Adversarial<br/>Agent: codex-rescue<br/>read-only contract"]
Disp --> G["Gemini × /ce:review<br/>Bash: gemini -p<br/>3-layer read-only"]
C1 --> S1[snapshot_git diff]
C2 --> S2[snapshot_git diff]
G --> S3[snapshot_git diff]
S1 --> M
S2 --> M
S3 --> M
M[Merge + Fingerprint<br/>file:line:severity:issue:fix-hash<br/>+0.15 per source · drop conf below 0.60]
M --> L{any findings?}
L -- yes --> K{fingerprint<br/>seen before?}
K -- yes --> E[ESCALATE<br/>human review]
K -- no --> F[Apply fix_code via Edit tool]
F --> I{iter < max?}
I -- yes --> Disp
I -- no --> FE{failOnError?}
FE -- true --> Ab([abort exit non-zero])
FE -- false --> E
end
subgraph Stage2["Stage 2: Deliver"]
D[summary · agreement rate · parse-error counts]
D --> AC{autoCommit?}
AC -- true --> GC[git commit]
AC -- false --> PC{postPrComment?}
GC --> PC
PC -- true --> SCR[Scrub secrets<br/>API keys · JWT · creds · bearer]
SCR --> GH[gh pr comment]
GH --> End([end])
PC -- false --> End
end
Pre --> Disp
L -- no --> D
E --> D
Why asymmetric (the "3 dispatches from 2 engines" question):
Codex Gemini
│ │
├── correctness framing (ARP) │
├── adversarial framing (ARP) ─┴── /ce:review
│ (multi-persona pipeline,
│ P0–P3 tiering, internal)
▼ ▼
2 dispatches 1 dispatch
└────────── 3 perspectives merged by fingerprint ──────────┘
Codex has no built-in multi-persona reviewer, so ARP supplies its dual-framing discipline directly. Gemini already ships /ce:review (the compound-engineering pipeline) — running ARP-side dual-framing on top would just duplicate work, so ARP delegates to it.
| Single-Engine Agent | ARP |
|---|---|
| One LLM, single framing | Codex × 2 framings + Gemini × /ce:review → 3 parallel perspectives |
| Suggests, leaves fix to human | Auto-fixes inline, bounded loop (cap 10) |
| Loops forever on unfixable bugs | Fingerprint kill switch escalates stuck findings |
| Commits without review | Opt-in commit and PR comment |
| Blind cost | Agreement telemetry for both → single-engine tuning |
/plugin marketplace add onchainyaotoshi/agent-review-pipeline
/plugin install agent-review-pipeline@agent-review-pipeline
/reload-plugins
Review the current branch, asymmetric 3-dispatch by default:
/arp
Review a specific PR:
/arp 42
npx claudepluginhub onchainyaotoshi/agent-review-pipeline --plugin agent-review-pipelineExpose your SiYuan wiki as native MCP tools — search, get, journal, ADR, push.
Multi-agent code review for Claude Code — parallel review by Codex (GPT-5.5), Gemini 3.1 Pro, and five Claude specialist subagents (one run blind), then verified synthesis.
Multi-lens code review pipeline: deep review (Claude or Codex), automated fix loop, interactive walkthrough, manual promote, external-finding injection.
Automated code review for pull requests using multiple specialized agents with confidence-based scoring
Agent skills for roborev code review, fix, refine, and response workflows.
AI-powered code review in Claude Code, powered by CodeRabbit
AI-powered code review analysis — Run three-level AI analysis and implement-review-fix loops directly in your coding agent. Works standalone, no server required.