A Git-native multi-agent pipeline that builds, reviews, fixes, verifies, and optionally arbitrates code changes before delivery.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install adversarial-code-loop
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install adversarial-code-loop using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Adversarial Code Loop?
Adversarial Code Loop orchestrates a sequential BUILD → REVIEW → (FIX → VERIFY)^N → ARBITER workflow for code that needs independent model review. A developer agent writes and commits the implementation, a separate reviewer inspects the real Git diff, and fixer and verifier roles iterate until findings are resolved or the loop is rejected.
Each run uses an isolated loop/
Adversarial Code Loop Use Cases
- Require an independent AI code review before merging a critical change.
- Validate security-sensitive code involving authentication, authorization, secrets, or payments.
- Review well-scoped multi-file refactors of up to approximately 15 files.
- Pair different models or providers for builder, reviewer, fixer, verifier, and arbiter roles.
- Add build and test gates before approval or merge.
- Resume an interrupted pipeline from its persisted state.json file.
- Preserve rejected work on a branch for manual inspection and recovery.
- Generate review evidence and retrospective failure logs for engineering teams using Openclaw Skills.
How Adversarial Code Loop Works
- The pipeline detects or initializes a Git repository, records the parent branch and branch-point SHA, stashes dirty files, bootstraps Git identity, and creates an isolated loop branch.
- The DEV or BUILDER agent implements the supplied specification. The orchestrator stages the resulting files and creates a build commit.
- An optional build command runs as a quality gate.
- The REVIEW or CRITIC agent examines the cumulative Git diff from the branch point and returns structured findings with severity, file, line, evidence, and verdict data.
- The DEV or FIXER agent addresses the findings, and the orchestrator commits each fix as a separate round.
- The VERIFY agent checks every finding against the current diff and marks it resolved, rejected, or disputed.
- Steps 5 and 6 repeat until all findings are settled, approval is reached, or --max-loops is exhausted.
- An optional ARBITER or JUDGE resolves remaining disputes after the final review cycle.
- An optional test command runs before delivery. Approved work is evidence-tagged and squash-merged, while rejected work receives a [REJECTED] marker and remains on its loop branch.
- The pipeline restores the original stash on every exit path and writes final.json as the authoritative machine-readable result.
Adversarial Code Loop Setup
The skill supports Linux and macOS and requires the adversarial-common sibling repository, which provides the shared subprocess, provider, JSON, and Git engine.
Install the skill and its dependency with:
curl -fsSL https://raw.githubusercontent.com/chpomob/adversarial-code-loop/main/scripts/install.sh | bash
Or install from an existing checkout:
bash scripts/install.sh
Both methods place adversarial-code-loop and adversarial-common side by side under ~/.hermes/skills. Set a custom target with the installer argument or HERMES_HOME.
Run a basic loop with a required specification and project directory:
python3 ~/.hermes/skills/adversarial-code-loop/scripts/adversarial_loop.py \
--spec /tmp/spec.md \
--workdir /path/to/project
Configure model commands with flags or the supported environment variables:
export ACL_DEV_CMD='codex exec --skip-git-repo-check --sandbox workspace-write'
export ACL_REVIEW_CMD='pi --provider zai --model glm-5.2'
export ACL_ARBITER_CMD='pi -p --provider gemini --model gemini-3-pro'
Add build and test gates, bounded retries, and a longer timeout when needed:
python3 ~/.hermes/skills/adversarial-code-loop/scripts/adversarial_loop.py \
--spec /tmp/spec.md \
--workdir /path/to/project \
--feature secure-auth \
--build-cmd 'cargo build' \
--test-cmd 'cargo test' \
--max-loops 3 \
--timeout 1800
Use --no-merge when reviewing the branch manually, and --resume to continue a run from its artifact state. Only ACL_DEV_CMD, ACL_REVIEW_CMD, and ACL_ARBITER_CMD are currently read from the environment; workdir, timeout, loop count, and output directory must be passed as flags. Do not run concurrent loops against the same worktree.
Adversarial Code Loop Data Schema & Taxonomy
Artifacts are written below
| File | Lifecycle stage | Purpose |
|---|---|---|
| state.json | Git setup | Stores completed phases, loop number, branch, branch-point SHA, stash ID, and findings for resume support. |
| 00_spec.txt | Build | Verbatim copy of the implementation specification. |
| 01_build.json | Build | Build result and commit SHA. |
| 01_build_gate.json | Build gate | Result of the optional --build-cmd command. |
| 02_review.json | Review | Findings array and REQUEST_CHANGES, APPROVE, or REJECT verdict. |
| 03_fix_ |
Fix | Result metadata for each fix round. |
| 04_verdict_ |
Verify | Per-finding resolved, rejected, or disputed statuses and verdict. |
| 05_arbiter.json | Arbitration | Arbiter decision and any recorded conditions. |
| 06_test_gate.json | Test gate | Result of the optional --test-cmd command. |
| final.md | Completion | Human-readable summary and evidence-tag annotation. |
| final.json | Completion | Machine-readable contract containing verdict, reason, loops, branch, merged state, conditions, arbitration status, and artifact directory. |
Review findings use a structured schema with an id, severity of blocker, major, minor, or nit, file, integer line, summary, and evidence. Missing IDs receive deterministic auto_
Git metadata follows a predictable taxonomy: loop/
Adversarial Code Loop Advanced Features
- Multi-role orchestration with distinct BUILDER, CRITIC, FIXER, VERIFIER, and JUDGE personas.
- Provider-agnostic command configuration for Codex, pi, Claude through claude-tmux, DeepSeek, GLM, Gemini, and other compatible CLIs.
- Quota-aware provider selection through --provider-config, --force, and repeatable --force-provider role:alias options.
- Git-native isolation, cumulative diff review, stash restoration, branch preservation, squash merging, and annotated evidence tags.
- Resumability from state.json after interruptions, timeouts, or partial phase completion.
- Optional build and test gates that convert failed validation into a controlled rejection.
- Robust JSON parsing that handles raw JSON, markdown-fenced JSON, surrounding prose, and extracted arrays, with one retry for malformed review output.
- Privilege reduction for review commands by stripping known dangerous flags, while clearly documenting that this is not OS-level sandboxing.
- Configurable per-subprocess timeouts for long-thinking models and multi-phase execution.
- English-only internal specs, personas, findings, verdicts, and commit metadata to keep role handoffs consistent.
- Automatic retrospective logging for diagnosing provider, timeout, Git, and workflow failures.
- Compatibility with Openclaw Skills workflows that need auditable, repeatable, multi-agent code quality controls.
SKILL.md
Loading
Related Openclaw Skills
Livestock Aggressive Behavior Detection
Analyzes livestock and poultry images or continuous barn videos to identify aggressive behavior, rate its intensity, and generate alert reports.
Multi-Research
Openclaw Skills multi-research delivers multi-dimensional A-share analysis, risk debate, and weighted scoring for faster, more stable investment decisions.
REST API Reference Free Edition
Openclaw Skills provides a free REST API reference for fast, secure integration across AI/ML, payments, and communication services.
API Scaffold Generator Pro
Openclaw Skills for fast, enterprise-grade API scaffolding across multiple stacks, DDD, microservices, OpenAPI, and deployment automation.
Session Management for Claude Code
Session Management gives Claude Code a durable, structured memory system for resuming work, checkpointing progress, closing sessions, and reviewing daily project health.
SocQ SEO Research
SocQ SEO Research gives AI agents a reliable workflow for collecting public keyword, ranking, intent, and organic search data.
Featured*
