# Coordinate AI Agents with Agentic Jujutsu

Concurrent AI agents can create conflicting changes and repeat failed workflows. This skill documents Agentic Jujutsu methods for coordination, trajectory learning, and integrity checks.

## Install

```bash
npx skillstore add ruvnet/agentic-jujutsu
```

## Metadata

- Status: approved
- Slug: ruvnet-agentic-jujutsu
- Skillstore revision: r2
- Version status: missing
- Tree hash: 28b157084309b96154aa3ac61374e601a4d851c51aa256029853399ed4f89585
- Author: ruvnet
- GitHub username: ruvnet
- License: MIT
- Repository: https://github.com/ruvnet/claude-flow/tree/main/.claude/skills/agentic-jujutsu
- Ref: ebdfe608f5de2b66ff37ab4af12af8ac4f5e8006
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: confirmation\_required
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, network
- Quality score: 38
- Quality tier: warning
- Public page: https://skillstore.pages.dev/skills/ruvnet-agentic-jujutsu
- Manifest: https://skillstore.pages.dev/api/skills/ruvnet-agentic-jujutsu/manifest

## Capabilities

- Initialize JjWrapper and perform status, commit, log, diff, branch, and rebase operations.
- Record task trajectories with operations, outcome scores, and critiques for later learning.
- Retrieve learned suggestions, discovered patterns, similar trajectories, and progress statistics.
- Coordinate multiple wrapper instances for concurrent development, review, and testing workflows.
- Generate and verify SHA3-512 fingerprints and configure documented trajectory encryption.
- Inspect operation counts, success rates, durations, and recent command history.

## Use Cases

- Coordinate Concurrent Development: Coordinate independent code changes across multiple AI agents while recording outcomes for later workflow suggestions.
- Improve Deployment Workflows: Track deployment trajectories, compare prior outcomes, and review suggested operations before running them.
- Check Repository Integrity: Generate integrity fingerprints and evaluate encryption workflows without assuming documented guarantees are independently verified.

## Prompt Templates

### Inspect Repository State

```
Use Agentic Jujutsu to inspect this repository and summarize status. Do not modify files or run installation commands without approval.
```

### Record a Task Trajectory

```
Track my task as a trajectory, record relevant version-control operations, and ask for an honest outcome score before finalizing.
```

### Coordinate Parallel Work

```
Plan independent branches for three agents, identify overlap, and propose a concurrent commit sequence using Agentic Jujutsu.
```

### Evaluate Learned Operations

```
Review ReasoningBank recommendations for this deployment. Validate every operation, flag destructive steps, and request approval before execution.
```

## Limitations

- Requires the external agentic-jujutsu NPM package and a compatible JavaScript runtime.
- It does not implement wrapper methods or verify package behavior, benchmarks, or cryptographic guarantees.
- Several examples depend on undefined deployment, review, and task-execution helpers.
- Learned recommendations reflect recorded trajectories and may contain unsafe or outdated operations.

## Best Practices

- Use specific task descriptions and record honest outcome scores with concise critiques.
- Review learned suggestions against repository state and require approval before destructive or networked operations.
- Redact secrets from commands and trajectory context before storing operation history.

## Anti Patterns

- Do not execute unpinned external packages without verifying provenance and obtaining approval.
- Do not run learned operations automatically or treat confidence scores as authorization.
- Do not inflate success scores or omit failures, because distorted trajectories weaken later suggestions.

## Security Audit

- Audited at: 2026-07-23T18:32:42.862\+00:00
- Summary: Most external-command alerts are false positives caused by Markdown fences, JavaScript template literals, and inline code. The setup command executes an NPM package, while examples automatically execute learned recommendations and record complete command strings. Publication should require safeguards for package installation, operation approval, and command-log redaction.

## Stats

- Views: 299
- Downloads: 10
- Favorites: 0
- Popularity score: 0