Mentu Plugin for Claude Code

The Commitment Ledger — accountability for AI agents.

Installation

From Marketplace (Recommended)

1. Add the Mentu marketplace:

   /plugin marketplace add mentu-ai/mentu-plugin
   

2. Install the plugin:

   /plugin install mentu
   

That's it! The plugin is now active.

Local Development

git clone https://github.com/mentu-ai/mentu-plugin
claude --plugin-dir ./mentu-plugin

Prerequisites

• mentu-cli: Install globally with npm install -g mentu-cli
• Mentu workspace: Run mentu init in your project

Commands

Command	Description
/mentu:mentu	Main command with subcommand dispatch
/mentu:commit "message"	Quick-create commitment
/mentu:claim cmt_xxx	Claim commitment
/mentu:submit cmt_xxx	Submit for review
/mentu:status	Show current state

How It Works

+---------------------------------------------------------------------+
|                    MENTU WORKFLOW                                    |
+---------------------------------------------------------------------+

SessionStart -----> Inject claimed commitments into context
                   "You have 2 open commitments..."

PostToolUse  -----> Capture Edit/Write/Bash as evidence
(automatic)        "Captured mem_xxx: Modified auth.ts"

Stop -------------> Enforce commitment closure
(if enabled)       "Cannot stop: cmt_xxx still open"

SessionStart Hook

When you start Claude Code, the plugin checks for your claimed commitments and injects them into the conversation context. This ensures agents resume work where they left off.

PostToolUse Hook

Every Edit, Write, or Bash operation is automatically captured as evidence in the Mentu ledger. No manual tracking required.

Stop Hook

When --mentu-enforcer is active, agents cannot stop until their claimed commitments are closed with evidence.

Team Setup

Add to your project's .claude/settings.json:

{
  "plugins": {
    "marketplaces": ["mentu-ai/mentu-plugin"],
    "installed": ["mentu@mentu-ai"]
  }
}

Environment Variables

Variable	Description
MENTU_ACTOR	Your actor identity (e.g., agent:claude-code)
MENTU_ENFORCER	Set to true to enforce commitment closure

Skill: mentu-workflow

The plugin includes a skill that auto-activates when you say things like:

• "track this task"
• "create a commitment"
• "add to my tasks"

It automatically captures -> commits -> claims -> submits.

Agents

Agent	Purpose
commitment-validator	Validates evidence sufficiency before closure
triage-agent	Classifies observations into actionable items

The Three Rules

1. Commitments trace to memories — every obligation has an origin
2. Closure requires evidence — proving done, not marking done
3. Append-only — nothing edited, nothing deleted

Recovery

When a commitment is rejected:

# 1. View rejection reason
mentu show cmt_xxx

# 2. Restore code (Claude Code)
# Press Esc twice or use /rewind

# 3. Address feedback and resubmit
mentu submit cmt_xxx --summary "Addressed feedback: ..."

Tip: /rewind recovers Edit/Write changes. Bash operations require manual cleanup.

Git-Backed Recovery

For long-running tasks or CI/CD, use git branches instead of /rewind:

# On claim: create branch
git checkout -b mentu/cmt_xxx

# During work: commit at logical points
git add -A && git commit -m "Add feature X"

# On submit: merge and cleanup
git checkout main
git merge --squash mentu/cmt_xxx

When to use which:

Situation	Use
Undo last few edits	/rewind
Bash broke something	git revert
Start over	git reset --hard

See skills/mentu-git-workflow/SKILL.md for full workflow.

Links

• Mentu Documentation
• Protocol Specification
• Report Issues

License

MIT

---

Agents start warm, every action leaves evidence, validation scales with risk.