Coach

Deprecated — retired in favor of retro-skill. This repository is archived and read-only.

Coach detected friction signals via runtime hook heuristics (regex over corrections, exit codes, tone) and recorded them to events.sqlite for later analysis. In practice, retro does one LLM pass over the actual session (and cross-session via memory/session scans) and judges friction semantically — providing a strictly better signal than pre-recorded heuristic proxies. Retro no longer consumes Coach's events; it analyzes sessions directly.

• Use netresearch/retro-skill instead (/retro, /retro "<problem>", cross-session sweeps).
• The content below is kept for historical reference only and is no longer maintained.

Self-improving learning system for Claude Code that detects friction signals and proposes rule updates.

🔌 Plugin Type: Feature Plugin

This is a Feature Plugin - it provides active automation via hooks and slash commands that run during your session, unlike Skill Plugins which only provide reference knowledge.

What makes it a Feature Plugin:

• Hooks that detect friction signals in real-time
• Slash commands for reviewing and approving changes
• Session analysis and cross-repo learning

Coach follows the Agent Skills specification and includes a skill component (SKILL.md), but extends it with automation features.

Supported Platforms:

• ✅ Claude Code (Anthropic)
• ✅ Other platforms supporting hooks and commands

Features

• Signal Detection: Automatically detects user corrections, tool failures, repeated instructions, and tone escalation
• Skill Update Suggestions: Detects when users supplement skills with additional guidance and proposes skill updates
• Outdated Tool Detection: Identifies deprecated tools and outdated dependencies from command output
• LLM-Assisted Generation: Uses Claude Haiku to generate specific, actionable learning candidates
• Transcript Analysis: Analyzes full session transcripts at session end for comprehensive learning
• Cross-Repo Learning: Tracks patterns across repositories and proposes promotion to global rules
• Proactive Scanning: /coach scan checks for outdated CLI tools and project dependencies
• Approval Workflow: All changes require explicit user approval

Installation

Install via Claude Code plugin marketplace:

/plugin marketplace add netresearch/claude-code-marketplace

Then install Coach from the plugin list. The plugin auto-configures itself on first use.

Slash Commands

Command	Description
/coach status	Show system status and statistics
/coach review	Show pending learning proposals
/coach approve <id>	Approve and apply a proposal
/coach reject <id>	Reject a proposal with reason
/coach edit <id>	Edit a proposal before approving
/coach promote <id>	Promote project rule to global
/coach scan	Scan for outdated tools and dependencies
/coach init	Initialize the coach system

How It Works

1. Hooks detect friction signals as you work:

   • UserPromptSubmit: Detects corrections and repetition in user messages
   • PostToolUse: Captures command failures with exit codes
   • Stop: Runs full aggregation at session end

2. Signals are stored in ~/.claude-coach/events.sqlite

3. Candidates are generated from patterns in signals

4. Review candidates with /coach review and approve/reject

5. Rules are added to your CLAUDE.md files

Configuration

The plugin auto-configures hooks. For manual configuration or customization, see hooks/hooks.json.

Stable Hook Paths

Coach automatically maintains stable hook paths that survive plugin version updates. Hooks run asynchronously for improved performance.

How it works:

1. On first hook execution, Coach auto-installs ~/.claude-coach/bin/coach-run launcher
2. Settings.json hooks are automatically upgraded to use the stable launcher with --async mode
3. The launcher dynamically resolves the current plugin version at runtime
4. Async mode spawns scripts in background - hooks return immediately without blocking Claude Code
5. Future plugin updates work seamlessly - no user action required

Manual recovery (if needed): If hooks break after an update, running /coach init will repair them.

License

This project uses split licensing:

• Code (scripts, workflows, configs): MIT
• Content (skill definitions, documentation, references): CC-BY-SA-4.0

See the individual license files for full terms.