name: orchestration
description: MANDATORY - Load this skill first. Routes to bootstrap + domain rules.
impact: CRITICAL
version: 3.0.0
invokes:
- rlm-process

Clorch -- Claude Orchestration

   ─────────────────◆─────────────────
           ░█████╗░██╗░░░░░░█████╗░██████╗░░█████╗░██╗░░██╗
           ██╔══██╗██║░░░░░██╔══██╗██╔══██╗██╔══██╗██║░░██║
           ██║░░╚═╝██║░░░░░██║░░██║██████╔╝██║░░╚═╝███████║
           ██║░░██╗██║░░░░░██║░░██║██╔══██╗██║░░██╗██╔══██║
           ╚█████╔╝███████╗╚█████╔╝██║░░██║╚█████╔╝██║░░██║
           ░╚════╝░╚══════╝░╚════╝░╚═╝░░╚═╝░╚════╝░╚═╝░░╚═╝
   ─────────────────◆─────────────────

Step 1: Load Bootstrap (ALWAYS)

Read bootstrap.md first. It contains:
- Role detection (Orchestrator vs Worker)
- Iron Law + Iron Claw (UNBREAKABLE)
- Tool ownership table
- Worker preamble templates
- Memory recovery protocol

Workers: Load bootstrap.md, execute task, return.

Step 2: Detect Task Type

Step 3: Load Rules JIT

Rule Index

Domain References

Auto-Learning Protocol

After any agent completes a non-trivial task, automatically capture learnings to build institutional memory.

Worker-Level Learning

Workers now auto-capture learnings via the preamble template. The orchestrator
should ALSO capture its own coordination-level learnings (decisions about which
agents to use, workflow routing choices, patterns in user requests).

Learning happens at TWO levels:
1. Worker level - What the agent learned during execution (codebase patterns, solutions, errors)
2. Orchestrator level - What the coordinator learned about task routing and workflow effectiveness

When to Capture

Capture learnings after:
- Successful implementations (new features, refactorings)
- Architectural or design decisions
- Error resolution or debugging
- Discovery of codebase patterns
- Failed approaches (to avoid repeating)

Skip for: Simple file reads, quick lookups, trivial one-line fixes.

What to Extract

For each learning capture:
1. What was the task? (Brief description)
2. What approach was used? (Strategy, tools, techniques)
3. What decisions were made and why? (Key choices and rationale)
4. What worked / what didn't? (Outcomes, gotchas, surprises)
5. Key file paths and patterns (With line numbers if relevant)

Capture Format

cd /Users/lexudo/.claude/opc && PYTHONPATH=. uv run python scripts/core/store_learning.py \
  --session-id "<task-identifier>" \
  --type <TYPE> \
  --content "<what was learned>" \
  --context "<task context>" \
  --tags "tag1,tag2,tag3" \
  --confidence high|medium|low

Learning Types

Capture Rules

• Default behavior: Auto-learning is ON by default. No opt-in needed.
• Timing: Capture AFTER agent returns results, not during execution
• File paths: Include absolute paths with line numbers in content
• Tags: Always tag with the skill/workflow that triggered the task (e.g., "orchestration", "kraken", "build-workflow")
• Confidence: Use high for verified solutions, medium for working approaches, low for experimental findings

Example

cd /Users/lexudo/.claude/opc && PYTHONPATH=. uv run python scripts/core/store_learning.py \
  --session-id "hook-auto-build-impl" \
  --type WORKING_SOLUTION \
  --content "TypeScript hooks require npm install in .claude/hooks/ before build.sh compiles TS to JS in dist/. The build.sh script handles compilation automatically. Hook errors often come from missing dependencies or path issues." \
  --context "Implementing auto-build for TypeScript hooks in hook development workflow" \
  --tags "hooks,typescript,build,orchestration" \
  --confidence high

───◆─── Routing Ready ───◆───