ccw-chain

CCW Chain Orchestrator

Chain-based workflow orchestrator using chain_loader for progressive step loading and LLM-driven decision routing.

Discovery

1. chain_loader list — list all chains with triggers, entries, and descriptions
2. Match user intent to chain triggers.task_types / triggers.keywords
3. chain_loader inspect — preview chain node graph and available entries
4. chain_loader start — begin from default entry, named entry (entry_name), or any node (node)

Execution Protocol

When chain_loader delivers a step node with a skill/command doc:

1. Read the loaded doc content to understand the skill's purpose and interface
2. Assemble the Skill call: Skill(skill_name, args)
   • First step: args = "${analysis.goal}"
   • Subsequent steps: args = "" (auto-receives session context)
   • Special args noted in step name (e.g., --bugfix, --hotfix, --plan-only)
3. Propagate -y: If auto mode active, append -y to args
4. Execute: Skill(skill_name, args) — blocking, wait for completion
5. Advance: chain_loader done to proceed to next step

const autoYes = /\b(-y|--yes)\b/.test($ARGUMENTS);

function assembleCommand(skillName, args, previousResult) {
  if (!args && previousResult?.session_id) {
    args = `--session="${previousResult.session_id}"`;
  }
  if (autoYes && !args.includes('-y') && !args.includes('--yes')) {
    args = args ? `${args} -y` : '-y';
  }
  return { skill: skillName, args };
}

Auto Mode (-y / --yes)

• D1 Clarity Check: always choose "Clear" (skip clarification)
• Confirmation: skip, execute directly
• Error handling: auto-skip failed steps, continue pipeline
• Propagation: -y injected into every downstream Skill call

Delegation Protocol

When chain_loader returns delegate_depth > 0:

1. Continue normal execution (read content, assemble Skill, execute)
2. On returned_from_delegate: true, resume parent chain context
3. Variables received from child chain are available for subsequent steps