Orchestrate Plan Execution

Execute features sequentially using sub-agents with verification.

Before Starting

Read ${CLAUDE_PLUGIN_ROOT}/skills/planning/SKILL.md for system overview.

Input

• Plan directory: $ARGUMENTS (e.g., dev/plans/my-feature/)

Workflow

1. Create/switch to plan branch: git checkout -b plan/<plan-name> or switch if exists

2. State reconciliation: Run recovery script

   bun run ${CLAUDE_PLUGIN_ROOT}/verification/reconcile-state.ts $ARGUMENTS
   

3. Load context files: Read manifest.jsonl, context.md, constraints.md from plan directory

4. Execute features sequentially:

   For each pending feature:

   a. Update status to in_progress in manifest and Beads

   b. Spawn sub-agent with:

   • Feature prompt content
   • Context from constraints.md
   • Verification requirements from ${CLAUDE_PLUGIN_ROOT}/skills/planning/reference/quality-gates.md

   c. Wait for sub-agent completion

   d. Verification sub-agent: Spawn verification agent to audit the work

   • Prompt: "Verify feature <ID> implementation against requirements"
   • Must confirm: tests pass, build succeeds, changes committed

   e. Mechanical verification: Run verify-feature.ts script

   f. If verified: Mark completed in manifest, close Beads task

   g. If failed: Mark failed, report to user, ask how to proceed

   h. Sync DevOps status (if configured)

5. Final validation: Run full test suite, sync DevOps

6. Create PR: Uses dual-repo support (GitHub or Azure DevOps)

   bun run ${CLAUDE_PLUGIN_ROOT}/verification/create-pr.ts $ARGUMENTS plan/<plan-name>
   

State Output (every response)

Output current state at the end of EVERY response:

---
## 🎯 Orchestration State
| Metric | Value |
|--------|-------|
| Current Feature | <current-id> - <status> |
| Progress | <completed> / <total> |
| Branch | plan/<name> |
| Last Verification | <PASS/FAIL> |

### Pending Features
- <next-id>: <title>

Constraints

• Execute ONE feature at a time
• NEVER trust sub-agent claims - always verify mechanically
• NEVER merge to main - create PR only
• Safe to re-run at any point (idempotent)

Output

• Committed code for each feature
• Updated manifest.jsonl with completion status
• Closed Beads tasks
• Pull request for review