Run

Run

Autonomous execution loop for TenantFlow.

Usage

• /run --plan — decompose project into milestones and tasks, show the plan, don't execute
• /run — start or resume execution from current state
• /run --task m1-t3 — execute a single specific task
• /run --milestone m2 — execute all tasks in milestone 2

Execution Loop

0. Reconcile State (always first)

• Read .claude/workspace/orchestrator_state.json if it exists
• Read .claude/tasks.json
• Scan .claude/workspace/ for result files
• If a task is marked done in tasks.json but has no result file → mark as review
• If a result file exists but tasks.json shows pending → validate and mark done if criteria pass
• Write reconciled state to orchestrator_state.json

1. Generate dashboard

• Copy dashboard template to .claude/workspace/dashboard.html

2. Pick next ready task

• Find the highest-priority task with status pending and all dependencies in status done

3. Assign to agent

• Use CLAUDE.md agent routing to select the correct agent

4. Validate result

• Read the agent's result file from .claude/workspace/[task-id].result.md
• Check each acceptance criterion — pass or fail
• If all pass → mark task done, log to progress.log
• If any fail → enter self-healing pipeline

5. Self-healing pipeline (on failure)

• Simple failure → retry once with hint
• Structural failure → invoke problem-solver
• Problem-solver pipeline: refine → split → reassign → escalate

6. Milestone boundary

• Auto-advance if auto_advance_milestones is true and all tasks pass

7. Completion check

• Check project completion criteria from CLAUDE.md