version: "1.0.0" name: executing-plans description: The Orchestrator's high-level checklist for loading, reviewing, and tracking multi-step implementation plans. license: MIT metadata: ciel-version: 1.0.0 ciel-extension: ciel.yaml

CIEL ADAPTATION: Executing Plans (Plan Lifecycle Management)

This skill dictates how the Orchestrator manages the high-level lifecycle of a written implementation plan (generated by the make-plan skill). It ensures that plans are not blindly executed, but rather critically reviewed and systematically tracked.

Integration Context

Adapted from ~/.agents/skills/executing-plans/. In CIEL, the actual execution of individual tasks within a plan is handled by subagent-driven-development or dispatching-parallel-agents. This skill (executing-plans) acts as the outer loop: the state management and review mechanism for the entire epic.

The Execution Lifecycle

Phase 1: Load and Critical Review (Mandatory Checkpoint)

Before dispatching the first sub-agent, the Orchestrator MUST:

1. Read the complete plan file.
2. Critical Analysis: Analyze the plan for missing dependencies, circular logic, or unaddressed technical debt.
3. Escalation: If the plan has critical gaps or architectural flaws, the Orchestrator MUST halt and escalate to the human user (ask_user) before beginning execution. Do not attempt to silently patch a fundamentally broken plan.

Phase 2: State Tracking (The Checkbox Protocol)

As tasks are delegated to sub-agents (via subagent-driven-development), the Orchestrator maintains the global state:

1. Log the initiation of each task in the MemPalace Diary (mempalace_diary_write).
2. Mark tasks as IN_PROGRESS.
3. Upon successful verification of a sub-agent's work (Spec + Quality reviews passed), mark the task as COMPLETED.

Phase 3: Blocker Protocol

STOP executing immediately when:

• A sub-agent reports a fundamental blocker that requires architectural changes to the original plan.
• The verification-loop fails consistently across multiple sub-agent retries.
• Action: Return to Phase 1 (Review) and propose an updated plan to the human user. Do not force through blockers via endless sub-agent loops.

Phase 4: Epic Completion

Once all tasks are marked COMPLETED and the final global verification-loop passes, transition to the finishing-a-development-branch skill to handle PR creation and cleanup.

Anti-Patterns

• Blind Execution: Dispatching sub-agents to execute a plan without first critically analyzing it for structural flaws.
• Silent Plan Mutation: Modifying the high-level plan mid-execution without explicit human approval.
• Working on Main: Starting implementation directly on the main or master branch.