executing-plans

name	executing-plans
description	Use when implementing a written plan task-by-task — the per-task cycle, commit discipline, and escalation rules that an execute-plan runbook orchestrates around.

Executing Plans

## Runbook-Orchestrated Skill This runbook requires the plan's **artifact URI** as `PlanPath` — an `rd://…` value produced by `write-plan`, **not** a filesystem path. Resolve `PlanPath` first (ask the user if it is not already known). Load the execution protocol *before* starting, so you can handle delegated tasks the moment they appear:

Load and follow the bundled running-runbooks skill before starting the runbook.

Stage 1 does not provide Codex hook parity for delegated child dispatch. If this workflow reaches a delegated step that requires automatic child-agent orchestration, stop and report that Stage 2 delegation hook support is required.

Then start it:

• Delegated pipeline (write-plan → execute-plan): PlanPath is inherited automatically — no explicit supply needed.
• Standalone: pass the plan's artifact URI through the dedicated artifact channel — rundown run rundown:execute-plan --artifacts PlanPath=<rd://… plan URI> (e.g. --artifacts PlanPath=rd://artifacts/<ctx>/<run>/plan.json). PlanPath is declared REQUIRED and consumed as an ARTIFACTS input by the runbook, so it must arrive as an rd:// URI, not a filesystem path. Discover the alias with rundown artifact ls, then read the uri field of rundown artifact uri PlanPath (JSON is the default; agents do not add --text).

Capture the run id from rundown run: it is printed at start and echoed as runbookId on every event. This runbook delegates the implementer, so every orchestrator command must carry --run <rd_…>.

Implement a written plan one task at a time, holding each task to its own tests and committing as you go. This skill is the context an execution runbook orchestrates: how to do each task well. The execute-plan runbook owns the sequence (implement → review → verify) and the gates; this skill owns the craft of a single task.

Rundown orchestrates workflow; it does not store craft. Keep the cycle here, not in the runbook.

When to Use

• Implementing a plan produced by writing-plans, whether driven by the execute-plan runbook or by hand.
• Resolving review findings against an already-implemented plan.

When NOT to Use

• Writing the plan — use writing-plans.
• Authoring or sequencing the runbook that drives execution — use writing-runbooks.

The Per-Task Cycle

Work the plan's tasks in order. For each task:

1. Follow its bite-sized steps exactly. A well-written task is TDD-shaped: write the failing test, run it red, implement the minimum, run it green.
2. Run the verifications the task specifies. Do not skip them.
3. Commit per the task's commit block before starting the next task.
4. Keep moving. Do not pause between tasks to check in — execute the whole plan. Stop only to escalate (below).

Commit Discipline

One commit per task, staging exactly the files the task's commit.files lists, with the task's commit.message. Frequent, atomic commits keep the work bisectable and give the review and verify gates a clean history to act on.

Review and Verify Gates

The execute-plan runbook reviews the implemented changes and runs npm run verify after implementation, looping a fix step until both are clean. Your responsibility is to make those gates reachable: leave the tree building, tests passing for the work you did, and changes scoped to the plan. When a gate sends work back (via address-review), resolve the recorded error-level findings without expanding scope.

When to Stop and Escalate

Stop and ask rather than guess when:

• A dependency, file, or symbol the plan references does not exist.
• A task's instruction is ambiguous or contradicts the codebase.
• A verification fails repeatedly and the fix is unclear.
• The plan has a gap that prevents starting a task.

Never start implementation on main/master without explicit consent.

Reference

• writing-plans — produces the plan this skill executes
• running-runbooks — executing the driving runbook
• Stage 1 delegated child dispatch note: automatic Codex child-agent orchestration requires Stage 2 hook support
• composing-runbooks.md — how execute-plan composes the leaves