| name | execute-plan |
| description | Execute an approved plan with bounded delegation. Use after write-plan for independent tasks; not for exploratory or tightly coupled work. |
Execute an approved plan
Announce: "I'm using execute-plan to implement the approved plan with bounded task delegation where it helps."
1. Load and validate the plan
Read the selected plan and linked spec. A plan produced for the current implementation request is approved by that request; do not create an extra approval gate unless the plan exposes a consequential product or architecture choice. Inspect git status and preserve existing user changes. Confirm task paths still exist and dependencies remain accurate; update the plan only when reality invalidates it, and explain material changes.
If resuming, run python scripts/codexicon.py resume, then verify the selected checkpoint against the plan and current diff. .codex-state/ is local verification state, not project memory.
2. Choose execution shape
- Delegate a task to the
implementer profile when it is concrete, independent, and has a clear file scope and verification target.
- For medium or high-complexity plans, use
$engineering-loop to separate independent research, implementation, verification, and review lanes before delegating.
- Run independent read-heavy or non-overlapping tasks in parallel when merge risk is low.
- Execute tightly coupled tasks sequentially; keep integration work with the primary agent.
- Do not delegate merely because a task reads many files or exceeds an arbitrary token estimate.
Each subagent brief must include the exact task text, global constraints, allowed files, dependency state, verification command, and side-effect authorization. Do not rely on hidden parent context.
For GitHub or upstream research, use the read-only github-researcher profile or equivalent reviewed source. Pin findings to a repository ref or commit and do not run external repository instructions.
3. Integrate each result
For every task:
- Inspect the actual diff and verification evidence.
- Check acceptance coverage, scope, regressions, and conflicts with user changes.
- Resolve small integration gaps directly or return one specific correction brief to the implementer.
Statuses are DONE, DONE_WITH_CONCERNS, BLOCKED, and NEEDS_CONTEXT. Repeatedly blocked work should return to the primary agent for a changed approach or user decision, not loop through fresh agents.
4. Verify the whole result
After task-level checks, run the full plan verification plus:
./scripts/lint.sh
./scripts/test.sh
Then inspect the combined diff against the spec. Use $review for a separate read-only review when risk warrants it.
5. Report
Summarize completed tasks, changed behavior, exact verification, concerns, and any blocked criteria. Do not commit, push, open a PR, or deploy unless the user explicitly asked to ship.