| name | execute-plan |
| description | Use when the user asks to execute, run, or implement a plan. Executes tasks from a plan file and updates checkboxes. |
Execute Plan
Execute the plan file provided.
Input
Path to plan file (e.g. docs/ai/plans/{feature-name}.md)
Hard Constraints — non-negotiable
- Do NOT explore the codebase broadly before starting
- Do NOT spawn parallel agents or run parallel discovery
- Do NOT read files not listed in the details file for the current phase
- Read file content only when executing the specific task that needs it
- Max 5 files read per task
Startup
- Read plan file
- Check for ## Enrich Summary section:
- If MISSING -> Inline mode (see Inline Mode section)
- If PRESENT -> Enriched mode (continue steps 3-5)
- Read total file count from ## Enrich Summary
- If plan has 2+ phases AND total files > 10 -> ask human:
"This plan touches {N} files across {X} phases.
Run all phases in one session, or one phase at a time? (all / phase-by-phase)"
Store the answer as {run_mode}.
- If plan has 1 phase OR total files <= 10 -> set {run_mode} = all, begin immediately
Enriched Mode — Process Per Phase
- Read docs/ai/plans/{name}-phase-{N}-details.md
- For each task:
a. Identify files listed in details file for this task
b. Open only those files
c. Find 1 similar existing file as pattern reference if needed
d. If pattern not found in CLAUDE.md or codebase -> ask, do not assume
e. Implement
f. Mark task done in plan file: [ ] -> [x]
Phase Transition
- If {run_mode} = all -> proceed to next phase immediately, no checkpoint
- If {run_mode} = phase-by-phase -> after each phase print:
"Phase {N} complete.
Done: [list what was implemented]
Files changed: [list files]
Reply 'continue' to proceed to Phase {N+1}, or give feedback."
Wait for explicit 'continue' before proceeding.
If feedback given -> address it, then re-output the checkpoint message.
Inline Mode (no Enrich Summary)
- Plan has 1 phase and <= 5 tasks — execute immediately, no questions
- Allowed to grep by task keywords to locate relevant files (max 3 files read per task)
- No details file required — use task descriptions as sole guide
- No phase checkpoint
After All Phases Complete
Write summary to docs/ai/summaries/{feature-name}.md
Summary Format
## Done
- [what was implemented, mapped to AC numbers]
## Not Done / Blocked
- [what was skipped and why]
## Verified
- AC1 [symbol] / AC2 [symbol] ...
## Not Verified
- AC3 [symbol] [reason]
Summary Rules
- AC pass only if: test passed (show output) OR human explicitly confirmed
- AC fail if: not tested yet, test skipped, or test failed
- Never self-judge AC as passed - evidence required
- If no test exists for an AC -> "No test coverage"
Blocked Rule
If blocked (open question, ambiguous pattern, missing context):
- Stop immediately
- Report: what the blocker is, which task it affects, what info is needed
- Do not guess or assume - wait for human input
