Execute Plan

Glob('{epic}/tasks/*.md')

Read each task's frontmatter. Build the dependency graph. Find tasks where:

• status: draft (not yet started)
• All entries in depends_on reference tasks with status: done

Pick the first executable task by ID order.

If no tasks are executable but some remain, report what's blocking (which dependencies aren't done) and stop the loop.

Step 3: Self-Review Task

Read the full task. Evaluate:

• Inputs clear? Critical files listed, acceptance criteria testable?
• Still relevant? Does this task still make sense given tasks already completed?
• Right-scoped? Completable in one session? If too big, split it and log the decision.

If the task needs changes, Edit it. Add an entry to its ## Decisions section:

## Decisions

- **[2026-03-25]** Rewritten: [what changed and why]

Then Edit frontmatter: status: approved.

Step 4: Execute Task

Load context proportionally — don't over-read:

• Read critical files listed in the task
• Read AGENTS.md of affected packages/apps (only the ones the task touches)
• If the task modifies a file, Read it first

Implement using Edit, Write, and Bash as needed.

Run verification checks based on what the task produces:

If verification fails, fix and retry. If it fails twice, stop — add a note to the task explaining what failed and why.

If the task produced code changes, create a PR via /github-create-pr. Include:

• Task reference (epic + task ID)
• What was implemented
• Verification results

Edit task frontmatter: status: done.

Step 5: Update Tracking

Edit the epic's AGENTS.md task table — update the task's status column.

Step 6: Loop or Complete

More executable tasks? → Back to Step 2.

All tasks done?

• Edit README.md: add a ## What Was Delivered section summarizing what was built, with PR links.
• Edit AGENTS.md frontmatter: status: done.

Tasks remaining but all blocked?

• Report: which tasks are pending, what dependencies they're waiting on.
• Stop the loop. The human may need to unblock (merge PRs, make decisions).

Boundaries

• Self-approve tasks within approved epics. Log every approval and rewrite.
• Cannot change epic scope. README.md goals and Definition of Done are the human's domain. If a task conflicts with the epic's DoD, stop and ask.
• Cannot merge PRs. Create them. The human merges.
• Stop on double failure. If a task fails verification twice, stop the loop and report.
• One task at a time. Don't batch-execute. Finish one, update tracking, then pick the next.

Dry Run

If the user says "dry run" or "show me what you'd do":

1. Find the epic (Step 1)
2. Scan all tasks (Step 2)
3. Self-review each executable task (Step 3) — describe changes you'd make
4. For each task, describe what you'd implement and how you'd verify it
5. Don't make any changes

This lets the human validate the agent's understanding before real execution.

Edge Cases

• No approved epics — report "no approved epics, run /review-plan first."
• Task already in-progress without changes — may be stale from a crashed session. Ask the human: resume or reset to draft?
• Circular dependencies — report the cycle, don't attempt execution.
• Task depends on task in another epic — cross-epic dependencies are valid. Check the other epic's task status.