chunk-execute

name	chunk-execute
description	Run a chunk's full plan → implement → complete cycle in the current session. Use /chunk-execute to run a chunk inline. Use ve orch inject to delegate to a background agent.

Tips

• The ve command is an installed CLI tool, not a file in the repository. Do not search for it - run it directly via Bash.

Instructions

1. Determine the target chunk. If a chunk name was provided as an argument, use that. Otherwise, run ve chunk list --current to find the currently IMPLEMENTING chunk. We will refer to the chunk directory below as <chunk directory>.

2. Plan phase guard. Check if <chunk directory>/PLAN.md already has substantive content beyond the template skeleton. Look for an ## Approach section that contains actual implementation details (not just HTML comments or placeholder text).

   • If the plan is still a bare template: invoke /chunk-plan to create the plan. Wait for it to complete before proceeding.
   • If a plan already exists with real content: report "Plan already exists, skipping /chunk-plan" and proceed to the next step.

3. Implement phase. Invoke /chunk-implement to execute the plan.

4. Error gate. If implementation encounters errors (test failures, build errors, or issues that prevent the chunk from being considered complete), STOP and report the error to the operator. Do NOT proceed to the review phase. The operator may want to intervene or adjust the plan.

5. Review loop. After successful implementation, enter the review loop. Track the current iteration number starting at 1, with a maximum of 5 iterations.

   LOOP START:

   a. Run review. Invoke /chunk-review and read its output to determine the review decision (APPROVE, FEEDBACK, or ESCALATE). The reviewer will call the ReviewDecision tool and output a YAML decision block — use the decision: field to determine the outcome.

   b. If APPROVE — The review is clean. Exit the loop and proceed to the complete phase (step 6).

   c. If ESCALATE — STOP execution entirely. Report the escalation reason and questions to the operator. Do NOT proceed to complete. The operator must intervene.

   d. If FEEDBACK — Issues were found that need fixing:

   1. Check the iteration count. If this is iteration 5 (the maximum), STOP and report to the operator: "Review found issues but max iterations (5) reached. Manual intervention required." List the remaining issues.

   2. Write the review issues to <chunk directory>/REVIEW_FEEDBACK.md in this format so /chunk-implement can address them:

      # Review Feedback (Iteration N)
      
      The following issues were identified during review. Address each one.
      
      ## Issue 1: [concern]
      - **Location**: [file:line]
      - **Concern**: [what's wrong]
      - **Suggestion**: [how to fix]
      - **Severity**: [severity]
      
      ## Issue 2: [concern]
      ...
      

      Populate each issue from the FEEDBACK decision's issues array.

   3. Invoke /chunk-implement to address the feedback. It will read REVIEW_FEEDBACK.md and fix each issue (or defer/dispute with documented rationale).

   4. Error gate (re-implementation). If re-implementation encounters build/test errors, STOP and report to the operator.

   5. Increment the iteration count and go back to LOOP START.

6. Complete phase. Invoke /chunk-complete to finalize code references, run overlap analysis, and transition the chunk to its final status.

7. Summary. Report the final status of the chunk execution:

   • Which phases ran (plan, implement, review, complete)
   • Whether any phases were skipped (e.g., plan already existed)
   • How many review iterations were needed (1 = clean first pass)
   • The chunk's final status
   • Any issues encountered along the way