execute

name	execute
description	Use when the user wants to run an implementation plan. Triggers on "실행해줘", "구현해줘", "시작해줘", "execute", or after plan approval.

Execute — Plan-Based Implementation with Auto-Ralph

Overview

Process all TODOs from a plan by dispatching executor agents with category routing. Automatically links ralph for persistent completion.

Activation

When this skill is invoked, ralph is automatically activated. The execution loop continues until all TODOs are complete and final verification passes.

Process

1. Load Plan

Read the plan from the vault path provided via skill args (e.g., --plan-path {VAULT_PATH}).

Path resolution:

1. If --plan-path is provided in args → Read the plan directly using Read tool from the vault path
   • The vault path format is an absolute path, e.g.: agmo-everywhere/plans/[Plan]-제목.md
   • Resolve to full path: {AGMO_VAULT_ROOT}/{vault_path} (e.g., {AGMO_VAULT_ROOT}/agmo-everywhere/plans/[Plan]-제목.md)
2. If no path provided → invoke agmo:vault-search to find the most recent plan for the current project
3. If Read fails (file not found) → invoke agmo:vault-search as fallback

Parse all TODOs into a work queue.

Classify each TODO:

• Category: quick / standard / complex (determines model)
• Dependencies: which TODOs must complete first
• Parallelizable: can this run concurrently with others?

2. Execute TODOs

For each TODO (respecting dependency order):

1. Dispatch executor agent with appropriate category
   - quick (haiku): simple file changes, config updates
   - standard (sonnet): feature implementation, most work
   - complex (opus): complex logic, architectural changes

2. Executor completes the TODO and reports results

3. Auto-trigger agmo:verification
   - If FAIL → collect failure context (approach taken + verification judgment)
   - Retry with accumulated feedback (up to 3 attempts):
     Each retry prompt MUST include ALL prior failure context:
     ```
     이전 시도 #{N}:
     - 접근법: {what was changed}
     - 실패 원인: {verification judgment summary}
     - 금지: 위 접근법 재시도
     ```
     (2nd retry includes attempt #1; 3rd retry includes attempts #1 + #2)
   - If 3 failures → auto-trigger agmo:debugging (include all failure contexts)
   - If PASS → proceed to step 4

4. Tag-based post-processing
   - Check the TODO's Tags field
   - If Tags contain `frontend` or `ui`:
     → Invoke agmo:accessibility with the TODO's changed files
     → If CRITICAL issues found → dispatch executor to fix → re-verify from step 3
     → If MINOR only → log summary, mark TODO complete
   - If Tags do not contain `frontend` or `ui`:
     → Mark TODO complete, proceed to next

Completion Status

When a TODO is marked complete (after step 3 PASS and step 4 post-processing), report one of the following statuses:

Status	When to use
DONE	All acceptance criteria met, no known issues
DONE_WITH_CONCERNS	Functionally complete but has known limitations or warnings
BLOCKED	Cannot proceed due to missing dependency or external blocker
NEEDS_CONTEXT	Requires clarification or additional context from the user

Output format per TODO:

Status: DONE | DONE_WITH_CONCERNS | BLOCKED | NEEDS_CONTEXT
Evidence: [Brief summary of how completion was verified]

For DONE_WITH_CONCERNS, also include:

Concerns:
  - [Concern description]
Follow-up: [Recommended next steps]

For BLOCKED, also include:

Reason: [Why it is blocked]
Blocker: [What needs to be resolved]
Next-Steps: [What must happen to unblock]

For NEEDS_CONTEXT, also include:

Question: [What clarification is needed]
Options: [Possible approaches, if applicable]

Behavior by status:

• DONE: Proceed to the next TODO immediately.
• DONE_WITH_CONCERNS: Log the concerns, report to the user, then proceed to the next TODO. Do not halt execution.
• BLOCKED: Report the blocker to the user and wait. Do not auto-proceed to the next TODO.
• NEEDS_CONTEXT: Ask the user the clarifying question and wait for their response. Do not auto-proceed.

Connection to verification:

• Verification PASS + no concerns → DONE
• Verification PASS + known limitations or warnings → DONE_WITH_CONCERNS
• Verification FAIL after 3 retries → debugging invoked; status resolved after debugging resolves
• Dependency missing or external blocker encountered → BLOCKED
• Acceptance criteria ambiguous before or during execution → NEEDS_CONTEXT

3. Parallel Execution

If multiple TODOs are independent (no dependency, no shared files, and no import/export dependencies between their target files):

• Invoke agmo:parallel to dispatch multiple executors concurrently
• Each executor gets exclusive file ownership
• Integration phase after parallel batch completes

4. Final Verification

After all TODOs are complete:

• Dispatch architect agent for final verification
• Verify: build passes, tests pass, all acceptance criteria met
• If any failure → address it and re-verify

5. Completion

When architect approves final verification:

1. Invoke agmo:save-impl to save implementation summary to Obsidian
2. If significant learnings occurred, suggest agmo:wisdom recording
3. Report completion to user

TDD Integration

If the plan specifies TDD:

• Before each TODO, invoke agmo:tdd to enforce RED → GREEN → REFACTOR
• Executor writes failing test first, then implementation

Error Handling

Situation	Action
TODO acceptance criteria unclear	Ask user for clarification
3 consecutive verification failures	Invoke agmo:debugging (3-fix limit)
Dependency deadlock	Report to user, suggest plan revision
Build broken mid-execution	Stop, fix build, then resume
Tags missing from TODO	Log warning, attempt fallback by file extension (.tsx/.jsx/.css/.html → treat as frontend). If ambiguous, skip post-processing