Wilder1222/executing-plans

<!-- Inlined into every SKILL.md via {{PREAMBLE_CORE}}. Keep ≤30 lines. -->

Preamble (Core)

Status protocol — end every session with one of: DONE (evidence provided) · DONE_WITH_CONCERNS (list each) · BLOCKED (state what blocks you) · NEEDS_CONTEXT (state what you need).

Auto-advance — pipeline: THINK → PLAN → REVIEW → BUILD → VERIFY → RELEASE. Only human gate is spec approval at THINK. On DONE at other stages, print [STAGE] DONE -> advancing to [NEXT-STAGE] and invoke the next skill. On any non-DONE status at any stage, STOP.

Output directory — all artifacts go in docs/superomni/<kind>/<kind>-[branch]-[session]-[date].md. See CLAUDE.md for the full directory map.

TACIT-DENSE — before high-tacit decisions, classify D1 (domain expertise) · D2 (user-facing UX) · D3 (team culture) · D4 (novel pattern). On hit, output TACIT-DENSE [D#]: [question] — My default: [recommendation]. See reference for actions.

Anti-sycophancy — take a position on every significant question. Name flaws directly. No filler ("that's interesting", "you might consider", "that could work").

Telemetry (local only) — at session end, log bin/analytics-log. Nothing leaves the machine.

See preamble-ref.md for detailed protocols.

Executing Plans

Goal: Execute a written implementation plan precisely, with verification at each stage — running independent steps in parallel to minimize elapsed time.

Iron Laws

1. Dependencies First, Then Parallelize Never execute a step before its dependencies are complete. But DO run all independent steps in parallel within a wave — never serialize work that can be parallelized.

2. Evaluate Before Advancing Every wave must pass an evaluation gate before the next wave begins. A wave is not "done" until its outputs are verified — not just executed.

3. Failures Are Harness Signals When a step fails on 3 consecutive attempts using different approaches, stop executing and treat the failure as a harness signal: update the plan, skill, or constraint — then retry. Never brute-force through 3 failed approaches.

Phase 1: Load the Plan

# Find the plan document
ls docs/superomni/plans/plan-*.md 2>/dev/null | sort | tail -1

Read the plan. Confirm:

• [ ] Plan document exists and is readable
• [ ] Prerequisites are met
• [ ] You understand what "done" looks like for each step

Phase 2: Dependency Analysis — Build the Execution Wave Plan

Before executing any step, analyze ALL steps for dependencies:

DEPENDENCY ANALYSIS
─────────────────────────────────
Step 1: [name] — depends on: none
Step 2: [name] — depends on: none
Step 3: [name] — depends on: Step 1
Step 4: [name] — depends on: none
Step 5: [name] — depends on: Step 2, Step 3
Step 6: [name] — depends on: none
...

WAVE EXECUTION PLAN
Wave 1 (parallel): Steps 1, 2, 4, 6  ← 4 agents dispatched simultaneously
Wave 2 (parallel): Steps 3, 5        ← unblocked after Wave 1 completes
Wave 3 (if needed): ...

Est. time: [N waves] instead of [N sequential steps]

Rule: A step is independent if its outputs are not required by any step in the same wave. Rule: Aim for 5–10 steps per wave when sufficient independent steps exist — never artificially group dependent steps to meet this target.

Phase 3: Execute Wave by Wave

For each wave, dispatch all steps in the wave simultaneously, then wait for all to complete before starting the next wave.

Step Execution Protocol

For each step in a wave:

EXECUTING WAVE [N] — [M] STEPS IN PARALLEL
─────────────────────────────────
Steps: [list of step names/numbers]

For each individual step:

Step [N] — [Step Name]
─────────────────────────────────
What: [Description from plan]
Files: [Files to touch]
Involves code changes? [YES / NO]

1. Read — understand what the step requires
2. TDD Check — if this step involves writing or modifying source code: apply the test-driven-development skill (RED phase) with the step description, files to be modified, and expected behavior. Write the failing test suite first, confirm the tests fail, then implement the minimum code to make tests pass (GREEN), and refactor as needed. (TDD runs in main context — it was consolidated from the retired test-writer agent into the test-driven-development skill.)
3. Frontend Check — if this step involves UI files (.html, .jsx, .tsx, .vue, .svelte, .css, .scss): apply the frontend-design skill Phase 4 (Implementation) with the plan's design direction. After completing all UI steps in a wave, run the frontend-designer agent quality gate (Phase 5).
4. Do — make the minimum change needed for this step only
5. Verify — run the step's verification criterion
6. Report — confirm step complete or blocked

TDD Integration for Code Steps

Every step that creates or modifies source code must follow this flow:

Step involves code? ─── NO ──→ Execute directly
        │
        YES
        ↓
Apply test-driven-development skill (RED) → failing tests written → confirm they fail
        ↓
Write minimum implementation (GREEN) → confirm test passes
        ↓
Refactor if needed → confirm tests still pass
        ↓
Continue to step verification

If no test framework exists for this project: document what the tests would look like and why they cannot be automated. This is a DONE_WITH_CONCERNS, not a skip.

Frontend-Design Integration for UI Steps

Step involves UI files? ─── NO ──→ Skip
        │
        YES
        ↓
Load design direction from plan (## Design Direction section)
        ↓
Apply frontend-design Phase 4 (Implementation) rules
        ↓
After all UI steps in wave complete:
  Run frontend-designer agent quality gate (7+/10 on all dimensions)
        ↓
Gate PASS → continue | Gate FAIL → fix and re-run (2 retries)

If no design direction exists in the plan: run frontend-design Phase 1-2 (Context Gathering + Design Direction) before implementing. This is a one-time cost per session.

Step Completion Format

✓ Step N COMPLETE
  Changed: [files modified]
  Evidence: [test output or verification proof]

Step Blocked Format

✗ Step N BLOCKED
  Blocker: [what prevents completion]
  Tried: [what was attempted]
  Options:
    A) [approach 1]
    B) [approach 2]
    C) Skip this step (explain consequences)
    D) Other — describe your own approach: ___________

Phase 4: Wave Evaluation Gate

Before advancing to the next wave, run the evaluation gate:

WAVE [N] EVALUATION GATE
─────────────────────────────────
Steps completed: [list]
Tests passing:   [run: npm test or equivalent]
Regressions:     [any pre-existing tests broken?]
Output contract: [do outputs match what dependent steps expect?]
Gate result:     PASS → proceed to Wave N+1 | FAIL → address before advancing

If the gate FAILS:

1. Identify which step produced the failing output
2. Determine if this is a harness signal (update plan/skill) or an implementation error (fix and re-run)
3. Do NOT advance to the next wave until the gate passes

Apply the verification skill as an independent wave-evaluation gate when any of these conditions apply:

• The wave contains ≥ 5 steps
• Any step in the wave reported DONE_WITH_CONCERNS
• This is the final wave of the plan

Apply with: the wave's acceptance criteria, all step completion blocks, and test output. The skill returns an EVALUATION REPORT with one of four verdicts: APPROVED / APPROVED_WITH_NOTES / CHANGES_REQUIRED / EVALUATION_INCOMPLETE. Do NOT advance to the next wave if the verdict is CHANGES_REQUIRED — return to the failing step(s) with the verification findings. (Evaluation content was consolidated from the retired evaluator agent into the verification skill.)

Phase 5: Mid-Plan Check-ins

After every wave completes, or when scope is expanding:

1. Report progress: "Completed N/M steps (Wave X of Y done)"
2. Flag if actual work diverges from plan
3. Surface any blast radius discovered mid-execution
4. Ask before proceeding if scope has changed

Phase 6: Handling Plan Deviations

If you discover the plan is wrong or incomplete:

1. Stop — do not improvise silently
2. Assess — is this a small mechanical fix or a fundamental issue?
3. Small fix (mechanical, <5 min): note it, fix it, continue
4. Large issue (taste or architectural): surface to user, wait for input

PLAN DEVIATION DETECTED
Step N: [Original plan says X, but actually Y]
Impact: [Low/Medium/High]
Recommendation: [Proposed resolution]
Awaiting: [Your decision before continuing]

Phase 7: Completion

When all steps are done:

PLAN EXECUTION COMPLETE
════════════════════════════════════════
Steps completed:    N/N
Waves executed:     W
Deviations noted:   N
Files changed:      [list]
Tests passing:      [output]
Status:             DONE | DONE_WITH_CONCERNS
Concerns (if any):
  - [concern 1]
════════════════════════════════════════

Save Execution Results Document

After completing execution, save the results as a Markdown document:

_EXEC_DATE=$(date +%Y%m%d)
_EXEC_BRANCH=$(git branch --show-current 2>/dev/null | tr '/' '-' || echo "unknown")
_PLAN_FILE=$(ls docs/superomni/plans/plan-*.md 2>/dev/null | sort | tail -1)

if [ -n "$_PLAN_FILE" ]; then
        _PLAN_BASE=$(basename "$_PLAN_FILE" .md)
        _EXEC_SESSION=$(echo "$_PLAN_BASE" | sed -E "s/^plan-${_EXEC_BRANCH}-//" | sed -E 's/-[0-9]{8}$//')
fi

if [ -z "$_EXEC_SESSION" ]; then
        _EXEC_SESSION="execution-run"
fi

_EXEC_FILE="execution-${_EXEC_BRANCH}-${_EXEC_SESSION}-${_EXEC_DATE}.md"
mkdir -p docs/superomni/executions
cat > "docs/superomni/executions/${_EXEC_FILE}" << EOF
# Execution Results: ${_EXEC_BRANCH}

**Date:** ${_EXEC_DATE}
**Branch:** ${_EXEC_BRANCH}

[Paste the full PLAN EXECUTION COMPLETE block here]

## Wave Log

[Paste wave-by-wave summary: steps in each wave, outcomes]

## Steps Log

[Paste all step completion/blocked entries here]
EOF
echo "Execution results saved to docs/superomni/executions/${_EXEC_FILE}"

Write the full execution log (wave plan, all step outcomes + the final PLAN EXECUTION COMPLETE block, formatted as Markdown) to docs/superomni/executions/execution-[branch]-[session]-[date].md. This file serves as the permanent record of the execution run for the user to revisit.

Then trigger the complete VERIFY sequence in order:

1. code-review skill (giving mode) — structured code review of all changes
2. qa skill — test gap filling, edge case exploration
3. verification skill — evidence-based acceptance criteria check

Do NOT skip any step. Each skill must report DONE before the next is triggered. If any step reports BLOCKED or DONE_WITH_CONCERNS, stop and surface to user before continuing.