Announce: “I’m using dev-implement (Phase 5) to orchestrate implementation.”

Load shared enforcement:

Read ${CLAUDE_SKILL_DIR}/../../references/constraints/dev-common-constraints.md.

Where This Fits

Main Chat (you)                    Task Agent
─────────────────────────────────────────────────────
dev-implement (this skill)
  → dev-ralph-loop (per-task loops)
    → dev-delegate (spawn agents)
      → Task agent ──────────────→ follows dev-tdd
                                   uses dev-test tools

Main chat orchestrates. Task agents implement.

Contents

• Prerequisites
• Implementation Strategy Choice
• The Iron Law of Delegation
• The Process (Sequential)
• Sub-Skills Reference
• If Max Iterations Reached
• Agent Team Implementation (Parallel)
• Test Gap Validation Gate (MANDATORY)
• Phase Complete

Implementation (Orchestration)

Do NOT start implementation without these:

1. .planning/SPEC.md exists with final requirements
2. .planning/PLAN.md exists with chosen approach
3. User explicitly approved in /dev-design phase
4. .planning/PLAN.md Testing Strategy section is COMPLETE (all boxes checked)

If any prerequisite is missing, STOP and complete the earlier phases.

Check .planning/PLAN.md for: files to modify, implementation order, testing strategy.

Pre-Flight Testing Check (MANDATORY)

Before starting ANY task, verify .planning/PLAN.md Testing Strategy:

[ ] Framework specified (not empty, not “TBD”)
[ ] Test Command specified (runnable command)
[ ] First Failing Test described (specific test name)
[ ] Test File Location specified (actual path)

If ANY box is unchecked → STOP. Go back to design phase.

This is your LAST CHANCE to catch missing test strategy before writing code. </EXTREMELY-IMPORTANT>

Implementation Strategy Choice

After prerequisites pass, check PLAN.md for parallelization potential:

Skip this choice when:

• PLAN.md has fewer than 4 tasks
• All tasks are dependent (every task is after N with no independent groups)
• CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS is not available

Otherwise, ask the user:

AskUserQuestion(questions=[{
  "question": "How should we implement the tasks in PLAN.md?",
  "header": "Strategy",
  "options": [
    {"label": "Sequential (Default)", "description": "One ralph loop per task, complete N before N+1. Safest, no merge conflicts."},
    {"label": "Agent team (parallel)", "description": "Spawn teammate per independent task group. Faster for 4+ independent tasks. Requires reconciliation."}
  ],
  "multiSelect": false
}])

If Sequential: Proceed to The Process below (current behavior).

If Agent team: Skip to Agent Team Implementation (Parallel).

YOU CANNOT WRITE IMPLEMENTATION CODE WITHOUT A FAILING TEST FIRST.

This is not a suggestion. This is the workflow. Every task follows:

1. READ the test description from PLAN.md
2. WRITE the test file
3. RUN the test → SEE RED (failure)
4. ONLY THEN write implementation
5. RUN the test → SEE GREEN (pass)

Rationalization Prevention (Implementation Phase)

If you catch yourself thinking these, STOP IMMEDIATELY:

If you wrote code without a failing test first, DELETE IT and start over. </EXTREMELY-IMPORTANT>

MAIN CHAT MUST NOT WRITE CODE. This is not negotiable.

Main chat orchestrates. Subagents implement. If you catch yourself about to use Write or Edit on a code file, STOP.

If you’re about to edit code directly, STOP and spawn a Task agent instead.

Rationalization Prevention

These thoughts mean STOP—you’re rationalizing:

The Meta-Rationalization

If you’re treating these rules as “guidelines for complex work” rather than “invariants for ALL work”, you’ve already failed.

Simple work is EXACTLY when discipline matters most—because that’s when you’re most tempted to skip it. </EXTREMELY-IMPORTANT>

Context Monitoring

Before starting each task, check context availability:

Thresholds:

At Warning level: After the current task completes (don't abandon mid-task), invoke: Read ${CLAUDE_SKILL_DIR}/../../skills/dev-handoff/SKILL.md and follow its instructions.

At Critical level: Stop immediately. Invoke dev-handoff before context is exhausted. A degraded handoff is better than no handoff.

Why: A 10-task implementation phase with 20% context remaining produces garbage for the last 5 tasks. Better to handoff cleanly and resume fresh than to push through with degraded output.

The Process

For each task N in PLAN.md:
    1. Determine loop type:
       - Visual task? → discover and read skills/visual-verify/SKILL.md via cache lookup
       - Standard task? → discover and read skills/dev-ralph-loop/SKILL.md via cache lookup

    2. Inside loop: spawn Task agent
       → discover and read skills/dev-delegate/SKILL.md via cache lookup

    3. Task agent follows TDD (dev-tdd) using testing tools (dev-test)
       Visual tasks: also render output and vision-check with look-at

    4. Verify tests pass (+ visual check passes for visual tasks), output promise

    5. Move to task N+1, start NEW loop

**Cache lookup pattern for all paths above:**Read ${CLAUDE_SKILL_DIR}/../../TARGET/PATH and follow its instructions.

Visual Task Detection

If a PLAN.md task involves rendered visual output, use visual-verify instead of plain ralph-loop. Visual-verify adds render → look-at → fix steps inside each iteration.

Signals a task is visual: task mentions "render", "slide", "chart", "figure", "layout", "UI", "screenshot", "visual", "diagram", or produces any file meant to be seen by humans (PNG, PDF, SVG).

Read ${CLAUDE_SKILL_DIR}/../../skills/visual-verify/SKILL.md and follow its instructions.

Step 1: Start Ralph Loop for Each Task

REQUIRED SUB-SKILL:

Read ${CLAUDE_SKILL_DIR}/../../skills/dev-ralph-loop/SKILL.md and follow its instructions.

Key points from dev-ralph-loop:

• ONE loop PER TASK (not one loop for feature)
• Each task gets its own completion promise
• Don’t move to task N+1 until task N’s loop completes

Step 2: Inside Loop - Spawn Task Agent

REQUIRED SUB-SKILL:

Read ${CLAUDE_SKILL_DIR}/../../skills/dev-delegate/SKILL.md and follow its instructions.

Key points from dev-delegate:

• Implementer → Spec reviewer → Quality reviewer
• Task agent follows dev-tdd protocol
• Task agent uses dev-test tools

Step 3: Verify and Complete (MANDATORY - DO NOT SKIP)

After Task agent returns, you must personally verify (not trust the agent’s report):

3a. Read the Actual Code

Read the implementation file(s) the agent claims to have written.
Compare to SPEC.md requirements line by line.

• Code matches spec (not a different approach)
• No substitutions (e.g., spec says IPC, code uses DOM = FAIL)

3b. Check Test Reality

Read the test file(s). Look for .skip(), mock-only tests, or tests that don’t call real code.

• Tests EXECUTE code (not grep/mock-only)
• Tests are NOT skipped (SKIP ≠ PASS)
• Integration tests exist and run (not just unit tests)

3c. Run Tests Yourself

Actually run the test command. Read the output.

• Test command runs without error
• Tests actually pass (not “66 pass, 0 fail” with 50 skipped)
• Test output shows real assertions (not just “test exists”)

3d. Verify Real Integration (FOR EXTERNAL SYSTEMS)

If the feature integrates with an external system (Electron app, API, database),
you MUST verify it works against the real system, not just mocks.

• External system is actually running
• Feature actually works (not just “code runs without error”)
• Output is visible in the external system

If ANY check fails → REJECT the work. Do NOT mark task complete.

Rationalization Prevention (Verification Phase)

If ALL pass → output the promise. If ANY fail → iterate.

Task Summary (MANDATORY after each task)

After a task passes review, append a structured summary to LEARNINGS.md:

## Task N: [task description]

---
task: N
status: completed
implements: [REQ-01, REQ-03]
affects: [src/auth/, tests/test_auth.py]
key-files:
  created: [list of new files]
  modified: [list of changed files]
deviations: {r1: 0, r2: 1, r3: 0, r4: 0}
---

One-liner: [SUBSTANTIVE summary — not "Task complete" but "JWT refresh rotation with 7-day expiry using jose library"]

Changes: [what was added/modified and why]
Test: [test command and result]

One-liner rule: Must be SUBSTANTIVE. Good: "Added rate limiting middleware with sliding window at 100 req/min". Bad: "Implemented task 3" or "Done".

Deviation Rules (CRITICAL)

You WILL discover unplanned work during implementation. Apply these rules automatically and track all deviations.

Priority: Rule 4 (STOP) > Rules 1-3 (auto) > unsure → Rule 4 Edge cases: missing validation → R2 | null crash → R1 | new table → R4 | new column → R1/2

Rule 4 Format

When you encounter an architectural deviation, STOP and present:

⚠️ Architectural Decision Needed
- Current task: [task name]
- Discovery: [what prompted this]
- Proposed change: [modification]
- Why needed: [rationale]
- Impact: [what this affects]
- Alternatives: [other approaches]
Proceed with proposed change? (yes / different approach / defer)

Documenting Deviations

All deviations tracked per task:

[Rule N - Category] Title

• Found during: Task X
• Issue: [description]
• Fix: [what was done]
• Files modified: [list]
• Verification: [how confirmed]

End each task summary with: Total deviations: N auto-fixed (R1: X, R2: Y, R3: Z). Impact: [assessment].

Why Skipping Hurts the Thing You Care About Most

The protocol is not overhead you pay. It is the service you provide.

Sub-Skills Reference

Failure Recovery Protocol

Pattern from oh-my-opencode: After 3 consecutive implementation failures, escalate.

3-Failure Trigger

If you attempt 3 implementations and ALL fail tests:

Iteration 1: Implement approach A → tests fail
Iteration 2: Implement approach B → tests fail
Iteration 3: Implement approach C → tests fail
→ TRIGGER RECOVERY PROTOCOL

Recovery Steps

1. STOP all further implementation attempts

   • No more “let me try a different approach”
   • No guessing or throwing code at the problem

2. REVERT to last known working state

   • git checkout <last-passing-commit>
   • Or revert specific files
   • Document what was attempted in .planning/RECOVERY.md

3. DOCUMENT what was attempted

   • All 3 approaches tried
   • Test failures for each
   • Why each approach failed
   • What this reveals about the problem

4. CONSULT with user BEFORE continuing

   • “I’ve tried 3 approaches. All fail tests. Here’s what I’ve learned...”
   • Present test failure patterns
   • Request: requirements clarification, design input, or different strategy

5. ASK USER for direction

   • Option A: Re-examine requirements (may need /dev-clarify)
   • Option B: Try completely different design (may need /dev-design)
   • Option C: Investigate why tests fail (may need /dev-debug)
   • Option D: User provides domain knowledge

NO PASSING TESTS = NOT COMPLETE (hard rule)

Recovery Checklist

Before continuing after multiple failures:

• All 3 approaches documented with test failures
• Pattern in failures identified (same tests? different errors?)
• Current code reverted to clean state
• User consulted with specific question
• Clear direction from user before proceeding

Anti-Patterns After Failures

DON’T:

• Keep trying “just one more thing”
• Make larger and larger changes
• Skip TDD “to get it working first”
• Suppress test failures (“I’ll fix them later”)
• Blame the tests (“tests are wrong”)

DO:

• Stop and analyze the failure pattern
• Revert to clean state
• Document what each approach revealed
• Consult user with specific findings
• Get clear direction before continuing

Example Recovery Flow

Loop 1: Implement with synchronous approach → Tests timeout
Loop 2: Implement with async/await → Tests hang
Loop 3: Implement with promises → Tests fail assertion

→ RECOVERY PROTOCOL:
1. STOP (no loop 4)
2. REVERT: git checkout HEAD -- src/feature.ts tests/
3. DOCUMENT in .planning/RECOVERY.md:
   - Pattern: All async implementations cause timing issues
   - Tests expect synchronous behavior
   - Hypothesis: Requirements may need async, tests don’t handle it
4. ASK USER:
   “I’ve tried 3 async implementations. All cause timing issues.
    Tests expect synchronous behavior.

    This suggests either:
    A) Feature should actually be synchronous (simpler)
    B) Tests need updating for async behavior

    Which direction should I take?”

When to Trigger Recovery

Trigger after 3 failures when:

• Same test keeps failing despite different approaches
• Different tests fail in pattern (suggests wrong approach)
• Tests pass locally but fail in CI
• Implementation works but breaks unrelated tests

Don’t wait for max iterations - trigger early when pattern emerges.

If Max Iterations Reached

Ralph exits after max iterations. Still do NOT ask user to manually test.

Main chat should:

1. Summarize what’s failing (from LEARNINGS.md)
2. Report which automated tests fail and why
3. Ask user for direction:
   • A) Start new loop with different approach
   • B) Add more logging to debug
   • C) User provides guidance
   • D) User explicitly requests manual testing

Never default to “please test manually”. Always exhaust automation first.

No Pause Between Tasks

Post-Promise Checklist (mandatory, same response)

1. Update PLAN.md - Mark task [x] complete
2. Log to LEARNINGS.md - What was done
3. Start next task’s ralph loop - No waiting

Valid Stopping Points (only these three)

1. ALL tasks in PLAN.md are marked [x] complete
2. You hit a blocker requiring user input (state exactly what you need)
3. User explicitly interrupted

The promise signals task completion. After outputting promise, update PLAN.md, then IMMEDIATELY start next task’s loop.

Pausing between tasks is procrastination disguised as courtesy.

Task Transition Gate (MANDATORY)

After each task’s ralph loop completes:

1. Update PLAN.md — mark completed task [x]
2. Append to LEARNINGS.md — what was accomplished, test command, exit code
3. Check for blockers — dependencies from task N needed for N+1?
4. If clear → IMMEDIATELY spawn ralph loop for task N+1
5. If blocked → Ask user EXACTLY what’s missing (not "I’m blocked")

Violations to catch:

• "Let me check with user if they want me to continue" → NO, continue automatically
• "Should I move to task N+1?" → NO, you’re supposed to move
• "Let me summarize what we learned" → NO, move to task N+1

Pausing > 30 seconds between tasks means you’ve stopped. You shouldn’t have. </EXTREMELY-IMPORTANT>

Agent Team Implementation (Parallel)

For parallel implementation using agent teams, read the full protocol:

Read ${CLAUDE_SKILL_DIR}/../../skills/dev-implement/references/agent-team-protocol.md and follow its instructions.

When to use: User explicitly requests parallel implementation, OR 4+ independent tasks in PLAN.md.

Key rules:

• Each teammate gets a self-contained prompt with full context
• Main agent coordinates, does NOT implement directly
• Reconcile results after all teammates complete
• Fall back to sequential if fewer than 3 tasks

Exit Gate

Checkpoint type: human-verify (all tasks pass tests — machine-verifiable)

Test Gap Validation Gate (MANDATORY)

This gate validates that every requirement in SPEC.md has corresponding test coverage. TDD ensures task-level coverage; test gap ensures requirement-level coverage. They are different checks.

Invoke test gap Validation

Read ${CLAUDE_SKILL_DIR}/../../skills/dev-test-gaps/SKILL.md and follow its instructions.

Gate Conditions

Must produce .planning/VALIDATION.md before proceeding to review.

Re-validation After Gap Fixes

If test gap reports implementation bugs (escalations):

1. Spawn ralph loops ONLY for the specific failing requirements
2. After fixes, re-invoke dev-test-gaps to re-validate
3. Repeat until VALIDATION.md status is validated
4. Max 2 re-validation cycles. After that, escalate to user.

Rationalization Prevention

Phase Complete

REQUIRED SUB-SKILL: After ALL tasks complete with passing tests AND test gap validation passes:

Read ${CLAUDE_SKILL_DIR}/../../skills/dev-review/SKILL.md and follow its instructions.

Do NOT proceed until automated tests pass for every task AND .planning/VALIDATION.md status is validated.