Implement Phase

Implement Phase

Execute a single phase from an implementation plan with comprehensive quality gates. This skill is designed to be called by implement-plan but can also be invoked directly.

CRITICAL: Orchestrator Pattern (MANDATORY)

THIS SESSION IS AN ORCHESTRATOR. YOU MUST NEVER IMPLEMENT CODE DIRECTLY.

What This Means

Enforcement

⛔ VIOLATION: Using Write/Edit/NotebookEdit tools directly
⛔ VIOLATION: Creating files without spawning a subagent
⛔ VIOLATION: Fixing code without spawning a subagent
⛔ VIOLATION: Running implementation commands directly

✅ CORRECT: Task(subagent): "Create the AuthService at src/auth/..."
✅ CORRECT: Task(subagent): "Fix the lint errors in src/auth/..."
✅ CORRECT: Task(subagent): "Run npm test and report results..."

Why Orchestration?

1. Context preservation - Main session retains full plan context
2. Parallelization - Independent tasks run concurrently
3. Clean separation - Orchestration logic separate from implementation
4. Better error handling - Failures don't pollute main context

Subagent Spawning Pattern

Task (run_in_background: true): "Create [file] implementing [feature].

Context: Phase [N] - [Name]
Requirements:
- [Requirement 1]
- [Requirement 2]

RESPONSE FORMAT: Be concise. Return only:
- STATUS: PASS/FAIL
- FILES: created/modified files
- ERRORS: any issues (omit if none)
Write verbose output to logs/[task].log"

Subagent Communication Protocol (CRITICAL)

Subagents MUST be concise. Context preservation is paramount.

Every subagent prompt MUST include the response format instruction. Verbose responses waste orchestrator context.

Required Response Format Block (include in EVERY subagent prompt):

RESPONSE FORMAT: Be concise. Return ONLY:
- STATUS: PASS/FAIL
- FILES: list of files created/modified
- ERRORS: brief error description (omit if none)

DO NOT include:
- Step-by-step explanations of what you did
- Code snippets (they're in the files)
- Suggestions for next steps
- Restating the original task

For large outputs, WRITE TO DISK:
- Test results → logs/test-[feature].log
- Build output → logs/build-[phase].log
- Error traces → logs/error-[task].log
Return only: "Full output: logs/[filename].log"

Good vs Bad Subagent Responses:

❌ BAD (wastes context):
"I have successfully created the SummaryAgentService. First, I analyzed
the requirements and determined that we need to implement three methods:
summarize(), retry(), and handleError(). I created the file at
src/agents/summary-agent/summary-agent.service.ts with the following
implementation: [300 lines of code]. The service uses dependency
injection to receive the OllamaService. I also updated the module file
to register the service. You should now be able to run the tests..."

✅ GOOD (preserves context):
"STATUS: PASS
FILES: src/agents/summary-agent/summary-agent.service.ts (created),
       src/agents/summary-agent/summary-agent.module.ts (modified)
ERRORS: None"

Disk-Based Communication for Large Data:

Architecture

implement-plan (orchestrates full plan)
    │
    └── implement-phase (this skill - one phase at a time)
            │
            ├── 1. Implementation (subagents)
            ├── 2. Exit Condition Verification (build, lint, unit tests)
            ├── 3. Automated Integration Testing (Claude tests via API/Playwright)
            ├── 4. Code Review (code-review skill)
            ├── 5. ADR Compliance Check
            ├── 6. Plan Synchronization
            ├── 7. Prompt Archival (if prompt provided)
            └── 8. Phase Completion Report

Design Principles

Single Responsibility

This skill does ONE thing: execute a single phase completely and correctly.

Extensibility

The phase execution pipeline is designed as a sequence of steps. New steps can be added without modifying the core logic. See Phase Steps.

Quality Gates

Each step is a gate. If any gate fails, the phase cannot complete.

Composability

This skill orchestrates other skills (code-review, adr) and can be extended to include more.

Mandatory Exit Conditions

These conditions are NON-NEGOTIABLE. A phase cannot complete until ALL are satisfied.

Why Recommendations Are Mandatory

❌ WRONG: "It's just a recommendation, we can fix it later"
❌ WRONG: "PASS_WITH_NOTES is good enough"
❌ WRONG: "We'll address it in the next phase"

✅ CORRECT: "Recommendations are blocking issues"
✅ CORRECT: "Only clean PASS allows phase completion"
✅ CORRECT: "Fix it now or the phase cannot complete"

The Clean Baseline Principle requires:

• Each phase ends with zero outstanding issues
• The next phase inherits a clean codebase
• Technical debt is not accumulated across phases
• Recommendations, if worth noting, are worth fixing

Input Context

When invoked, this skill expects:

Plan Path: [path to plan file]
Phase: [number or name]
Task ID: [task_id from implement-plan's TaskList]
Prompt Path: [optional - path to pre-generated prompt from prompt-generator]
Changed Files: [optional - auto-detected if not provided]
Skip Steps: [optional - list of steps to skip, e.g., for testing]
TDD Mode: [enabled/disabled - from plan metadata, CLI flag, or global settings]
Coverage Threshold: [percentage - default 80%, applies when TDD mode enabled]

Prompt Integration

If a Prompt Path is provided (from prompt-generator skill):

1. Read the prompt file - Contains detailed orchestration instructions
2. Use prompt as primary guidance - Follows established patterns and conventions
3. Plan file as reference - For exit conditions and verification steps
4. Archive on completion - Move prompt to completed/ subfolder

# Prompt provides:
- Detailed orchestration workflow
- Subagent delegation patterns
- Specific task breakdowns
- Error handling guidance

# Plan provides:
- Exit conditions (source of truth)
- Success criteria
- Dependencies

Phase Execution Pipeline

⚡ AUTO-CONTINUE RULES (READ THIS FIRST)

┌─────────────────────────────────────────────────────────────────────────┐
│                     AUTOMATIC CONTINUATION ENGINE                        │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│   RULE 1: After completing ANY step, IMMEDIATELY start the next step.   │
│   RULE 2: Do NOT output "waiting for input" or ask to continue.         │
│   RULE 3: Do NOT summarize and stop. Summarize and CONTINUE.            │
│   RULE 4: The ONLY valid stop point is after Step 8 completion.         │
│                                                                          │
│   EXECUTION ALGORITHM:                                                   │
│                                                                          │
│   current_step = 1                                                       │
│   while current_step <= 8:                                               │
│       result = execute_step(current_step)                                │
│       if result == PASS:                                                 │
│           current_step += 1          # AUTO-CONTINUE                     │
│       elif result == FAIL:                                               │
│           fix_and_retry(current_step) # Stay on step, fix, retry         │
│       elif result == BLOCKED:                                            │
│           return BLOCKED             # Only valid early exit             │
│   return COMPLETE                    # Only stop here                    │
│                                                                          │
└─────────────────────────────────────────────────────────────────────────┘

WHAT TO DO AFTER EACH STEP

On FAIL at any step: Fix the issue, re-run the SAME step, get PASS, then continue.

NEVER DO THESE

❌ "Step 2 complete. Let me know when you want to continue."
❌ "Verification passed. Would you like me to proceed to Step 3?"
❌ "Code review done. Waiting for your input."
❌ "I've completed the exit conditions. What's next?"
❌ Outputting results without immediately starting the next step
❌ Asking permission to continue between steps 1-7

ALWAYS DO THESE

✅ "Step 2 PASS. Executing Step 3: Integration Testing..."
✅ "Code review PASS. Now checking ADR compliance..."
✅ "Exit conditions verified. Running integration tests now..."
✅ Immediately invoke the next step's tools/skills after reporting status
✅ Chain steps together without pause

Continuous Execution Details

The entire pipeline (Steps 1-8) MUST execute as one continuous flow.

After EACH step completes (including skill invocations), IMMEDIATELY proceed to the next step WITHOUT waiting for user input.

Pause Points (ONLY these):

DO NOT PAUSE after:

• Implementation complete → Continue to Step 2
• Exit conditions pass → Continue to Step 3
• Integration tests pass → Continue to Step 4
• Code review returns PASS → Continue to Step 5
• ADR compliance returns PASS → Continue to Step 6
• Plan sync complete → Continue to Step 7
• Prompt archived → Continue to Step 8
• Any successful step completion → Continue to next step
• Fix loop completes with PASS → Continue to next step

Fix Loops (internal, no user pause):

• Verification fails → Fix code, re-run verification, expect PASS
• Integration tests fail → Fix code, re-run tests, expect PASS
• Code review returns PASS_WITH_NOTES → Fix notes, re-run Step 4, expect PASS
• Code review returns NEEDS_CHANGES → Fix issues, re-run Step 4, expect PASS
• Any step has fixable issues → Spawn fix subagents, re-run step

Continuous Flow Example:

Step 1: Implementation → PASS
        ↓ (immediately)
Step 2: Exit Conditions → PASS
        ↓ (immediately)
Step 3: Automated Integration Testing → PASS
        ↓ (immediately)
Step 4: Code Review Skill → PASS_WITH_NOTES
        ↓ (fix loop - spawn subagents to fix notes)
        → Re-run Code Review → PASS
        ↓ (now continue)
Step 5: ADR Compliance → PASS
        ↓ (immediately)
Step 6: Plan Sync → PASS
        ↓ (immediately)
Step 7: Prompt Archival → PASS
        ↓ (immediately)
Step 8: Completion Report → Present to user
        ↓ (NOW wait for user confirmation)

Goal: Clean PASS on all steps. PASS_WITH_NOTES means there's work to do.

Blocking Elements (ONLY Valid Reasons to Stop)

A blocking element is something YOU cannot fix autonomously.

Do NOT stop for fixable issues. Only stop when you genuinely cannot proceed without user intervention.

Valid Blocking Elements:

NOT Blocking (fix these yourself):

Blocker Protocol:

When you hit a genuine blocker:

⛔ BLOCKED: [Brief description]

Phase: [N] - [Name]
Step: [Current step]
Blocker Type: [Permission | Infrastructure | Credentials | External | Ambiguous | Destructive]

Details:
[Specific details about what failed and why]

What I Need:
[Specific action required from user]

Options:
A) [Resolve the blocker and continue]
B) [Skip this verification and proceed with risk]
C) [Abort phase]

Resume After Blocker:

Once the user resolves the blocker, resume from the blocked step (not from Step 1).

Step Completion Checklist (MANDATORY)

Before reporting phase complete, ALL steps must be executed.

Use this checklist internally. If any step is missing, execute it before completing:

PHASE COMPLETION VERIFICATION:
- [ ] Step 1: Implementation - Subagents spawned, work completed
- [ ] Step 2: Exit Conditions - Build, runtime, unit tests all verified
- [ ] Step 3: Integration Testing - YOU tested via API calls or Playwright
- [ ] Step 4: Code Review - Achieved PASS (not PASS_WITH_NOTES)
- [ ] Step 5: ADR Compliance - Checked against relevant ADRs
- [ ] Step 6: Plan Sync - Work items verified, phase status updated
- [ ] Step 7: Prompt Archival - Archived or explicitly skipped (no prompt)
- [ ] Step 8: Completion Report - Generated and presented

⛔ VIOLATION: Stopping before Step 8
⛔ VIOLATION: Waiting for user input between Steps 1-7
⛔ VIOLATION: Reporting "phase complete" with unchecked steps
⛔ VIOLATION: Proceeding with PASS_WITH_NOTES without fixing notes
⛔ VIOLATION: Asking user to "manually test" instead of testing yourself

Self-Check Protocol:

After invoking a skill (like code-review), ask yourself:

1. Did the skill complete? → Check the result status
2. Did it return PASS? → CONTINUE to next step immediately
3. Did it return PASS_WITH_NOTES? → Spawn fix subagents, re-run step, expect PASS
4. Did it return NEEDS_CHANGES? → Spawn fix subagents, re-run step, expect PASS
5. Am I at Step 8? → If no, execute next step immediately
6. Did I test the feature myself? → If no, go back to Step 3

The goal is always a clean PASS. PASS_WITH_NOTES is not "good enough" - fix the notes.

Progress Tracker (MANDATORY OUTPUT)

After EVERY step, you MUST output a Progress Tracker before doing ANYTHING else.

This is not optional. The Progress Tracker forces explicit acknowledgment of state and next action.

Format (output after each step completes):

┌─────────────────────────────────────────┐
│ PROGRESS: Step [N] → Step [N+1]         │
├─────────────────────────────────────────┤
│ ✅ Step 1: Implementation    [DONE/SKIP]│
│ ✅ Step 2: Exit Conditions   [DONE/SKIP]│
│ ✅ Step 3: Integration Test  [DONE/SKIP]│
│ ✅ Step 4: Code Review       [DONE/SKIP]│
│ ⏳ Step 5: ADR Compliance    [CURRENT]  │
│ ⬚ Step 6: Plan Sync         [PENDING]  │
│ ⬚ Step 7: Prompt Archival   [PENDING]  │
│ ⬚ Step 8: Completion Report [PENDING]  │
├─────────────────────────────────────────┤
│ NEXT ACTION: [Describe what you do next]│
└─────────────────────────────────────────┘

Rules:

1. Output this tracker IMMEDIATELY after each step completes
2. Mark the CURRENT step you are about to execute
3. The NEXT ACTION must describe executing the next step (not waiting for user)
4. If NEXT ACTION says anything other than executing a step, you are VIOLATING the protocol

Example - After Step 4 Code Review Returns PASS:

┌─────────────────────────────────────────┐
│ PROGRESS: Step 4 → Step 5               │
├─────────────────────────────────────────┤
│ ✅ Step 1: Implementation    [DONE]     │
│ ✅ Step 2: Exit Conditions   [DONE]     │
│ ✅ Step 3: Integration Test  [DONE]     │
│ ✅ Step 4: Code Review       [DONE]     │
│ ⏳ Step 5: ADR Compliance    [CURRENT]  │
│ ⬚ Step 6: Plan Sync         [PENDING]  │
│ ⬚ Step 7: Prompt Archival   [PENDING]  │
│ ⬚ Step 8: Completion Report [PENDING]  │
├─────────────────────────────────────────┤
│ NEXT ACTION: Check ADR compliance now   │
└─────────────────────────────────────────┘

⛔ VIOLATION Examples:

• Not outputting the Progress Tracker after a step
• NEXT ACTION: "Waiting for user confirmation" (before Step 8)
• NEXT ACTION: "Let me know if you want me to continue"
• NEXT ACTION: "Please manually verify the feature works"
• Skipping to Step 8 without completing Steps 5-7

Execution Contract (READ BEFORE STARTING)

⚠️ THIS IS A BINDING CONTRACT. VIOLATION = FAILURE.

Before executing ANY step, internalize these rules:

┌─────────────────────────────────────────────────────────────────────────┐
│                         EXECUTION CONTRACT                               │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│  I WILL execute Steps 1-8 as ONE continuous operation.                   │
│  I WILL NOT stop between steps (except Step 8 or BLOCKED).               │
│  I WILL invoke the next step's tools IMMEDIATELY after each step.        │
│  I WILL output a Progress Tracker after EVERY step.                      │
│  I WILL test the feature MYSELF in Step 3 (not ask the user).            │
│  I WILL NOT stop after Step 4 (code review) - there are 4 more steps.    │
│  I WILL NOT ask the user if they want me to continue.                    │
│  I WILL NOT ask the user to manually verify anything.                    │
│  I WILL only stop at Step 8 after presenting the Completion Report.      │
│  I WILL only stop early for genuine BLOCKING elements I cannot fix.      │
│                                                                          │
│  AFTER EACH STEP OUTPUT:                                                 │
│  ├── Output status (PASS/FAIL)                                           │
│  ├── Output Progress Tracker                                             │
│  └── IMMEDIATELY execute next step (no pause, no waiting)                │
│                                                                          │
└─────────────────────────────────────────────────────────────────────────┘

If you find yourself about to stop before Step 8, RE-READ this contract.

Step 1: Implementation

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ⏩ ENTERING STEP 1 - This is part of a continuous 8-step pipeline.            ║
║    When this step completes with PASS, IMMEDIATELY execute Step 2.            ║
║    Do NOT stop. Do NOT wait for user input.                                   ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Responsibility: Execute all tasks in the phase using subagent delegation.

REMINDER: You are an orchestrator. Spawn subagents for ALL implementation work.

Process:

1. Read phase requirements and tasks from plan (orchestrator reads)
2. Read coding standards from docs/standards/CODING_STANDARDS.md (if exists)
3. Identify independent tasks for parallelization
4. SPAWN test subagents FIRST (verification-first)
5. SPAWN implementation subagents (include coding standards reference)
6. Monitor subagent progress and handle blockers
7. Collect results and changed files list from subagent responses

CODING STANDARDS: All subagent prompts MUST reference coding standards. Include size limits (services <500 lines, controllers <30 lines/method), interface requirements (DTOs typed), and forbidden patterns (no console.log, no empty catch blocks). See docs/standards/CODING_STANDARDS.md.

Subagent Spawning Examples:

# Writing tests (FIRST - verification-first pattern)
Task (run_in_background: true): "Write unit tests for SummaryAgentService.

Context: Phase 5b-ii - SummaryAgent Service
Location: agentic-core/src/agents/implementations/summary-agent/

Test scenarios:
- Successful summarization
- Retry with feedback
- Error handling

RESPONSE FORMAT: STATUS, FILES created, test count. Write output to logs/."

# Implementation (AFTER tests exist)
Task (run_in_background: true): "Implement SummaryAgentService.

Context: Phase 5b-ii - SummaryAgent Service
Requirements from plan: [list requirements]
Must pass the tests at: [test file path]

CODING STANDARDS (MANDATORY):
- Services: <500 lines, single responsibility
- Interfaces: Required for DTOs and response types
- Errors: Domain exceptions, no empty catch blocks
- Logging: Use project logger, no console.log
Ref: docs/standards/CODING_STANDARDS.md

RESPONSE FORMAT: STATUS, FILES created/modified, ERRORS if any."

# Verification
Task (run_in_background: true): "Run build and test verification.

Commands: npm run build && npm run lint && npm test
Report: PASS/FAIL per command, error details if any.
Write full output to logs/verify-phase-5b-ii.log"

What You Do vs What Subagents Do:

Output:

IMPLEMENTATION_STATUS: PASS | FAIL
FILES_CREATED: [list]
FILES_MODIFIED: [list]
TEST_RESULTS: [summary]
ERRORS: [if any]
SUBAGENTS_SPAWNED: [count]
─────────────────────────────────────────────────
⚡ NEXT_STEP: EXECUTE STEP 2 NOW (verification-loop)

CRITICAL: The output MUST include NEXT_STEP: EXECUTE STEP 2 NOW. This is not optional. When you see this in the output, you MUST immediately invoke the verification-loop skill without waiting for user input.

Gate: Implementation must PASS to proceed.

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ✅ STEP 1 COMPLETE → IMMEDIATELY EXECUTE STEP 2 NOW                           ║
║    Do NOT output results and wait. Do NOT ask "shall I continue?"             ║
║    Your next action MUST be: invoke the verification-loop skill               ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Step 2: Exit Condition Verification (verification-loop)

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ⏩ ENTERING STEP 2 - This is part of a continuous 8-step pipeline.            ║
║    When this step completes with PASS, IMMEDIATELY execute Step 3.            ║
║    Do NOT stop. Do NOT wait for user input.                                   ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Responsibility: Verify all exit conditions using the comprehensive 6-check verification-loop.

verification-loop is the DEFAULT exit condition verification. It provides comprehensive validation that goes beyond basic build/test checks.

Process:

1. Read exit conditions from plan
2. Invoke verification-loop skill with phase context
3. verification-loop executes 6 checks:
   • Check 1: Build - Compilation, bundling, artifact generation
   • Check 2: Type - Type checking, interface compliance
   • Check 3: Lint - Code style, static analysis
   • Check 4: Test - Unit tests, integration tests, coverage
   • Check 5: Security - Dependency audit, secret scanning
   • Check 6: Diff - Review changes, detect unintended modifications
4. Aggregate results and report

Invocation:

Skill(skill="verification-loop"): Verify Phase [N] implementation.

Context:
- Plan: [plan file path]
- Phase: [N] ([Phase Name])
- Changed Files: [list of files modified in this phase]

Execute all 6 verification checks and return structured result.

Output:

VERIFICATION_LOOP_STATUS: PASS | FAIL
CHECKS_COMPLETED: 6/6
CHECK_RESULTS:
  BUILD: PASS | FAIL
  TYPE: PASS | FAIL
  LINT: PASS | FAIL
  TEST: PASS | FAIL
  SECURITY: PASS | FAIL
  DIFF: PASS | FAIL
FAILED_CHECKS: [list if any]
EVIDENCE: logs/verification-loop-phase-N.log
─────────────────────────────────────────────────
⚡ NEXT_STEP: EXECUTE STEP 3 NOW (Integration Testing)

CRITICAL: The output MUST include NEXT_STEP: EXECUTE STEP 3 NOW. This is not optional. When you see this in the output, you MUST immediately execute Step 3 without waiting for user input.

Gate: ALL 6 verification checks must PASS to proceed.

On Failure: Spawn fix subagents for failed checks, re-run verification-loop, repeat until all pass or escalate.

Disabling verification-loop (not recommended):

# In plan metadata - only for special cases
phase_config:
  verification_loop: false  # Falls back to basic exit conditions

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ✅ STEP 2 COMPLETE → IMMEDIATELY EXECUTE STEP 3 NOW                           ║
║    Do NOT output results and wait. Do NOT ask "shall I continue?"             ║
║    Your next action MUST be: run integration tests (Step 3)                   ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Step 3: Automated Integration Testing

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ⏩ ENTERING STEP 3 - This is part of a continuous 8-step pipeline.            ║
║    When this step completes with PASS, IMMEDIATELY execute Step 4.            ║
║    Do NOT stop. Do NOT wait for user input.                                   ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Responsibility: Verify the implementation works end-to-end through automated testing performed by YOU, not the user.

YOU are the tester. Do not ask the user to manually verify. Use tools to test the system yourself.

For UI Testing: Use the browser-verification-agent - spawn ONE agent per test scenario for context preservation. The agent wraps Playwright MCP and returns structured evidence.

Process:

1. Determine the testing approach based on implementation type:
   • Backend/API: Use curl, httpie, or spawn subagent to make API calls
   • Frontend/UI: Spawn browser-verification-agent for each test scenario
   • CLI tools: Execute commands and verify output
   • Libraries: Write and run integration test scripts
2. Spawn testing subagents for each verification scenario (ONE test per agent for UI)
3. Capture results and any failures
4. On failure: spawn fix subagents, re-test

Testing by Implementation Type:

Subagent Examples:

# API Testing (general-purpose subagent)
Task: "Test the new /api/users endpoint.

Make these API calls and report results:
1. POST /api/users with valid payload - expect 201
2. POST /api/users with invalid email - expect 400
3. GET /api/users/:id - expect 200 with user data
4. GET /api/users/nonexistent - expect 404

RESPONSE FORMAT: STATUS, test results summary, ERRORS if any."

# UI Testing (browser-verification-agent) - ONE test per agent spawn
Task(subagent_type="browser-verification-agent"): "Verify login with valid credentials.

base_url: http://localhost:3000
test_description: Navigate to /login, enter 'test@example.com' in email field,
                  enter 'password123' in password field, click Login button
expected_outcome: URL changes to /dashboard, welcome message visible
session_context: fresh"

Task(subagent_type="browser-verification-agent"): "Verify login with invalid credentials.

base_url: http://localhost:3000
test_description: Navigate to /login, enter 'test@example.com' in email field,
                  enter 'wrongpassword' in password field, click Login button
expected_outcome: Error message 'Invalid credentials' is displayed, URL stays on /login
session_context: fresh"

UI Testing Response Format (from browser-verification-agent):

STATUS: PASS | FAIL | FLAKY | BLOCKED
SCREENSHOT: logs/screenshots/2026-01-23-143022-login-test.png
OBSERVED: [what actually happened]
EXPECTED: [echo of expected_outcome]
ERRORS: [if any]

Aggregated Output (for Step 3 completion):

INTEGRATION_TEST_STATUS: PASS | FAIL
TESTS_RUN: [count]
TESTS_PASSED: [count]
TESTS_FAILED: [count]
FAILURE_DETAILS: [if any]
EVIDENCE: [log files, screenshots]
─────────────────────────────────────────────────
⚡ NEXT_STEP: EXECUTE STEP 4 NOW (Code Review)

CRITICAL: The output MUST include NEXT_STEP: EXECUTE STEP 4 NOW. This is not optional. When you see this in the output, you MUST immediately invoke the code-review skill without waiting for user input.

Gate: Integration tests must PASS to proceed.

On Failure:

1. Analyze failure root cause
2. Spawn fix subagents
3. Re-run failed tests
4. Repeat until pass or hit blocking element

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ✅ STEP 3 COMPLETE → IMMEDIATELY EXECUTE STEP 4 NOW                           ║
║    Do NOT output results and wait. Do NOT ask "shall I continue?"             ║
║    Your next action MUST be: invoke the code-review skill                     ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Step 4: Code Review

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ⏩ ENTERING STEP 4 - This is part of a continuous 8-step pipeline.            ║
║    When this step completes with PASS, IMMEDIATELY execute Step 5.            ║
║    Do NOT stop. Do NOT wait for user input.                                   ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Responsibility: Validate implementation quality across all dimensions.

Process:

1. Invoke code-review skill with phase context
2. Provide: plan path, phase number, changed files
3. Receive structured review result

Output:

CODE_REVIEW_STATUS: PASS | PASS_WITH_NOTES | NEEDS_CHANGES
BLOCKING_ISSUES: [count]
RECOMMENDATIONS: [list]
─────────────────────────────────────────────────
⚡ NEXT_STEP: EXECUTE STEP 5 NOW (ADR Compliance)

CRITICAL: When CODE_REVIEW_STATUS is PASS, the output MUST include NEXT_STEP: EXECUTE STEP 5 NOW. This is not optional. When you see this in the output, you MUST immediately check ADR compliance without waiting for user input.

Gate: Code review must be PASS to proceed. PASS_WITH_NOTES is NOT acceptable.

⚠️ MANDATORY: All Recommendations Must Be Fixed

This is a non-negotiable exit condition. PASS_WITH_NOTES means there are recommendations that MUST be addressed before the phase can complete.

• Recommendations are NOT optional suggestions
• Recommendations are NOT "nice to have"
• Recommendations are blocking issues that must be resolved
• The only acceptable code review status is PASS

On PASS_WITH_NOTES or NEEDS_CHANGES:

1. Spawn fix subagents to address ALL issues (blocking issues AND recommendations)
2. Re-run code review
3. Repeat until clean PASS (max 3 retries)
4. Escalate to user only if max retries exhausted

Why are recommendations mandatory?

• Recommendations indicate pattern violations, missing tests, or technical debt
• Leaving them unfixed accumulates debt that compounds across phases
• The "Clean Baseline Principle" requires each phase to end clean
• Future phases inherit our mess if we don't fix it now
• Consistency: if it's worth noting, it's worth fixing

⚠️ CRITICAL TRANSITION POINT - DO NOT STOP HERE ⚠️

After code-review skill returns, you MUST continue. This is the #1 failure point.

• Code review returned PASS? → Output Progress Tracker → Execute Step 5 NOW
• Code review returned PASS_WITH_NOTES? → Fix issues → Re-run → Get PASS → Execute Step 5
• DO NOT report to user and wait. DO NOT ask if they want to continue.
• The phase is NOT complete. You have 4 more steps to execute.

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ✅ STEP 4 COMPLETE → IMMEDIATELY EXECUTE STEP 5 NOW                           ║
║    Do NOT output results and wait. Do NOT ask "shall I continue?"             ║
║    Your next action MUST be: check ADR compliance                             ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Step 5: ADR Compliance Check

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ⏩ ENTERING STEP 5 - This is part of a continuous 8-step pipeline.            ║
║    When this step completes with PASS, IMMEDIATELY execute Step 6.            ║
║    Do NOT stop. Do NOT wait for user input.                                   ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Responsibility: Ensure architectural decisions are followed and documented.

Process:

1. Read docs/decisions/INDEX.md to identify relevant ADRs
2. Check implementation against applicable ADRs
3. Identify any new architectural decisions made during implementation
4. If new decisions found, invoke adr skill to document them

Output:

ADR_COMPLIANCE_STATUS: PASS | NEEDS_DOCUMENTATION
APPLICABLE_ADRS: [list]
COMPLIANCE_RESULTS: [per-ADR status]
NEW_DECISIONS_DOCUMENTED: [list of new ADR numbers, if any]
─────────────────────────────────────────────────
⚡ NEXT_STEP: EXECUTE STEP 6 NOW (Plan Synchronization)

CRITICAL: The output MUST include NEXT_STEP: EXECUTE STEP 6 NOW. This is not optional. When you see this in the output, you MUST immediately verify plan synchronization without waiting for user input.

Gate: ADR compliance must PASS to proceed.

On NEEDS_DOCUMENTATION: Invoke adr skill for each undocumented decision.

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ✅ STEP 5 COMPLETE → IMMEDIATELY EXECUTE STEP 6 NOW                           ║
║    Do NOT output results and wait. Do NOT ask "shall I continue?"             ║
║    Your next action MUST be: verify plan synchronization                      ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Step 6: Plan Synchronization

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ⏩ ENTERING STEP 6 - This is part of a continuous 8-step pipeline.            ║
║    When this step completes with PASS, IMMEDIATELY execute Step 7.            ║
║    Do NOT stop. Do NOT wait for user input.                                   ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Responsibility: Verify work items completed and update plan status.

Process:

1. Verify all work items for this phase were completed
2. Add ADR references if new ADRs were created
3. Note any deviations from original plan
4. Mark phase status as complete (add ✅ to phase header)

Note: Per ADR-0001, plans are specification documents. Progress is tracked via Task tools, not by modifying checkboxes in the plan file.

Output:

PLAN_SYNC_STATUS: PASS | FAIL
WORK_ITEMS_VERIFIED: [count]
DEVIATIONS_NOTED: [count]
ADR_REFERENCES_ADDED: [count]
─────────────────────────────────────────────────
⚡ NEXT_STEP: EXECUTE STEP 7 NOW (Prompt Archival)

CRITICAL: The output MUST include NEXT_STEP: EXECUTE STEP 7 NOW. This is not optional. When you see this in the output, you MUST immediately archive the prompt (or skip if none) without waiting for user input.

Gate: Plan sync must complete successfully.

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ✅ STEP 6 COMPLETE → IMMEDIATELY EXECUTE STEP 7 NOW                           ║
║    Do NOT output results and wait. Do NOT ask "shall I continue?"             ║
║    Your next action MUST be: archive the prompt (if provided)                 ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Step 7: Prompt Archival

╔═══════════════════════════════════════════════════════════════════════════════╗
║ ⏩ ENTERING STEP 7 - This is part of a continuous 8-step pipeline.            ║
║    When this step completes with PASS, IMMEDIATELY execute Step 8.            ║
║    Do NOT stop. Do NOT wait for user input.                                   ║
╚═══════════════════════════════════════════════════════════════════════════════╝

Responsibility: Archive the phase prompt to the completed folder (if prompt was provided).

Process:

1. Check if a prompt file was used for this phase
2. If yes, move to completed/ subfolder:

   # Create completed folder if it doesn't exist
   mkdir -p docs/prompts/completed
   
   # Move the prompt file
   mv docs/prompts/phase-2-data-pipeline.md docs/prompts/completed/
   

3. Log the archival

Output:

PROMPT_ARCHIVAL_STATUS: P

…(truncated)