Implementing Specs

Purpose

Orchestrates specification implementation by invoking existing wrangler skills (writing-plans, implementing-issue) rather than reimplementing their logic. Focuses on workflow coordination, subagent dispatch, and compliance verification.

When to Use

• Implementing specifications (SPEC-XXXXXX files)
• User-facing features requiring verification gates
• Need systematic audit trail through GitHub PR
• Want to leverage existing planning and implementation skills

Prerequisites

• Specification file exists in .wrangler/specifications/
• gh CLI installed and authenticated
• Git worktree configured (session_start will create it)
• Bash 4.0+ (for scripts)
• jq command-line JSON processor

Architecture: Skill Orchestration

This skill orchestrates existing wrangler skills rather than reimplementing their logic:

Phase 1 (INIT): Uses session_start MCP tool (session tools) Phase 2 (PLAN): Invokes writing-plans skill to create MCP issues Phase 3 (REVIEW): Validates planning coverage before execution Phase 4 (EXECUTE): Dispatches subagents per issue following implementing-issue skill Phase 5 (VERIFY): LLM-based compliance audit Phase 6 (PUBLISH): GitHub PR finalization

Benefits of this approach:

• No duplicated planning logic (writing-plans is source of truth)
• Direct coordination of subagents for each issue
• Modular and maintainable (changes to planning flow in one place)
• Testable (each skill can be tested independently)

Workflow Phases

INIT → PLAN → REVIEW → EXECUTE → VERIFY → PUBLISH → COMPLETE

---

Phase 1: INIT

Initialize session, create worktree, and establish context.

Objective

Create isolated worktree using MCP session tools and verify environment.

Actions

1. Start session - Call session_start MCP tool

   session_start(specFile: "{SPEC_FILE}")
   

   Capture response:

   • SESSION_ID = response.sessionId
   • WORKTREE_ABSOLUTE = response.worktreePath
   • BRANCH_NAME = response.branchName
   • AUDIT_PATH = response.auditPath

2. Verify worktree - Ensure on correct branch

   cd {WORKTREE_ABSOLUTE} && \
     echo "=== WORKTREE VERIFICATION ===" && \
     echo "Directory: $(pwd)" && \
     echo "Git root: $(git rev-parse --show-toplevel)" && \
     echo "Branch: $(git branch --show-current)" && \
     test "$(git branch --show-current)" = "{BRANCH_NAME}" && echo "VERIFIED" || echo "FAILED"
   

   If verification fails, STOP and report error.

3. Log phase complete

   session_phase(
     sessionId: SESSION_ID,
     phase: "init",
     status: "complete"
   )
   

Outputs

• Session ID for all subsequent phases
• Worktree absolute path for subagent context
• Branch name for PR creation

Quality Gate

Worktree must exist and be on correct branch.

---

Phase 2: PLAN

Create implementation plan using writing-plans skill.

Objective

Break specification into implementable tasks tracked as MCP issues.

Actions

1. Log phase start

   session_phase(sessionId: SESSION_ID, phase: "plan", status: "started")
   

2. Invoke writing-plans skill on worktree

   Use Task tool to dispatch planning subagent:

   Tool: Task
   Description: "Create implementation plan for {SPEC_FILE}"
   Prompt: |
   You are creating an implementation plan for a specification.
   
   ## Working Directory Context
   
   **Working directory:** {WORKTREE_ABSOLUTE}
   **Branch:** {BRANCH_NAME}
   **Session ID:** {SESSION_ID}
   
   ## Specification
   
   Read and analyze: {SPEC_FILE}
   
   ## Your Job
   
   1. Read the specification file completely
   2. Invoke the writing-plans skill to break it down into tasks
   3. The writing-plans skill will:
      - Create MCP issues for each task (issues_create)
      - Optionally create plan file if architecture context needed
      - Return list of issue IDs created
   4. Return to me:
      - Total task count
      - List of issue IDs created (e.g., ["ISS-000042", "ISS-000043"])
      - Any blockers or clarification needed
   
   IMPORTANT: Let writing-plans handle all planning logic. Your job
   is to invoke it and report back the results.
   

3. Capture issue IDs

   Parse planning subagent response for:

   • ISSUE_IDS = list of created issue IDs
   • TASK_COUNT = number of tasks created

4. Create GitHub PR with overview (not full plan)

   cd "{WORKTREE_ABSOLUTE}" && \
   gh pr create \
     --title "feat: {SPEC_TITLE}" \
     --body "Implements {SPEC_FILE}. See .wrangler/issues/ for task details." \
     --draft \
     --base main \
     --head "{BRANCH_NAME}"
   

   Capture:

   • PR_URL = PR URL from output
   • PR_NUMBER = PR number from output

5. Log phase complete

   session_phase(
     sessionId: SESSION_ID,
     phase: "plan",
     status: "complete",
     metadata: {
       issues_created: ISSUE_IDS,
       total_tasks: TASK_COUNT,
       pr_url: PR_URL,
       pr_number: PR_NUMBER
     }
   )
   

6. Save checkpoint

   session_checkpoint(
     sessionId: SESSION_ID,
     tasksCompleted: [],
     tasksPending: ISSUE_IDS,
     lastAction: "Created implementation plan with {TASK_COUNT} tasks",
     resumeInstructions: "Continue with execute phase, issues: {ISSUE_IDS}"
   )
   

Outputs

• Wrangler issues created (ISS-XXXXXX files in .wrangler/issues/)
• Local plan file (if writing-plans created it for architecture context)
• GitHub PR created (draft mode)
• Issue IDs list for execution phase

Quality Gate

Planning succeeds and issues are created. If planning fails or returns blockers, ESCALATE to user.

Key Design Decision

Why GitHub PR shows overview, not full plan:

• PR is for stakeholders/reviewers (high-level context)
• MCP issues are source of truth (complete implementation details)
• Optional plan file (if created) provides architecture reference
• This keeps PR descriptions concise while maintaining full traceability

---

Phase 3: REVIEW

Validate planning completeness BEFORE execution starts.

Objective

Catch planning gaps early before wasting time on execution.

Actions

1. Log phase start

   session_phase(sessionId: SESSION_ID, phase: "review", status: "started")
   

2. Read plan file coverage analysis (if plan file was created)

   Read .wrangler/plans/YYYY-MM-DD-PLAN_<spec>.md and extract:

   • "Acceptance Criteria Coverage" section
   • Coverage summary (% covered)
   • List of AC with no implementing tasks

3. Validate coverage

   Coverage	Status
   >= 95%	✅ PASS (automatic approval)
   < 95%	⚠️ NEEDS ATTENTION (user decision)

4. If coverage < 95%:

   Present coverage report to user:

   REVIEW Phase: Planning Coverage Gap Detected
   
   Coverage: X% (Y/Z acceptance criteria covered)
   
   Missing AC:
   
   - AC-XXX: [description]
   - AC-YYY: [description]
     ...
   
   Options:
   a) Auto-create missing tasks (Recommended)
   b) Proceed anyway (risks VERIFY failure)
   c) Abort and replan from scratch
   
   Your decision?
   

5. If auto-create chosen:

   For each uncovered AC:

   • Create MCP issue with implementation details
   • Add satisfiesAcceptanceCriteria: ["AC-XXX"] metadata
   • Add to ISSUE_IDS list

   Update plan file with new tasks and re-calculate coverage.

6. Update session checkpoint

   session_checkpoint(
     sessionId: SESSION_ID,
     tasksCompleted: [],
     tasksPending: ISSUE_IDS, // updated list
     lastAction: "REVIEW phase complete, coverage: X%",
     resumeInstructions: "Continue with execute phase"
   )
   

7. Log phase complete

   session_phase(
     sessionId: SESSION_ID,
     phase: "review",
     status: "complete",
     metadata: {
       coverage_percentage: X,
       supplemental_tasks_created: N
     }
   )
   

Outputs

• Validated plan with >= 95% AC coverage (or user approval to proceed)
• Updated ISSUE_IDS list (if supplemental tasks created)
• Session checkpoint with review results

Quality Gate

Advisory gate - offers to fix gaps, doesn't block arbitrarily:

• Coverage >= 95%: Automatic PASS → EXECUTE
• Coverage < 95%: User decision required → EXECUTE or ABORT

Skip Condition

If spec has no explicit acceptance criteria section:

• Skip REVIEW phase
• Proceed directly to EXECUTE
• Log warning: "Skipping REVIEW (no AC in spec)"

---

Phase 4: EXECUTE

Implement all tasks by dispatching subagents for each issue.

Objective

Execute all implementation tasks with TDD and code review, coordinated directly by this orchestrator.

Actions

1. Log phase start

   session_phase(sessionId: SESSION_ID, phase: "execute", status: "started")
   

2. Execution loop - For each issue in ISSUE_IDS:

   a) Mark issue in progress

   issues_update(id: ISSUE_ID, status: "in_progress")
   

   b) Dispatch implementation subagent

   Use Task tool to dispatch a subagent that follows the implementing-issue skill:

   Tool: Task
   Description: "Implement {ISSUE_ID}: {issue_title}"
   Prompt: |
   You are implementing a single issue.
   
   ## Working Directory Context
   
   **Working directory:** {WORKTREE_ABSOLUTE}
   **Branch:** {BRANCH_NAME}
   
   ### MANDATORY: Verify Location First
   
   Before any work, run:
   
   ```bash
   cd {WORKTREE_ABSOLUTE} && \
     echo "Directory: $(pwd)" && \
     echo "Branch: $(git branch --show-current)" && \
     test "$(pwd)" = "{WORKTREE_ABSOLUTE}" && echo "VERIFIED" || echo "FAILED - STOP"
   ```
   

   ALL bash commands MUST use: cd {WORKTREE_ABSOLUTE} && [command]

   Issue

   {ISSUE_ID}: {issue_title}

   Read the full issue with: issues_get(id: "{ISSUE_ID}")

   Your Job

   Follow the implementing-issue skill:

   1. Read the issue and understand requirements
   2. Follow TDD (practicing-tdd skill) - write failing test, implement, refactor
   3. Request code review (requesting-code-review skill)
   4. Fix any Critical/Important issues (2 attempts each)
   5. Verify all tests pass
   6. Commit your work
   7. Report back:
      • Implementation summary
      • Test results (all passing)
      • TDD Compliance Certification
      • Commit hash
      • Any blockers encountered

   If requirements are unclear, STOP and report the blocker.

   
   **c) Verify subagent response**
   
   Check for:
   - Location verification passed (VERIFIED)
   - Implementation summary provided
   - All tests passing
   - TDD Compliance Certification
   - Work committed
   
   **d) Handle blockers**
   
   If subagent reports a blocker:
   - Unclear requirements: ESCALATE to user immediately
   - Failed after 2 fix attempts: ESCALATE to user
   - Git conflicts: ESCALATE to user
   - Missing dependencies: Try to resolve, escalate if cannot
   
   **e) Save checkpoint**
   

   session_checkpoint( sessionId: SESSION_ID, tasksCompleted: [...completed_ids], tasksPending: [...remaining_ids], lastAction: "Completed {ISSUE_ID}: {issue_title}", resumeInstructions: "Continue with next issue or proceed to VERIFY if all done" )

   
   **f) Mark issue complete**
   

   issues_mark_complete(id: ISSUE_ID)

3. Log phase complete

   session_phase(
     sessionId: SESSION_ID,
     phase: "execute",
     status: "complete",
     metadata: {
       tasks_completed: TASK_COUNT,
       total_commits: N
     }
   )
   

Outputs

• All issues implemented with TDD
• Code reviewed for each issue
• Checkpoints saved after each issue
• PR description updated with progress

Quality Gate

All issues complete. If any issue cannot be completed, escalate to user with blocker details and pause session.

---

Phase 5: VERIFY

Run compliance audit using LLM-based verification.

Objective

Verify all acceptance criteria met using intelligent extraction (not brittle scripts).

Actions

1. Log phase start

   session_phase(sessionId: SESSION_ID, phase: "verify", status: "started")
   

2. Extract acceptance criteria from spec (LLM-based)

   Read specification and use LLM to extract:

   • All acceptance criteria (AC-001, AC-002, etc.)
   • E2E test requirements
   • Manual testing checklist items

   Why LLM not script:

   • Handles varied spec formats gracefully
   • Can infer criteria from prose descriptions
   • Won't break if spec structure differs slightly
   • More intelligent than regex parsing

3. Verify each criterion has evidence

   For each acceptance criterion:

   • Search for corresponding test
   • Search for implementing code
   • Search for relevant commits
   • Calculate compliance percentage

4. Run fresh test suite

   cd "{WORKTREE_ABSOLUTE}" && npm test 2>&1 | tee ".wrangler/sessions/{SESSION_ID}/final-test-output.txt"
   

   Capture:

   • TEST_EXIT_CODE = exit code
   • TESTS_TOTAL = total test count
   • TESTS_PASSED = passing test count

5. Check git status

   cd "{WORKTREE_ABSOLUTE}" && git status --short
   

   Capture:

   • GIT_CLEAN = true if output is empty

6. Calculate compliance percentage

   COMPLIANCE = (criteria_met / total_criteria) * 100
   

7. Log phase complete

   session_phase(
     sessionId: SESSION_ID,
     phase: "verify",
     status: "complete",
     metadata: {
       compliance_percentage: COMPLIANCE,
       tests_exit_code: TEST_EXIT_CODE,
       tests_total: TESTS_TOTAL,
       tests_passed: TESTS_PASSED,
       git_clean: GIT_CLEAN
     }
   )
   

Outputs

• Compliance report (X% complete)
• Test results (all passing or failures documented)
• Git status (clean or uncommitted changes documented)

Quality Gate & Self-Healing

Goal: Achieve >= 95% compliance through autonomous remediation.

Self-Healing Workflow:

1. Initial Compliance Audit

   • Run compliance check
   • Calculate compliance percentage
   • Categorize gaps

2. Gap Categorization

   For each unmet acceptance criterion, classify:

   • AUTO_FIX_TEST: Missing test coverage
   • AUTO_FIX_DOC: Missing documentation
   • AUTO_FIX_EDGE: Missing edge case handling
   • SEARCH_RETRY: Evidence likely exists but not found
   • FUNDAMENTAL_GAP: Core requirement not implemented (planning failure)

3. Autonomous Remediation (max 3 iterations)

   For AUTOFIX* gaps:

   • Create supplemental MCP issue with specific requirement
   • Execute using implementing-issue skill
   • Commit changes
   • Re-run compliance audit
   • Loop until compliance >= 95% OR 3 iterations reached OR no more auto-fixable gaps

   Log each remediation iteration:

   session_phase(
     sessionId: SESSION_ID,
     phase: "verify-remediation",
     status: "started",
     metadata: {
       iteration: N,
       gaps_to_fix: GAP_COUNT,
       gap_types: ["AUTO_FIX_TEST", ...]
     }
   )
   

4. Quality Gate Decision

   Compliance	Behavior
   >= 95%	✅ PASS → Document any gaps-fixed in PR, proceed to PUBLISH
   90-94%	⚠️ PASS WITH WARNINGS → Document minor gaps in PR, proceed to PUBLISH
   < 90%	❌ FAIL → Escalate to user with detailed gap analysis

5. Escalation Format (if < 90%)

   VERIFY Phase: Compliance Below Threshold
   
   Current compliance: X%
   Quality gate: 90% required
   
   Self-healing attempted:
   
   - Created and executed Y supplemental tasks
   - Fixed Z auto-fixable gaps
   - Remaining gaps: W
   
   Gap Analysis:
   
   - FUNDAMENTAL_GAP: [list core requirements not implemented]
   - SEARCH_RETRY failures: [list evidence not found after retries]
   
   Options:
   
   1. Review and approve current implementation (partial delivery)
   2. Let me create additional tasks for remaining gaps
   3. Abort session and revisit planning
   
   Your decision?
   

6. Document self-healing in PR (if any remediation occurred)

   Append to PR body:

   ### Compliance Self-Healing
   
   During verification, the following gaps were auto-fixed:
   
   - Created ISS-XXXXX: Added missing test for feature X
   - Created ISS-XXXXX: Added documentation for API Y
   - Created ISS-XXXXX: Added edge case handling for Z
   
   Final compliance: 96%
   

Critical Rules:

• NEVER escalate for missing tests/docs/edge cases - auto-fix them
• ONLY escalate for fundamental requirement gaps or after self-healing exhaustion
• Document all self-healing activity in PR description

Also block if:

• TEST_EXIT_CODE != 0 - tests failing (cannot proceed)
• GIT_CLEAN == false - uncommitted changes (cannot proceed)

---

Phase 6: PUBLISH

Finalize PR and mark ready for review.

Objective

Update PR with final summary and mark ready for merge.

Actions

1. Log phase start

   session_phase(sessionId: SESSION_ID, phase: "publish", status: "started")
   

2. Generate final PR description

   Create comprehensive PR body from audit data:

   ## Summary
   

References

For detailed information, see:

• references/detailed-guide.md - Complete workflow details, examples, and troubleshooting