oh-my-speckit

name	oh-my-speckit
description	MANDATORY workflow for GitHub Speckit projects. Enforces Spec-First → Clarify → Research → Evaluate → Design → Plan → TDD → Review, with optional Constitution and quality checks. Activates when working in projects with .github/speckit/ directory.
mcp	["context7"]

Oh-My-Speckit Workflow - Specification-Driven Development

[BLOCKING] EXECUTION DIRECTIVES

These directives are NOT suggestions. They are MANDATORY execution requirements.

IMPORTANT: How to Execute Speckit Commands

Speckit commands are prompt template files. To execute them:

1. Read the command file from project directory:

   • Primary: .opencode/commands/speckit.{command}.md
   • Fallback: .claude/commands/speckit.{command}.md

2. Process special syntax in the file:

   Syntax	Action
   $ARGUMENTS	Replace with user's feature description
   @filepath	Read that file and insert contents
   !command	Execute shell command, insert stdout

3. Execute ALL instructions in the processed file exactly as written

IMMEDIATE ACTION REQUIRED

Upon loading this skill, you MUST:

1. Check project structure first using Glob/Read tools
2. If .github/speckit/ missing → Read and execute .opencode/commands/oms-init.md
   • If .opencode/commands/oms-init.md and .claude/commands/oms-init.md are both missing, run global /oms-init and follow its prompts.
3. If constitution missing → Read and execute .opencode/commands/speckit.constitution.md
4. Otherwise → Read and execute .opencode/commands/speckit.specify.md

NO TEXT BEFORE TOOL CALLS - Start with file checks immediately.

WHAT "MUST" MEANS

Directive	Meaning
MUST	Non-negotiable. Failure = workflow broken.
MUST NOT	Forbidden. Doing this = workflow broken.
SHOULD	Strong recommendation. Skip only with explicit user override.

---

Critical Guardrails

• NO TodoWrite in Phases 0–5: These phases require user input and explicit approvals.
• Approval Gate: After Phases 0–3.6, STOP and wait for explicit user approval.
• Do NOT auto-advance: If approval is missing, do not continue to the next phase.
• Research before Evaluate: Phase 3.5 (Research) MUST complete before Phase 3.6 (Evaluate).

Workflow Overview (Hybrid)

0. Initialize     → Setup Speckit structure (oms-init)
1. Constitution   → (MANDATORY for new projects, first-time only)
2. Spec-First     → Define WHAT (speckit.specify)
3. Clarify        → Resolve unknowns (speckit.clarify)
3.5 Research      → Market/competitive analysis (librarian) [NEW]
3.6 Evaluate      → Feature add/remove/modify assessment (speckit-architect) [NEW]
4. Design Session → Refine HOW (brainstorming)
5. Plan           → Create detailed plan (speckit.plan + writing-plans)
6. (Optional)     → Analyze / Checklist quality pass
7. Tasks          → speckit.tasks
8. TDD Execution  → /orchestrate → parallel workers → TDD (MANDATORY)
9. Review & Update→ Verify & update spec (requesting-code-review)

When to Use This Skill

Use only when explicitly invoked by name.

DO NOT activate when:

• Simple bug fixes (use systematic-debugging instead)
• Documentation updates only
• Refactoring without behavior change

---

Quick Mode (OPTIONAL - User Must Enable)

When to Use Quick Mode

Quick Mode is for small features, internal tools, or time-sensitive work where full Research/Evaluate phases are overkill.

Eligible scenarios:

• Internal tools with no external users
• Bug fixes that require spec tracking
• Small features (< 1 day implementation)
• Prototypes or MVPs
• User explicitly requests "quick mode" or "빠른 모드"

NOT eligible:

• User-facing features
• Security-related changes
• Payment/financial features
• Features requiring market differentiation

How to Enable Quick Mode

User must explicitly say one of:

• "Use quick mode"
• "빠른 모드로 해줘"
• "Skip research and evaluate"
• "Research/Evaluate 생략해줘"

Quick Mode Workflow

0. Initialize     → (same)
1. Constitution   → (same, first-time only)
2. Spec-First     → (same)
3. Clarify        → (same)
3.5 Research      → ⏭️ SKIPPED (Quick Mode)
3.6 Evaluate      → ⏭️ SKIPPED (Quick Mode)
4. Design Session → (same)
5. Plan           → (same)
6-10              → (same)

Quick Mode Spec Frontmatter

When Quick Mode is enabled, add to spec frontmatter:

---
status: clarified
mode: quick
quick_mode_reason: [user's reason for enabling quick mode]
skipped_phases: [research, evaluate]
---

Quick Mode Guardrails

Even in Quick Mode:

• ✅ Spec is STILL required
• ✅ Clarify phase is STILL required
• ✅ TDD is STILL required
• ✅ Tests are STILL required
• ❌ Cannot skip spec creation
• ❌ Cannot skip clarify (edge cases matter!)

Reverting from Quick Mode

If during implementation you discover:

• Competitors exist that should be analyzed
• Feature scope expanded beyond "small"
• Security implications found

Then: STOP and revert to full workflow

1. Notify user: "This feature needs full Research/Evaluate. Reverting from Quick Mode."
2. Execute Phase 3.5 and 3.6
3. Continue from there

Phase 0: Initialize (MANDATORY, FIRST-TIME ONLY)

👉 EXECUTE NOW (if structure missing)

Read and execute the speckit init command file:

Read: .opencode/commands/oms-init.md (or .claude/commands/oms-init.md)
Then: Follow ALL instructions in that file

If both files are missing:

• Run global /oms-init and follow its prompts

---

REQUIRED for all new projects. Setup Speckit directory structure.

1. Check if Speckit structure exists:

   • .github/speckit/ directory
   • .opencode/designs/ and .opencode/plans/ directories

2. If structure missing:

   • Read .opencode/commands/oms-init.md (or .claude/commands/oms-init.md)
   • Execute all instructions in that file
   • If both files are missing, run global /oms-init and follow its prompts
   • Get approval before proceeding to Phase 1

3. If structure exists:

   • Verify structure is complete
   • Skip to Phase 1

Phase 1: Constitution (MANDATORY, FIRST-TIME ONLY)

👉 EXECUTE NOW (if constitution missing)

Read and execute the constitution command file:

Read: .opencode/commands/speckit.constitution.md
Then: Follow ALL instructions in that file

---

REQUIRED for all new projects. Set guiding principles before any feature work.

1. Check if constitution exists:

   • .github/speckit/constitution.md
   • .specify/memory/constitution.md

2. If constitution missing:

   • Read .opencode/commands/speckit.constitution.md
   • Execute all instructions in that file
   • Interview user for project principles
   • Get approval before proceeding to Phase 2

3. If constitution exists:

   • Read and apply it to all feature work
   • Skip to Phase 2

Phase 2: Spec-First (MANDATORY)

[BLOCKING] EXECUTION REQUIRED

MUST: Read and execute .opencode/commands/speckit.specify.md MUST NOT: Skip reading the command file and make up your own spec format.

---

Phase 2A: Initial Draft

Before ANY code or design work:

1. Delegate to Spec Writer Agent:

   task(
     subagent_type="speckit-spec-writer",
     description="Create initial spec",
     prompt="Read .opencode/commands/speckit.specify.md and execute it for: [user's feature request]"
   )
   

2. Read and Execute the Speckit command file:

   • Read .opencode/commands/speckit.specify.md
   • Process any @filepath references (read those files too)
   • Execute all instructions exactly as written

3. Create comprehensive initial draft:

   • Analyze user requirements
   • Review existing codebase context
   • Create structured spec document following the template
   • Emphasize deep reasoning and completeness

4. Show draft to user:

   • Present the initial spec
   • Collect direct feedback for refinement

Phase 2B: Iterative Refinement

1. Continue with Spec Writer Agent:

   task(
     subagent_type="speckit-spec-writer",
     description="Refine spec",
     prompt="Refine spec based on user feedback: [feedback]"
   )
   

2. Rapid feedback cycles (5-10+ rounds expected):

   • Quickly incorporate user feedback
   • Make focused edits and clarifications
   • Favor speed and iteration over perfection

3. Spec approval:

   • Continue cycles until user explicitly approves
   • Save final draft to .github/speckit/specs/{feature-name}/spec.md
   • STOP here until user approves

Phase 3: Clarify (MANDATORY)

[BLOCKING] EXECUTION REQUIRED

MUST: Read and execute .opencode/commands/speckit.clarify.md MUST: Delegate to speckit-architect for validation.

---

Architect: speckit-architect

Validate and clarify the draft:

1. Delegate to Architect Agent:

   task(
     subagent_type="speckit-architect",
     description="Validate and clarify spec",
     prompt="Read .opencode/commands/speckit.clarify.md and execute it for spec at: [spec path]"
   )
   

2. Read and Execute the Speckit command file:

   • Read .opencode/commands/speckit.clarify.md
   • Execute all instructions exactly as written

3. Deep Validation:

   • Find ALL edge cases not covered
   • Identify ALL ambiguities
   • Add missing non-functional requirements
   • Perform risk analysis
   • Generate clarifying questions

4. Refinement (1-2 rounds expected):

   • Present findings to user
   • Get answers to clarifying questions
   • Update spec with clarifications
   • Ensure spec is crystal clear

5. Completion:

   • Ensure ALL uncertainties resolved
   • Update spec status to clarified
   • Proceed to Phase 3.5 (Research)

Phase 3.5: Market Research (MANDATORY)

[BLOCKING] EXECUTION REQUIRED

MUST: Execute market/competitive research before evaluation. MUST: Use librarian agent for research (runs in background).

---

Research Agent: librarian

Research the market landscape for the clarified spec:

1. Launch Research Agent (Background):

   background_task(
     agent="librarian",
     description="Market research for [feature name]",
     prompt="Research market landscape for this feature spec:
   
     Feature: [feature name]
     Spec path: [spec path]
   
     Research and report:
     1. **Competitive Analysis**: What similar solutions exist? (3-5 competitors)
        - Their approach, strengths, weaknesses
     2. **Market Trends**: Current trends in this space (2025)
     3. **Best Practices**: Industry standards and patterns
     4. **Potential Differentiators**: How could this feature stand out?
     5. **Risks/Concerns**: Common pitfalls to avoid
   
     Output a structured research.md report."
   )
   

2. Wait for research completion

3. Present research.md to user:

   • Show competitive landscape
   • Show market trends
   • Allow user to note any irrelevant findings
   • Save to .github/speckit/specs/{feature-name}/research.md

4. Proceed to Evaluate phase

Phase 3.6: Feature Evaluation (MANDATORY)

[BLOCKING] EXECUTION REQUIRED

MUST: Evaluate spec against research findings. MUST: Delegate to speckit-architect (Opus) for deep reasoning.

---

Architect: speckit-architect

Evaluate the spec with research context:

1. Delegate to Architect Agent:

   task(
     subagent_type="speckit-architect",
     description="Evaluate spec against market research",
     prompt="Evaluate this feature spec against market research.
   
     Spec: [spec path]
     Research: [research.md path]
   
     Provide evaluation:
     1. **Features to ADD**: What's missing that competitors have or market expects?
     2. **Features to REMOVE**: What's over-engineered or unnecessary?
     3. **Features to MODIFY**: What needs adjustment based on best practices?
     4. **Priority Assessment**: Which features are must-have vs nice-to-have?
     5. **Differentiation Strategy**: How to stand out from competitors?
   
     For each suggestion, provide:
     - Rationale (why)
     - Evidence (from research)
     - Impact (high/medium/low)
   
     Output structured evaluation.md report."
   )
   

2. Present evaluation.md to user:

   • Show feature recommendations (add/remove/modify)
   • Show priority assessment
   • Show differentiation opportunities

3. User Decision:

   • For EACH suggestion: Accept or Reject
   • User may provide reasoning for rejections
   • Update spec based on accepted suggestions

4. Finalize Spec:

   • Apply accepted changes to spec
   • Update spec status to evaluated
   • Save evaluation decisions to .github/speckit/specs/{feature-name}/evaluation.md
   • STOP and confirm with user before proceeding to Design

Phase 4: Design Session

Architect: speckit-architect

ONLY after spec is evaluated and finalized (Phase 3.6 complete):

1. Delegate to Architect Agent:

   task(
     subagent_type="speckit-architect",
     description="Design architecture",
     prompt="Design architecture for: [feature name]. Follow brainstorming skill methodology."
   )
   

2. Design Session Questions:

   • "What's the overall architecture?"
   • "Which existing components/modules can be reused?"
   • "What new files/components are needed?"
   • "What are the key technical decisions?"
   • "What could go wrong?"

3. Create design document:

   • Save to .opencode/designs/{feature-name}-design.md
   • Include:
     • Architecture diagram (text-based)
     • Component structure and responsibilities
     • Key technical decisions with rationale
     • Risk mitigation strategies
     • Data flow design

4. Update spec with design:

   • Add "## Design" section to spec
   • Link to design document
   • Update status to designed
   • Get user approval before proceeding

Phase 5: Implementation Plan

[BLOCKING] EXECUTION REQUIRED

MUST: Read and execute .opencode/commands/speckit.plan.md MUST: Delegate to speckit-architect for planning.

---

Architect: speckit-architect

ONLY after design is approved:

1. Delegate to Architect Agent:

   task(
     subagent_type="speckit-architect",
     description="Create implementation plan",
     prompt="Read .opencode/commands/speckit.plan.md and execute it for: [feature name]"
   )
   

2. Read and Execute the Speckit command file:

   • Read .opencode/commands/speckit.plan.md
   • Execute all instructions exactly as written

3. Create Detailed Plan:

   • Break design into concrete tasks
   • Map task dependencies
   • Define file structure
   • Specify API contracts
   • Plan test strategy
   • Define rollout approach

4. Link plan to spec:

   • Save to .opencode/plans/{feature-name}-plan.md
   • Update spec with "## Implementation Plan" section
   • Link to plan file
   • Update status to planned

5. CRITICAL DECISION POINT:

   • Present the plan to the user
   • Ask: "Do you approve this plan? If yes, I will proceed to implement ALL tasks autonomously."
   • WAIT for explicit approval
   • Once approved, DO NOT STOP until Phase 9 is complete

Phase 6: (Optional) Analyze / Checklist

Optional quality pass before task execution:

1. Run analysis (if needed):

   • Read .opencode/commands/speckit.analyze.md
   • Execute all instructions in that file

2. Run checklist (if needed):

   • Read .opencode/commands/speckit.checklist.md
   • Execute all instructions in that file

3. Skip to Phase 7 if not needed

Phase 7: Tasks (AUTO-EXECUTE)

[BLOCKING] EXECUTION REQUIRED

MUST: Read and execute .opencode/commands/speckit.tasks.md

---

Main Sisyphus

Triggered IMMEDIATELY after Plan approval:

1. Announce autonomous execution mode:

   • "Engaging autonomous implementation. Stand by for completion."

2. Read and Execute the Speckit tasks command:

   • Read .opencode/commands/speckit.tasks.md
   • Execute all instructions exactly as written

3. Skill Mapping (Smart Context):

   • Review the generated tasks.md
   • /speckit.tasks command now auto-generates skill tags (see .opencode/command/speckit.tasks.md)
   • Verify skill tags are present: → /skill-name at end of task descriptions
   • Common skill mappings:

     Task Type	Skill Tag
     Browser/E2E testing	→ /playwright
     UI/UX components	→ /frontend-ui-ux
     Quality checks	→ /speckit.checklist
     Complex debugging	→ /systematic-debugging

   • Optional: add dependencies with (id: X) and (depends_on: A,B)

   중요: Phase 8에서 태스크 실행 시, → /skill-name 표기를 만나면:

   1. Read .opencode/commands/{skill-name}.md 또는 해당 스킬 파일
   2. Execute 그 파일의 지시사항을 따름

4. Proceed to Phase 8 immediately

Phase 8: TDD Execution via Orchestration (AUTO-EXECUTE, MANDATORY)

[BLOCKING] PARALLEL EXECUTION REQUIRED

MUST use /orchestrate for ALL task execution. No sequential fallback.

⛔ FORBIDDEN ACTIONS

• ❌ Implementing tasks directly without /orchestrate
• ❌ Calling task(subagent_type="speckit-developer") directly for implementation
• ❌ Writing code yourself in this phase
• ❌ Skipping orchestration "because it's simpler"

✅ REQUIRED ACTION

IMMEDIATELY execute this command:

/orchestrate execute .opencode/plans/<feature-name>-plan.md

If /orchestrate skill is not loaded, first load it:

Read: ~/.claude/skills/orchestrate/SKILL.md
Then execute the orchestration command above

---

Orchestration Flow

1. Spawn parallel workers via /orchestrate:

   /orchestrate execute .opencode/plans/<feature-name>-plan.md
   

   • LLM analyzes tasks and groups them dynamically
   • Each worker group spawns as tmux session
   • Workers execute in parallel

2. Each Worker calls appropriate subagent:

   Worker → task() → speckit-developer (or other agent based on domain)
   

3. speckit-developer handles task (with sub-delegation):

   Developer가 태스크를 받으면 도메인에 따라:

   • UI/Frontend → task() → frontend-ui-ux-engineer
   • Documentation → task() → document-writer
   • Complex Analysis → task() → oracle
   • Backend/API/Testing → 직접 TDD 구현

   하나의 태스크에서 여러 서브에이전트 호출 가능. 예: "로그인 기능" = UI(위임) + API(직접) + 문서(위임)

4. TDD Cycle (for direct implementation):

   1. Load Skill: 태스크에 → /skill-name이 있으면:
      • Read 해당 스킬의 command 파일 또는 SKILL.md
      • Follow 그 파일의 지시사항
   2. RED: Write failing test
   3. GREEN: Write minimal code to pass
   4. REFACTOR: Clean up while keeping tests green
   5. Mark Complete: Check [x] in tasks.md

5. Result Collection:

   • Orchestrator collects results from all workers
   • Synthesizes and detects conflicts
   • Reports completion

Why Mandatory Orchestration?

• 일관된 실행 모델: 항상 같은 방식으로 동작
• 병렬 처리: 태스크가 1개여도 worker 1개로 실행
• 서브에이전트 체인 활용: Worker → Developer → 전문 에이전트
• 토큰 폴백 지원: 각 에이전트별 모델 폴백 설정 적용

Phase 8.5: Global Refactoring (AUTO-EXECUTE)

Optimizer: speckit-optimizer

Project-Wide Cleanup:

1. Delegate to Optimizer Agent:

   task(
     subagent_type="speckit-optimizer",
     description="Global refactoring",
     prompt="Perform global refactoring after implementation. DRY, cleanup, optimize."
   )
   

2. Execute Cleanup (leveraging 400k context):

   • DRY: Extract duplicated logic from multiple files into shared helpers
   • Structure: Verify file organization and move files if needed
   • Clean: Remove unused imports, console logs, and dead code
   • Style: Ensure consistent naming conventions across modules
   • Performance: Optimize algorithms and queries where beneficial

3. Verify Integrity:

   • Run full test suite (npm test)
   • Must pass 100% before proceeding
   • Verify no regressions introduced

Phase 9: Review & Update (AUTO-EXECUTE)

Developer: speckit-developer

Triggered IMMEDIATELY after refactoring complete:

1. Delegate to Developer Agent:

   task(
     subagent_type="speckit-developer",
     description="Review and update spec",
     prompt="Review implementation against spec and update status to 'implemented'"
   )
   

2. Self-review checklist:

   • All acceptance criteria met (from spec)
   • All tests pass
   • Code follows project conventions
   • No unrelated changes
   • Plan tasks all completed

3. Update spec status:

   • Change status to implemented
   • Add implementation date
   • Link to PR (if created)
   • Add actual implementation notes

Phase 10: Systematic Verification & Quality Score (AUTO-EXECUTE)

Step 10.1: E2E Testing (Conditional)

If project has E2E tests or UI components:

1. Delegate to E2E Runner Agent:

   task(
     subagent_type="e2e-runner",
     description="Run E2E tests",
     prompt="Run E2E tests for implemented features using Playwright"
   )
   

2. E2E Runner Responsibilities:

   • Run Playwright tests for critical user flows
   • Validate UI interactions work correctly
   • Check for visual regressions
   • Report flaky tests

3. Skip if: No E2E tests exist or feature is backend-only

---

Step 10.2: Security Review (Conditional - SaaS/Auth)

If feature involves: auth, payments, user data, API endpoints:

1. Delegate to Security Reviewer Agent:

   task(
     subagent_type="security-reviewer",
     description="Security audit",
     prompt="Security audit for: [feature name]. Check OWASP Top 10."
   )
   

2. Security Review Checklist (OWASP Top 10):

   • No SQL/NoSQL injection vulnerabilities
   • Input sanitization on all user inputs
   • No hardcoded secrets/credentials
   • Proper authentication checks
   • Authorization on all endpoints
   • XSS prevention (output encoding)
   • CSRF protection
   • Secure session management

3. Critical Vulnerability Protocol:

   • 🔴 CRITICAL found → STOP, fix immediately, re-review
   • 🟠 HIGH found → Fix before completion
   • 🟡 MEDIUM found → Document, fix in follow-up
   • 🟢 LOW/None → Proceed

4. Skip if: Feature has no security implications (pure UI, docs, etc.)

---

Step 10.3: Quality Scoring

Optimizer: speckit-optimizer

Final Quality Gate with Scoring:

1. Delegate to Optimizer Agent:

   task(
     subagent_type="speckit-optimizer",
     description="Quality verification",
     prompt="Final verification, quality scoring, and debugging if needed"
   )
   

2. Run Full Test Suite:

   • Ensure no regressions
   • Verify all tests pass

3. Calculate Quality Score (0-100):

   Category	Max Points	Criteria
   Unit Tests	25	Pass rate: (passed/total) * 25
   Code Coverage	25	Coverage %: (coverage/100) * 25
   Integration Tests	15	Pass rate: (passed/total) * 15
   E2E/Browser Tests	15	Pass rate: (passed/total) * 15
   Spec Compliance	10	All acceptance criteria met: 10, partial: 5
   Code Quality	10	No lint errors: 5, No type errors: 5

   Minimum Passing Score: 80/100

4. Quality Score Calculation Script:

   # Run tests and capture results
   TEST_RESULT=$(npm test 2>&1)
   PASSED=$(echo "$TEST_RESULT" | grep -oP '\d+(?= passed)' | head -1 || echo 0)
   FAILED=$(echo "$TEST_RESULT" | grep -oP '\d+(?= failed)' | head -1 || echo 0)
   TOTAL=$((PASSED + FAILED))
   
   # Calculate unit test score
   if [ $TOTAL -gt 0 ]; then
     UNIT_SCORE=$((PASSED * 25 / TOTAL))
   else
     UNIT_SCORE=0
   fi
   
   # Get coverage (if available)
   COVERAGE=$(npm run coverage 2>&1 | grep -oP 'All files[^\d]*\K\d+' || echo 0)
   COVERAGE_SCORE=$((COVERAGE * 25 / 100))
   
   # Calculate total
   TOTAL_SCORE=$((UNIT_SCORE + COVERAGE_SCORE))
   echo "Quality Score: $TOTAL_SCORE/100"
   

5. Score Actions:

   • Score >= 80: Proceed to completion
   • Score 60-79: Warning, ask user if acceptable
   • Score < 60: BLOCK completion, require fixes

6. If bugs found:

   • Use systematic debugging (scientific method)
   • Fix root causes, not symptoms
   • Add tests to prevent recurrence
   • Re-calculate quality score

7. Report Completion:

   • ONLY NOW stop and report to user
   • Include quality score breakdown
   • Provide summary of changes
   • List test results
   • Note any deviations from plan

   Example Output:

   ## Implementation Complete
   
   **Quality Score: 87/100**
   - Unit Tests: 23/25 (92% pass rate)
   - Coverage: 21/25 (84% coverage)
   - Integration: 15/15 (100%)
   - E2E: 13/15 (87%)
   - Spec Compliance: 10/10
   - Code Quality: 5/10 (minor lint warnings)
   
   All acceptance criteria met. Ready for review.
   

[BLOCKING] Subagent Delegation Rules

MUST delegate to specialized subagents. Direct execution is FORBIDDEN.

Mandatory Agent Assignments

Phase	Agent	Invocation	MUST Use
Phase 2 (Spec)	speckit-spec-writer	task(subagent_type="speckit-spec-writer", ...)	YES
Phase 3 (Clarify)	speckit-architect	task(subagent_type="speckit-architect", ...)	YES
Phase 3.5 (Research)	librarian	background_task(agent="librarian", ...)	YES
Phase 3.6 (Evaluate)	speckit-architect	task(subagent_type="speckit-architect", ...)	YES
Phase 4-5 (Design/Plan)	speckit-architect	task(subagent_type="speckit-architect", ...)	YES
Phase 8-9 (TDD/Review)	speckit-developer	task(subagent_type="speckit-developer", ...)	YES
Phase 8.5, 10.3 (Refactor/Verify)	speckit-optimizer	task(subagent_type="speckit-optimizer", ...)	YES
Phase 10.1 (E2E)	e2e-runner	task(subagent_type="e2e-runner", ...)	Conditional
Phase 10.2 (Security)	security-reviewer	task(subagent_type="security-reviewer", ...)	Conditional

Domain-Specific Delegation (via speckit-developer)

When speckit-developer receives a task, it auto-delegates based on domain:

Task Domain	Keywords	Delegate To
Frontend/UI	UI, styling, component, layout, animation, CSS	frontend-ui-ux-engineer
Database/Schema	database, schema, migration, SQL, model design	speckit-architect
Documentation	docs, README, API docs, guide	document-writer
Complex Analysis	architecture, security audit, performance analysis	oracle
Backend/API/Testing	(default - developer handles)	Self

Note: This delegation happens automatically within speckit-developer. You don't need to call these agents directly from the main workflow.

Why Delegation is MANDATORY

1. Token Limit Fallback: Subagents have automatic model fallback configured
2. Domain Optimization: Each agent uses a model optimized for its domain
3. Context Isolation: Prevents context pollution between phases
4. Parallel Execution: Multiple subagents can run concurrently

FORBIDDEN Actions (Main Agent)

• MUST NOT: Write spec content directly (delegate to spec-writer)
• MUST NOT: Create design documents directly (delegate to architect)
• MUST NOT: Write implementation code directly (delegate to developer)
• MUST NOT: Perform refactoring directly (delegate to optimizer)

ALLOWED Actions (Main Agent)

• MAY: Read files for context
• MAY: Run bash commands for status checks
• MAY: Coordinate between subagents
• MAY: Verify subagent output quality
• MAY: Update todo lists and progress tracking

Agent Location: ~/.config/opencode/agents/ or ~/.claude/agents/

Skill Tags in Tasks

Tasks in tasks.md may have skill tags at the end:

- [ ] T007 [US1] Create login page → /playwright

Execution Rule: When encountering → /skill-name:

1. Parse the skill name (e.g., playwright)
2. Read the skill's command file: .opencode/commands/{skill-name}.md or ~/.claude/skills/{skill-name}/SKILL.md
3. Execute the task following that file's instructions

Available Skills for Task Mapping

Skill	Purpose	How to Execute
/playwright	Browser automation, E2E tests	Read .opencode/commands/playwright.md
/frontend-ui-ux	UI component design	task(subagent_type="frontend-ui-ux-engineer", ...)
/speckit.checklist	Quality validation	Read .opencode/commands/speckit.checklist.md
/systematic-debugging	Complex debugging	Read ~/.claude/skills/systematic-debugging/SKILL.md
/orchestrate	Parallel worker execution	Read ~/.claude/skills/orchestrate/SKILL.md

Critical Rules

NEVER:

• ❌ Skip spec creation
• ❌ Skip clarify before design
• ❌ Skip tasks before TDD
• ❌ Start coding before spec approval
• ❌ Write code before tests (except when TDD doesn't apply)
• ❌ Mark task complete without verification
• ❌ Change spec status without completing phase

ALWAYS:

• ✅ Ask before proceeding to next phase
• ✅ Update spec at each phase
• ✅ Run tests after every change
• ✅ Commit after each task
• ✅ Show progress to user
• ✅ Prefer scripts for deterministic data extraction to reduce context cost
• ✅ Archive reusable prompts in prompts/ (or .opencode/prompts/) and reference them

Verification Before Completion

Before declaring "done", verify:

# 1. All tests pass
npm test  # or appropriate test command

# 2. Spec is updated
grep "status: implemented" .github/speckit/specs/{feature-name}/spec.md

# 3. Plan is complete
grep "\[x\]" .opencode/plans/{feature-name}-plan.md | wc -l

# 4. No uncommitted changes
git status

Example Session

User: "Add user authentication"

You:

1. ✅ /oms-init (first-time only - setup structure)
2. ✅ /speckit.constitution (first-time only - project principles)
3. ✅ /speckit.specify → spec created/updated
4. ✅ /speckit.clarify → clarifications recorded
5. ✅ /speckit.research → market/competitive research completed
6. ✅ /speckit.evaluate → feature evaluation, accept/reject decisions
7. ✅ Use superpowers:brainstorming → Design session
8. ✅ Create .opencode/designs/user-auth-design.md
9. ✅ /speckit.plan → plan created
10. ✅ (Optional) superpowers:writing-plans enhancement
11. ✅ (Optional) /speckit.analyze / /speckit.checklist
12. ✅ /speckit.tasks → tasks approved
13. ✅ Use superpowers:test-driven-development
14. ✅ Execute each task: RED → GREEN → REFACTOR
15. ✅ Use superpowers:requesting-code-review
16. ✅ Update spec status to implemented
17. ✅ Done!

File Structure

project/
├── .github/
│   └── speckit/
│       ├── specs/
│       │   ├── user-auth/
│       │   │   ├── spec.md           ← WHAT (requirements)
│       │   │   ├── research.md       ← Market/competitive research
│       │   │   └── evaluation.md     ← Feature evaluation decisions
│       │   └── payment/
│       │       └── spec.md
│       ├── decisions/
│       │   └── auth-strategy.md
│       └── constitution.md           ← optional, first-time only
│
├── .opencode/
│   ├── designs/
│   │   └── user-auth-design.md       ← HOW (architecture)
│   └── plans/
│       └── user-auth-plan.md         ← TASKS (implementation)
│
├── src/
│   └── ... (code)
└── tests/
    └── ... (tests)

Troubleshooting

"User wants to skip clarify" → Explain: "Clarify is required before design to avoid rework."

"Feature is too small for full workflow" → Still create minimal spec. If truly trivial (1-2 line change), keep phases compact.

"Spec already exists but is outdated" → Update spec first, then re-run clarify.

"User wants to code immediately" → Firmly but politely insist: "Let me create a quick spec first. It will save time."

Integration with Superpowers

This skill orchestrates other Superpowers skills:

oh-my-speckit (meta-skill)

    ├─ uses: oms-init (initialization)
    ├─ uses: brainstorming (design)
    ├─ uses: writing-plans (planning)
    ├─ uses: test-driven-development (implementation)
    ├─ uses: requesting-code-review (review)
    ├─ uses: systematic-debugging (verification)
    ├─ uses: spec-cache (optimization)
    └─ manages: Spec lifecycle

Priority: This skill takes precedence over individual skills when Speckit directory is present.

Status Lifecycle

draft → clarified → researched → evaluated → designed → planned → tasked → implemented → verified

Track status in spec frontmatter. Never skip statuses.