superbuild（superbuild 搜索）

Superbuild: Plan Execution Engine

Execute implementation plans one phase at a time with strict quality enforcement, test verification, and conventional commit generation.

Overview

Superbuild is a rigid execution engine for implementation plans. It enforces:

• Phase-by-phase execution (no skipping ahead)
• Definition of Done verification before phase completion
• Test presence and passing verification
• Linter/formatter/typechecker enforcement
• Conventional commit message generation per phase

This is NOT a planning skill. Use superplan to create plans, then superbuild to execute them.

Reference Index - MUST READ When Needed

References contain detailed templates and patterns. Read BEFORE you need them.

DO NOT SKIP REFERENCES. They contain exact templates and patterns that are NOT duplicated here.

Critical Workflow

┌─────────────────────────────────────────────────────────────────────┐
│                      SUPERBUILD EXECUTION FLOW                       │
│                     (REPEAT FOR EACH PHASE)                          │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  1. INGEST PLAN    │  User provides plan document path              │
│         ↓          │  NO PLAN = EXIT (ask user, then exit if none)  │
│  2. READ PHASES    │  Output ALL phases with estimates              │
│         ↓          │  IF context high → suggest compact first       │
│  2.5 PLAN REVIEW   │  Check for ambiguity, missing deps, test gaps  │
│         ↓          │  IF concerns → PAUSE and ask user              │
│  3. EXECUTE PHASE  │  One phase at a time (or parallel if marked)   │
│         ↓          │  USE SUB-AGENTS for parallel phases            │
│  4. ENFORCE DOD    │  Tests exist? Tests pass? Linter? Formatter?   │
│         ↓          │  ALL must pass → continue. ANY fail → STOP     │
│  5. UPDATE PLAN    │  Check off tasks, update status in plan file   │
│         ↓          │  ⚠️  THIS HAPPENS AFTER EVERY PHASE            │
│  6. COMMIT MSG     │  Generate conventional commit (NEVER git ops)  │
│         ↓          │  User handles all git operations               │
│  7. FUNCTIONAL TEST│  Explain how to test. Offer integration script │
│         ↓          │  NEVER auto-create scripts. ALWAYS ask first   │
│  8. STOP           │  Full stop. Suggest compact. Wait for user.    │
│         ↓          │  OVERRIDE: --build-all flag continues          │
│  9. AUTO-COMPACT   │  BUILD-ALL ONLY: Run /compact at CHECKPOINT    │
│                    │  Parse focus from CHECKPOINT, compact, continue│
│                                                                     │
│  ════════════════════════════════════════════════════════════════   │
│  Steps 3-8 repeat for EACH PHASE. Plan updates after EVERY phase.  │
│  BUILD-ALL: Steps 3-9 repeat automatically with auto-compact.       │
│  ════════════════════════════════════════════════════════════════   │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

Step 1: Ingest Plan

REQUIRED: Plan document must be provided.

Plan Header Recognition

Plans created by superplan include this header:

> **For Claude:** Execute this plan using `/superbuild` skill.

This confirms the plan is in the expected format with:

• Parallelizable phases with poker estimates
• Definition of Done per phase
• 5-step TDD micro-structure per task

I'll help you execute your implementation plan.

Please provide the plan document:
1. Path to plan file (e.g., docs/plans/001-feature/plan.md)
2. Paste the plan content directly

Which would you prefer?

If no plan provided after asking: EXIT immediately.

I cannot execute without a plan document.

To create a plan, use the `superplan` skill first:
  /superplan

Then come back with the completed plan.

[EXIT - No further action]

NO EXCEPTIONS. Do not improvise. Do not create plans on the fly. Do not proceed without a plan document.

Step 2: Read All Phases

After ingesting the plan:

1. Output all phases with their estimates and dependencies
2. Check context usage - if high, suggest compacting first

PLAN LOADED: [Feature Name]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

| Phase | Name | Est. | Depends On | Parallel With | Status |
|-------|------|------|------------|---------------|--------|
| 0 | Bootstrap | 5 | - | - | ⬜ |
| 1 | Setup | 3 | 0 | - | ⬜ |
| 2A | Backend | 8 | 1 | 2B, 2C | ⬜ |
| 2B | Frontend | 5 | 1 | 2A, 2C | ⬜ |
| 2C | Tests | 3 | 1 | 2A, 2B | ⬜ |
| 3 | Integration | 5 | 2A,2B,2C | - | ⬜ |

Total: 29 points | Parallel phases: 2A, 2B, 2C

⚠️  Context Usage Advisory
If context is high, consider compacting before continuing.
Large plans consume significant context per phase.

Ready to execute Phase 0?

Step 2.5: Critical Plan Review

Before executing, review the plan for concerns:

1. Ambiguous tasks - Any task unclear about what to do?
2. Missing dependencies - Are required files/APIs/packages identified?
3. Test gaps - Does each task have testable acceptance criteria?
4. Risky changes - Any destructive operations or breaking changes?

If Concerns Exist

⚠️  PLAN REVIEW - Concerns Identified
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

1. [Concern 1 - what and why]
2. [Concern 2 - what and why]

Please clarify before I proceed.
[EXECUTION PAUSED]

If No Concerns

Proceed to Step 3 (Execute Phase).

Step 3: Execute Phase

TDD Micro-Structure Enforcement

Plans created by superplan use a 5-step TDD micro-structure per task:

1. Write failing test
2. Run test, verify failure (MANDATORY - must see it fail)
3. Write minimal implementation
4. Run test, verify pass (MANDATORY - must see it pass)
5. Stage for commit

STOP. Read superplan/references/TASK-MICROSTRUCTURE.md and superplan/references/TDD-DISCIPLINE.md NOW before executing any task.

TDD Enforcement Rules

When executing tasks, enforce this order STRICTLY:

If test passes immediately: The test is wrong. It tests existing behavior, not new behavior. Fix the test or you are not doing TDD.

If you cannot explain why the test failed: You do not understand what you are testing. Stop and clarify requirements.

Sequential Phases

Execute one at a time. Do not proceed to next phase until current is COMPLETE.

Parallel Phases

For phases marked "Parallel With", MUST use sub-agents or parallel Task tool calls.

EXECUTING PARALLEL PHASES: 2A, 2B, 2C
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Launching 3 parallel sub-agents...

[Sub-agent 2A: Backend implementation]
[Sub-agent 2B: Frontend implementation]
[Sub-agent 2C: Test implementation]

Each sub-agent MUST return:
- Implementation status
- Definition of Done checklist status
- Conventional commit message

CRITICAL: Each sub-agent returns its commit message. Main agent MUST bubble up ALL commit messages to user.

Sub-Agent Result Verification

CRITICAL: Do NOT trust sub-agent claims without independent verification.

When parallel phases complete:

1. Collect all sub-agent results
2. For EACH sub-agent claiming success:

- Run npm test (or equivalent) to verify tests pass - Run npm run lint to verify linter passes - Verify plan file was actually updated

1. Only after independent verification, output commit messages

Sub-agent reports are CLAIMS, not EVIDENCE. Fresh verification required.

Step 4: Enforce Definition of Done

EVERY phase must pass ALL quality gates before completion.

Fresh Verification Requirement

NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE.

The Verification Gate (run for EVERY check):

1. IDENTIFY - Which command proves this claim?
2. RUN - Execute the command NOW (not from memory)
3. READ - Examine FULL output, count failures/errors
4. VERIFY - Does output match the claim?
5. THEN CLAIM - Only state success after steps 1-4

Invalid Evidence:

• Previous run output (even from 2 minutes ago)
• Assumptions based on code changes
• Partial verification ("linter passed so types probably pass")
• Agent/sub-agent reports without independent verification

Red Flag Language - HALT Immediately

If you find yourself using these words about verification:

Any hedging language = missing verification. Stop and run the actual check.

Quality Gate Checklist

DEFINITION OF DONE - Phase [X]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

[ ] Tests exist for new code
[ ] All tests pass (new AND existing)
[ ] Linter passes ([detected linter])
[ ] Formatter passes ([detected formatter])
[ ] Type checker passes ([detected checker])
[ ] No new warnings introduced
[ ] Plan document updated (checkboxes, status)  ← BEFORE commit message

Enforcement Rules

STOP means STOP. Do not proceed. Do not offer to fix automatically. Ask user to fix and re-run.

Failure Output Format

When any check fails, output:

⛔ DEFINITION OF DONE FAILED
━━━━━━━━━━━━━━━━━━━━━━━━━━━
Issue: [Missing tests | Tests failing | Linter errors | etc.]
[Details of what failed]
Please fix, then tell me to continue.
[EXECUTION HALTED]

STOP. Read ENFORCEMENT-GUIDE.md NOW when any check fails - contains failure templates and stack-specific commands.

Step 5: Update Plan Document (EVERY PHASE)

⚠️ MANDATORY: This step executes after EVERY phase, not just at the end.

After ALL quality gates pass, BEFORE generating commit message, UPDATE THE PLAN FILE:

1. Check off completed tasks (- [] → - [x])
2. Update phase status in overview table (⬜ → ✅)
3. Mark DoD items complete (- [] → - [x])

PLAN DOCUMENT UPDATED - Phase [X]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

File: [plan-file-path]

Updates applied:
- Tasks: X/X items checked [x]
- DoD: X/X items checked [x]
- Status: ⬜ → ✅

The plan now reflects Phase [X] completion.

Read PLAN-UPDATES.md for checkbox patterns and error handling.

WHY EVERY PHASE:

• Plan survives context compaction (conversation may not)
• Progress visible to anyone reading the plan
• Enables clean handoff between sessions
• Creates audit trail in git history

DO NOT SKIP THIS STEP. If you find yourself generating a commit message without updating the plan first, STOP and update the plan.

Step 6: Generate Conventional Commit

After Definition of Done passes, generate commit message.

CRITICAL: OUTPUT ONLY. NEVER run git commands.

Read COMMIT-FORMAT.md for CLI-safe characters and HEREDOC patterns. Read EXECUTION-CONTROL.md for parallel phase commit output format.

Step 7: Functional Testing Instructions

After commit message, explain how to functionally test the phase.

Provide step-by-step manual verification instructions. Offer integration test script only if applicable - ALWAYS ask, NEVER auto-create.

Read EXECUTION-CONTROL.md for output format and script offer template.

Step 8: Stop Execution

FULL STOP after each phase (unless --build-all override).

1. Output phase completion summary
2. Show progress table
3. Suggest context compaction
4. Wait for explicit "Continue to Phase X" instruction

Post-compaction behavior: Complete in-progress phase only, then STOP. Todo list is NOT authorization to continue.

Build-all override: Only if user explicitly requests. Warn about risks first.

Read EXECUTION-CONTROL.md for output templates and compaction behavior.

Step 9: Auto-Compact in BUILD-ALL Mode

CRITICAL: In BUILD-ALL mode, automatic context compaction is MANDATORY.

When running with --build-all flag, execute /compact automatically after each phase CHECKPOINT to preserve context for remaining phases.

Auto-Compact Trigger

Plans created by superplan include CHECKPOINT markers at the end of each phase:

- [ ] **CHECKPOINT: Run `/compact focus on: Phase N complete, [key artifacts], Phase N+1 goals`**

BUILD-ALL Auto-Compact Workflow

┌─────────────────────────────────────────────────────────────────────┐
│                    BUILD-ALL MODE: AUTO-COMPACT                      │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  1. COMPLETE PHASE   │  Execute all tasks, pass quality gates       │
│          ↓           │                                              │
│  2. UPDATE PLAN      │  Check off tasks, mark CHECKPOINT done       │
│          ↓           │                                              │
│  3. GENERATE COMMIT  │  Output conventional commit message          │
│          ↓           │                                              │
│  4. DETECT CHECKPOINT│  Find CHECKPOINT task in completed phase     │
│          ↓           │                                              │
│  5. EXTRACT FOCUS    │  Parse focus directive from CHECKPOINT       │
│          ↓           │                                              │
│  6. RUN /COMPACT     │  Execute: /compact focus on: [extracted]     │
│          ↓           │  ⚠️  NO USER INTERVENTION REQUIRED           │
│  7. CONTINUE         │  Proceed to next phase automatically         │
│                                                                     │
│  ════════════════════════════════════════════════════════════════   │
│  Steps 1-7 repeat for EACH PHASE until plan complete.               │
│  ════════════════════════════════════════════════════════════════   │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

CHECKPOINT Parsing

Extract the focus directive from the CHECKPOINT task:

Input:  - [ ] **CHECKPOINT: Run `/compact focus on: Phase 1 complete, auth models created, Phase 2 needs endpoints`**

Parsed: focus on: Phase 1 complete, auth models created, Phase 2 needs endpoints

Auto-Compact Execution

Execute /compact with the parsed focus directive:

AUTO-COMPACT TRIGGERED - Phase [X] Complete
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Running: /compact focus on: [parsed focus directive]

Preserving:
- Phase [X] completion status
- Key artifacts: [list from focus]
- Phase [X+1] goals: [from focus]

[compact executes]

Context refreshed. Continuing to Phase [X+1]...

No CHECKPOINT Found

If a phase lacks a CHECKPOINT (legacy plan or manual creation):

⚠️  NO CHECKPOINT FOUND - Phase [X]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

This phase has no CHECKPOINT marker for auto-compact.

Options:
1. Continue without compact (risk context exhaustion)
2. Pause for manual /compact
3. Add CHECKPOINT to plan and retry

Recommendation: Option 2 for safety

[EXECUTION PAUSED - Awaiting instruction]

Read EXECUTION-CONTROL.md for detailed auto-compact patterns.

Rationalizations to Reject

Red Flags - STOP Immediately

If you catch yourself thinking any of these, STOP:

• "This is taking too long, let me skip ahead"
• "The user seems impatient, let me batch phases"
• "Tests can come after the feature works"
• "Linting is just style, not critical"
• "I'll generate all commit messages at the end"
• "The plan doesn't explicitly require tests"
• "The sub-agent confirmed it passes"
• "I already verified this earlier"
• "This is obviously correct"
• "I wrote the code first, but I'll test it now"
• "The test passed on the first try"
• "This feature is too simple for TDD"
• "Let me just get it working, then add tests"

All of these = violation of superbuild protocol.

Quality Commands by Stack

Read ENFORCEMENT-GUIDE.md for stack-specific commands (JS/TS, Python, Go, Rust).

Summary: The Iron Rules

1. No plan = No execution - Exit if plan not provided
2. One phase at a time - Unless parallel phases (use sub-agents)
3. All quality gates must pass - No exceptions
4. Update plan after EVERY phase - Check off tasks, update status
5. Generate commit message - Never run git commands
6. Explain functional testing - Ask before writing scripts
7. Full stop after phase - Unless --build-all override
8. Suggest compact - Context management is critical
9. Fresh verification required - Run checks NOW, not from memory
10. Auto-compact in BUILD-ALL - Parse CHECKPOINT, run /compact, continue automatically

Superbuild is rigid by design. The enforcement protects code quality. Do not rationalize around it.