build（构建）

Build - Context-Isolated Feature Development

When this skill is invoked, IMMEDIATELY output the banner below before doing anything else. Pick ONE tagline at random — vary your choice each time. CRITICAL: Reproduce the banner EXACTLY character-for-character. The first line of the art has 4 leading spaces — you MUST preserve them.

{tagline}

⠀   ██╗██████╗ ██╗   ██╗██╗██╗     ██████╗
   ██╔╝██╔══██╗██║   ██║██║██║     ██╔══██╗
  ██╔╝ ██████╔╝██║   ██║██║██║     ██║  ██║
 ██╔╝  ██╔══██╗██║   ██║██║██║     ██║  ██║
██╔╝   ██████╔╝╚██████╔╝██║███████╗██████╔╝
╚═╝    ╚═════╝  ╚═════╝ ╚═╝╚══════╝╚═════╝

Taglines:

• ⚙️ Compiling hopes and dreams...
• 🏗️ Bob the Builder has entered the chat!
• 🤖 Assembling the Voltron of code!
• 🏭 Feature factory: ONLINE
• ☕ Hold my coffee, I'm building!
• 📦 Some assembly required...
• 🧱 Bricks, mortar, and semicolons!
• 🏎️ Let's see what this baby can do!

Output Formatting

After the banner, display parsed input:

┌─ Input ────────────────────────────────────────
│  {Field}:  {value}
│  Flags:    {parsed flags or "none"}
└────────────────────────────────────────────────

Pre-flight results:

── Pre-flight ───────────────────────────────────
  ✅ {dep}           {version or "found"}
  ⚠️ {dep}           not found → {fallback detail}
  ❌ {dep}           missing → stopping
──────────────────────────────────────────────────

Stage/phase headers: ━━ {N} · {Name} ━━━━━━━━━━━━━━━━━━━━━━━━━

Status icons: ✅ done · ❌ failed · ⚠️ degraded · ⏳ working · ⏭️ skipped

Execute a detailed design/plan through the full feature-dev lifecycle with maximum context isolation. Every heavy stage runs in a subagent so the primary conversation only accumulates structured reports.

Stage prompts: references/stage-prompts.md Report budgets: references/report-contracts.md Agent selection: references/architecture-notes.md Project detection: references/project-detection.md

Flags

Parse optional flags from the request:

• --skip-questions: Skip Stage 2 (clarifying questions)
• --skip-review: Skip Stage 5 (code review)
• --no-ship: Stop after Stage 8 docs update
• --parallel-impl: Split implementation into parallel agents when independent

Pre-flight

Before starting, check all dependencies in this table:

For each row, in order:

1. Run the Check command (for cli/npm) or test file existence (for agent/skill)
2. If found: continue silently
3. If missing: apply Resolution strategy

- stop: notify user with Detail, halt execution - url: notify user with Detail (install link), halt execution - install: notify user, run the command in Detail, continue if successful - ask: notify user, offer to run command in Detail, continue either way (or halt if required) - fallback: notify user with Detail, continue with degraded behavior

1. After all checks: summarize what's available and what's degraded
2. Capture PLAN (the user's argument) and FLAGS
3. Load project context — invoke /prime to load domain-specific context (CLAUDE.md, specs, memory). If /prime is unavailable, fall back to manually scanning CLAUDE.md and specs/ directory.
4. Detect project type using references/project-detection.md to populate PROJECT_CONFIG (language, test_runner, test_setup)
5. Check for outstanding items from previous work:

- Query persistent tasks via TaskList for incomplete items - Search CLAUDE.md for a "Known Issues" or "Open Questions" section - Search memory (if available) for recent unresolved items

1. If outstanding items found, present via AskUserQuestion: "Found {count} outstanding items from previous work:" {numbered list with summary of each} "Address any of these before starting the build?" Options:

- "Yes, let me choose which ones" → present each; options: "Incorporate into this build" / "Skip for now" / "Explain more" Items marked "incorporate" get appended to the PLAN as additional requirements for Stage 1 to explore. - "No, proceed with the build" → continue normally

Plan Resolution

Before Stage 1, resolve the PLAN argument into content:

1. File detection — If the argument contains / or ends with .md, .yaml, .json, or .txt, treat it as a file path:

- Try reading the path as-is - If not found, try specs/{arg} - If found, use file content as PLAN - If not found at any location, treat the original argument as free-form text

1. Free-form text — If not a file path (or file not found), use the argument verbatim as PLAN
2. Display — In the Input box, show the resolved source:

- File: Plan: {file path} ({line count} lines) - Text: Plan: inline ({word count} words)

Stage 1: Explore

Launch feature-dev:code-explorer (fallback: general-purpose):

Task(
  subagent_type: "feature-dev:code-explorer",
  description: "Explore codebase for build plan",
  prompt: <read from references/stage-prompts.md#stage-1>
)

Substitute {PLAN} into the prompt. Parse EXPLORE_REPORT. Extract questions for Stage 2.

Stage 2: Clarifying Questions

Skip if --skip-questions or no questions found.

Runs on the primary thread (requires user interaction).

1. Review EXPLORE_REPORT questions and potential_issues
2. Present questions to user via AskUserQuestion
3. Store answers as CLARIFICATIONS

Stage 3: Architecture Design

Launch feature-dev:code-architect (fallback: general-purpose):

Task(
  subagent_type: "feature-dev:code-architect",
  description: "Design implementation architecture",
  prompt: <read from references/stage-prompts.md#stage-3>
)

Substitute {PLAN}, {EXPLORE_REPORT}, {CLARIFICATIONS}. Parse ARCH_REPORT. Present approach_summary to user for confirmation.

If rejected, incorporate feedback and re-run.

Stage 4: Implementation

If --parallel-impl and ARCH_REPORT has independent parallel_groups, launch multiple general-purpose subagents in parallel.

Otherwise launch one general-purpose subagent:

Task(
  subagent_type: "general-purpose",
  description: "Implement plan",
  prompt: <read from references/stage-prompts.md#stage-4>
)

Substitute {PLAN}, {ARCH_REPORT}, conventions, {PROJECT_CONFIG.test_runner}. Parse IMPL_REPORT(s). If any failed, assess retry or abort.

Stage 5: Code Review

Skip if --skip-review.

Launch 3 feature-dev:code-reviewer subagents in parallel (fallback: general-purpose):

1. Simplicity & DRY
2. Bugs & Correctness
3. Conventions & Integration

Prompts in references/stage-prompts.md#stage-5.

Consolidate reports. Present only critical and high severity findings. Ask: "Fix these now, or proceed as-is?"

Stage 6: Fix Review Findings

Only if Stage 5 found issues AND user wants them fixed.

Launch general-purpose subagent:

Task(
  subagent_type: "general-purpose",
  description: "Fix review findings",
  prompt: <read from references/stage-prompts.md#stage-6>
)

Stage 7: Verify

Launch Bash subagent (haiku):

Task(
  subagent_type: "Bash",
  model: "haiku",
  description: "Run verification tests",
  prompt: <read from references/stage-prompts.md#stage-7>
)

Substitute {PROJECT_CONFIG.test_runner} and {PROJECT_CONFIG.test_setup}.

If tests fail:

• First failure: launch general-purpose agent to fix, retry
• Second failure: report to user and stop

Stage 8: Update Progress Documentation

Skip if EXPLORE_REPORT has no source_docs.

Launch general-purpose subagent:

Task(
  subagent_type: "general-purpose",
  description: "Update progress documentation",
  prompt: <read from references/stage-prompts.md#stage-8>
)

If --no-ship: Stop here and present final summary.

Stage 9: Ship

Invoke the /ship skill:

/ship {approach_summary from ARCH_REPORT}. Files: {files from IMPL_REPORT}

Stage 10: Debrief

Always runs on the primary thread (requires user interaction).

1. Scan all stage reports for unresolved items:

- EXPLORE_REPORT: potential_issues not addressed by implementation - ARCH_REPORT: risks with deferred mitigations - REVIEW_REPORT: medium/low findings not fixed in Stage 6 - TEST_REPORT: warnings, skipped tests, flaky results - DOCS_REPORT: docs_skipped items - IMPL_REPORT: issues_encountered that were worked around

1. Compile into DEBRIEF_ITEMS (see references/stage-prompts.md#stage-10). Categorise each as: unresolved_risk, deferred_fix, open_question, assumption, or tech_debt.
2. If no items found, skip to Final Report.
3. Present numbered summary via AskUserQuestion grouped by category. Each item shows: [category] summary (effort). Options:

- "Fix now" → create a task list of resolution activities for each item; present for user confirmation, then work through them - "Create tasks for future sessions" → use TaskCreate for each item as a persistent task, with category as prefix and suggested action as description - "Note and continue" → acknowledge items without formal tracking; log to memory (if exists) or as source file comments. No further action. - "Let me choose per item" → present each individually with full description, evidence, and impact. Options per item: "Fix now" / "Create task" / "Explain more" / "Note and continue". "Explain more" reads source files cited in evidence, provides expanded context, then re-presents the item for decision.

1. After resolution, include debrief summary in the Final Report.

Final Report

┌─ Build · Report ───────────────────────────────
│
│  ✅ Build complete
│
│  📋 Plan:      {first line of plan}
│  🏗️ Approach:  {approach_summary}
│
│  📝 Changes
│     Files modified:  {count}
│     Files created:   {count}
│     Tests:           {passed}/{total} ✅
│     Docs updated:    {count or "none"}
│
│  🔍 Review
│     Findings addressed: {count fixed} / {count found}
│
│  📊 Debrief: {count resolved} / {count surfaced}
│     {list of created tasks}
│
│  🔗 Links
│     PR:  {pr_url}
│     CI:  {merge_commit}
│
│  ⚡ Next steps
│     {debrief items or "none — all clear"}
│
└─────────────────────────────────────────────────

If any stage failed, report the failure point and what was accomplished.

Rollback

If implementation succeeds but later stages fail:

• Tests fail: fix agent attempts repair, then reports to user
• Review critical: user decides fix or proceed
• Ship fails: code is still committed locally; user can manually push
• Never silently revert completed implementation work