spec-code-pipeline

Spec Code Pipeline

Use this skill when the user asks for a complete development workflow, "开发工作流", "完整管线", or explicitly names:

Spec -> Matrix -> Tasks -> Plan -> Prompt -> Code

The workflow target is one module under module/{module}/. It turns an approved module spec into traceability, tasks, implementation plan, task prompt, and one bounded code execution.

Invocation

$spec-code-pipeline {module}
$spec-code-pipeline {module} --from matrix
$spec-code-pipeline {module} --stage prompt

Chinese trigger equivalents:

完整工作流 {module}
开发工作流 {module}
Spec 到 Code {module}

Canonical Flow

Spec -> Matrix -> Tasks -> Plan -> Prompt -> Code

Expanded repo gate sequence (每个 executor 后都接 四源团队评分 + 仲裁)：

spec
  -> [spec-team-score: claude + codex + copilot + rules-engine]
  -> pipeline-arbiter  (gate: composite_score = min(claude.score, codex.score, copilot.score, rules.score) >= 98 且无红线、无 LLM 低置信度、LLM 分差与 rules 异构分歧在阈值内)
  -> 自动翻转 Status: Approved
  -> matrix
  -> [matrix-team-score] -> pipeline-arbiter (gate)
  -> task-split
  -> [tasks-team-score] -> pipeline-arbiter (gate)
  -> task-planner
  -> [plan-team-score] -> pipeline-arbiter (gate)
  -> prompt-builder
  -> [prompt-team-score] -> pipeline-arbiter (gate)
  -> task-executor
  -> [code-team-score] -> pipeline-arbiter (gate)

每个阶段必须满足（唯一门禁）：

• composite_score = min(claude.score, codex.score, copilot.score, rules.score)
• composite_score >= 98
• 无任一评分源红线
• 无任一 LLM 平台低置信度，且三 LLM 平台分差不超过阈值
• rules.score 与三 LLM 中位数的异构分歧不超过阈值

Confidence 与分差属于 gate 条件，不能被忽略或人工豁免。失败一律自动路由回当前阶段 executor 或 scorer 修复；同阶段 3 次失败自动升级到上一阶段；上游再失败继续向上直到 spec。修复循环必须有界：默认 max_stage_attempts = 3，max_total_gate_failures = 18。达到全链路上限后输出 pipeline_blocked 和 retrospective，不得无限循环。

Workflow self-improvement is allowed only as bounded RSI: create docs/governance/improvements/{YYYYMMDD}-{slug}/SPEC.md, run the same Spec -> Matrix -> Tasks -> Plan -> Prompt -> Code gates, and obey CONSTITUTION.md §14 for protected files.

详见 docs/governance/STRUCTURAL-SCORING.md 与 docs/governance/scoring/ARBITER-PROTOCOL.md。

Stage Contract

Arbiter agent：pipeline-arbiter（三运行时均有等价实现，结果等价）。 状态目录由运行时决定：

• Claude：.omc/state/pipeline/{module}/{stage}/{scores/,verdict.json,attempts.json}
• Codex：.omx/state/pipeline/{module}/{stage}/{scores/,verdict.json,attempts.json}
• Copilot：.copilot/state/pipeline/{module}/{stage}/{scores/,verdict.json,attempts.json}

每个阶段在三 LLM 评分后、调用 arbiter 前，必须运行规则引擎评分：

python3 scripts/rule-scorer.py <stage> <module> --runtime <claude|codex|copilot>
# 默认读取 SPEC_PIPELINE_RUNTIME；未设置时写入 .omc/state/pipeline/<module>/<stage>/scores/rules.json

rules.json 是宪法 §14.4 要求的异构第 4 源；arbiter 期望 4 个 scores/*.json 文件齐全。

Execution Rules

1. Work one module at a time.
2. Do not code from a Draft spec.
3. Every stage must pass team-scoring gate (composite_score = min(claude.score, codex.score, copilot.score, rules.score) >= 98, no redline, no LLM low confidence, bounded LLM score spread, heterogeneous divergence <= 15) before advancing. This is the only gate.
4. Every stage uses an agent team: executor + 3 LLM scorers + rule-scorer.py + pipeline-arbiter.
5. Do not skip Matrix or Tasks before Plan.
6. Do not ask task-executor to implement more than one task unless the user explicitly requests a batch.
7. When using --stage, verify upstream artifacts AND upstream verdicts (verdict.json gate=pass) exist first.
8. When using --from, resume from that stage; downstream verdict.json and scores/ must be re-generated.
9. Update TRACEABILITY.md and task status after each code task.
10. Never bypass scoring gate by editing verdict.json manually. There is no human override path.
11. Enforce bounded RSI: after 3 failed attempts in one stage, escalate upstream; after 18 total gate failures, write pipeline_blocked plus PIPELINE-RETROSPECTIVE.md and stop automatic advancement.
12. Do not modify protected workflow/rubric/agent/arbiter files through normal repair routing. Open a workflow-improvement spec and apply CONSTITUTION.md §14.

Repair Routing

• Missing or incomplete spec structure: route to spec.
• Structural redline at any stage: route to that stage's executor.
• Broken FR/AC/TC chain: route to matrix.
• Overlarge or cross-module task: route to task-split.
• Ambiguous execution order: route to task-planner.
• Missing implementation context: route to prompt-builder.
• Code/test failure: keep scope in current task and route to task-executor.

State

Per-stage state under:

{state_root}/pipeline/{module}/{stage}/
├── scores/
│   ├── claude.json   ├── claude.md
│   ├── codex.json    ├── codex.md
│   ├── copilot.json  ├── copilot.md
│   └── rules.json
├── verdict.json      # arbiter output (gate=pass|fail)
└── attempts.json     # retry counter & escalation chain

Top-level pipeline state:

{state_root}/pipeline/{module}.json
{state_root}/pipeline/{module}/repair-budget.json
{state_root}/pipeline/{module}/pipeline_blocked.json
module/{module}/PIPELINE-RETROSPECTIVE.md

Final Report

End with:

• current stage and completed stages
• created or updated artifacts
• active task, if code started
• validation evidence
• blockers or remaining gates