技能验证门

其他 社区

技能解读

当前原文的自动摘译,尚未人工审校。

大门

在声称任何成功或表示满意之前:识别——什么命令证明了这一说法? RUN — 执行完整命令(新鲜的,未缓存) READ — 完整输出,检查退出代码,计数失败 VERIFY — 输出是否确实确认了声明?

什么才算是证据

声明 · 要求 · 不充分 测试通过 · 测试命令输出显示 0 次失败 · 上次运行,“应该通过” 构建成功 · 构建命令退出 0 · Linter 通过 Bug 修复 · 重现原始症状:现在通过 · “代码已更改,应该可以工作” 回归测试有效 · 红色(未修复失败)→ 绿色(修复后通过) · 测试通过一次

危险信号 — 停止并验证

如果您发现自己在思考这些问题,请停止:思考·该怎么做“现在应该可以”·进行验证“我有信心”·信心≠证据“就这一次”·没有例外

多提供商环境

在 Claude Octopus 工作流程中,验证尤其重要,因为: 提供者输出可能会产生幻觉 — Codex/Gemini/Copilot 可能会在没有证据的情况下宣称成功 共识 ≠ 正确性 — 三个模型达成一致并不意味着它们是正确的 合成文件可能已过时 — 检查时间戳,不要假设新鲜度 Orchestrate.sh 退出代码 0 ≠ 质量 — 脚本已运行,但它是否产生了良好的输出?

何时申请

始终之前: 提交代码 创建 PR 标记任务完成 进入下一个工作流程阶段

流狐档案 运行与权限由流狐静态检查推断,不代表作者声明或实际测试结果。
领域
通用
作者声明的智能体
未找到明确声明;不据此推断已兼容或已测试
静态检查
88 / 100 · 启发式扫描,不代表运行安全
运行环境证据
未执行;当前没有隔离运行测试证据
作者 / 版本 / 许可证
@nyldn · 未声明许可证
上下文用量评级
低消耗
接入复杂程度
需简单配置
是否需要外部接口密钥
不需要
兼容的系统
macOS · Linux · Windows
运行环境要求
未检出明确要求
检测到的文件与系统行为
  • 只读
  • 允许写入 / 修改
  • 命令执行
网络行为
仅限本地
安装命令数
无(仅作为资料)

需要注意: 未声明工具权限范围;实际权限由运行环境决定。

作者说明文件(原文) SKILL.md

> **Host: Codex CLI** — This skill was designed for Claude Code and adapted for Codex.
> Cross-reference commands use installed skill names in Codex rather than `/octo:*` slash commands.
> Use the active Codex shell and subagent tools. Do not claim a provider, model, or host subagent is available until the current session exposes it.
> For host tool equivalents, see `skills/blocks/codex-host-adapter.md`.


# Verification Gate

## The Iron Law

```
NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE
```

If you haven't run the verification command in this turn, you cannot claim it passes.

## The Gate

Before claiming any success or expressing satisfaction:

1. **IDENTIFY** — What command proves this claim?
2. **RUN** — Execute the full command (fresh, not cached)
3. **READ** — Full output, check exit code, count failures
4. **VERIFY** — Does output actually confirm the claim?
5. **ONLY THEN** — State the claim WITH evidence

Skip any step = the claim is unverified.

## What Counts as Evidence

| Claim | Requires | NOT Sufficient |
|-------|----------|----------------|
| Tests pass | Test command output showing 0 failures | Previous run, "should pass" |
| Build succeeds | Build command exit 0 | Linter passing |
| Bug fixed | Reproduce original symptom: now passes | "Code changed, should work" |
| Regression test works | Red (fail without fix) → Green (pass with fix) | Test passes once |
| Subagent completed task | `git diff` shows expected changes | Subagent says "done" |
| Requirements met | Line-by-line checklist against spec | Tests passing |
| Provider dispatch worked | Output contains expected content | No error ≠ success |

## Red Flags — STOP and Verify

If you catch yourself thinking any of these, STOP:

| Thought | What to do instead |
|---------|-------------------|
| "Should work now" | Run the verification |
| "I'm confident" | Confidence ≠ evidence |
| "Just this once" | No exceptions |
| "The linter passed" | Linter ≠ tests ≠ build |
| "The agent said it worked" | Verify independently |
| "It's a small change" | Small changes cause big bugs |

## Multi-Provider Context

In Claude Octopus workflows, verification is especially critical because:

- **Provider outputs can be hallucinated** — Codex/Gemini/Copilot may claim success without evidence
- **Consensus ≠ correctness** — three models agreeing doesn't mean they're right
- **Synthesis files may be stale** — check timestamps, don't assume freshness
- **orchestrate.sh exit code 0 ≠ quality** — the script ran, but did it produce good output?

After any multi-provider workflow:
```bash
# Verify synthesis file exists and is recent
ls -la ~/.claude-octopus/results/*-synthesis-*.md | tail -1

# Verify it has content (not just headers)
wc -l ~/.claude-octopus/results/*-synthesis-*.md | tail -1
```

## When to Apply

**ALWAYS before:**
- Committing code
- Creating PRs
- Marking tasks complete
- Moving to next workflow phase
- Reporting results to user
- Claiming a bug is fixed

**In orchestrate.sh workflows:**
- After `probe` (discover) — verify synthesis file exists
- After `grasp` (define) — verify consensus score meets threshold
- After `tangle` (develop) — verify tests pass, not just that code was written
- After `ink` (deliver) — verify review actually ran, not just that it was dispatched

## Examples

### Correct: Evidence-Based Claim
```
$ npm test
  ✓ user.create() saves to database (45ms)
  ✓ user.create() validates email (12ms)
  Tests: 2 passed, 2 total

All 2 tests pass. ← Claim backed by output.
```

### Incorrect: Claim Without Evidence
```
I've implemented the feature. It should work now. The tests should pass.
← No test was run. "Should" is not evidence.
```

### Correct: Regression Test Red-Green
```
1. Write test → run → FAIL (expected, proves test detects the bug)
2. Implement fix → run → PASS (proves fix works)
3. Revert fix → run → FAIL (proves test isn't false-positive)
4. Restore fix → run → PASS (final confirmation)
```

## Integration with Other Skills

This skill is referenced by:
- `flow-develop.md` — verification gate after implementation
- `flow-deliver.md` — verification gate before delivery
- `skill-code-review.md` — verify review findings before reporting
- `skill-tdd.md` — red-green cycle requires evidence at each step
- `skill-factory.md` — autonomous pipeline must verify at every phase

讨论

基于 GitHub Discussions。登录 GitHub 即可参与讨论、点赞、订阅更新。