Execution Harness

Agent 执行可靠性体系的导航入口。当用户询问 agent 为什么停了、怎么防 agent 提前退出、execution harness 是什么、40 个 pattern 有哪些、6 轴架构、hook 怎么配时匹配。不执行具体操作——各子 skill 分别处理。

Security Scan

The skill is internally coherent: it's a navigation/hub for a collection of execution-harness sub-skills and ships a set of bash hooks and tests; nothing in the bundle demands unrelated credentials or network exfiltration, but you should review the included scripts before enabling hooks because they read/write files and can run commands on your host.

Purpose & Capability

The declared purpose is a 'navigation entry' for execution reliability patterns and explicitly says it 'does not execute operations' itself. The package contains many bash hook scripts and tests and the SKILL.md gives example settings that call those scripts (e.g., bash execution-loop/scripts/ralph-stop-hook.sh). This is explainable (the hub points users to sub-skills and shipped scripts), but there's a mild mismatch between the 'not execute' wording and the fact that the repo includes runnable scripts and example configs that will execute them if the user wires them into ~/.claude/settings.json. Review the scripts before enabling.

Instruction Scope

Runtime instructions and examples direct agents (or operator's config) to run local bash scripts and to read/write session/transcript files under the user's home (e.g., ~/.openclaw/shared-context/sessions, .working-state, ~/.claude/settings.json). The scripts operate on local filesystem artifacts, tmux panes, run linters/tests, and manipulate state (git stash, write handoff files). They do not reference external endpoints or ask for secrets in the metadata, but they do have broad local-file access — expected for this purpose, but something to audit before enabling.

Install Mechanism

No install spec is provided (instruction-only metadata), which minimizes automatic remote code installation. The bundle includes scripts and tests but no automated download-from-untrusted-URL step. The user would need to place/clone these files and add hooks manually or via their agent config; that manual step is the security control point.

Credentials

The skill declares no required environment variables, credentials, or config paths. The included scripts expect runtime-provided variables (e.g., TRANSCRIPT_PATH, NC_SESSION, COMPACTION_EXTRACT_INTERVAL) typical for hooks; they do not request API keys or tokens. No unrelated service credentials are demanded.

Persistence & Privilege

The skill is not always-enabled and does not request elevated platform privileges, but the SKILL.md explicitly recommends editing ~/.claude/settings.json to add hooks that will run the shipped scripts. Enabling those hooks gives the scripts persistent operational presence (they'll run on hook events) and lets them write into user home directories (sessions, handoffs, working-state). This is expected for a hooks-based harness, but it increases the blast radius — inspect scripts and back up configs before enabling.

Assessment

This bundle is a coherent 'execution harness' and primarily operates on local files and hooks. Before installing or enabling the hooks, do the following: 1) Inspect the included bash scripts (scripts/*) to ensure they do only what you expect — pay attention to any 'git', 'rm', or tmux commands and to where files are written (HOME paths). 2) Backup your ~/.claude/settings.json and any repositories the harness might touch. 3) Run the scripts in a sandbox or container (or under a test HOME) to observe behavior. 4) Start by enabling only non-destructive hooks (e.g., read-only token-budget or context-usage) and review outputs before enabling hooks that run git stash, checkpoint/rollback, or send input to tmux. 5) If you run the test suite (pytest) do so in an isolated environment. 6) Only enable hooks if you trust the source or have audited the code; although there are no external endpoints or requested credentials in the metadata, scripts with filesystem and tmux access can still disrupt local workflows if misconfigured.

Execution Harness — 导航入口

40 patterns x 6 轴，让 Claude Code agent 把活干完。不是框架，不做模型调用——只管 hook 和脚本。

这个 skill 不执行操作。它帮你找到该用哪个子 skill。

你遇到什么问题？

问题	用哪个 skill	关键 pattern
Agent 做了一半就停了	execution-loop	1.1 Ralph Stop Hook — 阻止提前退出，5 个安全阀
"应该可以"但没跑测试	execution-loop	1.2 Doubt Gate — 检测投机语言，要求验证
cargo build 重试 12 次	tool-governance	2.1 Tool Error Escalation — 3 次提示、5 次 block
rm -rf 毁了未提交代码	tool-governance	2.3 Checkpoint + Rollback — 自动 git stash
压缩后忘了设计决策	context-memory	3.1 Handoff Documents — 决策写磁盘
Context 快满了还在读大文件	context-memory	3.4 Token Budget — 80%+ 禁止直读
限速后 tmux 挂死	error-recovery	5.1 Rate Limit Recovery — cron 扫描恢复
5 个 agent 编辑同一文件	multi-agent	4.3 File Claim and Lock — 10min TTL 排他锁
提交了编译不过的代码	quality-verification	6.4 Test-Before-Commit — commit 前跑测试

6 轴速查

轴	Skill	Pattern 数	核心能力
1	execution-loop	7	Ralph Stop Hook、Doubt Gate、Drift Re-anchoring
2	tool-governance	6	错误升级、权限否决、破坏性命令拦截
3	context-memory	8	Handoff 文档、Compaction 抢救、Token Budget
4	multi-agent	6	Coordinator/Fork/Swarm、文件锁、盲审分离
5	error-recovery	7	限速恢复、Crash 恢复、模型降级建议
6	quality-verification	6	编辑后 lint、commit 前测试、session 指标

最小配置

3 个 hook 解决最常见的 3 个问题：

// ~/.claude/settings.json
{
  "hooks": {
    "Stop": [{"hooks": [
      {"type": "command", "command": "bash execution-loop/scripts/ralph-stop-hook.sh"},
      {"type": "command", "command": "bash execution-loop/scripts/doubt-gate.sh"}
    ]}],
    "PreToolUse": [{"hooks": [
      {"type": "command", "command": "bash tool-governance/scripts/tool-error-advisor.sh"}
    ]}],
    "PostToolUse": [{"matcher": {"tool_name": "Write|Edit|MultiEdit"}, "hooks": [
      {"type": "command", "command": "bash quality-verification/scripts/post-edit-check.sh"}
    ]}]
  }
}

仓库

GitHub: lanyasheng/execution-harness | 90 tests | 依赖：bash、jq、python3、pytest