Back to skills
SKILL.md
Claude Code Harness
ASecurityUse when setting up a structured Plan→Work→Review→Release cycle for autonomous agents, integrating Claude Code with Codex/Cursor/OpenCode, enforcing spec-first development gates, or protecting against agent self-modification. Triggers on: 'claude-code-harness', 'harness cycle', 'agent release cycle', 'spec gate', 'plan work review release', 'harness-plan', 'harness-work', 'harness-review', 'harness-release', 'codex integration', 'cursor integration', 'opencode integration', 'multi-tool agent ...
- 10 stars
- 0 votes
- 0 copies
- 3 views
- Added September 9, 2026
Works with
Security analysis
100/100npx -y skills add yanacuti1121/Yana-AI --skill claude-code-harness --agent claude-codeAre you the author of Claude Code Harness?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/yanacuti1121-claude-code-harness)---
name: claude-code-harness
description: "Use when setting up a structured Plan→Work→Review→Release cycle for autonomous agents, integrating Claude Code with Codex/Cursor/OpenCode, enforcing spec-first development gates, or protecting against agent self-modification. Triggers on: 'claude-code-harness', 'harness cycle', 'agent release cycle', 'spec gate', 'plan work review release', 'harness-plan', 'harness-work', 'harness-review', 'harness-release', 'codex integration', 'cursor integration', 'opencode integration', 'multi-tool agent workflow', 'autonomous agent cycle', 'agent tự động release'."
---
# Claude Code Harness Skill
# Source: chachamaru127/claude-code-harness (MIT) — v4.15.0, 2.6k stars
# Tier: TIER 3 — PRODUCTIVITY
Framework biến agents thành autonomous delivery pipeline.
Plan→Work→Review→Release cycle tự chạy — human chỉ approve spec một lần.
---
## Khi nào dùng
- Muốn agents tự plan + implement + review + release mà không cần trigger từng bước
- Tích hợp nhiều AI tool (Claude Code + Codex + Cursor) trong cùng repo
- Cần verification gate độc lập (agent không tự review code mình viết)
- Bảo vệ harness rules khỏi bị agent tự sửa (v4.15.0 self-modification protection)
---
## 5 Commands cốt lõi
```
/harness-setup — khởi tạo spec.md + Plans.md, cấu hình tool tier
/harness-plan — draft spec từ intent → human approve contract
/harness-work — implement tasks từ approved spec (TDD)
/harness-review — independent review, không phải agent tự review
/harness-release — package evidence, prepare deployment
```
**Quan trọng:** `/harness-work` bị block nếu spec.md chưa được approve.
Human chỉ cần approve 1 lần → agents tự chạy phần còn lại.
---
## Tool Capability Matrix
| Capability | Claude Code | Codex CLI | OpenCode | Cursor |
|---|---|---|---|---|
| skill_loading | ✅ Full | ✅ Full | ◐ Partial | ◐ Partial |
| pre_use_guard | ✅ Full | ◐ Partial | ✗ No | ◐ Partial |
| post_use_gate | ✅ Full | ✅ Full | ◐ Partial | ✅ Full |
| memory_bridge | ✅ Full | ◐ Partial | ✗ Future | ✗ Future |
| review_artifact | ✅ Full | ✅ Full | ◐ Partial | ✅ Full |
> "The same capability name ≠ the same enforcement strength" — mỗi tool implement khác nhau.
---
## Cursor Integration (Planning ↔ Implementation)
Cursor làm PM side, Claude Code làm implementation side — cùng repo, không conflict.
```markdown
# Plans.md — shared source of truth
## Task: Add auth middleware
pm:依頼中 / cc:TODO ← Cursor tạo task
cc:WIP ← Claude Code đang làm
cc:完了 ← Claude Code xong
pm:確認済 ← Cursor approve
```
**Workflow:**
1. Cursor: draft `Plans.md` với acceptance criteria
2. Claude Code: `/harness-work` → implement → update marker `cc:完了`
3. Cursor: review + `/harness-release handoff`
**Rule:** Không để Cursor và Claude Code edit cùng 1 task block cùng lúc.
---
## Codex CLI Integration
```bash
# Codex-specific skills tách riêng khỏi shared skills
skills/ # shared — Claude Code + Codex đều dùng
skills-codex/ # Codex-only — instruction surface khác
# Sync sau khi edit skill
bash scripts/sync-skill-mirrors.sh
```
Codex không có `pre_use_guard` đầy đủ — safety dựa vào Go runtime guardrails thay vì hooks.
---
## OpenCode Integration
OpenCode tier: **internal-compatible** (partial support).
```yaml
# opencode.json config
{
"harness": {
"spec_gate": true,
"pre_use_guard": false, # OpenCode không support
"fallback": "go-runtime" # dùng Go guardrails thay thế
}
}
```
`memory_bridge` chưa có → facts không persist giữa sessions với OpenCode.
---
## Self-Modification Protection (v4.15.0)
```
Agent KHÔNG được sửa:
skills/ — skill definitions
harness-rules/ — gate logic
Plans.md — task contracts (chỉ update markers)
Nếu agent cố sửa → Go runtime block + log SELF_MOD_ATTEMPT
```
Tương đương `49-immutable-infrastructure-law.md` của YAMTAM nhưng apply cho harness layer.
---
## Anti-Fake-Pass Checks
```
❌ FAIL nếu /harness-work chạy trước khi spec.md được approve
❌ FAIL nếu agent tự review code mình vừa viết (phải dùng /harness-review độc lập)
❌ FAIL nếu Codex/OpenCode dùng feature đánh dấu ✗ trong capability matrix
✅ PASS khi: spec approved → work complete → review pass → release artifact generated
```
## See also
- `executing-plans` — plan execution đơn giản không có release cycle
- `spec-planner` / `spec-executor` — YAMTAM native planning pipeline
- `autonomous-agent-harness` — agent action space design
Attribution
Comments
Loading comments…
Newsletter
Get the weekly list of safe new skills, and the ones we pulled.