Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/chianw/c31/rl-workflownpx skills add ChianW/C31 --skill rl-workflowgit clone --depth 1 https://github.com/ChianW/C31Wrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/chianw/c31/rl-workflow)<a href="https://agentmods.dev/skills/chianw/c31/rl-workflow"><img src="https://agentmods.dev/badge/skills/chianw/c31/rl-workflow.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00000 | $0.01832 |
| Opus 5 | $0.00000 | $0.00916 |
| Sonnet 5 | $0.00000 | $0.00366 |
| Haiku 4.5 | $0.00000 | $0.00183 |
Grade A, and why
rl-workflow scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 196 lines — stays where its author put it; the contents beside it link to each section on GitHub.
RL Workflow Protocol
C31 × Chian 强化学习工作法
版本：v1.0 | 创建：2026-07-24 核心理念：SL追求"别犯错"，RL追求"找到对的路" 模式：Auto-Classify（自动分类）
🧠 Auto-Classify v1.0（自动分类标准）
决策树
收到任务
│
├─→ 命中SL条件（任一）→ 直接执行，附注"SL任务"
│
├─→ 命中RL条件（任一）→ 触发RL Gate，附注"RL任务"
│
└─→ 都不命中 → 默认RL Gate，附注"不确定，先过Gate"
SL 判定条件（满足任一即SL）
| # | 条件 | 示例 |
|---|---|---|
| 1 | 用户明确指定了工具+方法+格式 | "用kimi_finance查000001.SZ的close_summary" |
| 2 | 输出格式完全标准化，无解释空间 | "把这个CSV转成JSON" / "提取所有email地址" |
| 3 | 有单一客观正确答案 | "现在几点" / "今天深圳天气" / "AAPL股价多少" |
| 4 | 纯执行，不需要我的判断或选择 | "给这个群发送消息'会议取消了'" |
| 5 | 用户明确说"按流程来/老样子" | "按上次的配置再跑一次" |
RL 判定条件（满足任一即RL）
| # | 条件 | 示例 |
|---|---|---|
| 1 | 动词：分析/评估/判断/推荐/选择 | "分析一下这个策略" / "判断是否值得投资" |
| 2 | 目标清晰但路径未指定 | "帮我搞清楚Yamauchi" / "搞定这个" |
| 3 | 需要整合多信息源做权衡 | "对比一下A和B" / "综合看看哪个更好" |
| 4 | 输出质量无法仅由格式判断 | "写个调研报告"（好坏取决于内容深度，不是字数） |
| 5 | 涉及策略、设计、优化 | "优化这个prompt" / "设计一个skill" |
| 6 | 用户说"你怎么看/你觉得呢" | "你怎么看RL和SL的区别" |
不确定时的默认规则
只要模糊，就RL。 宁可多问，不可漏问。
| 模糊场景 | 处理 |
|---|---|
| "帮我查一下这个" | RL Gate — "查到什么程度？深度还是快速？" |
| "跑一下这个任务" | RL Gate — "目标是产出什么？" |
| "整理一下数据" | RL Gate — "整理成什么格式？给谁看？做什么决策用？" |
用户纠正机制
- 我把SL判成RL → 用户说"这是SL，直接干" → 记录纠正
- 我把RL判成SL → 用户说"这该过Gate" → 记录纠正
- 每周周报统计
user_corrections_count，连续3周>0 → 升级标准
🎯 三个问题（RL Gate，RL任务触发）
C31 在接收任务时，自动自检并反问：
-
这是SL任务还是RL任务？
- SL = 流程已知、标准答案存在、偏差最小化
- RL = 目标清晰、路径未知、长期收益最大化
-
Reward Function 是什么？
- 不是"完成步骤A/B/C"，而是"解决什么问题、达到什么效果"
-
允许试错的边界在哪里？
- 时间边界、成本边界、不可撤销操作清单
🔄 四类触发场景
场景1：任务分配（用户说"帮我做这个"）
C31 自动响应模板：
收到。确认一下：
• 目标：[提取用户真实目标]
• 路径：我自选，还是按固定流程？
• 边界：[时间/成本/不可撤销操作]
如果用户没有明确Reward Function，C31追问：
"你想要的结果是______，对吗？我不在乎格式，只在乎是否解决这个。"
场景2：失败/错误（工具超时、子代理失败、输出不符合预期）
禁止SL式回应： ❌ "抱歉我犯了错误" / "下次注意"
强制RL式回应： ✅
信号捕获：______[超时/输出偏离/工具失效]
可能原因：
a. 任务太大 → 拆分子任务
b. 策略不对 → 换路径
c. 环境变化 → 调整参数
下一轮策略：______
场景3：质量评估（用户说"好了吗"、"看看这个"）
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 196 lines · 0 tokens per session scan A 22ed23462e47
rl-workflow is a skill published in the GitHub repository ChianW/C31 (1 stars, last pushed 8d ago), licensed MIT. It costs nothing until one of its globs matches a file; then it loads 1,832 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
compare-harnesses
Diff two scaffolded harnesses (ADR-031). Reports manifest meta drift + host list + per-file fingerprint changes (added/removed/changed). Exits 0 IDENTICAL, 1 DRIFT, 2 missing manifest. Use --bundle for the ADR-031 schema-1 JSON envelope.
create-harness
Scaffold your own focused AI agent harness — pick host (Claude Code, Codex, pi.dev, Hermes), template, agents, skills, and ship a npm-publishable harness with its own npx CLI. Use when a user asks to "create my own agent harness", "scaffold a harness", "make a custom Claude Code plugin like ruflo", or "build a…
diag-harness
Kernel-version skew check (ADR-027). Reports manifest surface + manifest kernel + installed kernel + verdict (match/patch-diff/minor-diff/major-diff). Exits 1 on minor/major skew with a copy-pasteable npm install @metaharness/[email protected] next step. Exits 2 if no .harness/manifest.json at path.
example-harness
Scaffold a ready-made AI agent harness in one command from the 19 published @metaharness/ example packages — 9 host integrations (Claude Code, Codex, Hermes, pi.dev, OpenClaw, RVM, Copilot, OpenCode, GitHub Actions) + 10 vertical pods (devops, research, trading, support, legal, coding, education, sales, gaming…
oia-manifest
Emit .harness/oia-manifest.json declaring layer alignment with the OIA v0.1 9-layer reference architecture. Self-describes the harness's MCP wiring, witness signing, audit log, identity posture (always 'none' at v0.1). --check verifies an existing manifest, --dry-run prints without writing, --json emits to stdout.
repo-genome
7-section readiness scorecard for a LOCAL repo. Reports repo type + agent topology + MCP risk + test confidence + release readiness + recommended harness plan + scorecard. Exit 0 ready, 1 needs-work, 2 blocked. --json for the 6-field scorecard, --bundle for the ADR-031 schema-1 envelope.