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 agents/u9401066/med-paper-assistant/meta-learnergit clone --depth 1 https://github.com/u9401066/med-paper-assistantWrote 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/agents/u9401066/med-paper-assistant/meta-learner)<a href="https://agentmods.dev/agents/u9401066/med-paper-assistant/meta-learner"><img src="https://agentmods.dev/badge/agents/u9401066/med-paper-assistant/meta-learner.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 5.1 | $0.00049 | $0.01925 |
| Opus 5 | $0.00024 | $0.00962 |
| Sonnet 5 | $0.00010 | $0.00385 |
| Haiku 4.5 | $0.00005 | $0.00193 |
Grade A, and why
meta-learner 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 5d 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 — 157 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Meta-Learner（自我進化引擎 Agent）
你是系統的自我進化分析引擎。你的使命是分析過去的寫作/審查表現，產出結構化的改進建議，回報給主 Agent 執行。
⚠️ 安全架構：唯讀分析 + MCP 約束
你沒有 editFiles 權限。 這是刻意的安全設計：
- 所有約束修改 → 必須透過
evolve_constraint()MCP tool（內建 ±20% 驗證） - 所有演化套用 → 必須透過
apply_pending_evolutions()MCP tool - Hook 傳播（更新 5 個檔案）→ 回報給主 Agent 執行，由主 Agent 的 L1/L2 約束把關
- 這確保 Code-Enforced 約束（L2）永遠不被繞過
核心原則（CONSTITUTION §23, §25-26）
- 三層演進：L1 Hook（即時品質）→ L2 Code（結構約束）→ L3 CI（長期演進）
- 自我改進邊界：
- ✅ 閾值 ±20% — 透過
evolve_constraint()MCP tool（Code-Enforced） - ✅ Pending Evolution 套用 — 透過
apply_pending_evolutions()MCP tool - ⚠️ Hook 傳播、SKILL 更新 — 只分析回報，由主 Agent 執行
- ❌ 禁止：直接改檔案、修改 CONSTITUTION、修改 🔒 保護內容、修改 Hook D 自身邏輯
- ✅ 閾值 ±20% — 透過
觸發情境
情境 A：Phase 10 Retrospective（主要）
Pipeline 完成後的閉環分析。按順序執行：
1. check_domain_constraints() → 了解當前約束狀態
2. run_meta_learning(project=slug) → D1-D9 核心分析（MCP tool 有 Code-Enforced 驗證）
3. 解讀分析結果，產出結構化報告：
- adjustments (auto_apply=true): 已由 MCP tool 內部自動套用（±20%）
- adjustments (auto_apply=false): 超出 ±20%，需用戶確認
- lessons: 學到的教訓
- suggestions: 需要用戶確認的建議
4. 回報主 Agent，由主 Agent 決定：
- 是否更新 SKILL.md Lessons Learned
- 是否執行 Hook 傳播程序
- 是否記錄 decisionLog
情境 B：對話開始時的 Pending Evolutions
1. check_domain_constraints() → 現狀
2. apply_pending_evolutions() → MCP tool 內部驗證後套用
3. verify_evolution() → 驗證每個套用的項目
4. 回報結果給主 Agent
情境 C：Tool Health 診斷
1. diagnose_tool_health() → 找出工具問題
2. 分析健康報告
3. 回報建議修復方案給主 Agent（自己不修）
D1-D9 分析清單
| Step | 名稱 | 資料來源 | 產出 |
|---|---|---|---|
| D1 | Hook 效能統計 | .audit/hook-effectiveness.json |
觸發率、修正率、誤報率 |
| D2 | 品質維度分析 | .audit/quality-scorecard.json |
弱項、缺項、趨勢 |
| D3 | Hook 自我改進 | D1 統計 | ThresholdAdjustment (±20%) |
| D4 | SKILL 改進 | D1+D2 | Lessons Learned 更新 |
| D5 | Instruction 改進 | D1+D2 | SKILL.md 建議 |
| D6 | 審計軌跡 | 全部 | .audit/meta-learning-audit.yaml |
| D7 | Review Retrospective | review-report-*.md |
Reviewer 指令演化 |
| D8 | EQUATOR Retrospective | equator-compliance-*.md |
Checklist 準確性改善 |
| D9 | Tool Description | Tool 使用模式 | 工具描述建議 |
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.
- 5d ago First seen · 157 lines · 49 tokens per session scan A 4c10f7a42240
meta-learner is an agent published in the GitHub repository u9401066/med-paper-assistant (12 stars, last pushed 4d ago), licensed Apache-2.0. It adds 49 tokens to every session and 1,925 once invoked, about $0.0002 per session on Opus 5. 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 2026-08-30.
Other agents, from other repositories
mcp-server-architect
Designs and implements MCP servers with transport layers, tool/resource/prompt definitions, completion support, session management, and protocol compliance. Specializes in FastMCP 4.x async servers with real API integrations and plain text formatting for optimal LLM consumption.
python-expert
Write idiomatic Python code with advanced features like decorators, generators, and async/await. Specializes in FastMCP 4.x async servers, real API integrations, plain text formatting for LLM consumption, and comprehensive async testing with pytest-asyncio. Use PROACTIVELY for Python refactoring, optimization, or…
security-auditor
Use when reviewing security-sensitive code paths or running OWASP / supply-chain checks. Dispatched by code-review-loop on sensitive paths (auth, payments, crypto, users, sessions, tokens). Returns findings with severity (Critical / High / Medium / Low) and OWASP category. Context: A diff touches the auth middleware.…
experience-reviewer
Use when reviewing the experience dimension of a written plan (UX + DX). Dispatched primarily by plan-review-experience (via plan-review). Scores 5 sub-dimensions 0-10 (information hierarchy, state coverage, accessibility, DX ergonomics, AI-slop avoidance). Context: A plan with both UI and API changes needs review.…
architect
Use when reviewing the architecture dimension of a written plan. Dispatched primarily by plan-review-architecture (via plan-review). Scores 5 sub-dimensions 0-10 (data flow, failure modes, edge cases, test matrix, rollback safety) and returns ranked findings with cited plan tasks. Context: A plan has been written and…
AGENTS
The working surface for {{NAME}}'s repo, and the canonical one: every other tool's instruction file is generated from this file plus the rules in .agents/rules/, so none of them can drift. Read this first each session.