self-regulation-brake-system

name	self-regulation-brake-system
description	Brake system that auto-stops the agent after N repeated failures, prevents token burning on loops.
tags	["self-regulation","brake","efficiency","feedback","timeout"]
metadata	{"hermes":{"platform":["cursor","codex","claude-code"]},"triggers":["brake","fail limit","stop retry","regulation"]}
domain	agent-core
subdomain	safety
tokens	{"scan":210,"load":2340,"category":"detailed"}

Self-Regulation Brake System

Problem

Agent can burn thousands of tokens repeating the same failed action (GitHub login, OAuth, browser loops). The user only discovers this when they see the bill.

Solution: brake.sh

A lightweight shell script at ~/.hermes/scripts/brake.sh that:

• Tracks failure count per action name across sessions
• Exits with code 2 when limit (default: 3) is exceeded
• Resets on success

Integration Points

0. System-Level Injection（最強保證 — config.yaml prefill_messages_file）

Brake rules 已寫入 system instructions file，透過 config.yaml 嘅 prefill_messages_file 注入：

• System instructions file: ~/.hermes/system-instructions/self-regulation.md
• 呢個 file 每次新 session 開波會自動 inject 做 system instruction
• 因為係 config-level injection，Agent 無法跳過或忽略

1. session-start.sh hook

Already updated. Every new session shows brake status:

=== 🧠 SELF-REGULATION CONTEXT ===
⚠️  BRAKE SYSTEM IS ACTIVE

2. Agent memory

Injected as self-regulation rules. Ensures the agent knows about brakes even after context compaction.

3. Manual usage (agent self-enforces)

Whenever the agent encounters a failing loop:

# Before retrying, check if brake is triggered
bash ~/.hermes/scripts/brake.sh check "github-login-by-browser"
# exit 0 = safe, count = N
# exit 2 = STOP, ask user

# After a failure
bash ~/.hermes/scripts/brake.sh record "github-login-by-browser"

# After success
bash ~/.hermes/scripts/brake.sh reset "github-login-by-browser"

Brake Trigger Rules（hard-coded in memory，亦注入 system instructions）

Condition	Action
Same action fails ≥3 times	STOP → clarify user（用 clarify tool，唔好自己估）
5 min of no progress	STOP → clarify user
Third-party auth (GitHub/Google/OAuth)	First failure → STOP, don't retry
Complex multi-step auth	Delegate to subagent first
卡住咗唔出聲	違反 HARD RULE 2 — 失敗咗必須即刻報告，唔可以靜雞雞轉彎

🚩 CUA-Specific Auto-Stop Rules（2026-06-17 確立）

Terminal First 原則

做任何動作之前，先問「terminal 做唔做到」：

• terminal 做到嘅（開app、build、git、curl、cp/mv、改 code、compile）→ 用 terminal
• terminal 做唔到嘅（㩒GUI掣、睇畫面效果、OAuth login、系統對話框）→ 先用 CUA

CUA 紅旗自動停機制

CUA 操作見到以下任何一個紅旗，即停 CUA，轉用 terminal/其他方法：

紅旗	原因
同一 CUA 操作失敗 2 次	Pixel coord 唔準/AX tree 太複雜→唔會越試越準
Accessibility tree > 500 elements	DevTools/Chrome/Electron 內部→AX tree 大得滯，無用
Vision 座標 明顯唔合理	如話座標喺畫面範圍之外
畫面係 Canvas/WebGL 自繪	CUA 睇唔到入面啲掣

CUA 適用地區

CUA 只應該用喺以下場景：

• ✅ 睇畫面效果（vision screenshot）—— 純觀察，唔操作
• ✅ 㩒系統對話框（權限請求、save dialog）
• ✅ OAuth login（browser 登入流程）
• ✅ Drag & drop

實戰案例：DevTools 操作反面教材

❌ 失敗做法：
  用 CUA pixel click 操作 DevTools 內部 UI（頁面路徑selector）
  → AX tree 971 elements，pixel coord 錯晒，失敗 5+ 次
  → 浪費大量 token，0 成果

✅ 正確做法：
  改 code → terminal `bash boot.sh pre-commit` compile
  → 出錯直接改 code，唔使郁 DevTools 掣

點解要咁嚴？

因為 CUA 嘅 vision 同 AX tree 喺複雜 UI（DevTools、Electron app）入面非常唔可靠，而且每失敗一次就燒大量 token（screenshot + vision analysis + 再去試）。Terminal command 反而係 deterministic，成功失敗都 cheap 好多倍。

Architecture

Session Start
  ├─ session-start.sh
  │    └─ brake.sh status → context injection
  │
  Agent working
  ├─ action fails → brake.sh record
  ├─ before retry → brake.sh check
  │    ├─ count < 3 → retry
  │    └─ count ≥ 3 → BRAKE! → clarify user
  │
  Action succeeds → brake.sh reset

🧠 Why This Works（防止 Agent 忽視嘅設計原理）

呢個 system 嘅關鍵唔係 script 本身，而係 注入層級：

1. System-level（config.yaml prefill_messages_file） → 每次 session 開波 server 端自動注入，Agent 無法跳過
2. Session-level（session-start.sh hook） → 每次開工自動執行，輸出 HARD RULES 硬提示
3. Agent-level（memory + skills） → 長期記憶有記錄，context compaction 後仍然存在

點解 Agent 理論上可以「忽略」system instruction？

• 冇錯，某啲 LLM provider 嘅 system message 可以被 user message override
• 但 brake.sh 係一個實際存在嘅 script，唔係純文字指令
• agent 如果 brake.sh 卡咗 3 次但繼續做，係違反咗自己嘅 tool call 結果（exit code 2）
• 呢個係「validate by execution」嘅設計 — 規則唔係用講嘅，而係用 tool 嘅行為嚟強制執行

File Locations

• Script: ~/.hermes/scripts/brake.sh
• State dir: ~/.hermes/state/brake/ (persists across sessions)
• Hook: ~/.hermes/hooks/session-start.sh