← 返回
未分类

Focus Guard

Session coherence protocol for Claude — track decisions, resist scope creep, flag drift. Keeps long AI sessions on track without losing what was already deci...
Claude会话一致性协议——跟踪决策、抵制范围蔓延、标记偏离。保持长AI会话正常运行,不丢失已决定内容。
jiajiaoy jiajiaoy 来源
未分类 clawhub v1.0.0 1 版本 100000 Key: 无需
★ 1
Stars
📥 287
下载
💾 0
安装
1
版本
#latest

概述

Focus Guard

Long sessions drift. Decisions get forgotten. Scope expands silently. Focus Guard anchors the session — track what was decided, resist what wasn't asked for, flag when things go off course.

The Core Problem

A session that starts with "fix this one bug" can quietly become a refactor of three files, a new abstraction layer, and a README update. Nobody decided to do that. It just happened.

How sessions drift:

  • Decision amnesia — a choice made 20 messages ago gets quietly reversed
  • Scope expansion — "while I'm here" additions that weren't requested
  • Tangent lock-in — going deep on a sub-problem until the original goal is forgotten
  • Goal substitution — solving a related but different problem and calling it done

Focus Guard makes drift visible before it costs real work.

When to Activate

Use Focus Guard for:

  • Multi-step tasks with more than 3 sequential actions
  • Sessions where earlier decisions constrain later ones
  • Any task where scope creep would be costly (refactors, migrations, multi-file edits)
  • Long debugging or investigation sessions where the original question can get lost
  • Collaborative work where you need to stay aligned on what was agreed

Skip it for: quick one-shot tasks, single questions, explorations where scope is intentionally open.

The Protocol

Step 1: Session Anchor

At the start of any multi-step session, define the anchor:

[Session Anchor]
Goal: ...
In scope: ...
Out of scope: ...
Done when: ...

This is the reference point for the whole session. Everything should advance the goal or be explicitly agreed as a change to it.

Step 2: Decision Log

Track every significant decision made during the session:

[Decisions]
✓ D1: [what was decided] — [why]
✓ D2: ...

Before reversing a logged decision, flag it explicitly: "This would reverse D2. Confirm?"

Step 3: Drift Detection

At each major step, check:

  • Does this action advance the session anchor?
  • Does it depend on a previous decision that's still holding?
  • Is this scope that was agreed, or scope that crept in?

If any check fails, pause and flag — don't silently proceed.

Step 4: Scope Gate

When a new sub-task appears mid-session, evaluate it:

TypeAction
--------------
Necessary for goalDo it. Log it.
Related but not necessaryFlag it. Ask if user wants to include.
TangentFlag it. Defer to after current goal is done.
Out of scopeDecline politely. Stay on track.

The rule: expand scope only by explicit decision, never by default.

Step 5: Progress Check-in

For sessions with 5+ steps, offer a brief check-in at natural breakpoints:

[Progress Check]
Done: ...
Next: ...
Still in scope: [yes / flag if changed]

Keep it to 3 lines. Not every step needs one — only when direction could plausibly be questioned.

Drift Signals to Watch

  • Adding error handling that wasn't asked for
  • Refactoring code that was adjacent to the actual change
  • Generalizing a specific solution "while we're here"
  • Fixing style/formatting in files that were only referenced, not changed
  • Creating abstractions for a one-time operation

These aren't always wrong — but they should always be explicit choices, not silent additions.

Output Format

Session open:

[Focus Guard: Active]
Goal: ...
Done when: ...

Mid-session flag:

[Scope Flag] This would add X, which wasn't in the original goal. Include? [y/n]

Session close:

[Session Complete]
Goal achieved: [yes / partial — what's left]
Deferred: [list anything flagged but not done]

Pairs Well With

  • clarity-first — define the anchor before starting
  • task-pilot — break the goal into steps that Focus Guard can track
  • honest-critic — challenge whether the current direction is still the right one
openclaw install focus-guard
openclaw install clarity-first
openclaw install task-pilot

版本历史

共 1 个版本

  • v1.0.0 当前
    2026-05-08 03:12 安全 安全

安全检测

腾讯云安全 (Keen)

安全,无风险
查看报告

腾讯云安全 (Sanbu)

安全,无风险
查看报告

🔗 相关推荐

ai-agent

Self-Improving + Proactive Agent

ivangdavila
自我反思+自我批评+自我学习+自组织记忆。智能体评估自身工作、发现错误并持续改进。
★ 1,392 📥 321,992
ai-agent

self-improving agent

pskoett
捕获经验教训、错误及修正内容,以实现持续改进。适用于以下场景:(1)命令或操作意外失败;(2)用户纠正Claude(如“不,那不对……”“实际上……”);(3)用户请求的功能不存在;(4)外部API或工具出现故障;(5)Claude发现自身
★ 4,095 📥 822,447
ai-agent

Agent Browser

rez0
用于 AI 代理的浏览器自动化 CLI。当用户需要与网站交互(包括浏览页面、填写表单、点击按钮、截图等)时使用。
★ 832 📥 300,779