context-clean-up

使用时机：提示语境过于臃肿（回复缓慢、成本上升、文字记录嘈杂），并且您需要一份按顺序排列的罪犯名单 + 可逆计划。 不要在以下情况下使用：您想要自动删除或无人值守的配置编辑。 输出：仅审计报告（最严重的问题 + 3-8 个风险最低的修复 + 回滚注释）。不会自动应用任何更改。

## 概述（中文）

使用时机：提示语境过于臃肿（回复缓慢、成本上升、文字记录嘈杂），并且您需要一份按顺序排列的罪犯名单 + 可逆计划。
不要在以下情况下使用：您想要自动删除或无人值守的配置编辑。
输出：仅审计报告（最严重的问题 + 3-8 个风险最低的修复 + 回滚注释）。不会自动应用任何更改。

## 原文

# Context Clean Up (audit-only)

This skill identifies what is bloating prompt context and turns it into a **safe, reversible plan**.

## Contract

- **Audit-only by default.**
- No automatic deletions.
- No unattended config edits.
- No silent cron/session pruning.
- If you ask for changes, the skill should propose:
  1. exact change,
  2. expected impact,
  3. rollback plan,
  4. verification steps.

## Safety model

- No `exec` tool usage.
- No `read` tool usage.
- If you want file-level analysis, run the bundled script manually and paste the JSON.

## Quick start

- `/context-clean-up` → audit + actionable plan (no changes)

Optional manual report generation:

```text
python3 scripts/context_cleanup_audit.py --out context-cleanup-audit.json
```

Windows variant:

```text
py -3 scripts/context_cleanup_audit.py --out context-cleanup-audit.json
```

## What to measure (authoritative, not vibes)

When available, prefer **fresh-session `/context json` receipts** over subjective claims like “it feels leaner”.

High-signal fields:
- `eligible skills`
- `skills.promptChars`
- `projectContextChars`
- `systemPrompt.chars`
- `promptTokens`

If exact receipts are unavailable, fall back to ranked offenders + change scope, but label confidence lower.

## Common offender classes

1. **Tool result dumps**
   - oversized `exec` output
   - large `read` output
   - long `web_fetch` payloads

2. **Automation transcript noise**
   - cron jobs that say “OK” every run
   - heartbeat messages that are not alert-only

3. **Bootstrap reinjection bloat**
   - overgrown `AGENTS.md` / `MEMORY.md` / `SOUL.md` / `USER.md`
   - long runbooks embedded directly in `SKILL.md`

4. **Ambient specialist surface**
   - too many always-visible specialist skills that should be on-demand workers/subagents instead

5. **Summary accretion**
   - repeated summaries that keep historical detail instead of restart-critical facts only

## Recommended trim ladder (lowest-risk first)

### Phase 1 — Noise discipline
- Make no-op automation truly silent (`NO_REPLY` or nothing on success).
- Keep alerts out-of-band when possible.

### Phase 2 — Bootstrap slimming
- Keep always-injected files short.
- Move long guidance to `references/`, `memory/`, or external notes.

### Phase 3 — Ambient surface reduction
- Remove low-frequency specialist skills from always-on prompt surface.
- Prefer worker/subagent invocation for specialist flows.

### Phase 4 — Higher-risk changes
- Tool-surface or deeper runtime/config narrowing.
- Only propose with stronger rollback and explicit approval.

## Workflow (audit → plan)

### Step 0 — Determine scope
You need:
- workspace dir
- state dir (`<OPENCLAW_STATE_DIR>`)

Common defaults:
- macOS/Linux: `~/.openclaw`
- Windows: `%USERPROFILE%\.openclaw`

### Step 1 — Run the audit script

```text
python3 scripts/context_cleanup_audit.py --workspace . --state-dir <OPENCLAW_STATE_DIR> --out context-cleanup-audit.json
```

Interpretation cheatsheet:
- huge tool outputs → transcript bloat
- many cron/system lines → automation bloat
- large bootstrap docs → reinjection bloat

### Step 2 — Produce a fix plan
Include:
- top offenders
- lowest-risk fixes first
- expected impact
- rollback notes
- verification plan

### Step 3 — Verify
After changes:
- confirm automation is silent on success
- check context growth flattens
- if possible, compare fresh-session `/context json` before/after

## Important caveat

Many OpenClaw runtimes snapshot skills/bootstrap per session.
So skill/config slimming often **does not fully apply to the current session**.
Use a **new session** for authoritative verification.

## References

- `references/out-of-band-delivery.md`
- `references/cron-noise-checklist.md`