Context Clean Up

Context Clean Up

Runbook-style workflow to audit and (optionally) apply safe fixes that keep OpenClaw sessions lean.

Principle: the fastest way to lose to context overflow is letting recurring automation (cron/heartbeat/reporting) write long outputs back into the same interactive session transcript.

Quick start

• User command:
  • /context-clean-up → audit + actionable plan (no changes)
  • /context-clean-up apply → apply low-risk changes (with backups / reversible patches)

Workflow (audit → plan → apply)

Step 0 — Determine scope

1. Identify the OpenClaw workspace dir (usually current directory).
2. Identify the OpenClaw state dir (usually ~/.openclaw).

If unsure, run:

bash -lc 'echo "$HOME" && ls -ld ~/.openclaw'

Step 1 — Audit what is actually bloating context

Run the bundled audit script (short stdout; writes detailed JSON to file):

bash -lc 'cd "${WORKDIR:-.}" && python3 context-clean-up/scripts/context_cleanup_audit.py --out memory/context-cleanup-audit.json'

If the repo is not in the current workdir, adapt the path accordingly.

Interpretation:

• If you see huge entries under toolResult (exec/read/web_fetch): those are transcript bloat.
• If you see repeated System: Cron: lines: that is automation bloat.
• If workspace bootstrap docs are huge: that is reinjected rules bloat.

Step 2 — Plan fixes (batch, lowest-risk first)

Create a short plan with:

• Top offenders (largest N transcript entries)
• Noisiest cron jobs (frequent + non-empty output)
• Quick wins (reversible)

Use these standard levers:

Lever A — Make no-op cron jobs truly silent

Goal: cron jobs that do maintenance should output exactly NO_REPLY.

Heuristic:

• If a cron job is deliver=false, it should never output long text.
• If a cron job is a “heartbeat” or “harvester” and has no anomalies, it should output NO_REPLY.

Implementation pattern: update the job prompt to end with:

• Finally output ONLY: NO_REPLY

Lever B — Keep scheduled reports, but avoid transcript injection

If the user wants notifications but you still want a lean interactive session:

• Prefer out-of-band delivery from the isolated worker:
  1. send a message (Telegram/Slack/etc.) using the platform tool
  2. output NO_REPLY

This keeps the main session transcript cleaner while the user still receives the report.

Lever C — Keep workspace bootstrap context small and stable

If the injected bootstrap docs are large:

• move “rarely-needed” notes into memory/*.md or references/*.md
• keep only restart-critical rules in MEMORY.md
• keep persona files short (SOUL.md, USER.md, etc.)

Always create .bak.<date> backups before edits.

Step 3 — Apply (only when user asked for apply)

If the user ran /context-clean-up apply:

1. Patch noisy cron jobs (safe edits only):

• Convert success/no-op outputs to NO_REPLY
• Leave user-facing reports alone unless the user explicitly agrees

2. (Optional) Propose a bootstrap docs compaction PR:

• Only do this with explicit confirmation because it edits the user’s rules/persona.

Step 4 — Verify

• Confirm the next cron run no longer injects Cron: ok / Cron: HEARTBEAT_OK noise.
• Watch for compaction events in the session (context ratio should drop).

Notes / best-practice hints

• Telegram auto-delete helps your chat app, but OpenClaw still has its own local session logs; auto-delete alone usually does not shrink the model prompt.
• For long-running agents, pair this with a memory layer (e.g., openclaw-mem) so you can retrieve state on demand instead of dragging the full transcript forward.

References

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