Introduction
# OpenClaw Agent Optimization
Use this skill to tune an OpenClaw workspace for **cost-aware routing**, **parallel-first delegation**, and **lean context**.
## Quick Start (copy/paste)
1) **Full audit (safe, no changes):** > Audit my OpenClaw setup for cost, reliability, and context bloat. Output a prioritized plan with rollback notes. Do NOT apply changes.
2) **Context bloat / transcript noise:** > My OpenClaw context is bloating (slow replies / high cost / lots of transcript noise). Identify the top offenders (tools, crons, bootstrap files) and propose the smallest reversible fixes first. Do NOT apply changes.
3) **Model routing / delegation posture:** > Propose a model routing plan for (a) coding/engineering, (b) short notifications/reminders, (c) reasoning-heavy research/writing. Include an exact config patch + rollback plan, but do NOT apply changes.
## What you will get (output shape) - **Executive summary** (what matters + why) - **Top offenders / drivers** - Cost drivers - Context drivers - Reliability risks - **Options A/B/C** (tradeoffs made explicit) - **Recommended plan** (smallest change first) - **Exact change proposals** (patch snippets) + **rollback**
## Safety Contract (must follow) - Treat this skill as **advisory by default**, not autonomous control-plane mutation. - **Never** mutate persistent settings (e.g., `config.apply`, `config.patch`, `update.run`) without explicit user approval. - **Never** create/update/remove cron jobs without explicit user approval. - If an optimization reduces monitoring coverage, present options (A/B/C) and require the user to choose. - Before any approved persistent change, show: (1) exact change, (2) expected impact, (3) rollback plan.
## OpenClaw 2.9+ notes (skills + context) - Skills are snapshotted per session; if you install/update skills, start a **new session** (or wait for watcher refresh). - Prefer **short SKILL.md + references/** for long runbooks. Keep injected prompt text lean. - For risky / heavy skills, consider `disable-model-invocation: true` so they only run when explicitly invoked. - Gating matters: use `metadata.openclaw.requires` (bins/env/config) so skills don’t appear but fail at runtime. - Sandboxed runs don’t inherit host env; if a skill needs secrets in sandbox, set them via sandbox env config (not skill env).
## High-ROI optimization levers (typical wins)
### 1) Output discipline for automation Make maintenance loops **truly silent on success**: - Cron/heartbeat jobs should output exactly `NO_REPLY` unless something is wrong.
### 2) Separate “do the work” from “notify the human” If you want alerts but want the interactive session lean: - Send a short out-of-band alert (Telegram/Slack/etc.), then output `NO_REPLY`.
### 3) Prefer isolated agentTurn for autonomous background work If a job should execute *without* requiring attention, prefer: - `sessionTarget="isolated"` + `payload.kind="agentTurn"`
### 4) Hardening & guardrails - Use scripts-first for complex cron jobs (avoid fragile multi-line `bash -lc` quoting). - Add circuit breakers / global locks for heavy jobs.
### 5) Ops hygiene checklist - Snapshot backups: freshness threshold + retention + failure markers. - Heartbeat coverage: check model auth, disk/snapshot freshness, and **ClawHub CLI auth** (`npx clawhub whoami`) if you rely on publishing/installs.
## Workflow (concise) 1. **Audit rules + memory**: ensure rules are modular/short; memory keeps only restart-critical facts. 2. **Model routing**: confirm tiered routing (light / mid / deep) matches live config. 3. **Context discipline**: apply progressive disclosure; move large static data to references/scripts. - If transcripts are bloating, run `context-clean-up` (audit-only) to get a ranked offender list + plan. 4. **Delegation protocol**: parallelize independent tasks; use isolated sub-agents for long/noisy work. 5. **Heartbeat optimization (control-plane only)**: propose options A/B/C (coverage vs cost). 6. **Execution gate**: if user approves changes, apply the smallest viable change first, then verify and report.
## References - `references/optimization-playbook.md` - `references/model-selection.md` - `references/context-management.md` - `references/agent-orchestration.md` - `references/cron-optimization.md` - `references/heartbeat-optimization.md` - `references/memory-patterns.md` - `references/continuous-learning.md` - `references/safeguards.md`