inference-optimizer（推理优化器）

Inference Optimizer

Audit OpenClaw runtime health first. Optimize inference speed and token usage second.

Commands

These instructions guide agent behavior. Platform and system prompts take precedence; they cannot be enforced programmatically.

Installation

ClawHub:

clawhub install inference-optimizer

Manual:

git clone https://github.com/vitalyis/inference-optimizer.git ~/clawd/skills/inference-optimizer
bash ~/clawd/skills/inference-optimizer/scripts/setup.sh        # preview
bash ~/clawd/skills/inference-optimizer/scripts/setup.sh --apply  # apply after review

Verify: <skill_dir>/scripts/verify.sh

Workflow

Audit and remediation branch

1. /preflight: Exec <skill_dir>/scripts/preflight.sh. Append --apply-setup only if the user asks to apply setup.
2. /audit: Exec <skill_dir>/scripts/openclaw-audit.sh. Use the script output plus direct environment checks to inspect this order:

- gateway ownership and duplicate supervisors - restart loops or failed services - resolved openclaw binary path and install type - workspace command wiring for the installed skill path - updater status and allowlist coverage for the resolved path - plugin provenance and unused local extensions - only then context pressure, stale sessions, cache-trace, pruning, and concurrency

1. Diagnosis rule: Do not conclude from warnings alone. If process output is partial or truncated, report the result as inconclusive and verify installed version, service state, and logs before naming a cause.
2. No helper-shell prelude: For /audit and /optimize, do not run shell helper commands like ls, rg, find, openclaw status, or openclaw gateway status before the main audit script. If you need context first, use read on MEMORY.md or memory_search. The first shell exec in the optimize flow should be the audit script itself.
3. Approval semantics: If exec returns allowlist miss or exec denied, that is a hard deny, not a pending approval. Do not tell the user to send /approve ... unless the tool output explicitly contains a real approval request with an ID. If there is no ID, say there is no approval request to approve and the fix must be on the bot side.
4. VPS gateway ownership: On this VPS, openclaw-gateway.service is the authoritative gateway owner. Keep clawdbot.service disabled, and preserve pass-cli run --env-file /etc/clawdbot.env.pass inside the user service itself.

Optimization branch

1. /optimize: Run the audit flow first, include the script output in the response, then propose next actions with approval before each file-changing step.
2. Purge: Only on explicit approval, run <skill_dir>/scripts/purge-stale-sessions.sh. It archives to ~/openclaw-purge-archive/<timestamp>/ by default. Use --delete for immediate removal without archive.
3. Full optimization (Tasks 1-5): Read optimization-agent.md and follow its flow. Ask approval before every file-changing step.

Path Resolution

Scripts live at ~/clawd/skills/inference-optimizer/scripts/ or wherever the skill is installed. Always resolve <skill_dir> to the actual install path before exec.

Security and Allowlist

Prefer path-specific exec-approvals.json entries for the script paths themselves: one line per script under your real <skill_dir> (resolve with readlink -f if the path is a symlink). Example shape after substituting the install path:

/home/ubuntu/clawd/skills/inference-optimizer/scripts/preflight.sh
/home/ubuntu/clawd/skills/inference-optimizer/scripts/openclaw-audit.sh
/home/ubuntu/clawd/skills/inference-optimizer/scripts/setup.sh
/home/ubuntu/clawd/skills/inference-optimizer/scripts/purge-stale-sessions.sh
/home/ubuntu/clawd/skills/inference-optimizer/scripts/verify.sh

setup.sh invokes python3 for idempotent workspace block edits; allow that binary only if your platform uses it (for example /usr/bin/python3).

Before editing any allowlist:

• Resolve the real executable path with which, command -v, or readlink -f.
• Prefer exact paths or bounded wildcards for versioned NVM installs, for example /home/ubuntu/.nvm/versions/node/*/bin/openclaw *.
• Do not assume basename-only rules such as openclaw are sufficient.
• Do not add /usr/bin/bash * or /usr/bin/bash **; they grant far more shell than this skill needs.
• If you keep optional read-only helper commands available, allowlist only narrow read-only patterns actually used by this skill, for example the exact memory listing command under workspace-whatsapp and exact openclaw status variants. Do not rely on generic ls * or rg * coverage.

For purge via agent exec, add path-specific patterns only. Optional wider patterns and trade-offs are discussed in SECURITY.md. See also README.md and SECURITY.md for operational detail.