OpenClaw Reliability

OpenClaw Reliability is a reliability smoke-check and guarded repair plugin for OpenClaw. It diagnoses runtime, provider, tool, plugin, channel, and configuration noise, then turns those findings into human-readable reports, watch decisions, and safe repair flows.

中文说明见 README.zh-CN.md.

What it is for

Use this plugin when you want to know whether an OpenClaw runtime problem is truly a runtime failure or just provider noise, model-call pressure, caller input error, stale tool state, channel expiry, or approval governance.

It helps answer:

• Is the Gateway actually degraded, or is the signal short-window sampling noise?
• Are failures coming from providers, credentials, HTTP endpoints, tool inputs, plugin config, or channels?
• Should the user be notified, or should the signal stay quiet as known noise?
• Is there a safe read-only diagnostic action to run?
• Is a state-changing repair blocked until explicit human approval?

Philosophy

Reliability follows a conservative operating rule:

Diagnose first. Repair only what is safe, narrow, verified, and reversible.

The plugin separates:

1. Smoke checks — collect local, redacted health evidence.
2. Reports — summarize blockers, warnings, and next action.
3. Watch decisions — de-duplicate signals and stay quiet for non-actionable noise.
4. Repair plans — create advisory actions with evidence and anti-actions.
5. Watch repair — run only safe non-mutating actions automatically; keep state-changing actions guarded.
6. Cron runner — emit NO_REPLY or a ready-to-send Chinese alert for schedulers.

It is not a security boundary. Use openclaw-language-boundary for action policy.

What it can do

• Run passive reliability smoke checks.
• Sample Gateway health and sustained CPU.
• Classify provider timeout, HTTP error, missing config, tool input errors, approval outcomes, plugin runtime noise, source-only shadowing, channel expiry, and stale tool failures.
• Generate short human-readable reports.
• Generate JSON output for automation.
• De-duplicate watch notifications.
• Run safe read-only diagnostic repair actions.
• Run one guarded state-changing cleanup for stale tool-failure records only when explicitly allowlisted.
• Emit NO_REPLY for quiet scheduled checks.

What it will not do by default

• It will not restart Gateway.
• It will not patch config.
• It will not write credentials.
• It will not disable plugins.
• It will not delete files.
• It will not mutate cron jobs, channels, sessions, or external services.
• It will not upload logs.

Install

openclaw plugins install clawhub:openclaw-reliability@0.3.14
openclaw gateway restart

For the latest ClawHub release:

openclaw plugins install clawhub:openclaw-reliability
openclaw gateway restart

Common commands

From the installed package or source directory:

npm run smoke
npm run smoke:json
npm run report
npm run repair-plan
npm run repair-plan:json
npm run watch-check
npm run watch-check:json
npm run watch-repair
npm run watch-repair:json
npm run cron-runner
npm run cron-runner:json
npm run release:check

Useful examples:

npm run smoke -- --cpu-sample-seconds=15 --failure-freshness-minutes=60
npm run report -- --cpu-sample-seconds=15 --failure-freshness-minutes=60
npm run watch-repair:json -- --cpu-sample-seconds=5 --failure-freshness-minutes=60
npm run cron-runner -- --cpu-sample-seconds=5 --failure-freshness-minutes=60

Cron runner

cron-runner is designed for scheduled checks.

• If there is no actionable signal, it prints exactly:

NO_REPLY

• If a human should be notified, it prints a concise Chinese alert.

This reduces the scheduler's need to parse JSON or summarize with a model. The scheduler can simply run the command and forward stdout when it is not NO_REPLY.

Safe repair model

watch-repair automatically runs only actions whose repair plan says:

• wouldChangeSystem=false
• requiresConfirmation=false

State-changing actions require explicit allowlists such as:

npm run watch-repair -- --allow-state-change=tool_failure_stale

The currently implemented guarded state-changing action is cleanup-stale-tool-failures. It only removes stale non-healthy tool records, creates a backup, reruns smoke, and restores the backup if the post-check gets worse.

Relationship with Language Boundary

• openclaw-language-boundary protects action safety boundaries.
• openclaw-reliability diagnoses runtime reliability and guarded repair signals.

Use both together for a safer OpenClaw runtime: one controls action policy, the other watches operational health.

Current release

• Latest verified release: openclaw-reliability@0.3.14
• ClawHub release ID: rd7a6g0tky94f0sbfrxa27n6e18895e7
• Status at final verification: published, scanned clean, isolated install verified.