baseline-health

name	baseline-health
description	Use Baseline's local MCP tools to set up, run, inspect, accept, and schedule coding-agent health checks in Codex.

Baseline Health

Use this skill when a Codex workspace needs a local coding-agent health check, Good Baseline acceptance, drift comparison, or daily self-check through Baseline.

Requirements

• The baseline CLI must be installed and available on PATH.
• Install the CLI with curl -fsSL https://trackbaseline.com/install.sh | sh or from the GitHub Release assets before calling MCP tools.
• Verify baseline --version prints baseline 0.1.0, then run baseline doctor for read-only preflight before starting an eval.
• Baseline is local-first. Raw prompts, responses, local paths, and secrets must not be copied into cloud systems unless the operator explicitly enables redacted sync.

Workflow

1. Start with baseline --version and baseline doctor if the CLI has not been verified in this environment. baseline doctor is read-only and does not send probe messages.
2. Start the first run with baseline_setup or CLI baseline setup. This configures the local workspace, writes Baseline-owned local state, and starts the default target eval in the background when needed.
3. For later checks, call baseline_run. It returns quickly with run_status.run_id while the eval continues.
4. Poll baseline_report with that run id until the report is completed.
5. Show the operator the report and responses before asking for an accept/reject/defer decision.
6. Accept a Good Baseline only after explicit operator approval: baseline_accept with confirm: "accept <RUN_ID>", or CLI baseline accept <RUN_ID> --confirm "accept <RUN_ID>".
7. Use baseline_doctor only for read-only preflight. It must not send agent probes or create a Good Baseline candidate.
8. Use baseline_scrub_preview before enabling sync or sharing any text externally.

Daily Self-Check

1. Call baseline_schedule with action: "status" to inspect the configured schedule.
2. If the operator asks to install a daily check, call baseline_schedule with action: "install" and an at value such as 09:00.
3. If the operator asks to trigger the scheduled check now, call baseline_schedule with action: "run".

Recovery

• If MCP startup fails because baseline is missing, install the CLI or point Codex at a built binary, then re-run baseline --version. Do not add a new MCP tool for version or preflight; the advertised local tool count stays seven.
• If a lifecycle run is still running, keep polling baseline_report.
• If a lifecycle run failed before writing a result row, inspect the stdout and stderr paths in the report, then ask before rerunning with baseline_run using rerun_id or CLI baseline rerun <RUN_ID>.
• If OpenClaw logs show idleMs=60007, timeoutMs=60000, or turn_completion_idle_timeout, report that as the OpenClaw Codex idle watchdog and run setup before starting a fresh eval.
• If 401 Unauthorized appears with an __OPENCLAW_REDACTED__ placeholder, treat it as an auth/env configuration problem, not a timeout.