OpenClaw Model Router Skill

Deterministic model routing with a closed-loop production switch pipeline.

Why use this skill

• Explicit user intent routing (@codex, @mini)
• Time-window model policy (router.schedule.json)
• Safe switch flow with verification and rollback
• Concurrency lock to prevent race conditions
• Structured audit logs for operations and incident review

Core capabilities

1) Prefix routing

• @codex → openai-codex/gpt-5.3-codex
• @mini → minimax/MiniMax-M2.5
• Aliases: @c, @m

2) Scheduler routing

• Rule-based windows by day/time/priority
• Timezone-aware resolution (router.schedule.json.timezone)
• Supports overnight windows (e.g. 18:00 → 09:00)

3) Production switch loop

• schedule apply / schedule end
• Auth gate (auth.requiredEnv[])
• Switch + readback verify
• Failure classification (auth_expired, rate_limit, provider_drift, unknown)
• Optional rollback (safety.rollbackOnFailure)
• Lock file (safety.lockPath) to avoid concurrent switching
• Audit logs (router.log.jsonl + rotation)

Quick commands

# Validate config
node src/cli.js validate

# Route inspection
node src/cli.js route "@codex implement this" --json

# Scheduler
node src/cli.js schedule validate
node src/cli.js schedule resolve --at "2026-03-02T10:00:00+01:00" --json
node src/cli.js schedule apply --json
node src/cli.js schedule end --id workday_codex --json

Key config files

• router.config.json (prefix map, retry, auth, safety, logging, controller)
• router.schedule.json (timezone + rules)

Reliability checklist

• Deterministic mapping
• Idempotent switching behavior
• Execute only after switch verification
• Observable success/failure events
• Recoverability on failure paths

Tests

node --test