Intelligent Router — Core Skill

CORE SKILL: This skill is infrastructure, not guidance. Installation = enforcement. Run bash skills/intelligent-router/install.sh to activate.

What It Does

Automatically classifies any task into a tier (SIMPLE/MEDIUM/COMPLEX/REASONING/CRITICAL) and recommends the cheapest model that can handle it well.

The problem it solves: Without routing, every cron job and sub-agent defaults to Sonnet (expensive). With routing, monitoring tasks use free local models, saving 80-95% on cost.

MANDATORY Protocol (enforced via AGENTS.md)

Before spawning any sub-agent:

python3 skills/intelligent-router/scripts/router.py classify "task description"

Before creating any cron job:

python3 skills/intelligent-router/scripts/spawn_helper.py "task description"
# Outputs the exact model ID and payload snippet to use

To validate a cron payload has model set:

python3 skills/intelligent-router/scripts/spawn_helper.py --validate '{"kind":"agentTurn","message":"..."}'

❌ VIOLATION (never do this):

# Cron job without model = Sonnet default = expensive waste
{"kind": "agentTurn", "message": "check server..."}  # ← WRONG

✅ CORRECT:

# Always specify model from router recommendation
{"kind": "agentTurn", "message": "check server...", "model": "ollama/glm-4.7-flash"}

Tier System

SIMPLE fallback chain: anthropic-proxy-4/glm-4.7 → nvidia-nim/qwen/qwen2.5-7b-instruct ($0.15/M)

⚠️ ollama-gpu-server is BLOCKED for cron/spawn use. Ollama binds to 127.0.0.1 by default — unreachable over LAN from the OpenClaw host. The router_policy.py enforcer will reject any payload referencing it.

Tier classification uses 4 capability signals (not cost alone):

• effective_params (50%) — extracted from model ID or known-model-params.json for closed-source models
• context_window (20%) — larger = more capable
• cost_input (20%) — price as quality proxy (weak signal, last resort for unknown sizes)
• reasoning_flag (10%) — bonus for dedicated thinking specialists (R1, QwQ, Kimi-K2)

Policy Enforcer (NEW in v3.2.0)

router_policy.py catches bad model assignments before they are created, not after they fail.

Validate a cron payload before submitting

python3 skills/intelligent-router/scripts/router_policy.py check \
  '{"kind":"agentTurn","model":"ollama-gpu-server/glm-4.7-flash","message":"check server"}'
# Output: VIOLATION: Blocked model 'ollama-gpu-server/glm-4.7-flash'. Recommended: anthropic-proxy-6/glm-4.7

Get enforced model recommendation for a task

python3 skills/intelligent-router/scripts/router_policy.py recommend "monitor alphastrike service"
# Output: Tier: SIMPLE  Model: anthropic-proxy-6/glm-4.7

python3 skills/intelligent-router/scripts/router_policy.py recommend "monitor alphastrike service" --alt
# Output: Tier: SIMPLE  Model: anthropic-proxy-4/glm-4.7  ← alternate key for load distribution

Audit all existing cron jobs

python3 skills/intelligent-router/scripts/router_policy.py audit
# Scans all crons, reports any with blocked or missing models

Show blocklist

python3 skills/intelligent-router/scripts/router_policy.py blocklist

Policy rules enforced

1. Model must be set — no model field = Sonnet default = expensive waste
2. No blocked models — ollama-gpu-server/* and bare ollama/* are rejected for cron use
3. CRITICAL tasks — warns if using a non-Opus model for classified-critical work

Installation (Core Skill Setup)

Run once to self-integrate into AGENTS.md:

bash skills/intelligent-router/install.sh

This patches AGENTS.md with the mandatory protocol so it's always in context.

CLI Reference

# ── Policy enforcer (run before creating any cron/spawn) ──
python3 skills/intelligent-router/scripts/router_policy.py check '{"kind":"agentTurn","model":"...","message":"..."}'
python3 skills/intelligent-router/scripts/router_policy.py recommend "task description"
python3 skills/intelligent-router/scripts/router_policy.py recommend "task" --alt  # alternate proxy key
python3 skills/intelligent-router/scripts/router_policy.py audit     # scan all crons
python3 skills/intelligent-router/scripts/router_policy.py blocklist

# ── Core router ──
# Classify + recommend model
python3 skills/intelligent-router/scripts/router.py classify "task"

# Get model id only (for scripting)
python3 skills/intelligent-router/scripts/spawn_helper.py --model-only "task"

# Show spawn command
python3 skills/intelligent-router/scripts/spawn_helper.py "task"

# Validate cron payload has model set
python3 skills/intelligent-router/scripts/spawn_helper.py --validate '{"kind":"agentTurn","message":"..."}'

# List all models by tier
python3 skills/intelligent-router/scripts/router.py models

# Detailed scoring breakdown
python3 skills/intelligent-router/scripts/router.py score "task"

# Config health check
python3 skills/intelligent-router/scripts/router.py health

# Auto-discover working models (NEW)
python3 skills/intelligent-router/scripts/discover_models.py

# Auto-discover + update config
python3 skills/intelligent-router/scripts/discover_models.py --auto-update

# Test specific tier only
python3 skills/intelligent-router/scripts/discover_models.py --tier COMPLEX

Scoring System

15-dimension weighted scoring (not just keywords):

1. Reasoning markers (0.18) — prove, theorem, derive
2. Code presence (0.15) — code blocks, file extensions
3. Multi-step patterns (0.12) — first...then, numbered lists
4. Agentic task (0.10) — run, fix, deploy, build
5. Technical terms (0.10) — architecture, security, protocol
6. Token count (0.08) — complexity from length
7. Creative markers (0.05) — story, compose, brainstorm
8. Question complexity (0.05) — multiple who/what/how
9. Constraint count (0.04) — must, require, exactly
10. Imperative verbs (0.03) — analyze, evaluate, audit
11. Output format (0.03) — json, table, markdown
12. Simple indicators (0.02) — check, get, show (inverted)
13. Domain specificity (0.02) — acronyms, dotted notation
14. Reference complexity (0.02) — "mentioned above"
15. Negation complexity (0.01) — not, never, except

Confidence: 1 / (1 + exp(-8 × (score - 0.5)))

Config

Models defined in config.json. Add new models there, router picks them up automatically. Local Ollama models have zero cost — always prefer them for SIMPLE tasks.

Auto-Discovery (Self-Healing)

The intelligent-router can automatically discover working models from all configured providers via real live inference tests (not config-existence checks).

How It Works

1. Provider Scanning: Reads ~/.openclaw/openclaw.json → finds all models
2. Live Inference Test: Sends "hi" to each model, checks it actually responds (catches auth failures, quota exhaustion, 404s, timeouts)
3. OAuth Bypass: Providers with sk-ant-oat01-* tokens (Anthropic OAuth) are skipped in raw HTTP — OpenClaw refreshes these transparently, so they're always marked available
4. Thinking Model Support: Models that return content=None + reasoning_content (GLM-4.7, Kimi-K2, Qwen3-thinking) are correctly detected as available
5. Auto-Classification: Tiers assigned via tier_classifier.py using 4 capability signals
6. Config Update: Removes unavailable models, rebuilds tier primaries from working set
7. Cron: Hourly refresh (cron id: a8992c1f) keeps model list current, alerts if availability changes by >2

Usage

# One-time discovery
python3 skills/intelligent-router/scripts/discover_models.py

# Auto-update config with working models only
python3 skills/intelligent-router/scripts/discover_models.py --auto-update

# Set up hourly refresh cron
openclaw cron add --job '{
  "name": "Model Discovery Refresh",
  "schedule": {"kind": "every", "everyMs": 3600000},
  "payload": {
    "kind": "systemEvent",
    "text": "Run: bash skills/intelligent-router/scripts/auto_refresh_models.sh",
    "model": "ollama/glm-4.7-flash"
  }
}'

Benefits

✅ Self-healing: Automatically removes broken models (e.g., expired OAuth) ✅ Zero maintenance: No manual model list updates ✅ New models: Auto-adds newly released models ✅ Cost optimization: Always uses cheapest working model per tier

Discovery Output

Results saved to skills/intelligent-router/discovered-models.json:

{
  "scan_timestamp": "2026-02-19T21:00:00",
  "total_models": 25,
  "available_models": 23,
  "unavailable_models": 2,
  "providers": {
    "anthropic": {
      "available": 2,
      "unavailable": 0,
      "models": [...]
    }
  }
}

Pinning Models

To preserve a model even if it fails discovery:

{
  "id": "special-model",
  "tier": "COMPLEX",
  "pinned": true  // Never remove during auto-update
}

⚠️ Known Gap — Proactive Health-Based Routing (2026-03-04)

Current router is reactive not proactive:

• Fallback only fires AFTER a 429 is received
• No awareness of concurrent sessions on same proxy
• No cooldown tracking after rate-limit events

Needed improvements:

1. Track last-429 timestamp per provider → skip if within cooldown window
2. Track active concurrent spawns per provider → if >1 active, route to OAuth
3. Before spawning N parallel agents, check if single provider can handle N concurrent
4. Expose router.get_best_available(n_concurrent=2) API