Routing

Orch uses an LLM-as-classifier to route each task to the best agent. This is a non-agentic call (--print) — fast and cheap, no tool access needed.

How It Works

1. The Rust router (src/engine/router/) sends the task title, body, and labels to a lightweight LLM (default: claude --model haiku --print).
2. The router LLM returns JSON:

   {
     "executor": "claude",
     "complexity": "medium",
     "reason": "why this agent/model",
     "profile": {
       "role": "DeFi Protocol Engineer",
       "tools": ["Bash", "Edit", "Read"],
       "constraints": ["audit with sealevel attacks"]
     }
   }

3. Sanity checks run -- e.g. warns if a backend task gets routed to claude, or a docs task to codex. If the router fails repeatedly (configurable via max_route_attempts, default 3), it falls back to round-robin.
4. If the router fails, it falls back to config.yml's router.fallback_executor (default: codex).

Config

router:
  mode: "llm"               # "llm" | "round_robin" | "local", default round_robin
  agent: "claude"           # which LLM does the routing
  model: "haiku"            # fast/cheap model for classification
  timeout_seconds: 45       # values over 45 are lowered to 45
  max_route_attempts: 3     # LLM failures before falling back to round-robin
  max_tasks_per_tick: 1     # routing concurrency cap
  weighted_round_robin: false
  fallback_executor: "codex"
  allowed_tools:            # default tool allowlist
    - yq
    - jq
    - bash

To exclude an agent from routing entirely, remove it from the agents: discovery list (the router only considers installed, listed agents) — there is no separate disabled_agents knob.

Available Executors

Agents are auto-discovered from PATH. The router only picks from installed agents. A grok task also needs a grok model in the matching model_map tier. Take the id from grok models.

Complexity-Based Model Selection

The router assigns a complexity level (simple, medium, complex) which maps to agent-specific models:

model_map:
  simple:
    claude: "haiku"
    codex: "gpt-5.1-codex-mini"
  medium:
    claude: "sonnet"
    codex: "gpt-5.2"
  complex:
    claude: "opus"
    codex: "gpt-5.3-codex"
  review:
    claude: "sonnet"
    codex: "gpt-5.2"

The routing prompt is in prompts/route.md. The router only classifies — it never touches code or files.

When a task's stored model is no longer in its agent's model pools, orch replaces it with the agent's model for the task's complexity and records the reason in last_error.

Pre-emptive Routability

The router performs a pre-emptive routability check before attempting to invoke agents or the router LLM. This does two things:

• It excludes agents that are currently in cooldown or have been marked degraded by the health check (rate-limit signals). These checks are authoritative — if an agent is degraded, it will be skipped regardless of weight.
• When router.weighted_round_robin is enabled, an additional weight-based threshold is applied after the degraded check. Agents whose routing weight has decayed below this threshold are skipped proactively to avoid sending work to recently rate-limited agents. This weight-based exclusion only layers on top of existing cooldown/degraded logic and does not replace it.

The threshold is fixed at 0.3. No config key changes it.