Quality-first automatic model routing for Claude Code sub-agents. Every task gets the right model — default strongest, downgrade only when confidently trivial.
Most "model routers" optimize for cost and quietly drop quality on hard tasks. smart-dispatch inverts that: it never loses quality to a routing mistake. The only acceptable misjudgment is treating a simple task as hard (a little wasted spend) — never the reverse.
Before dispatching a sub-agent, smart-dispatch:
- Classifies the task with a cheap model (Haiku) →
{tier, confidence}. - Applies a quality-first policy: default
opus; downgrade only whentier ∈ {Trivial, Routine}ANDconfidence ≥ 0.8. - Dispatches the worker with the chosen model.
| Tier | Example | Model |
|---|---|---|
| Trivial | grep, list files, read config | haiku |
| Routine | clear-pattern edit, summarize, format | sonnet |
| Hard | design, debug, new code, architecture | opus |
| uncertain | anything fuzzy | opus (fallback) |
The router's own model suggestion is ignored — the policy re-derives the choice from tier + confidence alone.
claude plugin marketplace add dudupii/smart-dispatch
claude plugin install smart-dispatch@smart-dispatchOnce installed, routing is automatic and transparent: a PreToolUse hook intercepts every Agent tool call and rewrites the model in place via updatedInput — you never have to remember a command, and the model cannot bypass it by calling the Agent tool directly. If you name a model explicitly, smart-dispatch respects it and skips routing.
The hook uses conservative heuristics (it only downgrades read-only
Exploretasks). For higher-fidelity routing with a Haiku classifier, invoke/smart-dispatchexplicitly — both paths share the samesrc/decide-model.jspolicy and write to the same log.
These live in src/decide-model.js (the single source of truth; skills/smart-dispatch/SKILL.md mirrors them in prose):
DOWNGRADE_THRESHOLD(default0.8) — the confidence required to leave opus. Raise for more conservative routing (closer to all-opus); lower to downgrade more aggressively.BUDGET_FLOOR(default0.1) — only relevant to budget mode (the Workflow pro mode,workflows/batch-route.js): when remaining budget drops below this fraction, opus steps down to sonnet. Never escalates an already-downgraded task.- Router model — default Haiku (configured in
eval/run-eval.js). If eval shows false-downgrades, raise to Sonnet.
npm install # dev deps only (@anthropic-ai/sdk)
npm test # unit tests: policy, parser, metrics, dataset schema
ANTHROPIC_API_KEY=xxx npm run eval # live routing-quality eval over eval/dataset.jsonThe eval reports two numbers:
- falseDowngradeRate — Hard tasks routed below opus. Red line: ~0.
- savingsRate — spend vs an all-opus baseline. Target 0.3–0.5.
src/decide-model.js— the quality-first policy (single source of truth, fully unit-tested).src/classify-heuristic.js— conservative heuristic classifier used by the hook (read-onlyExploreonly).hooks/route.mjs+hooks/hooks.json— thePreToolUsehook that makes routing automatic.src/parse-router-output.js— defensive parser for the router agent's output.src/compute-metrics.js— false-downgrade + savings metrics.skills/smart-dispatch/SKILL.md— the shipped skill; mirrors the policy in prose.eval/— labeled dataset + harness that validates routing quality end-to-end.
The shipped plugin has zero runtime dependencies — the Anthropic SDK is dev-only, used solely by the eval harness.
workflows/batch-route.js is a Workflow for batch processing with cost control. It applies the same quality-first policy plus budget awareness: when remaining budget drops below BUDGET_FLOOR, opus tasks step down to sonnet (the only allowed downward override of opus). Hand it a task or an array of tasks as args; it routes each with Haiku, then executes each on the chosen model.
Caveat: workflow scripts run in a sandbox and cannot
importlocal modules, so the policy is inlined in the script.src/decide-model.jsremains the source of truth — keep them in sync. Running it spawns one sub-agent per task (multi-agent orchestration), so it spends tokens.
Every routing decision is shown inline (smart-dispatch → haiku (Trivial, conf 0.92)) and appended to a local log at ~/.smart-dispatch/log.jsonl — only tier, confidence, model, and a timestamp are recorded, never the task text.
See aggregate stats anytime:
npm run report # or the /smart-dispatch-report command in a sessionIt reports total decisions, model distribution, estimated savings vs all-opus, and how often budget mode downgraded opus. Override the log path with SMART_DISPATCH_LOG.
MIT.