Add Morph Model Router as a complexity classifier for routing decisions

[bhaktatejas922]

NadirClaw routes simple prompts to cheap/local models and complex ones to premium. The hard part is deciding "simple vs complex" reliably. Morph's Model Router does exactly that: ~50ms, returns difficulty (easy/medium/hard/needs_info) plus domain and ambiguity, trained on millions of coding prompts.

Instead of a length/keyword heuristic, NadirClaw could call /router/classify (or /router/multimodel with a cost_efficient policy) for the routing signal:

const { model } = await (await fetch("https://api.morphllm.com/v1/router/multimodel", {
  method: "POST",
  headers: { Authorization: `Bearer ${process.env.MORPH_API_KEY}`, "Content-Type": "application/json" },
  body: JSON.stringify({ input: prompt, allowed_providers: ["anthropic"], policy: "cost_efficient" }),
})).json();

Could sit behind a config flag as an optional routing strategy alongside the existing rules.

Docs: https://docs.morphllm.com/sdk/components/router

[doramirdor]

Thanks for the suggestion — this fits the existing extension surface well, and a config-flag-gated optional classifier is exactly the right shape.

Where it would slot in

NadirClaw's classifier abstraction lives in nadirclaw/classifier.py and is dispatched by get_binary_classifier() based on NADIRCLAW_CLASSIFIER (today: centroid / distilbert). The contract is small — a class with classify(prompt: str) -> Tuple[is_complex: bool, confidence: float]. A MorphRouterClassifier would call POST /v1/router/classify, map difficulty ∈ {easy, medium} → simple and {hard, needs_info} → complex, and use the model's returned confidence (or 1−ambiguity) as the score. That keeps the rest of the router untouched and lets nadirclaw report / dashboard / savings math all work without changes.

Things I'd want in the PR

1. Strictly opt-in. NADIRCLAW_CLASSIFIER=morph plus a required MORPH_API_KEY. Default stays centroid — we cannot make a remote round-trip a hard dependency for first-run UX.
2. Fail-closed to local. On HTTP error, timeout (>500ms budget), or missing key, log once and fall back to BinaryComplexityClassifier. Same pattern as compress_progressive's headroom fallback in feat(compression): headroom backend, progressive ladder, CCR, columnar packing #65.
3. Latency budget honored. ~50ms is fine on the LAN but tail latency from a CLI on residential internet is what bites. Suggest a MORPH_TIMEOUT_MS env (default ~200ms) and a small in-memory LRU on (prompt_hash → decision) for the run, since CLI sessions re-classify a lot of repeated prompts.
4. Cost tracking honesty. The savings report compares routed cost vs always-premium. A Morph call adds a few cents per 1k classifications — nadirclaw report should subtract that from "net saved" so users see the true figure. Open to suggestions on how Morph exposes that price.
5. Tests + offline mode. A unit test that monkeypatches urllib.request.urlopen for both the success and the fail-closed paths; nothing that hits the live API in CI.

If you want to take a swing at it, happy to review — start from BinaryComplexityClassifier and the dispatch block at classifier.py:340-373, and copy the env-flag plumbing pattern from how distilbert is wired in. If you'd rather we own the implementation, I can queue it; ping back here either way.

🤖 Automated triage reply via scheduled task (claude code, daily-check-github-status).