feat(auto-routing): Morph model router decisions for kilo-auto tiers

[shreybirmiwalmorph]

Summary

Adds per-prompt model routing for kilo-auto/* tiers via the Morph model router, running in shadow mode inside the auto-routing worker. The worker's response contract has carried decision: null since it shipped — this PR fills that slot with a real decision engine while leaving serving behavior untouched.

How it fits the existing architecture:

• Gateway (apps/web): the existing background mirror now includes a routing context for kilo-auto requests — the tier, the candidate models that tier may route among (new getMorphRouterCandidates, a product-owned set in auto-model/index.ts), and the model the static resolver actually picked. Optional field, so gateway/worker deploys never coordinate.
• Worker (services/auto-routing): new morph-router.ts calls POST /v1/router/multimodel with the tier's candidates (mapped Kilo public id ↔ Morph catalog id; unmapped candidates dropped), a tier-derived policy (frontier → capability_heavy, balanced → balanced), and the static pick as default_model so ambiguous prompts stay behavior-identical. Runs in parallel with the existing classifier; either side failing never affects the other.
• Caching: decisions reuse the per-conversation Durable Object cache, keyed by content hash + candidate-set/policy fingerprint, so multi-turn conversations stick to one model and config changes invalidate cleanly.
• Controls & telemetry: everything is behind a morph_router_enabled KV flag (default off, fail-closed). Decisions emit a sampled auto_routing_router_decision log line carrying the static pick vs. routed model, difficulty/confidence/domain, and latency — the dataset for deciding whether to serve these decisions later.

Morph only ever receives one bounded user-prompt prefix (≤1000 chars) — never the conversation, system prompt, tools, or any identifiers. Org/enterprise traffic is unaffected (the mirror already skips it).

Videos

Live end-to-end demo — wrangler dev + real Morph API: an easy prompt on the frontier tier routes to Sonnet instead of the static Opus pick, a heavyweight planning prompt confirms Opus (confidence 99%), the balanced tier picks across providers, and a repeat turn in the same conversation serves from the Durable Object cache in ~22 ms with no router call:

Full quality: demo.mp4

Verification

• Ran the worker locally (wrangler dev + local secrets-store values + morph_router_enabled=true in local KV) and sent realistic mirror payloads against the live Morph API: easy/hard/balanced prompts produced differentiated decisions, reverse-mapped to Kilo public ids (see video).
• Verified decision cache: a second request in the same conversation returned the identical decision in ~22 ms without a Morph call.
• Verified independence: with an invalid OpenRouter key the classifier errors while the router decision is still returned (and vice versa — a Morph 503 leaves classification untouched).
• Verified default-off: with the KV flag unset, no Morph request is made and decision stays null.

Visual Changes

N/A

Reviewer Notes

• Shadow mode only — nothing consumes decision yet. The intended follow-up is a flagged synchronous path in resolveAutoModel once the shadow logs show good agreement/cost numbers; happy to discuss rollout shape.
• New third-party data flow: bounded prompt prefixes go to Morph when the flag is on. Flag is off by default precisely so this can clear privacy/DPA review first. (Disclosure: I'm on the Morph team — shrey@morphllm.com.)
• The Kilo↔Morph model id map in morph-router.ts is deliberately small and explicit; candidates without a mapping are dropped before the call, and a tier with <2 routable candidates skips routing entirely (skipped:insufficient_candidates in logs).
• Per .specs/model-experiments.md, experimented public ids must not enter kilo-auto candidate sets — the new candidate map documents this constraint where it would be violated.
• MORPH_API_KEY needs to be added to the secrets store before enabling the flag; binding is already declared in wrangler.jsonc, and .dev.vars.example covers local dev.
• Full pnpm typecheck was not run (per AGENTS.md guidance); scripts/typecheck-all.sh --changes-only plus a targeted pnpm --filter web typecheck both pass.

🤖 Generated with Claude Code