wyuc 25cf58d19f
feat(model): optional per-stage LLM model routing (#745) (#747)
* feat(model): optional per-stage LLM model routing (#745)

Add a config-only stage -> model map consulted during model resolution,
falling back to DEFAULT_MODEL when unset (zero behavior change unless opted in).

- lib/server/model-routes.ts: parse/validate/cache MODEL_ROUTES env JSON;
  LLM_STAGES registry of routable stages; getStageModel(stage).
- resolveModel: resolution order x-model > stage route > DEFAULT_MODEL > builtin;
  thread optional `stage` through resolveModelFromHeaders/FromRequest.
- Wire each route's resolveModel call site to its stage; classroom-generation
  resolves generate-classroom and web-search-query-rewrite independently.
- .env.example: document MODEL_ROUTES; unit tests for both modules.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(model): address cross-review findings for per-stage routing (#745)

- classroom-generation: resolve the web-search-query-rewrite model lazily,
  only when explicitly routed and inside the web-search branch, with try/catch
  fallback to the classroom model. Fixes a regression where a misconfigured
  optional route (keyless/invalid provider) threw and aborted ALL classroom
  generation even with web search disabled; also avoids wasted resolution on the
  common no-search path and preserves classroom-model inheritance when unrouted.
- resolve-model: type the stage param as LlmStage across resolveModel/
  resolveModelFromHeaders/resolveModelFromRequest so a mistyped stage literal is
  a compile error instead of silently falling through to DEFAULT_MODEL.
- model-routes: simplify the cache to a process singleton (matching
  provider-config), dropping the raw-string-keyed cache and redundant STAGE_SET.
- .env.example: clarify x-model precedence and that a route to an unconfigured
  provider fails at request time (no startup validation), like a bad DEFAULT_MODEL.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(model): stage route takes precedence over client x-model (#745)

Cross-review (codex P2 + Claude tracer) found that the browser UI always sends
its saved model as x-model, so the previous x-model > stage order made
MODEL_ROUTES inert for normal UI traffic on exactly the heavy stages the RFC
targets (scene-content, quiz, pbl-chat, chat). Flip resolution to
stage route > x-model > DEFAULT_MODEL: a configured route is the operator's
deliberate per-stage choice and wins, while unrouted stages still honor the
client's x-model. Update docs and tests accordingly.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* style(model): prettier-format changed files for #745 CI

Run the repo formatter (pnpm check / prettier) on the changed routes, the new
model-routes module, and the new tests so `prettier . --check` (CI) passes;
also correct a stale precedence comment in the resolve-model test.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(model): isolate routed model from client connection params (#745)

When a stage route overrides the client's x-model and points at a different
provider, the client-sent apiKey/baseUrl/providerType (for the client's model)
must not bleed onto the routed provider — otherwise a routed Anthropic model
would be constructed with the client's OpenAI providerType/key and fail. A
routed model now resolves its connection params purely from server config, as
if no x-model was sent. Unrouted stages still honor the client params. Adds
tests covering both the drop (routed) and keep (unrouted) cases.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs(model): correct graceful-degradation comment in classroom (#745)

resolveModel does not throw on a missing key; clarify that a keyless route
surfaces later in callLLM (degraded by the outer try/catch), while only a
resolution-time failure is caught at the rewrite re-resolve. Comment-only.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(model): route scene-content per scene type (#745)

Extend MODEL_ROUTES with composite keys scene-content:<type> for the four core
scene types (slide/quiz/interactive/pbl). getStageModel now resolves composite
`a:b` stages most-specific-first, trimming `:` segments, so scene-content:<type>
falls back to the base scene-content route, then x-model, then DEFAULT_MODEL.
The scene-content route derives its stage from outline.type. Backward
compatible: with no composite key configured, behavior is unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(model): per-stage thinking effort + drop hardcoded gpt fallback (#745)

MODEL_ROUTES values may now be {model, effort} (not just a model string) to pin
the thinking effort per stage. Arbitration mirrors model routing: a routed stage
with effort set wins over the client's thinking (effort "none" disables); routed
with no effort uses the model's default and drops client thinking; unrouted
stages keep the client thinking. resolveModel is the single arbiter — body
thinking is threaded in via resolveModelFromRequest. classroom-generation passes
the resolved thinking into its callLLM calls so generate-classroom and
web-search-query-rewrite honor route effort too.

Also: removed the hardcoded `|| 'gpt-5.4-mini'` fallback in resolveModel — if no
model resolves (no route / x-model / DEFAULT_MODEL) it now throws instead of
silently picking a vendor default.

model-routes: getStageRoute returns {model, effort?}; getStageModel delegates to
it. Docs + tests updated.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* refactor(model): route value carries full ThinkingConfig (#745)

Per-stage routing accepts the full unified ThinkingConfig in the route value
({model, thinking:{mode,effort,level,enabled,budgetTokens,excludeReasoningOutput}})
instead of just an effort string. The route's thinking is passed through
resolveModel and normalized per the model's capability by callLLM, so
budgetTokens (qwen), level (Gemini), enabled/mode, etc. all work.

(qwen3.7-plus/max thinking capability now comes from upstream #753, so this no
longer registers them.)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* fix(model): honor routed thinking for chat-adapter; doc fixes (#745)

Cross-review (codex P2 + Claude) found chat-adapter was the only routable stage
that ignored the route's thinking: it took the routed model but built its own
thinkingConfig from the request body, so operator-pinned thinking never applied
and client thinking still leaked onto a routed model. Now pass the client
thinking into resolveModel and use the resolved thinkingConfig (route-pinned for
a routed stage, client otherwise; defaults to disabled for low-latency chat).

Also: clarify DEFAULT_MODEL is now required for server-side stages (the
hardcoded gpt-5.4-mini fallback was removed → resolveModel throws if nothing
resolves), and fix a stale {model,effort}→{model,thinking} doc comment.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-17 22:26:37 +08:00
..