MumuTW 20aa796d97 fix(sse): route task-aware defaults by intent, fix fitness pattern shadowing (#8602, #8603) (#8605) ...

* fix(sse): restore task-aware routing config on restart (#8601)

The T05 Task-Aware Smart Routing config was persisted to settings.taskRouting
by PUT /api/settings/task-routing but never read back, so it silently reverted
to enabled:false + the hardcoded default model map on every restart.

Two root causes, both fixed:

- No boot hydration existed. Adds hydrateTaskRoutingConfig(settings), wired into
  src/instrumentation-node.ts next to the Thinking-Budget restore (#5312). It
  accepts either the JSON string the route persists or an already-parsed object,
  and fails open on malformed values. applyRuntimeSettings does not cover this
  key, same as the Global System Prompt (#2470).

- The config lived in a plain module-level `let`, which is duplicated per module
  graph — a boot hydration would have landed on the instrumentation graph's copy
  and never reached the one src/sse/handlers/chat.ts reads. This is the exact
  break #5312 fix-A hit on the VPS. Moves the store to the globalThis pattern
  already used by thinkingBudget.ts and systemPrompt.ts.

Runtime stats are never restored from the persisted blob.

Note the hydration is wired into instrumentation-node.ts, not the unused
src/server-init.ts.

* docs(changelog): add fragment for #8604 task-routing boot restore

* fix(sse): route task-aware defaults by intent, guard fitness pattern order (#8602, #8603)

Two related defects in the hand-maintained model-quality tables.

#8602 — DEFAULT_TASK_MODEL_MAP hardcoded literal provider/model ids
(openai/gpt-4o, gemini/gemini-2.5-flash-lite, deepseek/deepseek-chat, ...).
Wrong twice over: the ids rotted by a generation or two, and applyTaskAwareRouting
overwrites body.model directly, so a literal target skipped auto-combo's 13-factor
scoring (quota, circuit-breaker health, cost, latency, stability), connection
cooldown and model lockout — hard-failing for any operator with no connection for
that provider. Refreshing the strings would only reset the rot clock, so the
defaults now name auto/* INTENTS that resolve against the operator's actually
connected backends:

  coding        -> auto/coding
  analysis      -> auto/reasoning
  vision        -> auto/vision
  summarization -> auto/chat:fast
  background    -> auto/chat:cheap

creative and chat stay pass-through. Operators can still pin a specific model via
PUT /api/settings/task-routing; only the shipped defaults change. No provider/model
literal remains in the module.

#8603 — the pattern-shadowing fix LANDED UPSTREAM while this PR was open
(9f5be229b, Train 1D). lookupStaticFitnessTable now ranks patterns longest-first,
so gpt-4o-mini no longer inherits gpt-4o's 0.9 and deepseek-v3.2 no longer inherits
deepseek-v3's 0.85. This PR therefore no longer changes that behaviour — the
upstream scan is kept verbatim.

What remains for #8603 is the regression guard. The resolution chain hits the DB
(user_override / arena_elo / models.dev tier) before reaching layer 4, so asserting
the ordering through getTaskFitness would depend on DB fixture state. The layer is
exposed as getStaticFitnessTableScore and pinned directly by
taskFitness-pattern-order-8603.test.ts (7 cases), so the guarantee survives future
edits to FITNESS_TABLE. Those 7 cases were written against this PR's original
implementation and pass unchanged against the upstream one — independent
confirmation that the two are behaviourally equivalent.

* docs(changelog): add fragment for #8605 task-routing intent + fitness order

2026-07-27 19:07:30 -03:00

13 KiB

Raw Permalink Blame History

View Raw