[KooshaPari]

Summary

Adds a dedicated speed-optimized routing mode that scores every provider×model pair in a combo against TTFT, tokens-per-second, E2E latency, p95 latency, circuit-breaker health, failure rate, and latency stability (std-dev) — and ranks them so the fastest reliable pair wins.

The runtime router (routerStrategy.ts → LatencyStrategyImpl) and the new MCP tool (omniroute_pick_fastest_model) now share a single pure scoring core at open-sse/services/autoCombo/speedRanking.ts → rankBySpeed(), so the user-facing preview and the live router pick the same winner.

What

• rankBySpeed(candidates, weights?) — pure scoring function. Per-candidate factor breakdown (ttft/tps/e2e/p95/health/reliability/stability), weighted composite, OPEN-circuit penalty, deterministic ordering. Tunable weights via SpeedRankingWeights (defaults bias toward perceived speed: TTFT 0.30, TPS 0.20, E2E 0.20, p95 0.10, health 0.08, reliability 0.07, stability 0.05).
• LatencyStrategyImpl refactored to delegate to rankBySpeed — the runtime router now benefits from the same rigorous scoring (was previously only p95LatencyMs + errorRate).
• New MCP tool omniroute_pick_fastest_model (schemas/tools.ts + tools/advancedTools.ts) — returns the full ranked list with per-factor scores and per-metric breakdown, not just a winner. Optionally applies the result to a target combo by switching its strategy to auto + autoRoutingStrategy: "latency". includeUnhealthy=true retains OPEN-circuit candidates (default drops them).
• 12 unit tests in open-sse/services/autoCombo/__tests__/speedRanking.test.ts covering: ranking stability, OPEN-circuit drop, weights effect, missing-telemetry neutrality, deterministic tie-breaks, malformed-input rejection, fast-but-flaky vs steady slower, apply-to-combo mutation.

Why

The current latency strategy is naive — it inverts p95LatencyMs against the pool max and adds 1 - errorRate, ignoring TTFT (the metric users actually feel), tokens-per-second (throughput), E2E latency, and latency stability (a flaky-fast provider can devastate UX). This PR brings the strategy up to par with how users actually think about "fast".

Test results

• All 88 related tests pass: 56 existing autoCombo.test.ts + 12 new speedRanking.test.ts + 8 existing routingStrategyTool.test.ts + 12 un-touched tests in adjacent files.
• 6 pre-existing failures in open-sse/mcp-server/__tests__/{cacheTools,dbHealthTool,essentialTools,httpAuthContext,toolSearch.catalog,toolSearch.tool}.test.ts are unrelated — caused by missing safe-regex dependency on ccrQuery.ts, present on upstream release/v3.8.44 before this PR.
• tsc --noEmit -p open-sse/tsconfig.json clean (only pre-existing baseUrl TS7 deprecation).
• eslint clean on all 6 modified files (only 3 pre-existing @typescript-eslint/no-explicit-any warnings in unrelated memoryTools registration).

Files

File	Change
open-sse/services/autoCombo/speedRanking.ts	new (336 lines) — pure scoring core
open-sse/services/autoCombo/routerStrategy.ts	refactor LatencyStrategyImpl → rankBySpeed; drop dead helpers
open-sse/mcp-server/schemas/tools.ts	add pickFastestModelInput / pickFastestModelOutput / pickFastestModelTool
open-sse/mcp-server/tools/advancedTools.ts	add handlePickFastestModel handler
open-sse/mcp-server/server.ts	register omniroute_pick_fastest_model
open-sse/services/autoCombo/__tests__/speedRanking.test.ts	new (226 lines) — 12 vitest cases
CHANGELOG.md	entry under [3.8.44] → ✨ New Features

Compatibility

• All added fields on ProviderCandidate (health, quotaRemainingPct, capacityScore, etc. used by the MCP input mapping) are optional — existing callers are unaffected.
• The MCP tool is purely additive.
• The runtime strategy is behaviorally compatible: same routerStrategy: "latency" value, same response shape from select(), just a more accurate winner.
• Default weights were tuned to match the previous LatencyStrategyImpl for the average case (TTFT-heavy for perceived speed, reliability multiplier prevents flakey-fast wins).

Closes #6010

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!