feat(llm): route search.ts classify/extract/synthesize through the LLM router

[HomenShum]

What

Track B of the LLM-router rollout. Wires the 7 hardcoded gemini-3.1-flash-lite-preview call sites in server/routes/search.ts through the shared planner-on-a-pool router (shared/llm/router.ts, merged in #460), so model choice is owned in one place and genuinely hard turns can escalate.

ADDITIVE + behavior-preserving. The floor of every classify/extract/synthesize pool is the same flash-lite model, and signals are derived cheaply + locally (query length -> inputChars, retrieved source count -> sourceCount, comparison branches -> multiEntity). A simple single-entity query routes to the exact same model as before — only long, analytical, multi-entity, or many-source turns escalate to gemini-3-flash-preview.

Call sites wired (server/routes/search.ts)

Site	Task class	Notes
classifyQueryWithLLM query classification	classify	single-candidate pool — guaranteed no-op, just centralizes the model id
agent_synthesize trace (synthesizeResults)	synthesize	surfaces chosen model + reason in the trace. The wire-level model for synthesizeResults itself lives in server/agentHarness.ts (out of scope); this routes + labels the search-side decision
why-this-team credibility enrichment	extract	structured extraction over synthesized result + local context
multi-entity comparison extraction	extract (multiEntity: true)	most likely branch to escalate
single-entity extraction	extract	source count = allSnippets.length
founder-direction extraction	extract	source count = genWebSnippets.length

Observability

The chosen model lands in each trace step's tool field, and the route reason ("<model> -- floor light (complexity N.NN)" / "<model> -- escalated to balanced (complexity N.NN)") is appended to the step's detail — matching the existing SearchTraceEntry shape exactly. Escalations are visible in "How we got this answer".

Reliability (.claude/rules/agentic_reliability.md)

• DETERMINISTIC — searchRouteSignals is a pure function (no Date/random), so routing is replay-safe. NaN/negative source counts are coerced to 0 (no NaN leak into the complexity score).
• The AbortController / Promise.race request-budget gates and the 4-layer grounding pipeline are untouched.

Tests

server/searchRouteLlmRouting.test.ts — scenario-based (founder bare-name lookup, investor head-to-head comparison, banker diligence teardown), asserting:

• the no-op floor for simple single-entity queries (the additive guarantee),
• escalation for hard multi-entity / analytical / many-source turns,
• classify never escalates regardless of complexity,
• determinism (identical inputs -> identical model).

Verification

• npx tsc --noEmit --pretty false — 0 errors
• npx vitest run server/searchRouteLlmRouting.test.ts shared/llm/router.test.ts — 21 passed
• npx vitest run server/searchRoute.test.ts — 25 passed (no regression)
• npm run build — clean

Do not enable auto-merge — for review.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
nodebench-ai	Ready	Preview, Comment	Jun 2, 2026 7:34am

[augmentcode]

🤖 Augment PR Summary

Summary: This PR completes “Track B” of the LLM-router rollout by routing the search pipeline’s Gemini calls through the shared deterministic router, centralizing model choice and enabling controlled escalation on hard turns.

Changes:

• Added deterministic signal derivation in server/routes/search.ts via searchRouteSignals() (query length, source count, multi-entity flag, analytical hint).
• Introduced routeSearchModel() to map task class + signals to a router decision (routeLLM) and emit a trace-friendly reason string.
• Rewired search route call sites that previously hardcoded gemini-3.1-flash-lite-preview (classify, extract branches, and search-side synthesize decision) to use the router-selected model.
• Enhanced trace observability by recording the chosen model in the trace step’s tool field and appending route reasoning to the step detail.
• Added scenario-based Vitest coverage (server/searchRouteLlmRouting.test.ts) to assert the behavior-preserving floor, escalation on hard multi-entity/analytical/many-source turns, and determinism.

Technical Notes: Routing remains replay-safe (pure + deterministic), and the classify pool remains a guaranteed no-op (single-candidate floor).

_{🤖 Was this summary useful? React with 👍 or 👎}

[augmentcode]

Review completed. 1 suggestion posted.

Comment augment review to trigger a new review at any time.

[augmentcode]

searchRouteSignals(query, 0) here only reflects the raw query, but the actual Gemini prompt includes synthesized.answer plus potentially large localContext, so routing may stay on the floor even when the input is long/complex. Consider deriving signals from the real prompt/context size (or passing a meaningful sourceCount) so routing decisions match the workload.

Severity: medium

_{🤖 Was this useful? React with 👍 or 👎, or 🚀 if it prevented an incident/outage.}