Trace-level LLM (GenAI) enrichment for OpenRouter OTel spans

[Makisuo]

Context

OpenRouter's Broadcast feature already ingests into Maple: it POSTs OTLP/HTTP-JSON traces to /v1/traces (documented in docs/openrouter-tracing.md) using the OTel gen_ai.* semantic conventions (gen_ai.operation.name, gen_ai.provider.name/legacy gen_ai.system, gen_ai.request.model, gen_ai.response.model, gen_ai.usage.input_tokens/output_tokens, gen_ai.usage.cost, gen_ai.response.finish_reasons). But those attributes landed in the generic SpanAttributes map with no LLM-aware rendering — this is roadmap item #13.

This PR closes that gap for trace-level LLM enrichment: open any trace with OpenRouter/LLM spans and immediately see which spans are LLM calls, on what model/provider, how many tokens, and at what cost — with prompts/completions inspectable inline and graceful handling of OpenRouter Privacy Mode. It is pure query-engine projection + frontend work — no ingest or storage changes.

What changed

• Projection — added the gen_ai.* keys to TREE_SPAN_ATTR_KEYS (packages/query-engine/src/ch/queries/errors.ts) so the waterfall/timeline/flow views can badge LLM spans without the per-span lazy detail fetch.
• Detector — new getGenAiInfo (packages/ui/src/lib/gen-ai.ts) normalizes gen_ai.* into one shape, coalescing gen_ai.system → gen_ai.provider.name and gen_ai.response.model → gen_ai.request.model, and tolerating Privacy Mode (absent prompt/completion/finish reasons). Detection keys on a real non-empty value, never key-presence (the trimmed-projection footgun).
• Cross-view badge + detail summary — a new genAiAdapter in the cloud-platform registry (packages/ui/src/lib/cloud-platforms/gen-ai.ts) lights up the sparkle/provider badge across every trace view and the span-detail summary block (provider, operation, model, tokens, cost, params, finish reason) with no component edits, since both are driven by getCloudPlatform.
• Filter alias — gen_ai.system ↔ gen_ai.provider.name now coalesce in attribute filters (packages/query-engine/src/traces-shared.ts), so a provider filter matches either spelling.
• Value formatting — gen_ai.request.model/response.model render as model chips, token counts are grouped, and gen_ai.usage.cost renders as currency in the attribute table (apps/web/src/components/attributes/commit-sha-attribute.tsx).
• Added SparklesIcon; updated docs/openrouter-tracing.md and roadmap Alerts #13.

Testing

• packages/ui — 14 new unit tests for the detector, adapter, and formatters (full OpenRouter span, provider coalescing, empty-value rejection, Privacy Mode, finish-reason parsing, bad-outcome flagging, no shadowing of DB/non-LLM spans). All 103 ui-lib tests pass.
• packages/query-engine — extended the span-hierarchy projection test to assert the gen_ai.* keys, and added a tracesTimeseriesQuery test asserting the gen_ai.provider.name/gen_ai.system coalescing compiles. All 105 ch tests pass.
• turbo typecheck clean across @maple/ui, @maple/query-engine, @maple/web; oxfmt + oxlint clean.
• Not run: full browser E2E (seed a gen_ai span → view enrichment); the pure logic is unit-covered and the React integration is typecheck-covered.

Follow-ups (deferred)

Aggregate "AI-first ops" surfaces — a dedicated AI page, token/cost dashboards + numeric-attribute aggregation, a gen_ai_usage_hourly rollup, MCP tools, and cost/budget alerts — are scoped in the plan for a phased follow-up (query-time first, typed rollup MV as a fast-follow).

🤖 Generated with Claude Code

https://claude.ai/code/session_01QdSppGz2eTovugvSVRQeBq

---

Generated by Claude Code

---

[pullfrog]

Your Pullfrog Router balance is exhausted.

You have a card on file but auto-reload is disabled, so runs paused once your balance went past the overdraft buffer.

Top up balance → · Enable auto-reload →

  ^{｜ ⚠️ this action is pinned to a commit SHA, which freezes the cleanup step — switch to @v0 or keep the SHA fresh with Dependabot ｜ Rerun failed job ➔ ｜ View workflow run ｜ via Pullfrog ｜ Using Claude Opus ｜ 𝕏}

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.