refactor(tracing): normalize agent spans in Rust

[yujonglee-berri]

TLDR

Problem this solves:

• Trace normalization is duplicated across Python and Rust
• Normalized fields lack a shared typed contract

How it solves it:

• Moves LangSmith, OpenInference, and GenAI normalization into Rust
• Validates Rust output with Pydantic at the Python boundary
• Checks normalized field mappings against live ClickHouse columns
• Generates dashboard trace types from the API response models

Merge preservation

#43975: the bounded OTLP receiver, shared resource identity, and paged span diagnostics remain covered by tracing and bridge tests

#44035: lifespan-owned tracing, access scoping, and shutdown draining pass the route, lifecycle, and storage tests

#44068: Lens trace access, setup polling, and investigations pass 56 focused dashboard tests

Review evidence

The native bridge returns consumed_attributes as a Python tuple at the current tip. Direct execution with the LangSmith export returned ('gen_ai.prompt', 'gen_ai.completion'); the bridge conversion test passes

User Flow

Before: a developer imports a LangSmith trace and inspects its agent and LLM spans

1. They POST the OTLP export to http://localhost:4097/v1/traces and receive HTTP 200 with {}
2. They GET /v1/traces/4bad42b84e9de3ba46fc870185f8f023 and see six spans, one LLM call, two tool calls, and 3332 input tokens
3. They GET /v1/traces/4bad42b84e9de3ba46fc870185f8f023/spans/8a6a1c31940d07af and see full tool call IDs and types

After: the same trace remains readable with the same fields while normalization runs in Rust

1. They POST the same OTLP export to http://localhost:4098/v1/traces and receive HTTP 200 with {}
2. They GET the same trace URL on port 4098 and see the same counts and request ID
3. They GET the same span URL on port 4098 and see the same full tool call payload

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally
• My PR passes all required CI/CD checks
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

Both runs used tests/test_litellm/tracing/fixtures/langsmith_deep_agent_export.json, local proxies, and separate PostgreSQL and ClickHouse databases

Before (a14af73)

1. curl -X POST http://127.0.0.1:4097/v1/traces -H 'Authorization: Bearer <local key>' -H 'Content-Type: application/json' --data-binary @tests/test_litellm/tracing/fixtures/langsmith_deep_agent_export.json returned HTTP 200 and {}
2. curl http://127.0.0.1:4097/v1/traces/4bad42b84e9de3ba46fc870185f8f023 -H 'Authorization: Bearer <local key>' returned HTTP 200 with span_count=6, llm_calls=1, tool_calls=2, input_tokens=3332, output_tokens=467, and LLM request ID chatcmpl-4077bb36-9380-4a3b-9481-245700cef09a
3. curl http://127.0.0.1:4097/v1/traces/4bad42b84e9de3ba46fc870185f8f023/spans/8a6a1c31940d07af -H 'Authorization: Bearer <local key>' returned HTTP 200 with full tool call IDs and types

After (246bf31)

1. The same POST on port 4098 returned HTTP 200 and {}
2. The same trace GET on port 4098 returned HTTP 200; the full JSON response matched Before
3. The same span GET on port 4098 returned HTTP 200; the full JSON response matched Before

Type

Refactoring

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[greptile-apps]

Confidence Score: 5/5

[High risk] Moves span normalization logic from Python to Rust.

The PR appears safe to merge; the remaining new finding is a non-blocking Python type-contract mismatch.

Findings

1.  Consumed attributes type mismatch ▶

Summary

The PR moves agent-span normalization into Rust, validates normalized fields at the Python boundary, checks their ClickHouse column mappings, and generates dashboard trace types from API response models.

• The latest revision changes how consumed-attribute arrays cross the Rust–Python boundary and updates their Python annotation.

_{Reviews (3) · Last reviewed commit: "fix(tracing): type consumed attributes a..."}

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_rust_trace_normalizers (246bf31) with main (d7e6154)¹}

Footnotes

1. No successful run was found on main (610576d) during the generation of this report, so d7e6154 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[devin-ai-integration]

Devin Review found 3 potential issues.

1 flag not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)

[codecov]

Codecov Report

❌ Patch coverage is 88.23529% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/rust_bridge/traces.py	84.61%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!

[yujonglee-berri]

@greptile-apps wdyt

[yujonglee-berri]

@greptile-apps The specific 3/5 rationale is not fair to this PR’s current diff. OTLP JSON identifier decoding was already on main; this PR leaves that code unchanged, and the existing hex ID fixture passes. This PR also changes no historical benchmark files. That work entered the merge base through #44057

[ishaan-berri]

can we add UI screenshots showing tracing still works ?