fix: normalize frozen Langfuse spans and update SDK test usage

[RasulOs]

OpenTelemetry 1.44 freezes span attributes before end callbacks, so Mobilerun's Langfuse preprocessor could fail with TypeError and leave screenshot data unconverted. Normalize a private export snapshot before the Langfuse SDK processes media, preserving the original ended span for other processors.

The internal tracer-provider adapter wraps the processor registered by the public Langfuse client. The SDK still owns one exporter, batching queue, and media-upload pipeline. Start-time metadata enrichment, initialization locking, registration checks, and failed-initialization guards remain in place. Tracing, flush, and shutdown delegate to the existing provider/processor. Span identity, parent, resource, timestamps, status, events, links, instrumentation metadata, and dropped-item counts are preserved.

The three OpenAI SDK regressions came from the shared usage fixture missing the newer required cache_write_tokens field. Supply zero when the installed schema defines that field. Keep httpx, SDK response validation, and all existing request-body, multimodal, tool-call, streaming, and token-count assertions. Production token accounting is unchanged.

Follow-up to #444. No public API, dependency pins, model defaults, credential storage, package version, README, Portal mapping, or agent-test-flows changes. Only two production Python files and two Python test files change.

Compatibility validation

Reproduced all four original failures on unchanged main 7ebede7 before editing: 4 failed in 13.74s (OpenTelemetry 1.44.0, OpenAI 2.54.0, Langfuse 4.15.1). All four now pass.

Four isolated Python 3.13.9 environments, all with Langfuse 4.15.1 and passing uv pip check:

OpenTelemetry SDK/API	OpenAI SDK	Focused tracing/provider/CLI suite	Full suite
1.40.0	2.30.0	175 passed, 3.44s	978 passed + 3 subtests, 15.42s
1.44.0	2.54.0	175 passed, 3.78s	978 passed + 3 subtests, 15.88s
1.40.0	2.54.0	175 passed, 20.54s	—
1.44.0	2.30.0	175 passed, 19.77s	—

Regression coverage uses real OpenTelemetry and Langfuse SDK processing with a capturing exporter and mocked upload I/O after the real media queue. Screenshot, chat, and completion spans each export once and produce one matching media-upload job. Processors before and after the wrapper receive the same unchanged original span. Additional coverage checks immutable/extended attributes, metadata and dropped counts, concurrent setup, initialization failures before/after registration, duplicate/foreign pipelines, and lifecycle delegation.

Commands:

python -m pytest tests/test_langfuse_v4.py tests/test_grok_api.py tests/test_phoenix_tracing.py tests/test_provider_defaults.py tests/test_openai_oauth_llm.py tests/test_grok_cli.py -q
python -m pytest -q
python -m black --check --diff .
python -m ruff check mobilerun/telemetry/langfuse_processor.py mobilerun/agent/utils/tracing_setup.py tests/test_langfuse_v4.py tests/test_grok_api.py
git diff --check

Black 26.5.1: 185 files unchanged. Relevant Ruff checks and git diff --check: passed. No test failures or warnings in the final runs.

Android 16 smoke

Real MobileAgent session on Medium_Phone_API_36.1 / emulator-5554, Android 16, Portal 0.7.25, saved Gemini OAuth, implicit gemini-3.8-flash-tiered default. Indexed tapping; coordinate tools disabled; maximum 15 steps / 180 seconds. Screenshot capture enabled with MOBILERUN_STREAM_SCREENSHOTS=true, Langfuse screenshots enabled, vision disabled.

Passed: 4 steps, 4 tool calls, 24.18 seconds. The agent opened Settings search, typed display, selected Display & touch, and completed. Independent device trees, dispatched actions, and the final screenshot confirmed Display & touch and Brightness level.

The real Langfuse SDK exported 36 unique spans, including 5 screenshot spans with matching media references and exactly 5 mocked upload jobs, one per screenshot span. Upload bytes matched the original screenshot bytes. One client/pipeline, original spans unchanged, and no normalization warnings.

Limits: trace export was captured locally; Langfuse authentication checks and upload network I/O were mocked. This validates SDK processing and queuing, not Langfuse cloud ingestion. The initial UI run passed in 4 steps / 25.57s but omitted screenshot capture; it was excluded from media validation and repeated with capture enabled. All scripts, credentials, screenshots, trees, and logs remain outside the repository. No release was prepared or published.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-10T14:48:40.540772Z	b6e8d45	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.