fix(media): don't crash when media block data is raw bytes

[devteamaegis]

What's broken

MediaManager._find_and_process_media builds a data URI for Anthropic- and Vertex-style media blocks with f"...;base64," + data["data"]. The branch guards only check that the data key exists, not that its value is a str. If a logged payload carries raw image bytes there, the str + bytes concatenation raises an uncaught TypeError. Because the caller runs outside the masking try/except, the crash propagates out of span/observation creation.

payload = {"role": "user", "content": [
    {"type": "base64", "media_type": "image/png", "data": b"\x89PNG ..."}]}
mm._find_and_process_media(data=payload, trace_id="t", observation_id="o", field="input")
# -> TypeError: can only concatenate str (not "bytes") to str

Why it happens

The format guards verify the dict shape but assume data["data"] is a base64 string.

Fix

Add an isinstance(data["data"], str) check to both branches, so non-string data falls through to normal pass-through handling instead of crashing.

Test

Processing a base64 block whose data is bytes passes through unchanged instead of raising.

Greptile Summary

This PR guards two branches in _find_and_process_media that build base64 data URIs by adding isinstance(data["data"], str) checks, preventing a TypeError crash when a logged payload carries raw bytes instead of a base64 string in the "data" field of Anthropic- or Vertex-style media blocks.

• Anthropic block (type == "base64"): now skips media processing and passes the dict through unchanged when data["data"] is not a str.
• Vertex block (type == "media"): same guard added for the mime_type-keyed variant.
• A new unit test covers both block shapes with bytes data and asserts no media is enqueued.

Confidence Score: 5/5

Safe to merge — the change is a minimal two-line guard that prevents a crash without altering any existing code path for valid string data.

Both guards touch only the condition that selects a code path; the string-data path is identical to before, and the bytes-data path now falls through to the existing generic dict recursion rather than throwing. The new test directly exercises both block shapes with raw bytes and confirms pass-through with no queued uploads.

No files require special attention.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[_process_data_recursively] --> B{isinstance data dict?}
    B -- No --> C[Other checks / pass-through]
    B -- Yes --> D{Anthropic block?\ntype==base64, media_type, data}
    D -- No --> E{Vertex block?\ntype==media, mime_type, data}
    D -- Yes --> F{isinstance data.data str?}
    E -- Yes --> G{isinstance data.data str?}
    F -- Yes --> H[Build base64 URI\nUpload media\nReturn copied dict with LangfuseMedia]
    F -- No --> I[Fall through to dict recursion\nPass-through unchanged]
    G -- Yes --> J[Build base64 URI\nUpload media\nReturn copied dict with LangfuseMedia]
    G -- No --> I
    E -- No --> K[Recurse into dict values]

Loading

_{Reviews (1): Last reviewed commit: "fix(media): don't crash when Anthropic/V..." | Re-trigger Greptile}

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[CLAassistant]

All committers have signed the CLA.