[ygyang11]

Problem

A Qwen-compatible endpoint (Aliyun MaaS compat-mode) rejects a request whose type: "file" part carries bytes that are not a valid PDF:

data: {"error":{"code":"invalid_parameter_error","param":null,"message":"Input file must be a valid PDF","type":"invalid_request_error"},"id":"chatcmpl-..."}

MEDIA_REJECTION_PHRASES has no phrase matching that message, so openai_provider raises a generic LLMError instead of LLMUnsupportedContentError. The agent-level self-heal (agent/base.py, which strips the offending tool-run attachments and retries once) therefore never fires and the whole run fails.

Evidence

Probed against the live endpoint with the exact wire shape the provider emits (_render_user_content file block):

• invalid PDF bytes in file_data → HTTP 400 at create(), for streaming and non-streaming alike; the streaming 400 body is the raw SSE line quoted above;
• the same request carrying a structurally valid minimal PDF → 200 OK.

So the endpoint supports file parts; only the payload was invalid, which is exactly the case the self-heal path exists for.

Change

• src/agent_harness/utils/media.py: add "must be a valid pdf" to MEDIA_REJECTION_PHRASES (placed in the existing file-payload cluster).
• tests/unit/test_harness/test_utils/test_media.py: one provider sample for the message above.

Matching stays narrow on purpose: "must be a valid pdf" covers the observed prefix variants without catching unrelated 400s. The matcher is a plain substring test on the lowercased error string, so wrappers such as data: {...} (streaming 400 bodies) or Error code: 400 - {...} need no separate sample.

Checks

• ruff check src tests — passed
• mypy src — passed
• pytest — 2770 passed, 3 skipped