fix: allow text files with non-multimodal models (#5137)

[devin-ai-integration]

Summary

Fixes #5137 — TextFile objects passed via input_files incorrectly raised "Model does not support multimodal input" for non-vision models (e.g. gpt-3.5-turbo, claude-sonnet-4.6).

The root cause was that _process_message_files rejected all files when the model wasn't multimodal, without distinguishing text files from binary files (images, PDFs, audio, video).

Fix: Text files are now inlined as plain text content in the message body instead of being rejected. Non-text files (images, PDFs, etc.) still correctly raise ValueError for non-multimodal models.

Files changed:

• base_llm.py — Rewrote _process_message_files to separate text vs non-text files; added _is_text_file() static helper
• llm.py — Same inlining logic applied to both sync and async _process_message_files
• crew.py / task.py — Updated is_auto_injected() to recognize text MIME types so text files don't unnecessarily require the read_file tool
• test_multimodal.py — 17 new tests

Review & Testing Checklist for Human

• Duplicated inlining logic across 3 locations (base_llm.py:648-703, llm.py:2056-2110, llm.py:2155-2209): Verify the logic is identical in all three copies. The async path (_aprocess_message_files) has no dedicated test coverage — consider whether that's acceptable.
• Mixed text+image behavior: When both a TextFile and an ImageFile are attached to a non-multimodal model, the text gets inlined as a side-effect before the ValueError is raised for the image. Confirm this is the desired behavior vs. raising immediately without modifying the message.
• MIME type lists not shared: _is_text_file() in base_llm.py and text_prefixes tuples in crew.py/task.py define the same set of text MIME types independently. A drift between these lists could cause inconsistent behavior. Consider whether a shared constant is warranted.
• Bare except Exception on read_text(): If a TextFile fails to read, it silently falls into the non-text bucket and may trigger the ValueError. Verify this fallback is appropriate vs. surfacing the read error.
• Manual E2E test: Pass a .txt or .json file via input_files to a crew using a non-multimodal model (e.g. gpt-3.5-turbo) and confirm the task completes successfully with the file content visible in the prompt.

Notes

• The 10 pre-existing test failures in test_multimodal.py (e.g. TestLiteLLMMultimodal::test_format_multimodal_content_image) are unrelated to this change — they fail on main as well due to the minimal test PNG not being processed by the current crewai_files library.

Link to Devin session: https://app.devin.ai/sessions/e46e8669f7a5459380d029d403270307

[devin-ai-integration]

Prompt hidden (unlisted session)

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring