[sepehrtaji-dev]

Summary

This PR fixes an edge case in GEPA request evidence selection where genuine user messages could be incorrectly excluded due to inconsistent trust metadata.

Previously, messages were withheld whenever "is_trusted" was false, even when "provenance" identified them as originating from the user. This could cause legitimate user requests to be missing from GEPA evidence.

Changes

• Retrieve "provenance" alongside "is_trusted" when selecting request evidence.
• Preserve messages with "provenance = "user"", even if their trust flag is false.
• Continue withholding untrusted messages originating from non-user sources.
• Add regression tests covering trusted and untrusted user messages, as well as untrusted non-user messages.

Testing

Added "tests/test_gepa_evidence_request_provenance.py" to cover the relevant cases. The test suite has not yet been executed in this environment, so CI validation is still needed.

Related to #747.

This change is intentionally limited to GEPA evidence selection and does not modify trust policies elsewhere.

[OAraLabs]

Thanks for digging into this. You're right that real requests go missing from GEPA evidence, and #747 points the same way. I can't take this approach, though. Today the rows with provenance="user" and is_trusted=0 aren't mostly mislabelled human turns: most are tool-result turns, which are stored as user-role messages and carry outside content (web pages, files, tool output). Letting them through would put that content in front of the model that writes skills. The is_trusted check is the only thing that keeps it out, so it needs to stay.

What would fix the missing requests is having _request skip tool-result rows and look for the newest message the user actually typed. After a tool round, the newest user-role row is usually a tool result, which is why the request comes back empty. Would you be up for reworking it that way, with a test where the newest user row is a tool result? The is_trusted=0 label on image-upload turns is a separate fix we'll handle under #747. Thanks again!

If you'd rather not, no problem: I'm happy to take it from here with you as co-author.

[sepehrtaji-dev]

Update: I pushed a follow-up commit ("befda6086783a90cbd13227f1ddcf1844169c19d") addressing the trust/provenance concern.

• Kept the "is_trusted" security check intact so untrusted tool output is never included as GEPA request evidence.
• When the newest "role='user'" row is an untrusted tool result, the code now recovers the latest earlier trusted user request instead of exposing the tool output or dropping the real request.
• Untrusted injected turns from non-user provenance remain withheld rather than falling back to an older request.
• Added regression tests for tool results, untrusted injected turns, missing recoverable requests, and normal trusted requests.

Validation shows:
Focused tests: 4 passed.
Ruff: passed.
The broader GEPA test suite could not be

collected in this environment because "croniter" is missing. CI checks are running on the new commit.

Please let me know if I could handle it!

Chatgpt 6 was used to fix this🤖

[OAraLabs]

Thanks for sticking with this. You got the important part right: keeping the trust check. I pushed one small commit on top to cover a last edge case (a tool round after an untrusted turn) plus a test for it, so there's nothing more you need to do. Merging once CI is green. Really appreciate it!