Natsumix/final-assignment-agent

Use focused Wikipedia queries for named topics

Handle self-contained evaluation tasks without search

Fallback to public GAIA attachment mirror

Analyze captionless YouTube videos with Gemini

Use transcripts as evidence for YouTube tasks

Require fallback web evidence for factual answers

Keep Gemini client alive during evaluation

Process course attachments with Gemini

Use Gemini 3.1 Flash Lite by default

Use Gemini 3 Flash for external runs

Checkpoint evaluation answers and tolerate search gaps

Limit Gemini evaluation to one call per question

Support Gemini for external agent runs

Add external course evaluation runner

Require complete founder names

Align agent output with exact-match evaluation

Route inference to a compatible provider

Add external final assignment agent

initial commit