I’ve now implemented the proposed “quality-gate + fallback” logic

[Eunchan24]

I’ve now implemented the proposed “quality-gate + fallback” logic in the Prepare Agent - please see the issues here:

https://github.com/HKUDS/AI-Researcher/compare/main…Eunchan24:AI-Researcher:main

What’s changed
1. URL validation & must-use pre-clone
• We read reference_codebases from the task JSON, filter out anything that isn’t a well-formed GitHub URL, and clone the rest up-front into reference_repos/.
2. Quality-gate on search results
• The LLM’s GitHub queries are now instructed to only consider repos with ≥ 10 stars, and last commit within the past 3 years.
3. Safe fallback
• If, after filtering search results, fewer than 5 repos remain, the agent automatically falls back to include the pre-cloned reference_codebases (deduped).
4. Prompt update only
• All of this is surfaced in the same instruction template (no tool-chain changes)—the LLM sees a “MUST-USE” list at the top and quality-gate rules in its criteria.

Why this matches our proposal
• Reproducibility: guaranteed inclusion of the user’s authoritative repos
• Precision: filters out low-quality or stale repos before the LLM even reasons
• Simplicity: all logic lives in the Prepare Agent prompt & pre-clone step; downstream tools remain untouched

Please take a look at the diff and let me know if you’d like any tweaks!