Reposcan

Tools

• reposcan_analyze_fragile_files Find the most breakage-prone files based on churn and fix-PR frequency. Use this when asked 'what file breaks most often?', 'what are the most fragile files?', 'which files are most risky to touch?', or 'what is the most changed file?'. Requires diffs in the dataset (transformers_with_diffs.jsonl). Reports both raw churn (lines added+deleted) and fix-PR frequency (files touched in bug/fix/regression PRs). Files high on both lists are the riskiest. Returns: str: Two ranked lists — most churned files and most fragile (fix-PR) files.
• reposcan_cluster_issues Group all issues into clusters and return the most common themes. Use this when asked 'what are the most requested features?', 'what are the most common complaints?', 'what do people keep asking about?', or any question about patterns and themes across the issue tracker. Returns: str: The largest issue clusters, each with representative titles and URLs.
• reposcan_count_issues_by_keyword Count and rank issues by how often they mention a keyword or regex pattern. Use this when asked 'which model has the most issues?', 'how many issues mention X?', or any counting/ranking question about issue content. Faster and more precise than semantic search for frequency questions. Returns: str: Ranked list of matched terms with their issue counts.
• reposcan_search_issues_and_prs Search the repo index for issues and PRs semantically related to a query. Use this to answer questions like "what issues does model X have?", "how was feature Y implemented?", "what broke in version Z?", or any question requiring specific issues/PRs. Rewrites the query internally to be dense with signal words for best results. Returns: str: Matching issues and PRs with titles, URLs, and full text, sorted by relevance.
• reposcan_search_source_files Find source code files semantically related to a query. Use this when asked 'where is X implemented?', 'which file handles Y?', 'what file is responsible for Z?', or any question about code location. Returns: str: Ranked list of source file paths most relevant to the query.

No tool matches.

As the server's own listing gives them. We never call a tool.