A tree-sitter code graph over a repository, answering questions from it with BM25 plus graph expansion.
repo2graph exposes 4 tools with minimal schema documentation and sparse descriptions. Tool names are action-oriented (query, answer, build, export) but descriptions lack depth, none explain WHEN to use each tool, what prerequisites exist, or how they chain together. Input schemas are present but incomplete: 'query' and 'answer' accept only a string with no enum/format constraints; 'build' has empty input {}; 'export' lacks format enum despite mentioning 'e.g., json, html'. No output schemas are documented. Error handling is absent, no guidance on retryability, user-fixable vs fatal errors, or recovery paths. The 'build' tool is destructive (WRITE risk) but has no confirmation pattern or dry-run option. Parameter descriptions are minimal (1-2 sentences) and do not state ranges, constraints, or dependencies. No tool composition guidance, unclear how query→answer→export chain together or what IDs flow between them.
Stream a grounded, citation-carrying answer from a packed context
Build or rebuild the code graph index for the repository
Export the code graph in various formats
Search the code graph and retrieve context for a question
build tool has empty input schema {} and no confirmation/dry-run pattern despite WRITE risk. Agents can trigger destructive index rebuilds without safeguards.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., does query return result IDs for answer to use?).
Tool descriptions are 25 - 40 characters, below the 50-char minimum for LLM selection. 'Search the code graph and retrieve context for a question' lacks WHEN to use, prerequisites, or how it differs from 'answer'.
export tool description mentions 'e.g., json, html' but format parameter has no enum constraint. LLMs may hallucinate invalid formats like 'xml' or 'csv'.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
No error handling guidance. Tools do not indicate retryability, user-fixable errors, or recovery paths. A failed query returns nothing, agent cannot self-correct.