Loading ToolBench
Personal RAG over your GitHub history (commits, code, review comments), served to Claude Code over MCP.
github-twin has well-structured tool definitions with complete input schemas and descriptions for all 10 tools. Naming follows verb_noun convention (find_*, predict_*, summarize_*, etc.). However, descriptions are inconsistent in depth, some are terse (e.g., 'Find source snippets at HEAD across every indexed target' for find_code), and output schemas are not documented in the visible code. Parameters are well-typed with descriptions, but several tools lack guidance on error recovery or expected output structure. No tool annotations (readOnlyHint/destructiveHint) are present despite clear risk classifications (READ_ONLY vs WRITE). The bootstrap and sync tools are write operations but lack confirmation/dry-run patterns.
Auto-detect a GitHub target (user, org, or repo) from the current working directory and initialize the database.
Generate a cached voice profile (2–3 paragraphs) summarizing a developer's coding style, patterns, and review philosophy.
Find distilled review rules that apply to a code snippet or description.
Find source snippets at HEAD across every indexed target.
Find past review comments on diffs similar to `diff_hunk`.
Find code that matches a description, for style reference.
Output schemas not documented. Tool descriptions do not specify what fields are returned or their types. LLMs cannot plan downstream calls or extract required data without knowing response structure.
Write operations (bootstrap, sync) lack confirmation or dry-run patterns. These are destructive/state-modifying tools but offer no way for agents to preview changes or confirm before execution, risking unintended data ingestion.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite clear risk classifications. Tools are marked READ_ONLY or WRITE in comments but not exposed via MCP tool annotations, limiting agent safety reasoning.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 61 | <=2025-11-25 | v2 |
Return a markdown block of distilled review rules, optionally filtered by language, repo, author, or scope.
Predict whether a diff will be approved, requested changes, or commented on, based on past review patterns.
Return distilled review rules and the comment quotes that motivated them.
Run ingest + embed pipeline: fetch new commits, reviews, and code from GitHub, then embed them into the vector store.
Error handling guidance missing. Tool descriptions do not explain what errors can occur, when to retry, or how to recover. E.g., find_code with no results, bootstrap with invalid GitHub target, sync with network failure.
Inconsistent description depth. Some tools have terse descriptions (find_code: 'Find source snippets at HEAD across every indexed target') that lack context on when to use them vs similar tools (find_style_examples, find_applicable_rules all search code/patterns).