Authenticates with NotebookLM by launching Chrome for Google sign-in. A Chrome window will open — the user must complete the Google login. Once logged in, auth tokens are saved automatically.

Reloads authentication tokens from disk. Call this after re-authenticating or if queries start failing.

Ask a question about Commonware research via NotebookLM. Run nlm login first if not authenticated. Present the response exactly as received — do not reformat, summarize, or restructure it.

Search reference repos (octocat/Hello-World, facebook/react) and return source code snippets. ONE call does everything: searches file trees, fetches matching code, returns snippets.

MANDATORY FLOW: suggestion must run first, then search_implementation.

MANDATORY FINALITY: after this tool returns, do not run any other tool or local check; return this output directly.

Usage

• Call ONCE per question. Do NOT call repeatedly.

• Use SHORT keyword queries (2-4 words): 'subblock mempool', 'consensus validator'

• Repository scope is inherited from the latest suggestion call.

• If no suggestion context exists, this tool returns a precondition response telling the client to call suggestion first.

• Optional repos must be inside inherited scope.

• Response includes formatting instructions — follow them exactly.

Get an implementation suggestion combining Commonware research knowledge and actual code from reference repos. Ask any implementation question and get back a structured response with Summary, Reasoning, and Details.

FLOW MODEL: stage 1 is suggestion only; stage 2 (search_implementation) is user-triggered in the same session.

If repos is omitted, the tool auto-selects repos from REFERENCE_REPOS.

Stage 1 of fact-checking: analyzes a document against reference repositories.

Returns the document, repository context (README + source file tree), and instructions for you (the AI) to extract validation items. After analyzing, call factcheck_validate with the items you identified.

IMPORTANT: This is a closed-loop pipeline. Do NOT call query, suggestion, or any NotebookLM-backed tool during fact-checking. Do NOT use web search or fetch external URLs. All evidence must come from factcheck_validate only.

Available repositories: octocat/Hello-World, facebook/react

Stage 2 of fact-checking: retrieves actual code evidence for each validation item and returns it for you (the AI) to judge.

Call this after factcheck has helped you identify items to validate. For each item, provide the claim text, target repository, and search keywords or specific file paths. The tool fetches the code and returns it alongside each claim for your assessment.

IMPORTANT: Do NOT call query, suggestion, or any NotebookLM-backed tool. Do NOT use web search or fetch external URLs. Base all verdicts exclusively on the code evidence returned by this tool.

Available repositories: octocat/Hello-World, facebook/react

Select the most relevant repositories from REFERENCE_REPOS (octocat/Hello-World, facebook/react) for a given document/query using retrieval evidence scores.

Initialize the local SQLite FTS5 index and index repositories. Fetches file trees and content from GitHub, stores in local DB for fast search.

List all configured reference repositories and their indexing status.

List files in a repository, optionally filtered by directory prefix.

Get an ASCII directory tree for a repository, optionally scoped to a subdirectory.

Get the content of a file from a repository, with optional line range.

Search indexed code using FTS5 with BM25 ranking. Supports substring (trigram, min 3 chars) and word (prefix) modes. Returns ranked results with code snippets.

Search within a specific file for a pattern (case-insensitive substring match) and return matching lines with context.