Gopher has 10 tools with clear naming (verb_noun pattern) and reasonable descriptions (avg ~150 chars). All tools have input schemas with type definitions and parameter descriptions. However, output schemas are undocumented, tools return JSON strings but the structure is not formally declared. Error handling is minimal; most tools return success strings without guidance on failure modes or recovery. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Parameter descriptions are adequate but lack format constraints (e.g., no regex for dot-notation paths, no bounds on limit integers). The digest_transcript tool has complex multi-step behavior (add/overwrite/refuse) but the response structure is only visible in code, not in formal schema documentation.
Delete a key by dot-notation path (e.g. 'projects.gopher.status').
Read a transcript, extract facts with the local model, and merge them into the context store. The destination is not a parameter. It is always the configured store, so this tool cannot be pointed at an arbitrary file. A fact that would overwrite an existing value is applied and recorded with what it replaced. A fact that would destroy a section is refused and reported, without stopping the rest of the batch.
Fetch a GitHub repo and return a Markdown digest with directory tree and top files.
Append a timestamped markdown entry to the diary. The entry body may contain its own markdown, headings included. Optional tag (e.g. 'GSOC', 'GATE') appears in the header.
Read the context store as JSON. With no argument this returns everything, which is fine while the store is small and wasteful once it is not. Pass a dot path to read one section, or use search_context when you know what you are looking for but not where it lives.
Output schemas undocumented. Tools return JSON strings (read_context, search_context, digest_transcript, etc.) but the structure of returned objects is not formally declared. LLMs cannot plan downstream operations without knowing what fields to expect.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Tools have explicit risk levels (READ_ONLY, WRITE, DESTRUCTIVE) in source but these are not exposed via MCP schema. Agents cannot distinguish safe from dangerous operations.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | <=2025-11-25 | v2 |
Return the last N entries from the diary (default 10).
Return the full contents of the digest log.
Find stored facts whose key path or value contains *query*. Case-insensitive substring match. Matches on the key come before matches on the value. Returns at most *limit* hits and says how many there were in total, so you know when you are seeing a slice.
Extract facts from a transcript and return them without writing. Runs the same validation the write path does, so what this shows is what a digest would actually store.
Write a value at a dot-notation path (e.g. 'projects.gopher.status'). Creates missing sections along the way. Refuses, rather than discarding data, if a key on the path already holds a value instead of a section, or if the final key holds a section that a value would replace. Delete the named key first if you meant to change its shape.
Parameter constraints missing. 'limit' in search_context has no bounds (could be 0, negative, or 1M+). 'prefix' and 'key' in context tools accept dot-notation but no regex pattern or format constraint is declared. 'transcript_path' in digest_transcript has no validation hint.
Error handling lacks recovery guidance. Tools return success strings ('Set key = value') but no error responses are documented. What happens if update_context tries to overwrite a section with a value? If digest_transcript fails to parse the transcript? Agents have no guidance on retry or fallback.
STDIO transport only. Server is not remotely accessible and cannot be used by hosted MCP clients. Limits deployment to local/CLI-only scenarios.