Loading ToolBench
Ask your teammates when they're not around — MCP server, skill, and session-capture hooks. Zero runtime dependencies.
MindSync defines 4 tools with clear, domain-specific descriptions (avg 180 chars, well above 34-char baseline). All tools have proper JSON Schema with typed parameters and descriptions. Naming follows verb_noun convention (ask_mind, get_file_authors, who_knows_about, escalate_question). However, output schemas are not formally documented, responses are text-only with no structured field definitions. Error handling is present but inconsistent: ask_mind returns NOT_CONFIGURED text; escalate_question lacks explicit error guidance. Parameter descriptions are strong (e.g., 'GitHub username of the teammate, e.g. atillacanuslu'), but no enums or format constraints are declared. Tool composition is sound, each has one responsibility and outputs chain naturally (ask_mind → escalate_question). Security is well-handled: API key injected via env, no secrets in params.
Ask the digital mind of a teammate (by GitHub username). It answers in their words from their own approved material, with citations, and declines when it has nothing — it never guesses. Use when you need a teammate's knowledge and they're not around. Always attribute the result as "<person>'s mind says…", never as the person themselves.
When a mind declined for lack of material, send the question to the real person's MindSync inbox. Their reply is offered into their own approval queue as highest-trust material, so their mind can answer it forever after. Use ONLY after ask_mind declined, only for questions worth a human's time, and only with the user's explicit consent — it interrupts a real person.
Find the main authors of a file, from your team's approved GitHub material (merged-PR file lists — not raw git blame). Direct read, instant, no model. Use before substantially changing a file someone else may own; if another member is the main author, consider ask_mind about intent before rewriting.
Rank teammates by how much of their approved material matches a topic. Direct text search, no model, instant. Use to find who to ask (via ask_mind) about a subsystem, decision, or technology.
Output schemas not formally documented. Responses are text-only with no structured field definitions (e.g., ask_mind returns citations array, confidence score, scope metadata, but LLM must parse from text). Agents cannot reliably extract structured data for downstream tool calls.
No input validation or constraint declarations. Parameters accept free-form strings with no enums, regex patterns, or length limits. E.g., 'person' param accepts any string; no validation that it's a valid GitHub username. LLMs may pass invalid values without guidance.
Error handling inconsistent and incomplete. ask_mind returns NOT_CONFIGURED text; escalate_question has no documented error cases. No guidance on retryability, user-fixable vs fatal errors, or recovery steps. Agents cannot distinguish transient failures from permanent ones.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | 2026-07-28+ | v2 |
No confirmation or dry-run pattern for escalate_question. This tool sends a real message to a teammate (WRITE operation) with no preview or confirmation step. Agents may escalate unintended questions, interrupting real people.
Missing pagination support. who_knows_about and get_file_authors return lists with no limit, offset, or total_count. Large result sets could exhaust context window. No guidance on result limits.