Loading ToolBench
A contract-enforced living context platform for AI coding assistants. Provides an MCP server with 13 tools for managing knowledge chunks, lessons learned, and project context.
RunawayContext has 13 tools with schemas and descriptions present, but quality is inconsistent. Tool names follow verb_noun convention (get_brief, search_chunks, propose_lesson_draft), which is good. However, descriptions are terse (10-50 chars for most tools), parameter descriptions are largely absent, and output schemas are not documented. The codebase shows a mini-validator for input schemas, but parameter-level descriptions are missing entirely. Error handling exists but lacks recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear read/write distinctions. This is a solid foundation but falls short of production-grade quality.
Verify audit log integrity (hash-chain validation).
Return the auto-generated brief for a project.
List lesson drafts in the inbox by status.
List registered specialist agents.
Log a fully-formed lesson (HR-2: project_tags required).
Advance a lesson through the maturity curve (scar → active → stable → internalized → superseded → archived).
Propose a knowledge chunk (HR-2: project required).
Parameter descriptions are missing entirely. Tools like search_chunks have 'query', 'project', 'limit' parameters with no descriptions explaining what each controls, expected format, or constraints. LLMs cannot infer parameter meaning from names alone.
Tool descriptions are too brief (10-50 chars). 'Return the auto-generated brief for a project' lacks context on WHEN to call this vs other tools, what the brief contains, or how it differs from search_chunks. Descriptions should be 50-200 chars with actionable guidance.
Output schemas are not documented. Tools return data but the response structure is not specified in the tool definition. LLMs cannot plan downstream calls or extract fields without knowing what fields exist in the response.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 49 | <=2025-11-25 | v2 |
Append a candidate lesson to the drafts inbox (HR-3: drafts only, never inserts into lessons_learned directly).
Regenerate the brief for a project (rebuilds active_lesson_ids and active_chunk_ids).
Full-text search over knowledge_chunks (active rows).
Search lessons_learned by title/what/why text.
Return install-wide statistics (row counts, storage, etc.).
Check the tier/plan of the install (free/pro/enterprise).
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). The Risk field in the metadata shows READ_ONLY vs WRITE, but this is not exposed in the MCP tool definition. Agents cannot distinguish safe tools from destructive ones without explicit annotations.
Error handling lacks recovery guidance. The mini-validator returns error strings but does not guide the LLM on what to do next (retry, ask user, call a different tool). Errors should be actionable.