Codemux exposes 13 tools across multiple React component files, but critical definition gaps severely limit production readiness. Most tools have minimal descriptions (under 50 chars), no visible input schema validation, and no documented output formats. Tool definitions appear to be inferred from component prop rendering rather than explicitly registered via a schema system. The codebase shows tool names and brief descriptions in UI components (ToolCallBodies.tsx, ToolCallStatus.tsx, tool-visuals.ts) but lacks formal MCP tool registration patterns, parameter type enforcement, and error recovery guidance.
Spawn a subagent
Execute bash commands
Edit file contents with diff-based modifications
Search for files matching a glob pattern
Search for patterns in files
Edit multiple files with diff-based modifications
Edit notebook cells
Read file contents with optional byte offset and limit
No documented input schemas. Tools like TodoWrite, NotebookEdit, Task, and Agent have empty input schemas ({}), making it impossible for LLMs or clients to know what parameters they accept.
Descriptions are too brief (20-40 chars for most tools). Rubric baseline for A+ tools is 194 chars avg (p10=34, p90=392). Examples: 'Execute bash commands' (24 chars), 'Search the web' (16 chars), 'Update task/todo items' (22 chars).
Parameter descriptions missing or incomplete. 'Read' accepts 'path' and 'file_path' (duplicate params), 'offset', 'limit' but no descriptions explain constraints (e.g. offset byte position, max limit). LLMs cannot infer whether offset is bytes or lines.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 30 | 2026-07-28+ | v2 |
Spawn a subagent task
Update task/todo items
Fetch and parse web content from a URL
Search the web for information
Write content to a file
No output schemas documented. Tools do not declare what they return. LLMs cannot plan chaining (e.g. after Grep returns matches, what fields identify the match for follow-up operations?).
Destructive tools lack confirmation or dry-run patterns. Bash and Write can delete files; Edit/MultiEdit can corrupt code. No pattern to require confirmation before execution.
Duplicate/ambiguous parameter names. Tools Read, Edit, MultiEdit, Write accept both 'file_path' and 'path', LLMs will guess which to use, risking parameter confusion.
No error guidance. If Bash fails, Grep finds no matches, or WebFetch times out, no recovery hints are documented. LLMs get no actionable error responses.
Tool definitions inferred from React components, not explicitly registered. Tools appear in UI rendering code (ToolCallBodies.tsx, ToolCallStatus.tsx) but no centralized schema registry or formal MCP server implementation is visible in the provided source.
Bash tool is DESTRUCTIVE but no security boundary documented. No permission checks, no scoping, no audit trail. An agent can execute 'rm -rf /' unchecked.
WebSearch and WebFetch have 'url' and 'query' parameters that could be confused. WebSearch should take 'query', not 'url'. Current naming invites parameter misuse.