A FastAPI-based backend for an AI agent system with LangGraph integration, supporting chat, workspace management, MCP servers, skills, memory, and terminal execution
SlotFlow exposes 20 tools with minimal documentation. All tools have descriptions (10-50 chars, below the 50-200 char baseline for LLM optimization), but most lack depth on WHEN to use them or prerequisites. Input schemas are present but sparse: most tools have 1-2 parameters with type and description, but no enums, ranges, or validation constraints. Output schemas are entirely undocumented, LLMs cannot predict what fields to expect. Error handling is absent from all tool definitions. Naming is verb-forward (workspace_read, sandbox_execute) but some names are ambiguous (agent_reach, skill_execute). No tool declares permissions, idempotency, or destructiveness. The codebase shows sophisticated context-window management (tool_output_offload.py) and dangling-call repair, but none of this safety logic is surfaced in tool definitions.
Reach out to external agents or services
List artifacts in the workspace
Navigate browser to a URL
Save context to archive
Convert documents between formats
Execute commands in Docker containers
Find available skills
Parse documents using MarkItDown
Retrieve information from long-term memory
Output schemas completely undocumented. LLMs cannot predict return types, field names, or structure for any of the 20 tools. This forces agents to guess what data is available and risks failed downstream tool chains.
Descriptions are too brief (10-50 chars vs. 50-200 char baseline). Most lack WHEN to use, prerequisites, or dependencies. E.g., 'Execute code in a sandboxed environment' does not explain language support, timeout, or output limits.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 44 | 2026-07-28+ | v2 |
Store information in long-term memory
Execute code in a sandboxed environment
Search for a specific skill
Execute a registered skill
View and analyze images
Fetch content from a web URL
Search for text patterns in workspace files
List contents of a workspace directory
Read file contents from the workspace
Search for files in workspace
Display directory tree structure of workspace
No error handling guidance in any tool definition. LLMs cannot determine if failures are retryable, user-fixable, or fatal. No recovery paths documented.
Destructive/write tools lack permission gates or confirmation patterns. sandbox_execute, docker_execute, memory_store, and context_archive_save can modify state but have no documented safeguards or dry-run options.
Ambiguous tool names reduce clarity. 'agent_reach' does not convey what happens (call external service? send message?). 'skill_execute' vs 'find_skills' + 'search_skill' creates redundancy and naming confusion.