DuraLang exposes 3 tools (dura__tool, dura__llm, dura__mcp) that are Temporal activity wrappers, not user-facing agent tools. Tool names violate verb_noun convention (dura__ prefix is framework-internal, not action-oriented). Descriptions exist but are terse (30-60 chars) and lack LLM-optimized guidance on WHEN to use each tool or what downstream actions they enable. Input schemas are present with type declarations but lack parameter descriptions, critical for LLM reasoning. No output schema documentation. Error handling is present in code (ValueError/TypeError caught) but not surfaced in tool descriptions. These are infrastructure tools, not domain tools, they lack the composition patterns needed for agent planning.

• Rename tools to action verbs: 'execute_tool', 'invoke_llm', 'call_mcp_tool'. This makes intent clear to LLMs without reading descriptions.
• Add parameter descriptions to all inputs. E.g., 'tool_name (string): Name of the registered LangChain tool to execute. Must match a tool created inside the @dura function.' Include constraints and examples.
• Document output schemas in tool descriptions. E.g., 'Returns ToolActivityResult with fields: output (str), tool_call_id (str), error (str|null).'
• Expand descriptions to 100-200 characters with WHEN/WHY guidance. E.g., 'Execute a registered LangChain tool within a Temporal activity. Use this to invoke tools that require durable retry semantics. Non-retryable errors (ValueError, TypeError, KeyError) are caught and returned as error results; retryable errors (network, timeout) trigger Temporal retry.'
• If these tools are meant for internal framework use only, consider removing them from the public tool list and documenting them as private activities. If they are meant for agent use, redesign them as domain-level tools that abstract Temporal complexity.