AgentClaw exposes 5 tools with significant definition gaps. Three tools (skill_curator, file_write, bash) have minimal descriptions (20-80 chars) and lack comprehensive parameter documentation. The two dynamic RAG/API tools (ks_*) are inferred from code patterns rather than explicit registration, making their actual schemas unverifiable. Input schemas are present but sparse, most parameters lack type constraints, enums, or validation guidance. No output schemas are documented. Error handling is absent from descriptions. The bash tool is particularly risky: no dry-run option, no exit-code guidance in the description, and no recovery hints for failures.

• Add explicit enum constraint to skill_curator.action parameter: {"enum": ["analyze", "status", "backup", "archive"]}. Update description to reference the enum.
• Expand bash tool description to 100+ chars covering: (1) what exit codes mean, (2) timeout behavior, (3) when to use dry-run, (4) recovery guidance for common failures. Add a 'dryRun' boolean parameter for safety.
• Document output schemas for all tools. For bash, specify: {"exitCode": number, "stdout": string, "stderr": string}. For file_write, specify: {"success": boolean, "path": string, "validationErrors": array}.
• For dynamic ks_* tools, explicitly register each tool variant with its own schema in the tool list, rather than relying on runtime templating. Include sample parameter objects in descriptions.
• Add error recovery hints to all tool descriptions. E.g., 'If file_write fails validation, check syntax with bash tool first.' This guides LLM multi-step planning.