Single tool 'swarm_command' has critical definition gaps. Description is present but generic (58 chars, below 10-char minimum for clarity). Input schema visible with 'command' (string) and 'args' (array) parameters, but 'args' lacks item type definition and both parameters have minimal descriptions. No output schema documented. Tool performs write operations (WRITE risk) but lacks error handling guidance, recovery paths, or confirmation patterns. No parameter constraints (enums, ranges, formats). Naming 'swarm_command' is generic and violates verb_noun convention, should be 'execute_swarm_command' or more specific action. No evidence of idempotency, rate limiting, or permission gates despite destructive capability.

• Rename tool to 'execute_swarm_command' or 'invoke_swarm_command' to follow verb_noun convention and signal action intent to LLM.
• Expand parameter descriptions: 'command' should explain valid swarm command names (with enum or pattern constraint); 'args' should specify that each element is a string argument, with examples of valid argument formats.
• Define output schema with typed fields: e.g. { success: boolean, output: string, exitCode: number, error?: string }. Document what each field means.
• Add error handling guidance: 'If command fails with exit code 1, check syntax. If exit code 127, command not found, try list_swarm_commands() first. Retryable errors: timeout, connection refused.'
• Implement confirmation pattern for destructive commands: add optional 'dry_run' boolean parameter (default false) so agent can preview changes before executing.