Two tools with partial schemas and minimal descriptions. claw_execute has a detailed input schema with 6 parameters (code, model, tools, maxTokens, sandbox, timeout) but description lacks actionable context about safety, reversibility, or error recovery. claw_undo has minimal schema (only executionId) and vague description. Neither tool documents output structure, error cases, or recovery guidance. No tool annotations (readOnlyHint/destructiveHint/idempotentHint). Naming is verb-based but descriptions fall short of LLM-optimization baseline (194 chars avg). Risk labels present but not integrated into descriptions.

• Expand claw_execute description to 150-250 chars: 'Execute code safely in an isolated sandbox with full audit trail. Supports Python, JavaScript, Bash. All executions are logged and can be undone via claw_undo. Specify allowed tools (read_file, write_file, run_command, git_commit, github_api, linear_api) to restrict capabilities. Returns executionId, status, stdout, stderr, and execution time.'
• Rewrite claw_undo description: 'Undo a previous code execution by executionId. Restores file state, reverses git commits, and clears side effects. Returns success status and list of reverted changes. Only works if execution completed; in-progress executions cannot be undone.'
• Add output schemas: claw_execute returns {executionId: string, status: 'success'|'error'|'timeout', stdout: string, stderr: string, duration_ms: number}. claw_undo returns {success: boolean, reverted_changes: string[], error?: string}.
• Add tool annotations to schema: claw_execute with destructiveHint=true (modifies files, runs commands), idempotentHint=false (side effects). claw_undo with idempotentHint=true (safe to retry).
• Constrain 'tools' parameter: change from free-form array to enum: ['read_file', 'write_file', 'run_command', 'git_commit', 'github_api', 'linear_api']. Add description: 'Whitelist of tools the code can invoke. Omit to deny all external tools.'