Five tools with complete Zod schemas and descriptions, but quality is uneven. spawn_codex has good parameter coverage (7 params, all typed and described), but descriptions are terse (10-60 chars, below the 50-200 char baseline for LLM optimization). check_budget, get_mux_status, decompose_task, and health_check have minimal descriptions (8-40 chars). No output schemas documented, LLMs cannot plan downstream calls. Error handling and recovery guidance absent. No tool annotations (readOnlyHint/destructiveHint). Naming is clear (verb_noun pattern), but descriptions fail to explain WHEN to use each tool or what to expect in responses.

• Expand tool descriptions to 80-150 chars. For spawn_codex: 'Spawn Codex CLI to execute a coding task in an isolated git worktree. Use for low-to-medium complexity tasks to optimize costs. Supports verification via tests, linting, or diff review. Returns task ID, branch name, and success status.'
• Document output schemas for all tools. E.g., spawn_codex returns {taskId: string, success: boolean, branchName: string, worktreePath: string, durationMs: number, exit: number, escalatedToClaude: boolean, escalationReason?: string}. This enables LLMs to chain tools correctly.
• Add tool annotations: mark spawn_codex with destructiveHint=true (modifies worktree); mark check_budget, get_mux_status, decompose_task, health_check with readOnlyHint=true.
• Add error recovery guidance to spawn_codex: 'If timeout occurs, retry with complexity=low or increase timeout. If escalatedToClaude=true, the task was too complex for Codex, consider breaking it into smaller subtasks.'
• Validate denyList parameter format in spawn_codex. Document: 'File patterns using glob syntax (e.g., *.log, src/private/**). Invalid patterns will be rejected.' Add server-side validation to reject shell metacharacters.