Strong foundation with clear naming, comprehensive descriptions, and proper schema definitions. All 6 tools have verb-noun names (bridge_status, codex_read, codex_run, codex_reply, codex_job_status, ask_chatgpt) and detailed descriptions (100-200 chars). Input schemas use Zod with type constraints and descriptions. Tool annotations present (readOnlyHint, destructiveHint, idempotentHint, openWorldHint). Key gaps: output schemas not documented in code; error handling guidance minimal; no pagination/limit enforcement visible for list operations; ask_chatgpt lacks enum constraints for model parameter.

• Document output schemas for all tools. Define ToolResult structure explicitly: {status: 'success'|'error', data?: {...}, jobId?: string, error?: string}. Include in tool descriptions or separate schema docs.
• Add enum constraint to ask_chatgpt 'model' parameter: z.enum(['gpt-4', 'gpt-4-turbo', 'gpt-5.5', ...]) with description listing supported models.
• Enhance error responses with recovery guidance. E.g., 'Job timed out. Call codex_job_status(jobId) with 2-5 second polling interval. If status remains running after 5 minutes, the job may be stuck.'
• Document codex_job_status polling contract: expected polling interval (e.g., 1-2 seconds), max retries before giving up, and how to distinguish transient vs permanent failures.
• Clarify ask_chatgpt 'reasoningEffort' parameter: enumerate valid values (e.g., 'low', 'medium', 'high') and which models support it (e.g., 'Only gpt-4-turbo and later').