Claude Code plugin for OpenAI Codex subagents, Codex Spark, parallel code review, and long-running Codex sessions through a daemonless MCP server.
Server has 12 tools with mixed quality. Core tools (codex_task, codex_task_group, codex_followup) have detailed schemas and descriptions (100-150 chars each), but lack critical error guidance and recovery patterns. Legacy tools (ask_codex, run_agent, run_agents) are minimally described. Debug tools are hidden behind env flags. Parameters are well-typed with enums (reasoning, sandbox, model_preset), but descriptions lack actionable constraints (e.g., timeout_ms has no min/max bounds). No tool declares permissions or security scopes. Output schemas are not documented in tool definitions. Error handling is absent, no guidance on retryability, user-fixable vs fatal errors, or recovery steps.
Legacy tool for running a Codex task. Deprecated in favor of codex_task. Hidden unless CODEX_SUBAGENTS_ENABLE_LEGACY_TOOLS=1.
Interactive tool selection helper to recommend the best Codex tool for a given task. Hidden unless CODEX_SUBAGENTS_ENABLE_DEBUG_TOOLS=1.
Diagnostic tool for troubleshooting Codex subagents issues. Returns detailed diagnostics about Codex binary, environment, and server state. Hidden unless CODEX_SUBAGENTS_ENABLE_DEBUG_TOOLS=1.
Export a debug bundle containing logs, diagnostics, and session state for troubleshooting. Hidden unless CODEX_SUBAGENTS_ENABLE_DEBUG_TOOLS=1.
Continue, steer, wait on, or cancel an existing Codex session. Modes: continue (add input), steer (change direction), wait (block until completion), cancel (stop execution).
Get current Codex subagents server status, queue depth, session limits, and diagnostics. Hidden unless CODEX_SUBAGENTS_ENABLE_DEBUG_TOOLS=1.
No error handling guidance. Tools lack recovery patterns, no indication of retryability, user-fixable vs fatal errors, or next steps on failure. E.g., codex_task timeout or session not found returns nothing actionable.
Output schemas not documented. Tool definitions do not specify what fields are returned or their types. LLMs cannot plan downstream calls or extract required data (e.g., session_id structure, result format).
Numeric parameters lack bounds. timeout_ms has no min/max; reasoning effort levels not constrained by enum; advanced options undocumented. LLMs can pass invalid values (negative timeouts, unknown reasoning levels).
Inferred effective spec: 2025-06-18+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 47 | 2025-06-18+ | v2 |
Run a single independent Codex subagent task with a description and prompt. Returns a session_id for background tasks or failures, or a completed result for foreground tasks.
Run multiple independent concurrent Codex tasks and return one combined response with rolled-up per-task findings.
Display comprehensive usage guide for Codex subagents tools, including best practices and tool selection guidance. Hidden unless CODEX_SUBAGENTS_ENABLE_DEBUG_TOOLS=1.
Wait for the first of multiple Codex sessions to complete without busy-polling.
Legacy tool for running a single agent. Deprecated in favor of codex_task. Hidden unless CODEX_SUBAGENTS_ENABLE_LEGACY_TOOLS=1.
Legacy tool for running multiple agents in parallel. Deprecated in favor of codex_task_group. Hidden unless CODEX_SUBAGENTS_ENABLE_LEGACY_TOOLS=1.
Legacy tools (ask_codex, run_agent, run_agents) have minimal descriptions (35 chars) and are marked deprecated but still exposed. Descriptions do not explain why they are deprecated or what to use instead.
No permission or scope declarations. Tools that execute code (codex_task, codex_task_group) do not declare required permissions (e.g., 'execute:code', 'write:filesystem'). Audit trail and least-privilege configuration impossible.