Codex subagent: crash on empty responses + streaming parser not capturing output

[brandontan]

Summary

Two issues with the Codex subagent in roundtable-ai v0.8.0:

Bug 1: list index out of range crash (easy fix)

server.py line 448 accesses agent_responses[-1] without checking if the list is empty:

# BUGGY (line 448)
final_response = summary if config.verbose else agent_responses[-1]

The other three agents (Claude, Cursor, Gemini) already have the guard:

# SAFE — used by claude_subagent, cursor_subagent, gemini_subagent
final_response = summary if config.verbose else (agent_responses[-1] if agent_responses else summary)

Fix: Add the same empty-list guard to the Codex path.

Bug 2: Streaming parser captures zero messages from Codex CLI v0.101.0

After fixing Bug 1, the Codex subagent no longer crashes but returns:

✅ Codex task completed successfully (no detailed output captured)

Meanwhile, running codex exec directly in the same directory works perfectly and returns a response.

The streaming message loop (lines 388–422) checks for message.role == "assistant" and other attributes, but none of the streamed messages from Codex CLI v0.101.0 match these conditions. The agent_responses list stays empty.

This suggests CodexCLI.execute_with_streaming() (from claudable_helper) returns messages in a format that doesn't match the categorization logic in server.py.

Environment

• roundtable-ai: 0.8.0
• codex-cli: 0.101.0
• macOS (Darwin 25.2.0, Apple Silicon)
• Python 3.11

Steps to Reproduce

1. Install roundtable-ai v0.8.0 with Codex enabled
2. Call codex_subagent with any simple instruction
3. Bug 1: Observe list index out of range error
4. After patching line 448, Bug 2: Observe empty "no detailed output captured" response

Expected Behavior

Codex subagent should return the actual response from Codex CLI, similar to how Claude and Gemini subagents work.