Codex runtime: surface precise unhealthy reason to users (rate limit / quota / auth)

[zylos01]

Summary

When activity-monitor marks the Codex runtime unhealthy, the channel reply to the user should explain the concrete reason instead of a generic unavailable/unhealthy message.

Claude runtime already has dedicated handling for usage-limit style failures. Codex should get an equivalent runtime-specific path for OpenAI/Codex CLI failures.

Goal

Improve Codex unhealthy-state user messaging so channel replies can distinguish at least:

• rate limit / TPM-RPM style throttling
• exhausted credits / insufficient_quota
• auth/login failures
• other API/network failures that are known but not quota-related

Scope

• Inspect existing Claude-side behavior and reuse the same activity-monitor / heartbeat surface where possible.
• Add Codex-specific failure pattern detection/classification.
• Persist enough reason metadata for the channel-side unhealthy reply to mention the real cause.
• Keep wording user-facing and actionable.
• Add tests for the new pattern classification and reply wording.

Notes

• We can start with pattern matching based on known Codex CLI / OpenAI error strings and refine after collecting real-world traces.
• Expected external examples include OpenAI/Codex CLI errors around Rate limit reached ... on tokens per min (TPM), insufficient_quota, and login/token exchange failures.