Selected model is at capacity. Please try a different model.

Environment & Subscription

• Application: OpenAI Codex Desktop App (macOS/Windows)
• Subscription Tier: ChatGPT Pro ($200/month)
• Models Affected: 5.5, 5.4

The Problem

As a paying Pro user ($200/mo), the Codex desktop application is currently delivering an absolutely unacceptable, broken experience. During multi-tasking workflows, the application repeatedly encounters the following backend failure:

"Selected model is at capacity. Please try a different model."

Instead of handling this backend overload gracefully, the Codex app reacts destructively:

1. Subscription Limits are Wasted: The agents send massive context payloads (worktrees, files, thread history). When the server rejects the completion due to capacity, the usage limits/quotas are still deducted from my Pro account. I am losing expensive limits for zero output.
2. Context and State are Permanently Lost: The agent thread breaks mid-execution. The context accumulated by the agent is dropped, corrupting the workflow execution state.
3. Zero Proactive Notification: The app does not check server health before initiating tasks, leading to mid-pipeline crashes.

Financial & UX Impact

Paying $200 per month implies enterprise-grade reliability or, at the very minimum, fault-tolerant software. Currently, I am paying to lose my context, burn my allowed requests on server-side failures, and manually restart broken pipelines. ### Expected Behavior (Urgent Fixes Required)

1. Pre-flight Checks & Graceful Pausing: Codex must verify model capacity before dispatching heavy prompt payloads. If the model is at capacity, the agent execution thread should PAUSE automatically and retry gracefully without breaking the pipeline or dropping context.
2. Quota/Limit Rollback: Failures resulting in 503 Service Unavailable or Model at capacity responses MUST NOT count against the user's daily/hourly Pro limits.
3. Proactive In-App Alerting: Display a persistent banner inside the Codex UI when models are degraded, rather than letting agents run blindly into brick walls and ruin hours of isolated worktree progress.

Please address how OpenAI plans to compensate Pro users for the wasted quotas and provide an immediate hotfix for context retention during outages.