codex exec --json emits non-fatal app-server stream lag as item.type="error"

Summary

In a long codex exec --json automation run, Codex completed successfully with a final assistant message, turn.completed, and process exit code 0. During the same run, the JSON stream also emitted item-level errors for app-server backpressure:

{"type":"item.completed","item":{"id":"item_115","type":"error","message":"in-process app-server event stream lagged; dropped 133 events"}}
{"type":"item.completed","item":{"id":"item_116","type":"error","message":"in-process app-server event stream lagged; dropped 96 events"}}

Because these are represented as item.type="error", downstream automation that treats structured Codex errors as fatal can misclassify a successful run as failed.

Observed Environment

• Codex CLI/app version in logs: 0.125.0
• Model: gpt-5.5
• Auth mode in logs: Chatgpt
• Transport in logs: responses_websocket
• Run mode: codex exec --json
• The run used verbose tracing (RUST_LOG=debug) and completed after a long automation task.

Expected Behavior

Non-fatal app-server backpressure should be distinguishable from fatal model/tool errors. Possible options:

• emit it as a warning notification/event instead of an item.completed with item.type="error";
• include a severity or fatal: false field;
• document this exact message as a non-fatal item error that clients should ignore when the turn completes.

The app-server README currently describes generic runtime warnings as non-fatal warning notifications, while error events are described as mid-turn errors that carry failure payloads. This stream-lag event behaves like a warning because the turn and process still succeed.

Actual Behavior

The stream lag messages appeared as item-level errors. A downstream runner had previously fixed a different bug by treating Codex error, turn.failed, and item.type="error" as fatal even when process exit code is 0. That was correct for unsupported-model and quota failures, but this app-server stream-lag event created a false failure after the task succeeded.

Minimal Consumer Reproduction

This is the problematic event sequence for consumers:

{"type":"thread.started","thread_id":"thread_issue_1696"}
{"type":"item.completed","item":{"id":"item_115","type":"error","message":"in-process app-server event stream lagged; dropped 133 events"}}
{"type":"item.completed","item":{"id":"item_419","type":"agent_message","text":"Done. PR is updated and ready for review."}}
{"type":"turn.completed","usage":{"input_tokens":26436275,"cached_input_tokens":26045952,"output_tokens":63343}}

A consumer cannot reliably tell from item.type="error" alone whether this should fail the automation.

Workaround

In our runner, we are filtering this exact item-level message as non-fatal:

/^in-process app-server event stream lagged; dropped \d+ events?$/i

We still treat top-level type="error" and turn.failed as fatal, and still treat other item-level errors as fatal.

Related Downstream Context

Downstream repository issue: link-assistant/hive-mind#1696

The issue was opened because a successful Codex run updated a pull request, pushed a commit, marked it ready, and confirmed CI, but the wrapper then posted a failure comment due to this item-level stream-lag event.