[macOS] Repeated Bad Request aborts desktop turns after successful commands and on simple follow-ups

What version of the Codex App are you using (From “About Codex” dialog)?

26.903.61454 (build 8378), verified from the installed app bundle metadata.

What subscription do you have?

ChatGPT Pro.

What platform is your computer?

macOS 26.6.2, Apple Silicon.

Darwin 25.6.0 arm64 arm

What issue are you seeing?

A long-running desktop conversation repeatedly aborts with only {"detail":"Bad Request"}. There are 13 failed turns between 2026-09-10 04:10:18 and 04:16:43 UTC. Some fail before an assistant reply; others fail immediately after successful local commands. Occasional simple replies succeed between failures, but the work cannot reliably continue.

Even asking the assistant to file these bug reports failed with the same error after a successful command.

Exact recorded task completion error:

{"last_agent_message":null,"error":{"message":"{\"detail\":\"Bad Request\"}","codex_error_info":"other"}}

Examples from the local session record (UTC):

Time	Turn	Evidence
04:12:43.183	01a08984-91be-7543-9a66-20da1af12f89	Simple “please check again” follow-up failed after 1,297 ms.
04:12:51.580	01a08984-b187-7840-a6a8-d6d90c565448	Asking why Bad Request appeared failed after 1,602 ms.
04:15:55.493	01a08987-5832-77d2-a112-c3e07b6538d0	Read-only shell inspection finished with exit code 0 at 04:15:54.373; the turn failed about 1.1 seconds later.
04:16:04.638	01a08987-a5fe-75a0-bae8-7c05a65cccb6	Another plain-text question failed after 1,002 ms.
04:16:43.611	01a08988-14af-7eb1-9f7e-5239a9ddb213	Assistant promised to file two issues; a command completed with exit code 0 at 04:16:42.675, then the turn failed with no final answer.

What steps can reproduce the bug?

Observed sequence; not a deterministic minimal reproduction:

1. Continue an existing, long-running desktop task with local shell and MCP tools available.
2. Ask it to inspect local application/MCP configuration. Read-only commands can finish successfully, then the turn aborts.
3. Send simple follow-ups such as “please check again”, “hello?”, or “why does Bad Request appear?”. Several also abort in approximately one second.
4. Ask it to report the failure on GitHub. The same failure interrupts that request.

Affected session: 01a0820c-21a3-7381-9817-5dd5952beede.
Model recorded for the failing turns: gpt-6-astra, reasoning effort low, default collaboration mode.
Last available token-count event: 149,004 input tokens / 258,400 model context window (approximately 58%; this is the last recorded usage, not proof of the exact failing request size).

What is the expected behavior?

Successful tool calls should allow the turn to continue. Plain-text follow-ups should receive a response. If a request cannot be accepted, show actionable diagnostic information and a recovery path instead of a generic error that leaves the authorized task unfinished.

Additional information

The issue was reported from a separate functioning task after the affected task could not finish reporting it. No failing endpoint, numeric HTTP status, or root cause has been established from the session error. Multiple MCP integrations were present, but there is no evidence identifying a particular MCP as the cause.

A separate issue will describe missing recent conversation history when returning to this same task. Whether the two symptoms share a cause is unknown.

Related symptom report: #43711 (Windows; this report provides a macOS occurrence, simple-message failures, and timestamps for correlation).

Only selected diagnostic excerpts are included; private project content, local configuration values, credentials, and media URLs are omitted.

Start with the affected session record 01a0820c-21a3-7381-9817-5dd5952beede and correlate the listed failed turns with the successful shell commands. Reproduce the observed follow-up and post-command failures while collecting the missing endpoint and HTTP status. Done means identifying the failing request or condition and providing actionable diagnostics and recovery instead of a generic Bad Request.