[Windows][Codex Desktop][Astra] Benign UI question repeatedly ends with misalignment_policy_violation before any tool call

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

26.901.6511.0 (Windows x64). The bundled Codex core recorded in the affected sessions was 0.153.4.

What subscription do you have?

API-key authentication. The local session metadata does not expose a separate subscription name.

What platform is your computer?

Microsoft Windows NT 10.0.19045.0 x64 (Windows 10 Pro, build 19045).

What issue are you seeing?

GPT-6 Astra repeatedly ends fresh Codex Desktop tasks with misalignment_policy_violation for a benign, read-only question about existing frontend behavior.

The prompt, with the private route generalized, was:

On the task tracking page (/admin/.../records), under what condition is the “Task Report” control clickable?

The first attempt ended after 5.4 seconds with no assistant response, tool call, file read, network request, or modification:

This request was blocked by our safety systems. Reason: Potentially unintended activity.
codex_error_info: misalignment_policy_violation

The same prompt was tried five times across two fresh task records. Every attempt ended with the same error after approximately 4.9-8.9 seconds. One retry emitted only a brief sentence saying it would inspect the click condition, then stopped before making any tool call.

All five clean reproductions used gpt-6-astra, medium reasoning. The sessions recorded approval_policy=never and sandbox=danger-full-access, but no consequential action was attempted before the block.

This is not a report about bypassing a valid safety finding. A separate local task reviewed an earlier finding involving database writes through a read-only integration; those stops may have been appropriate and are deliberately excluded from the false-positive reproduction above.

What steps can reproduce the bug?

1. Open Codex Desktop on Windows in an existing software repository.
2. Start a fresh task with GPT-6 Astra and medium reasoning.
3. Ask a simple read-only source question, for example: “On the task tracking page, under what condition is the Task Report control clickable?”
4. Observe that the task ends within seconds with misalignment_policy_violation, before any tool call or code inspection.
5. Retry or create another fresh task with the same prompt. In this occurrence, the result repeated five times.

Diagnostic IDs for OpenAI correlation:

• Session 01a07b36-5fa8-7712-bd9b-981b34610c69
  • Turn 01a07b36-afcd-7673-b516-4c871e396e36
  • Turn 01a07b37-70d4-77f2-8566-91fa8592efaf
• Session 01a07b39-e8a0-7e22-884f-f8e1835f0ea6
  • Turn 01a07b39-ec80-7362-a3d7-878b8e1c13b5
  • Turn 01a07b3a-7ab0-7451-b673-6936e73e3bd1
  • Turn 01a07b3d-8db8-7f83-9e91-a669e05e397b

UTC timestamps ranged from 2026-09-07 09:33:02 to 09:40:33.

What is the expected behavior?

A benign question about a UI control should proceed to read-only code inspection or request clarification. It should not terminate before the first tool call.

If a safety finding from another task is relevant, the UI should identify that relationship and provide an actionable review path. A fresh task should not be silently stopped because of unrelated prior context or state.

Additional information

A read-only scan of local Codex records found 12 distinct misalignment_policy_violation turns across six sessions between 2026-09-07 and 2026-09-08. This issue does not claim that all 12 were false positives. It reports only the five clean reproductions above, where the prompt was benign and the first stop occurred before any tool use.

There were also later stops during ordinary read-only frontend diagnosis in separate Astra tasks, which suggests the problem was not limited to one task record. Cross-task or account-level state is only a hypothesis; the client records cannot establish the backend cause.

Related reports:

• #43042: Astra safety checks during ordinary code questions
• #43192: repeated precautionary-message acknowledgement loop
• #43691: repeated Astra misalignment_policy_violation terminations

Official behavior reference: https://developers.openai.com/api/docs/guides/safety-checks/misalignment-monitoring

Private source code, credentials, hostnames, full prompts, and application data are intentionally omitted.

No source files, tests, or entry points are named. Start by reproducing the five clean cases using the listed steps and correlate the diagnostic session and turn IDs; done means a benign read-only task proceeds to inspection or presents an actionable review path instead of terminating silently.