[Windows] Root Codex thread changed reasoning effort during delegated work with unexpected quota impact

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

Codex Desktop originator, source vscode. The affected rollout metadata records bundled Codex CLI/runtime 0.153.4; the About-dialog package version was not captured.

What subscription do you have?

ChatGPT Pro. The local rate-limit snapshot reports plan_type: pro.

What platform is your computer?

Windows x64 with PowerShell. The exact Windows build was not captured in the rollout metadata.

What issue are you seeing?

A long-running Codex Desktop root task changed its effective reasoning effort during delegated work without a corresponding explicit model-picker action being recorded in the local rollout.

The affected task used model gpt-6-astra. Its rollout contains 74 thread_settings_applied records: 65 applied low and 9 applied max. The observed root-thread transitions were:

• low/default -> max/default at 2026-09-09T12:52:52.539Z
• max/default -> low/default at 2026-09-09T13:07:17.383Z
• low/default -> max/default at 2026-09-09T13:11:44.972Z
• max/default -> low/default at 2026-09-09T13:14:22.801Z

The user did ask Codex to accelerate work with the existing Luna worker swarm, so I cannot prove from the client log alone that every transition was unauthorized. However, there is no explicit user message in the corresponding intervals asking to change the root Astra thread to max. The recorded behavior is therefore ambiguous and needs product/server-side attribution: an app setting side effect, delegation/continuation state leak, automatic routing, or an unrecorded UI action.

This matters because an effort change can alter prompt-cache/reuse behavior and materially affect quota consumption. The same local rollout records the following diagnostic snapshot at 2026-09-09T13:14:13.540Z:

thread_token_usage:
  input_tokens:           584,669,424
  cached_input_tokens:    578,185,216
  output_tokens:                811,856
  reasoning_output_tokens:      184,124
  total_tokens:             585,481,280

rate_limits.primary:
  used_percent: 89
  window_minutes: 10080
  plan_type: pro

Around the effort transition, the seven-day Codex meter moved from 87% used at approximately 12:45Z, to 88% at approximately 12:52Z, and 89% at approximately 13:02Z. These figures are account-level snapshots; they do not prove that the entire change was caused by the effort transitions.

What steps can reproduce the bug?

1. On Windows, start a long-running Codex Desktop task with GPT-6 Astra at low reasoning effort.
2. Use delegated worker/subagent threads and continue the root task through tool-heavy work, resumptions, or interruptions.
3. Ask Codex to use the existing worker swarm, without explicitly selecting max for the root Astra thread.
4. Inspect the local rollout JSONL for thread_settings_applied events and correlate them with token_count and token_usage_record entries.
5. Observe whether the root thread changes from low to max and back, and whether the usage meter advances during the same interval.

This exact sequence was observed once in task/thread 01a07d9d-b649-76b1-9ba9-6a9b9e731f10; the independent reproduction rate is not yet known.

What is the expected behavior?

• A thread's selected model and reasoning effort should remain stable until the user explicitly changes them.
• Delegating to Luna workers should not silently mutate the root Astra thread's reasoning effort.
• If Codex intentionally escalates or normalizes effort automatically, the UI and task history should show the change, its reason, and its likely quota/cache impact.
• Reasoning-effort changes should preserve prompt-cache/reuse behavior where the selected model/provider supports that mechanism, or clearly disclose when a large re-read is expected.
• Usage should be auditable by root task, child task, model, reasoning effort, retries, and cache status.

Additional information

The local totals above are diagnostic counters, not a claim that raw local tokens equal billable usage. The requested investigation is a server-side reconciliation of the affected task and timestamp range, including which model/effort actually executed each request and which quota bucket was charged.

Please:

1. Confirm the trigger and effective execution settings for the four transitions above.
2. Determine whether any requests were billed at max or incurred a cache/reuse reset without an explicit root-thread selection.
3. Explain whether delegated work can change the parent thread's settings or whether this was a client/server state mismatch.
4. Provide per-task/per-model usage attribution for the affected interval.
5. If the audit confirms an unintended setting change, incorrect attribution, or avoidable cache-related charge, restore or credit the affected allowance. Account-specific compensation can be handled privately.

No account email, credentials, private prompts, repository paths, source files, or unredacted rollout content are included. Scoped request/response identifiers can be supplied privately if maintainers need them.

Related reports:

• #26930 — Codex Desktop reasoning level resets during delegations/continuations
• #35416 — changing reasoning level results in cache miss
• #41730 — long-running task silently switches reasoning effort
• #42996 — GPT-6 Astra reasoning-effort changes and cache preservation
• #43229 — Windows 0.153.4 high usage with V2 workers and effort-change cache misses
• #41220 — abnormal Codex usage/quota-accounting tracker

Start with the local rollout JSONL for task/thread 01a07d9d-b649-76b1-9ba9-6a9b9e731f10, correlating thread_settings_applied with token_count and token_usage_record around the four listed timestamps. Review related issues #26930, #35416, #41730, #42996, and #43229; done requires identifying the trigger, effective settings, cache behavior, and quota attribution.