GPT-5.6 Sol: apparent prior-turn context loss before compaction in Codex Desktop

Start by reproducing the multi-part follow-up scenario in Codex Desktop with gpt-5.6-sol while recording the context indicator and session log. Inspect the context assembly, telemetry, and compaction reporting paths to determine whether prior-thread content is truncated, summarized, routed, or reconstructed before compaction. Done means documenting the responsible behavior and whether effective per-turn context can be exposed without private prompt content.

What version of Codex are you using?

Codex Desktop via ChatGPT; bundled CLI reports codex-cli 0.145.0-alpha.18.

What subscription do you have?

ChatGPT-authenticated Codex session. Plan details omitted.

What platform are you using?

macOS.

Which model?

gpt-5.6-sol

What issue are you seeing?

In a continuing Codex thread, a follow-up that refers to a recently discussed item can appear to lose the semantic state of the immediately preceding turn. The agent may return to an earlier item instead of the item named or implied by the follow-up, despite that item having been addressed in the preceding exchange.

Tool/skill information sometimes appears to be re-read as well. I do not treat re-exposure to those instructions alone as proof of a defect; the reported issue is loss of task-relevant prior-turn detail.

Sanitized telemetry from an affected thread reports:

model_context_window: 258400
latest recorded input tokens: 152977
latest recorded output tokens: 142

The session log contains no explicit compaction event type. This is approximately 59% of the reported effective window, so the observed loss occurs before any recorded automatic-compaction event. This does not prove exactly what the model request retained; it is a request to investigate whether there is another per-turn truncation, summarization, routing, or state-reconstruction step.

Steps to reproduce

1. Start a new Codex Desktop thread using gpt-5.6-sol.
2. Give the agent a multi-part task with clearly distinct items (for example, A, B, and C).
3. Have the agent acknowledge or address more than one item.
4. Send a concise follow-up that focuses on B, using its terms or an unambiguous reference to the preceding discussion.
5. Repeat with several turns while the context indicator remains materially below the reported window and no compaction notice is shown.
6. Observe whether the agent responds to B or instead restarts/repeats an earlier item.

Expected behavior

Until a visible compaction or other context-management event occurs, the model should retain the task-relevant details of the immediately preceding turns and apply a focused follow-up to the intended item.

If Codex deliberately summarizes, omits, or reconstructs prior content before auto-compaction, that behavior should be observable in telemetry or the UI.

Actual behavior

Some focused follow-ups appear to be interpreted against incomplete or stale earlier context. This looks like loss of prior-turn state rather than ordinary disagreement or a single incorrect answer.

Related reports

This is related to, but distinct from, the documented GPT-5.6 Sol window-cap reports:

• #32806: 372,000 raw / 353,400 effective reportedly changed to 272,000 raw / 258,400 effective.
• #31860: GPT-5.6 Sol catalog cap versus the public API model specification.

The API comparison page currently lists GPT-5.6 Sol with a 1,050,000-token context window and a 128,000-token maximum output:
https://developers.openai.com/api/docs/models/compare

Request

Please investigate and clarify:

1. Whether a Codex Desktop turn can receive less prior-thread content than the reported model_context_window before automatic compaction.
2. Whether an internal summarization, truncation, routing, or state-reconstruction step can happen without a recorded compaction event.
3. Whether there is a supported way to inspect the effective context supplied to each turn without exposing private prompt content.
4. Whether this behavior is expected for gpt-5.6-sol under ChatGPT authentication.

Exact model name and public context claim

Codex Desktop offered and the active thread identifies the selected model as gpt-5.6-sol. This is the same canonical model named GPT-5.6 Sol on OpenAI's official API comparison page, which currently states:

Context window: 1,050,000
Max output tokens: 128,000

Source: https://developers.openai.com/api/docs/models/compare

The concern is therefore not a comparison between different model names or variants: Codex offered gpt-5.6-sol, while the official OpenAI page for GPT-5.6 Sol advertises a 1.05M-token window.

I recognize that the cited page is API documentation and may not by itself establish that every ChatGPT/Codex surface must expose the full API capacity. If Codex intentionally has a smaller product-specific window, it should be clearly disclosed in the Codex model picker, plan information, or model documentation before selection. The current server-reported effective window of 258,400 is only about 24.6% of the publicly stated 1.05M context window.

Practical user impact

After Codex reported the active effective window as 258400, I asked an AI assistant within the affected Codex thread which coding agent to use for a large, tightly connected codebase if the context limit is imposed server-side. It recommended using Claude Code as the primary agent because Anthropic documents 1M-context support in Claude Code for specified paid-plan models:
https://support.claude.com/en/articles/8606394-how-large-is-the-context-window-on-paid-claude-plans

This is not an official OpenAI or Codex position, endorsement, or admission. It is a practical user-impact observation: when the Codex product exposes only 258,400 effective tokens for a model publicly documented with a 1.05M-token window, generic tool-selection advice may rationally direct context-sensitive large-codebase work to a competing service. Clear Codex-specific disclosure and an explanation of the lower server-side limit would allow users to make that decision before selecting the model.