[Cursor extension] `model_reasoning_summary = "detailed"` resolves to `summary=none` for new turns

Start with the effective settings in ~/.codex/config.toml and trace the Cursor extension's turn-start configuration using the reported logs. Compare the configured model_reasoning_summary value with the resolved summary=none path and verify behavior through the documented reproduction steps. Done means new turns resolve to detailed and completed reasoning items contain summary parts.

What version of Codex are you using?

Codex IDE extension: openai.chatgpt-26.721.30844-win32-x64

What editor and platform are you using?

• Editor: Cursor
• Platform: Windows x64
• Authentication: ChatGPT account
• Model: gpt-5.6-sol

The exact Cursor version is not present in the captured logs.

What issue are you seeing?

The documented configuration setting:

model_reasoning_summary = "detailed"
model_supports_reasoning_summaries = true
hide_agent_reasoning = false
show_raw_agent_reasoning = true

is ignored when a new conversation/turn is created in the Cursor Codex extension. The extension explicitly resolves the effective turn setting to summary=none, and all returned reasoning-summary items contain zero summary parts.

This report concerns reasoning summaries, not access to raw chain-of-thought.

Expected behavior

A new turn should resolve the effective reasoning-summary setting to detailed (or at least a supported non-none value), and completed reasoning-summary items should contain readable summary parts.

Expected diagnostic shape:

Reasoning summary turn-start config resolved ... summary=detailed
Reasoning summary item completed ... summaryPartCount=>0

Actual behavior

A newly created conversation resolves to:

Conversation created conversationId=<redacted>
Reasoning summary turn-start config resolved concurrentReasoningSummariesFeatureOverrideEnabled=false conversationId=<redacted> requestApprovalPolicy=on-request requestApprovalsReviewer=auto_review requestPermissionProfile=null requestRuntimeWorkspaceRootCount=0 requestSandboxPolicyType=workspaceWrite requestSandboxWritableRootCount=2 resolvedApprovalPolicy=on-request resolvedApprovalsReviewer=auto_review summary=none useAppServerPermissionDefault=false

Every reasoning item then completes without summary content:

Reasoning summary item completed itemId=<redacted> summary=[] summaryPartCount=0 threadId=<redacted> turnId=<redacted>

Steps to reproduce

1. On Windows, install/use Codex IDE extension openai.chatgpt-26.721.30844-win32-x64 in Cursor.

2. Configure ~/.codex/config.toml with:

   model = "gpt-5.6-sol"
   model_reasoning_summary = "detailed"
   model_supports_reasoning_summaries = true
   hide_agent_reasoning = false
   show_raw_agent_reasoning = true
   

3. Fully close Cursor and terminate remaining Codex/app-server processes.

4. Reopen Cursor.

5. Create a completely new conversation.

6. Send a prompt that causes multiple reasoning/tool steps.

7. Inspect the Codex extension logs.

Reproducibility and diagnostics already tried

• Reproduces in a completely new conversation after a full restart.

• Adding model_supports_reasoning_summaries = true does not change the result.

• The model still emits multiple rs_... reasoning items, but each has summary=[] and summaryPartCount=0.

• Starting Cursor with:

  Cursor.exe --enable-proposed-api openai.chatgpt
  

  removes the separate chatSessionsProvider activation error, but does not change summary=none.

• The logs repeatedly show:

  Concurrent reasoning summaries feature override resolved featureOverride=null
  Concurrent reasoning summaries thread-start config resolved featureOverride=null threadStartConfig=null
  Reasoning summary turn-start config resolved concurrentReasoningSummariesFeatureOverrideEnabled=false ... summary=none
  

This suggests that the IDE extension's turn-start configuration or feature-gating path is overriding or failing to propagate model_reasoning_summary from the effective Codex configuration.

Additional notes

Other warnings in the logs (unsupported MCP resources/list, PowerShell shell snapshot, plugin prompt length, /settings/user 403) appear unrelated because summary=none is selected directly at turn start before the empty reasoning items are received.

Raw internal reasoning is not expected to be exposed for hosted models; the issue is specifically that the documented reasoning-summary setting does not take effect.