Windows Desktop: command executes in 0.3s but Codex task takes ~2 minutes, with reconnect stalls in real workspaces

Screenshot showing that Get-Location completed successfully with exit code 0 and a tool wall time of 0.3 seconds, while Codex reported 1 minute 58 seconds of total processing time.

<img width="1275" height="1650" alt="Image" src="https://github.com/user-attachments/assets/ee298635-c155-48f8-88bb-25416ef2b024" />

<img width="1371" height="420" alt="Image" src="https://github.com/user-attachments/assets/d9e3520b-6da4-4d2a-90cc-f4265087b279" />

Related but not identical to #32314, #31958, and #32315.

Environment

• OS: Windows 11
• Product: Codex Desktop
• Subscription: ChatGPT Pro
• Model primarily tested: GPT-5.6 Sol (High)
• Additional model tested: GPT-5.5
• Workspace: E:\StageR.v1.0
• Date observed: 2026-07-10 to 2026-07-11
• Time zone: UTC+8

Summary

Codex Desktop on Windows shows severe end-to-end latency around local tool calls even when the underlying command runner executes the command quickly.

In the latest isolated test:

• The actual Get-Location command completed in approximately 0.3 seconds.
• Codex took approximately 1 minute 58 seconds to complete the overall task.
• Windows CMD, PowerShell, both C: and E: drives, and the StageR workspace binding were all functioning normally.

In an existing real repository task, the same basic command stalled for more than ten minutes and the task showed Reconnecting 2/2.

This suggests that the underlying Windows command is not the slow component. The delay appears to occur elsewhere in the Codex execution path, such as model/tool scheduling, session state handling, context restoration, reconnect handling, or tool-result propagation.

Minimal command

Get-Location

Latest measured result

• Actual command-runner execution time: approximately 0.3 seconds
• Total Codex task duration: approximately 1 minute 58 seconds
• Windows CMD / PowerShell: normal
• C: and E: drive access: normal
• Workspace binding: normal

Real-workspace behavior

• Get-Location stalled for more than ten minutes in an existing task.
• The task showed Reconnecting 2/2.
• Starting a fresh task improved command execution, but the end-to-end delay remained unusually high.

Control tests completed

The following work normally outside Codex:

• Windows CMD
• PowerShell
• Git
• C: drive
• E: drive
• E:\StageR.v1.0 workspace access

The issue remained after:

• Restarting Codex
• Reinstalling Codex
• Testing under a newly created Windows user
• Testing empty and real workspaces
• Testing GPT-5.6 Sol and GPT-5.5

Expected behavior

A basic command such as Get-Location should complete close to its actual execution time, return its result promptly, and recover cleanly after an interrupted task.

Actual behavior

• A 0.3-second command can result in approximately 1 minute 58 seconds of total Codex processing time.
• Existing real-workspace sessions can stall for more than ten minutes.
• Reconnecting 2/2 may appear without successful recovery.
• Fresh tasks behave better than previously affected sessions.

Assessment

The evidence does not support Windows, PowerShell, CMD, Git, the disk, or the repository as the primary cause.

The likely failure area is within the Codex execution pipeline, potentially involving:

• model-to-tool scheduling
• task/session lifecycle
• context restoration or compaction
• runner session management
• IPC
• reconnect recovery
• result propagation

Request

Please investigate:

1. Time from model decision to tool-call dispatch
2. Time from tool-call dispatch to runner process start
3. Actual process execution time
4. Time from process completion to result delivery
5. Session restoration and context-compaction time
6. Reconnect and retry handling
7. Why a 0.3-second command produces approximately 1 minute 58 seconds of end-to-end latency

I can provide screenshots and additional sanitized timing details if needed.