What version of the IDE extension are you using?

26.5519.32039

What subscription do you have?

Recently, I have experienced a dual issue with the VSCode extension: a drastic, anomalous surge in quota consumption combined with a noticeable regression in model intelligence and task completion quality during my daily routines.

Which IDE are you using?

VS Code

What platform is your computer?

Linux 5.15.0-177-generic x86_64 x86_64

What issue are you seeing?

Describe the Bug

<img width="1475" height="276" alt="Image" src="https://github.com/user-attachments/assets/fb92a715-ebbd-44a4-92eb-dbef0f6742d4" />

Recently, I have experienced a dual issue with the VSCode extension: a drastic, anomalous surge in quota consumption combined with a noticeable regression in model intelligence and task completion quality during my daily routines.

1. Anomalous Quota Drain: My weekly and 5-hour fast quotas are being drained at an unprecedented speed (fully exhausted within 2 days of regular work). This strongly indicates that Prompt Caching is failing, forcing the system to re-evaluate the full workspace context or chat history on every single request.

2. Intelligence & Quality Degradation: Concurrently, the quality of task completion has significantly dropped. Previously, Codex could complete high-quality tasks autonomously with minimal guidance. Now, it seems to "lose track" of the project context mid-session, requiring frequent, multi-turn manual prompt adjustments and course-corrections to deliver acceptable results.

These two symptoms are likely highly correlated: the failure in context management/caching not only inflates token costs but also causes the model to lose the deep contextual awareness needed for high-quality generation.

What steps can reproduce the bug?

Steps to Reproduce

1. Open a workspace and initiate a coding task or daily routine.

2. Proceed with multi-turn interactions or continuous code generations within the same session.

3. Observe the swift depletion of the quota.

4. Note that the model frequently fails to retain previous instructions, requiring constant micro-management to get the output right.

What is the expected behavior?

Expected Behavior

• Prompt Cache should work seamlessly to optimize token usage.

• The model should maintain consistent high-quality reasoning and contextual awareness across multiple turns without needing redundant corrections.

Additional information

No response