Unclear Context Truncation Behavior in Long Sessions

[niksacdev]

Describe the bug

When the model context limit is reached, GitHub Copilot CLI displays a "Truncated" indicator in the status bar (e.g., claude-sonnet-4.5 (1x) • Truncated) but provides no information about how the context is being managed.
Current Behavior

✅ Copilot shows visual indicator when truncation occurs
❌ No indication of truncation strategy (FIFO/LIFO/summarization/sliding window)
❌ No user control or visibility into what context was dropped?
❌ Silent context loss leads to the assistant "forgetting" earlier instructions/code in some cases

Environment

Model: claude-sonnet-4.5
Platform: MacOS

Affected version

No response

Steps to reproduce the behavior

No response

Expected behavior

No response

Additional context

No response

[wired00]

Yeah wondering how to deal with this, any way to /compact ?

[AlexanderJohnston]

Sometimes I will open up a brand new terminal on a small (~500 LOC) project and ask it to add a single overload. Sometimes, GPT-5.1 is instantly truncated.

Why was it truncated? How is that even possible? What is it loading in the context window? There wasn't even enough raw bytes in the folder to fill out GPT-5.1's context window.

The process is too opaque for an end-user to make sense of it.

Here is my last truncation:

 Total usage est:       2 Premium requests
 Total duration (API):  33s
 Total duration (wall): 6m 2.161s
 Total code changes:    0 lines added, 0 lines removed
 Usage by model:
     gpt-5.1-2025-11-13   32.3k input, 1.7k output, 23.7k cache read (Est. 2 Premium requests)

How does that make any sense at all? It should be impossible to truncate on such a small request. It isn't even using a large cache. This has to be a bug.

Because of this, I never actually trust the model's context window.