Copilot CLI loops infinitely and fails apply_patch due to JSON-wrapping error

Describe the bug

While executing regular coding tasks that require file modifications, the Copilot CLI frequently fails to apply patches. The CLI enters a loop where it attempts to edit a file, fails with a JSON-wrapping error, and retries the exact same payload repeatedly.

This causes a task that should take 2 steps to take 30+ steps, resulting in massive latency, tool frustration, and high token consumption before it either randomly recovers or fails entirely. Many models are failing to adapt to apply_patch. I havent seen issue with sonnet 4.6 and gpt-sol but other models like sonnet 5, luna, terra are failing most of the times.

Error Message / Trace
When the CLI attempts to apply an edit, it outputs the following error:

✗ Edit 
 └ Failed to parse patch: The first line of the patch must be '*** Begin Patch' 
   The input begins with '{' and appears to be JSON-wrapped. apply_patch is a 
   freeform tool; pass raw patch text directly. Do not retry this payload.

Behavior Details
Infinite Loop: Once this error is triggered, the model consistently retries the same payload 10 to 30+ times in a row despite the warning "Do not retry this payload".

Temporary Recovery: Eventually, the model might format the tool call correctly by accident, but the error quickly resurfaces on subsequent edits.

CLI vs. VS Code: This behavior is highly localized to the CLI. Using the same prompts and models in the VS Code Copilot Chat interface works perfectly without any patch-parsing issues.

Affected version

GitHub Copilot CLI 1.0.80

Steps to reproduce the behavior

This is intermittent but seen in most of medium complexity prompts.

• Open Terminal ( I did on iterm)
• start interactive session (copilot) - i did not append any flag options ( i opened a session on TMUX)
• Trigger a medium complexity prompt which would need to change your sourcecode files.
• It should trigger the error - It generally happens mid sessions not in the first prompt itself. but yea this is intermittent behaviour

Expected behavior

No response

Additional context

• OS: Ubuntu server
  Tool: Github Copilot CLI (Latest Version)
  Comparison: Not observed in VS Code Copilot Chat (works successfully there)
  LLM Models tried : Both GPT-5.6-Luna and Claude-sonnet-5

No source files or tests are named. Start by reproducing the issue in Copilot CLI 1.0.80 on Ubuntu, using an interactive session and a medium coding prompt that edits files, then compare the CLI behavior with VS Code Copilot Chat. Done means apply_patch accepts the tool payload and the CLI stops retrying the same failed request.