How can I increase Copilot Cloud Agent context window to avoid token-limit failures?

[visirion07]

Select Topic Area

Question

Copilot Feature Area

Copilot Coding Agent

Body

I’m using a GitHub Copilot Cloud Agent inside a GitHub Actions runner, and I repeatedly hit this error when the agent builds a large prompt (diffs + file context + tool outputs):

CAPIError: 400 prompt token count exceeds the limit
code: model_max_prompt_tokens_exceeded

The current limit appears to be ~64K tokens. Modern models like Claude 3.5/4.5 or GPT-4.1 support 200K–1M context windows.

My question:
Is there any way to configure Copilot Cloud Agents to use a larger context window or select a model with higher token limits? For example:

• Switching the underlying model to one with 128k or higher context
• Increasing maxDiffSize / maxPromptTokens in the agent config
• Any GitHub setting or policy that enables a larger context model for cloud agents
• Any upcoming support for extended-context models

Here is the redacted log pattern:

CAPIError: token count > 64000
...
fallback completions succeed
...
forceExit is shutting down the process

Looking for guidance on removing or raising this 64K cap so large PRs and long diffs don’t cause the first completion to fail.

[liy8]

Unfortunately, I don’t think there’s a way to increase the Copilot Cloud Agent’s token limit right now. But if you use VS Code Insiders with Copilot Chat (GPT-4o), you can get a 128K token context window, i think.

[ryanmohib]

its likely not possible but here's some strategies i use personally so you can effectively manage the context you provide to the agent to avoid hitting those limits and get better results:
1-Provide Explicit, Relevant Context: Open only the files and select the code directly relevant to your current task. Irrelevant open tabs can consume valuable context space.
2-Keep Prompts Concise and Focused: Formulate your prompts clearly and avoid providing entire documents if only specific sections are needed.
3-Break Down Large Tasks: For complex projects or features, divide the work into smaller, more manageable tasks or steps. This keeps each interaction focused and within the token limits.

(Hope it helps even if a little)

[XrayX]

that's cool actually it may help a little honestly:)

[mecodeatlas]

Hi @visirion07 👋🏾
Just checking in - did @'s response help answer your question? If so, can you please mark their response as the answer? When you mark a question as an answer, GitHub will highlight the comment and replies to the comment to help visitors quickly find the answer.