VS Code Copilot with a local model via custom endpoint: agent mode stops halfway with `Sorry, no response was returned.` (fixed by lowering maxInputTokens)

Context

Source(s): https://www.v2ex.com/t/1225550 (fix confirmed by the original poster in the last reply) Status: reported working by source, not reproduced by us.

Problem

A local model (Qwen3.5-9B-Q4_0) was added to GitHub Copilot Chat in VS Code as an OpenAI-compatible custom endpoint. Ask mode worked, but in agent mode the reply often stopped halfway and Copilot showed "Sorry, no response was returned." The original poster had maxInputTokens set to 108000.

Error (redacted / normalized)

Sorry, no response was returned.

Applies when

VS Code with GitHub Copilot Chat, a custom language-model provider entry using "apiType": "chat-completions" and "toolCalling": true, pointing at a self-hosted OpenAI-compatible server. July 2026.

Confirm it is this problem

1. The error appears in agent mode while Ask mode works (reported by the original poster). Our inference: it gets more frequent as the conversation and tool output grow. → expect: the failure correlates with context size.
2. Compare the entry's maxInputTokens with <server_ctx>. (our inference) → expect: maxInputTokens is larger than what the server can really accept.

Not this when / boundary

A commenter said Qwen3.5-9B-Q4 still loops in agent mode because the model isn't capable enough; that is a different problem from this limit (our inference). Another commenter said the same problem happens when Copilot goes through a third-party relay service; that may have a different cause. If ask mode also fails, check the endpoint URL and key first (our inference).

Root cause

(our inference) Copilot uses maxInputTokens to decide how much history and tool output it may send. When that value is larger than the context the local server really serves, agent-mode requests eventually exceed the server's limit; the server errors out or truncates, and Copilot only shows its generic "no response" message. The original poster only reported that lowering the value fixed it.

Variables

<server_ctx>: the context length your local server actually loads the model with (for example the context size flag of llama.cpp or the num_ctx of Ollama). <max_out>: the maxOutputTokens value in the same entry.

Fix

1. Find <server_ctx> in your local server's launch settings or startup log.
2. In the Copilot custom model entry, lower maxInputTokens to below <server_ctx> minus <max_out> (our inference for the exact value; the source did not say what number they used). → expect: the original poster reported agent mode worked after reducing maxInputTokens.
3. Reload the VS Code window so the model list picks up the change, then start a new chat.

How to verify

Run an agent-mode task with several tool calls → expect: it completes without "Sorry, no response was returned."

What did not work

none reported

Rollback

Restore the previous maxInputTokens value.

Risks

A lower input limit means Copilot keeps less history and tool output per request, so long tasks may lose earlier context sooner.

Verified scope

VS Code GitHub Copilot Chat custom endpoint models, July 2026.

Becomes obsolete when

Copilot reads the real context length from the endpoint or shows the server's actual error instead of a generic message.