Claude models via the github-copilot provider stream in visible bursts — text arrives in chunks with pauses up to ~500-750ms between them. GPT models on the same provider stream continuously. Happens on every Claude model I've tried (sonnet-4.5, sonnet-4.6).
To rule out OpenCode being the cause, I posted the same prompt directly to the Copilot API using the same OAuth token. The burst pattern is already present in the raw SSE from /v1/messages. GPT endpoints (/chat/completions, /responses) don't have it.
The difference seems to come from how OpenCode routes each model: Claude goes through @ai-sdk/anthropic → POST /v1/messages (Anthropic shim), while GPT uses @ai-sdk/github-copilot → POST /chat/completions. The shim batches SSE frames in a way the OpenAI endpoint doesn't. The toolStreaming = false workaround in copilot.ts (for the shim rejecting eager_input_streaming) makes agentic turns worse since the whole tool-call JSON arrives as one chunk.
I have a small Python script that measures inter-chunk gaps directly against the Copilot API — happy to share if it'd be useful for confirming.
Steps to reproduce
- Set provider to
github-copilot, pick any Claude model (e.g. claude-sonnet-4.6)
- Send a prompt that produces a moderately long response
- Watch text arrive in bursts with noticeable pauses
- Switch to
gpt-4o or gpt-5.4 on the same provider — streams without gaps
OpenCode version: 1.14.30
OS: macOS arm64
Terminal: Ghostty
Claude models via the
github-copilotprovider stream in visible bursts — text arrives in chunks with pauses up to ~500-750ms between them. GPT models on the same provider stream continuously. Happens on every Claude model I've tried (sonnet-4.5, sonnet-4.6).To rule out OpenCode being the cause, I posted the same prompt directly to the Copilot API using the same OAuth token. The burst pattern is already present in the raw SSE from
/v1/messages. GPT endpoints (/chat/completions,/responses) don't have it.The difference seems to come from how OpenCode routes each model: Claude goes through
@ai-sdk/anthropic→POST /v1/messages(Anthropic shim), while GPT uses@ai-sdk/github-copilot→POST /chat/completions. The shim batches SSE frames in a way the OpenAI endpoint doesn't. ThetoolStreaming = falseworkaround incopilot.ts(for the shim rejectingeager_input_streaming) makes agentic turns worse since the whole tool-call JSON arrives as one chunk.I have a small Python script that measures inter-chunk gaps directly against the Copilot API — happy to share if it'd be useful for confirming.
Steps to reproduce
github-copilot, pick any Claude model (e.g.claude-sonnet-4.6)gpt-4oorgpt-5.4on the same provider — streams without gapsOpenCode version: 1.14.30
OS: macOS arm64
Terminal: Ghostty