BYOK `provider.maxOutputTokens` never reaches the upstream request body (chat completions & responses)

Start at the session.create JSON-RPC handling in the runtime and trace how ProviderConfig, model capabilities, and ProviderModelConfig reach the chat completions and responses request builders. Compare the generated bodies with the documented provider limits; done means the appropriate limit field is sent for each wire API, or the behavior and opt-in requirement are documented clearly.

When using a BYOK (OpenAI-compatible) provider, the max_output_tokens option documented on ProviderConfig is accepted and forwarded to the runtime over JSON-RPC, but the runtime never includes it in the request body sent to the upstream provider — under any configuration path I could find. The same applies to modelCapabilities.limits.maxOutputTokens and to ProviderModelConfig.maxOutputTokens.

The docs for ProviderConfig.maxOutputTokens say:

Overrides the resolved model's default max output tokens. When hit, the model stops generating and returns a truncated response.

As observed, no truncation limit is ever communicated to the provider, so the provider's own default applies instead.

Environment

• github-copilot-sdk (Python): 1.0.14
• Runtime bundle: CLI 1.0.85 (~/.cache/github-copilot-sdk/cli/1.0.85), also reproduced with the VS Code-bundled runtime
• OS: Linux x64
• Upstream: an OpenAI-compatible /v1/chat/completions endpoint (SenseNova), verified to accept and honor max_tokens

Repro

import asyncio
from copilot import CopilotClient

async def main():
    client = CopilotClient()
    session = await client.create_session(
        model="deepseek-flash",
        provider={
            "type": "openai",
            "base_url": "http://127.0.0.1:39221/v1",  # local recording proxy
            "api_key": "sk-...",
            "max_output_tokens": 2345,
        },
        available_tools=[],
    )
    await session.send("hi")
    await session.disconnect()
    await client.stop()

asyncio.run(main())

The proxy sits between the SDK and the real provider and records request bodies verbatim (no TLS issues, plain HTTP hop).

Evidence

1. The Python SDK does forward the option. Tracing the session.create JSON-RPC request shows:

{
  "method": "session.create",
  "params": {
    "model": "deepseek-flash",
    "provider": {
      "type": "openai",
      "baseUrl": "http://127.0.0.1:39221/v1",
      "apiKey": "sk-...",
      "maxOutputTokens": 2345
    }
  }
}

2. The upstream request body does not contain it. Captured POST /v1/chat/completions (the runtime uses an embedded OpenAI JS client — User-Agent: OpenAI/JS 5.20.1):

{
  "model": "deepseek-flash",
  "messages": [...],
  "reasoning_effort": "high",
  "temperature": 1
}

No max_tokens, max_completion_tokens, or max_output_tokens — the field is silently dropped by the runtime.

Variants tested

Configuration path	wire API	max_tokens in upstream body
provider["max_output_tokens"]	completions	no
provider["max_output_tokens"] + model_id="gpt-5.4" + wire_model	completions	no
model_capabilities=ModelCapabilitiesOverride(limits=ModelLimitsOverride(max_output_tokens=...))	completions	no
named providers: providers=[{...}] + models=[{... max_output_tokens: ...}]	completions	no
provider["max_output_tokens"]	responses	no

For the responses wire API the body carries model/instructions/input/tools/reasoning/store/include — still no max_output_tokens, even though that is the exact parameter name the OpenAI Responses API uses.

Expected behavior

At minimum, one of:

1. provider.maxOutputTokens (and models[].maxOutputTokens) is translated into the provider-appropriate request field (max_completion_tokens/max_tokens for completions, max_output_tokens for responses); or
2. The docs clarify that these are runtime-internal limits only (never sent to the provider), and an explicit opt-in is added for putting them on the wire.

Workaround

Run a local OpenAI-compatible proxy between the SDK and the provider that injects the missing field into the request body (e.g. inject max_tokens: 4096 when absent). Verified working with SenseNova.