BYOK provider with openai wireApi "responses" — maxOutputTokens not applied

[lxwl-fht]

BYOK provider with openai wireApi "responses" — maxOutputTokens not applied

Summary

When using a BYOK ProviderConfig with type: "openai" and wireApi: "responses", the maxOutputTokens field set on the provider is not forwarded to the OpenAI Responses API request body. Output gets truncated at the OpenAI default (~1024 tokens).

SDK version

@github/copilot-sdk@1.0.0

Reproduction

const provider: ProviderConfig = {
  type: "openai",
  baseUrl: "...",
  apiKey: "...",
  maxOutputTokens: 64000,
  maxPromptTokens: 936000,
  wireApi: "responses",
};

const sessionConfig: SessionConfig = {
  provider,
  model: "<openai-model>",
  modelCapabilities: {
    limits: {
      max_context_window_tokens: 1_000_000,
      max_prompt_tokens: 936_000,
      // Note: ModelCapabilitiesOverride.limits has no max_output_tokens field
    },
  },
  // ...
};

Expected

OpenAI Responses request body should include max_output_tokens: 64000 (or the user-supplied value).

Actual

• max_output_tokens is absent / undefined in the outbound request
• Response is truncated at ~1024 tokens (OpenAI default)
• No way to override: ModelCapabilitiesOverride.limits does not expose max_output_tokens, and ProviderConfig.maxOutputTokens is ignored on the responses wire path

Works correctly

• Anthropic provider — Provider.maxOutputTokens is honored
• OpenAI with wireApi: "chat" (untested — please confirm)

Impact

BYOK users on OpenAI Responses cannot configure output length. Long-output scenarios (code generation, reports) are silently truncated.

Possible fixes

1. Forward ProviderConfig.maxOutputTokens to the Responses API request body
2. Or add max_output_tokens to ModelCapabilitiesOverride.limits

[github-actions]

Investigation findings

I investigated this report by examining the relevant SDK source files:

• nodejs/src/types.ts — public SDK type definitions
• nodejs/src/generated/rpc.ts — auto-generated RPC schema (source of truth for CLI protocol)
• nodejs/src/client.ts — session creation / provider forwarding
• nodejs/test/e2e/session_config.e2e.test.ts and test/snapshots/session_config/

---

What I found

This is a confirmed bug with two distinct components:

1. SDK type gap: ModelCapabilities.limits is missing max_output_tokens

types.ts defines ModelCapabilities.limits as:

limits: {
    max_prompt_tokens?: number;
    max_context_window_tokens: number;
    vision?: { ... };
};

max_output_tokens is absent. Since ModelCapabilitiesOverride = DeepPartial<ModelCapabilities>, users cannot set modelCapabilities.limits.max_output_tokens at the SDK level — there's no TypeScript field for it.

Notably, the underlying RPC protocol schema (rpc.ts, the generated contract with the CLI) does include max_output_tokens in ModelCapabilitiesOverrideLimits (line 4860) and ModelCapabilitiesLimits (line 4711). The public SDK surface simply hasn't been updated to expose it.

2. CLI behavior: ProviderConfig.maxOutputTokens not forwarded on the responses wire path

The SDK correctly passes config.provider (including maxOutputTokens) verbatim to the CLI via JSON-RPC (client.ts line 1124). However, an existing E2E test comment (session_config.e2e.test.ts, lines 345–348) already acknowledges the gap:

"the CLI does not echo it as max_tokens on the OpenAI-style wire request, so we don't assert on it directly"

That note is about the chat completions path (where the parameter is max_tokens). For the wireApi: "responses" path, the equivalent parameter is max_output_tokens — and there is currently no test asserting it is forwarded on that path. This is a CLI-side bug, but it is directly observable through the SDK's BYOK provider configuration.

---

Summary

Component	Issue
SDK (types.ts)	ModelCapabilities.limits missing max_output_tokens; ModelCapabilitiesOverride can't express it
CLI (wireApi: "responses")	ProviderConfig.maxOutputTokens not included as max_output_tokens in the outbound Responses API request

The reporter's workaround options (forwarding maxOutputTokens to the Responses request, or adding max_output_tokens to ModelCapabilitiesOverride.limits) are both valid fixes. Labelling as bug.

Generated by Bug Handler for issue #1595 · sonnet46 1.6M · ◷