Assistant message "refusal": null in outbound chat-completion request breaks strict OpenAI-compatible providers (e.g. Gemini)

Start with the Copilot SDK/CLI outbound message serialization and reproduce the issue using the minimal TypeScript session against the Gemini-compatible endpoint. Inspect the captured request in ~/.copilot/logs/process-*.log, then verify that the assistant message no longer contains a null refusal field and that the second request completes successfully after a tool call.

TL;DR

@github/copilot-sdk@1.0.8 (bundled @github/copilot@1.0.73) sends an assistant message containing a literal JSON "refusal": null field in the outbound messages[] array when using a BYOK OpenAI-compatible provider. OpenAI's own API accepts null here, so nobody notices. But Google Gemini's OpenAI-compatible endpoint (generativelanguage.googleapis.com/v1beta/openai) strictly rejects it with 400 Bad Request ("Value is not a string: null"), and this error body is again swallowed by the CLI's retry logic, surfacing only CAPIError: 400 400 400 Bad Request with no indication of the real cause.

This is the same failure class as #1129 (copilot_mcp_server_name leaking into tools[]), but a different field/message, and #1129's fix does not cover this case — confirmed still failing on the latest published versions (SDK 1.0.8, CLI 1.0.73) as of 2026-07-21.

Environment

• @github/copilot-sdk: 1.0.8 (latest stable on npm at time of testing)
• @github/copilot (bundled darwin-arm64 binary): 1.0.73 (latest stable)
• Node.js: v24.16.0
• OS: macOS 15.7.7 (Darwin 24G720)
• Provider: BYOK, type: "openai", baseUrl: "https://generativelanguage.googleapis.com/v1beta/openai/"
• Model: gemini-2.5-flash
• MCP server: local stdio MCP server (Atlassian Jira MCP), at least one tool call in the conversation

Reproduction

Minimal SDK-driven session with any BYOK OpenAI-compatible provider that does strict schema validation, once the conversation reaches a second turn after a tool call has been executed (i.e. once an assistant message with tool_calls needs to be replayed back to the model):

import { CopilotClient, approveAll } from "@github/copilot-sdk";

const client = new CopilotClient();
const session = await client.createSession({
  onPermissionRequest: approveAll,
  model: "gemini-2.5-flash",
  provider: {
    type: "openai",
    baseUrl: "https://generativelanguage.googleapis.com/v1beta/openai/",
    apiKey: process.env.GEMINI_API_KEY!,
  }
});

await session.sendAndWait({ prompt: "Please call a tool, then respond." });

Expected

A normal multi-turn chat: tool call executes, result is sent back, model produces a final answer.

Actual

The first request (before any tool call) succeeds. The second request (after the tool result is appended to history) fails:

CAPIError: 400 400 400 Bad Request

Root cause

Captured the outbound request body via debug logging (~/.copilot/logs/process-*.log, logLevel: 'debug'). The second request's messages[] array contains an assistant message shaped like:

{
  "role": "assistant",
  "content": "...",
  "refusal": null,          // ← non-standard-for-Gemini null value
  "tool_calls": [ ... ]
}

Replaying the exact captured request body directly against Gemini via curl reproduces the failure:

{
  "error": {
    "code": 400,
    "message": "Value is not a string: null",
    "status": "INVALID_ARGUMENT"
  }
}

Removing only the "refusal": null key from that same assistant message and re-sending the identical request returns 200 OK with a valid completion. So the non-standard null value for refusal is the sole cause.

Verified by diffing:

1. Captured request as-is (with "refusal": null) → 400
2. Same request with refusal key removed → 200

Related — second, distinct bug with gemini-3.1-pro-preview

Using model gemini-3.1-pro-preview instead fails on the same kind of second-turn request with a different Gemini-side error:

{
  "error": {
    "code": 400,
    "message": "Function call is missing a thought_signature in functionCall parts. This is required for tools to work correctly...",
    "status": "INVALID_ARGUMENT"
  }
}

Gemini 3.x requires the thought_signature value (returned by Gemini in extra_content.google.thought_signature on the assistant's tool_calls[].function) to be echoed back verbatim on the next request. The CLI does appear to attempt round-tripping this value (it is present in the outbound extra_content.google.thought_signature of the replayed tool_calls entry), but Gemini still rejects it — suggests either a serialization mismatch (e.g. wrapping/escaping) or the signature isn't being preserved correctly end-to-end. Not fully root-caused to the same file/line level as the refusal: null issue above, but flagging it here since it's the same "strict-provider vs. Copilot's OpenAI-compat serializer" problem class, on the same code path.

Workaround

None found yet on the client side

Related issues

• #1129 — same failure class (unknown/non-conformant field breaking strict OpenAI-compatible providers), different field, confirmed fixed for that specific field but not for this one.