fix(core): tolerate multiple reasoning_opaque values from Copilot - #52190

Fixes the AI_InvalidResponseDataError reported in #51466 (also #17011, #20434).

Problem

Copilot models that use interleaved thinking — Claude Opus 5/5.5 and Fable 5.1 — emit a fresh signed reasoning_opaque before each tool call, so one streamed response legitimately carries several of them. The chat stream parser treats the second value as a protocol violation:

if (delta.reasoning_opaque) {
  if (reasoningOpaque != null) {
    throw new InvalidResponseDataError({
      data: delta,
      message: "Multiple reasoning_opaque values received in a single response. Only one thinking part per response is supported.",
    })
  }
  reasoningOpaque = delta.reasoning_opaque
}

The throw aborts the turn mid-stream. It reproduces on any multi-tool-call turn with thinking enabled, which is why it shows up as "opencode stops working during normal work" rather than as a specific action.

Why first-wins is the right behaviour

The transport carries exactly one reasoning_opaque per assistant message, and the outbound converter already reconciles multiple parts with a first-wins rule rather than erroring — convert-to-openai-compatible-chat-messages.ts:

if (partOpaque && !reasoningOpaque) {
  reasoningOpaque = partOpaque
}

So the parser was strictly less tolerant than the converter feeding it. This change makes the two consistent: keep the first signature, ignore subsequent ones.

Changes

• openai-compatible-chat-language-model.ts: replace the throw with a first-wins guard, with a comment explaining why multiple values are expected rather than exceptional.
• copilot-chat-model.test.ts: add a multipleReasoningOpaque fixture (interleaved thinking → two tool calls, two signatures) and a regression test asserting no stream error, first-wins metadata, and that both tool calls still arrive.

Verification

• The new test fails on dev with exactly the reported error, and passes with the fix.
• bun test test/github-copilot/ → 36 pass, 0 fail.
• Prettier clean; oxlint warning counts unchanged on both touched files (5 and 3, same as baseline).

Scope / limitations

This is deliberately minimal — it stops a crash and does not attempt to preserve every signature. Full multi-signature fidelity for interleaved thinking would need a wire-format change, since the assistant message has a single reasoning_opaque field; that seems worth treating as separate work. If maintainers would prefer last-wins instead of first-wins, that is a one-line change, but first-wins is what keeps the parser aligned with the existing converter and leaves all current tests green.