fix(llm): preserve nested OpenAI stream error details #36118

What to build

Decode and preserve OpenAI Responses stream errors whose details arrive under a top-level error object. The current event schema recognizes top-level code/message and response.error, but strips { type: "error", error: { code, message } }, reducing actionable failures such as context overflow to OpenAI Responses stream error.

Regression test

The red protocol-level regression test is committed here:

• Nested stream error regression test

On current origin/v2, the fixture decodes to:

provider-error: OpenAI Responses stream error
classification: undefined

It should retain context_length_exceeded: prompt too long and classify it as context-overflow.

Run from packages/llm:

bun test test/provider/openai-responses.test.ts --test-name-pattern "top-level error field"

Incident evidence and remaining uncertainty

A clean GPT-5.5 Fast request succeeded. The affected session history failed consistently after a provider switch, and a history-boundary bisect localized the first failure to one assistant step containing six web-search results (168,234 characters of tool output). History immediately before that step succeeded.

The installed build did not retain the raw SSE error event, so this test proves a real diagnostic-loss bug matching the observed fallback but does not yet prove which upstream validation rejected that history. Preserve or safely instrument the raw recognized error shape while fixing this so the underlying rejection can be classified.

Acceptance criteria

• Top-level nested OpenAI error details survive schema decoding.
• Context-overflow codes/messages receive the existing context-overflow classification.
• Truly detail-free error events retain the stable fallback.
• Diagnostics distinguish decoded provider failures from genuinely empty stream errors without logging prompts or credentials.
• The linked regression test passes.

Blocked by

None - can start immediately.