[BUG] xAI Responses API path replays persisted Anthropic thinking blocks as plain assistant text

[shaun0927]

Problem (one or two sentences)

When a persisted conversation contains Anthropic thinking blocks and that history is replayed through the xAI Responses API path, Roo converts the hidden reasoning into ordinary assistant text ([Thinking] ...) instead of preserving or safely dropping it.

Context (who is affected and when)

This affects users who switch providers or resume existing conversations where earlier turns were produced by an Anthropic reasoning model and later turns are sent through the xAI provider.

Reproduction steps

Tested locally on current main at cb83656718.

1. Install dependencies in the repo.
2. Run the following script from the repo root:

import { convertToResponsesApiInput } from './src/api/transform/responses-api-input'

const messages = [
  {
    role: 'assistant',
    content: [
      { type: 'thinking', thinking: 'SECRET_CHAIN_OF_THOUGHT', signature: 'sig-1' },
      { type: 'text', text: 'visible answer' },
    ],
  },
] as any

console.log(JSON.stringify(convertToResponsesApiInput(messages), null, 2))

3. Observe that the transform outputs:

[
  {
    "type": "message",
    "role": "assistant",
    "content": [
      {
        "type": "output_text",
        "text": "[Thinking] SECRET_CHAIN_OF_THOUGHT"
      }
    ]
  },
  {
    "type": "message",
    "role": "assistant",
    "content": [
      {
        "type": "output_text",
        "text": "visible answer"
      }
    ]
  }
]

Expected result

Persisted Anthropic thinking blocks should either:

• be dropped for the xAI Responses replay path, or
• be preserved only in a provider-safe/native hidden reasoning format.

They should not be flattened into normal assistant-visible text.

Actual result

thinking is converted into a normal assistant output_text entry prefixed with [Thinking], which changes the semantics of the conversation history and can leak previously hidden reasoning into later prompt context.

Variations tried (optional)

Related local source inspection on current main:

• src/core/task/Task.ts:904-924 persists Anthropic reasoning as type: "thinking"
• src/api/providers/xai.ts:93-104 routes xAI history through convertToResponsesApiInput(messages)
• src/api/transform/responses-api-input.ts:55-63 converts thinking into [Thinking] ... text

App Version

main (cb83656718)

API Provider (optional)

xAI

Model Used (optional)

Grok / xAI Responses path

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

This seems best categorized as a provider interoperability / reasoning round-trip bug rather than a general crash. The regression appears adjacent to the xAI Responses migration in #11961 / #11962, but I could not find an existing issue that tracks this exact replay behavior.

[roomote-v0]

I've analyzed this issue and here's my proposed implementation plan:

1. Modify src/api/transform/responses-api-input.ts (lines 55-63): Change the thinking case in the assistant content block switch statement to silently drop/skip Anthropic thinking blocks instead of converting them into [Thinking] ... output_text entries. The case will simply break without pushing anything to the output array.

2. Add test coverage in src/api/transform/__tests__/responses-api-input.spec.ts: Add tests to verify that:

   • Anthropic thinking blocks with content are dropped (not converted to output_text)
   • Anthropic thinking blocks with empty content are dropped
   • Mixed messages containing both thinking and text blocks only emit the text parts

This approach ensures that persisted Anthropic reasoning blocks are safely stripped when replaying history through the xAI Responses API path, preventing hidden chain-of-thought from leaking into later prompt context.

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.

[shaun0927]

I did not open a PR for this yet because the fix direction seems policy-sensitive: I can see at least two plausible behaviors (drop unsupported hidden reasoning for xAI replay, or preserve it in a provider-specific hidden format if available). Once maintainers confirm the preferred fallback, I can follow up with a narrow patch and regression test.

[shaun0927]

Claiming. I prepared a minimal follow-up patch that keeps the scope narrow: it stops convertToResponsesApiInput() from flattening persisted Anthropic thinking blocks into assistant-visible text and adds a regression test for the replay behavior.

[shaun0927]

Follow-up PR opened: #12131