Claude thinking model fails with 400 INVALID_ARGUMENT when session history contains signed thinking blocks

[Inrstning]

Pre-submission checklist

• I have searched existing issues for duplicates
• I have read the 📋 Troubleshooting Guide
• I have read the README installation instructions

Model used

antigravity-claude-opus-4-5-thinking

Exact error message

Request contains an invalid argument.
[Debug Info]
Requested Model: claude-opus-4-5-thinking
Effective Model: claude-opus-4-5-thinking
Project: extreme-elf-b6v41
Endpoint: https://daily-cloudcode-pa.sandbox.googleapis.com/v1internal:streamGenerateContent?alt=sse
Status: 400
Followed by:
Failed to process error response

Bug description

When using claude-opus-4-5-thinking or claude-sonnet-4-5-thinking models through Antigravity, subsequent requests in an existing session fail with 400 INVALID_ARGUMENT. New sessions work fine, but any session with prior thinking block responses cannot continue.

• opencode version: (run opencode --version)
• antigravity plugin: opencode-antigravity-auth@beta (v1.15.8)
• OS: macOS

1. Start a new session with claude-opus-4-5-thinking model
2. Send any message → receives successful response with thinking blocks
3. Send another message in the same session → 400 error
4. OR: Try to run /compact on the session → 400 error

• Models affected: claude-opus-4-5-thinking, claude-sonnet-4-5-thinking

Steps to reproduce

1. Start a new session with claude-opus-4-5-thinking model
2. Send many message → receives successful response with thinking blocks
3. Send another message in the same session → 400 error
4. OR: Try to run /compact on the session → 400 error

Did this ever work?

Worked before, now broken (regression)

Number of Google accounts configured

1

Reproducibility

Always (100%)

Plugin version

opencode-antigravity-auth@beta (v1.15.8)

OpenCode version

opencode --version 1.1.50

Operating System

MACOS 26.2 (25C56)

Node.js version

node --version v22.14.0

Environment type

Standard (native terminal)

MCP servers installed

No response

Debug logs (REQUIRED)

Root Cause Analysis
The session history sent to the API contains signed thinking blocks in this format:
{
  role: model,
  parts: [
    {
      thought: true,
      text: 用户在问...,
      thoughtSignature: RXFRRENrZ0lDeEFDR0FJcVFCcEZXWkNFU0k4Yi...
    },
    {
      functionCall: {...}
    }
  ]
}
When this history is sent back to the API (either for continuation or compact), the backend rejects it with INVALID_ARGUMENT.
Key observation: The same session works for the first message (no history) but fails on all subsequent messages (history includes signed thinking blocks).
Debug Log Evidence
From ~/.config/opencode/antigravity-logs/antigravity-debug-*.log:
[10:48:47.750Z] Response 400 Bad Request (6869ms)
[10:48:47.750Z] Response Body (400): {
  "error": {
    "code": 400,
    "message": "Request contains an invalid argument.",
    "status": "INVALID_ARGUMENT"
  }
}
[10:48:47.753Z] Note: Failed to transform Antigravity response
[10:48:47.755Z] Error: Error: THINKING_RECOVERY_NEEDED
    at transformAntigravityResponse (request.js:1235:47)

Configuration (optional)

Compliance

• I'm using this plugin for personal development only
• This issue is not related to commercial use or TOS violations

Additional context

No response

[github-actions]

👋 Thanks for opening this issue!

Thank you for the detailed report. I've analyzed this issue and identified the root cause.

Root Cause Analysis:
The error occurs because session history contains signed thinking blocks from previous responses that include a thoughtSignature field (Gemini's signed thinking format). When this history is sent back to the API for subsequent requests, the Claude endpoint rejects it with INVALID_ARGUMENT because Claude models don't expect Gemini-specific thoughtSignature fields in the conversation history.

Code Investigation:
The plugin has sanitization logic in cross-model-sanitizer.ts (stripGeminiThinkingMetadata()) that's supposed to remove thoughtSignature and thinkingMetadata fields when targeting Claude models. This is called in request.ts at line 709 via sanitizeCrossModelPayloadInPlace(). However, based on your debug logs, it appears this sanitization isn't catching all instances of the thoughtSignature field in the session history.

Why this is a regression:
This likely worked before because either (1) the API was more lenient with unknown fields, or (2) the session history format changed in recent OpenCode versions to include these signed thinking blocks.

Immediate Workarounds:

1. Start a new session (/new or opencode --new) when switching to/from thinking models
2. Use non-thinking variants (e.g., claude-opus-4-5 without -thinking) for multi-turn conversations
3. Avoid /compact on sessions that used thinking models

Technical Details:
The affected code path is in sanitizeCrossModelPayloadInPlace() which recursively sanitizes message contents and parts. The thoughtSignature field at contents[*].parts[*].thoughtSignature should be deleted but appears to be persisting in some edge cases, particularly when thought: true is also present.

This issue is specific to the thinking models (claude-opus-4-5-thinking, claude-sonnet-4-5-thinking) because they include signed thinking blocks in responses that get stored in session history.

I'll flag this for the maintainers to review the cross-model sanitizer logic.

---

This is an automated response. A maintainer will review your issue soon.

[ndycode]

devs pls wake up

[anilcancakir]

I have same issue and i disabled dcp plugin in opencode, it works fine. I guess, the DCP is changing request and it makes error. Do you have DCP plugin can you check?

[WangShayne]

same issue

[Sawyer1ShOt]

same