multiple reasoning_opaque values received in a single response. Only one thinking part per response is suported #51466

Activity

1. github-actions

   assigned

   kitlangton

   on Sep 26, 2026
2. github-actions commented on Sep 26, 2026

   

   github-actionsBot

   on Sep 26, 2026 – with GitHub Actions

   Contributor

   More actions

   ---

   This issue might be a duplicate of existing issues. Please check:

   • AI_InvalidResponseDataError: Multiple reasoning_opaque values in single response (github-copilot + claude-sonnet-4-6) #17011: Reports the same reasoning_opaque error with the GitHub Copilot provider and Claude.
   • Multiple reasoning_opaque values received in a single response. Only one thinking part per response is supported #20434: Reports the same error while using Opus; it was also closed as a duplicate of AI_InvalidResponseDataError: Multiple reasoning_opaque values in single response (github-copilot + claude-sonnet-4-6) #17011.
3. xiangyang-95 commented on Sep 28, 2026

   

   xiangyang-95

   on Sep 28, 2026

   More actions

   I met the same problem as well using opencode with Opus 5.5. This only start happening a few days ago.

   👍React with 👍8Reacted by Camilo Sampedro, mgreuelrwe, Karli Ots, arughrat, Luca Soato, Jean-Sébastien Eude, wzhkdi and johntws
4. param087 commented on Sep 30, 2026

   

   param087

   on Sep 30, 2026

   More actions

   Root cause and a fix: #52190.

   This is interleaved thinking. Claude Opus 5/5.5 and Fable 5.1 on Copilot emit a fresh signed reasoning_opaque before each tool call, so one streamed response legitimately carries several. openai-compatible-chat-language-model.ts treats the second one as a protocol violation and throws, aborting the turn — which is why it presents as "stops working during normal work" rather than tying to a specific action. Any multi-tool-call turn with thinking on will do it.

   Worth noting for #17011, which was closed as not-reproducible: the trigger is multiple tool calls in a single assistant response, not a particular prompt. That is probably why it resisted repro.

   The parser is also stricter than the converter that feeds it — convert-to-openai-compatible-chat-messages.ts already handles multiple opaque values with a first-wins rule (if (partOpaque && !reasoningOpaque)) instead of erroring. The PR makes the parser consistent with that, and adds a fixture-based regression test that fails on dev with exactly this error and passes with the fix.

   Repro as a unit test, no account or network needed:

   data: {"choices":[{"index":0,"delta":{"reasoning_text":"..."}}]}
   data: {"choices":[{"index":0,"delta":{"tool_calls":[...],"reasoning_opaque":"opaque-first"}}]}
   data: {"choices":[{"index":0,"delta":{"reasoning_text":"..."}}]}
   data: {"choices":[{"finish_reason":"tool_calls","index":0,"delta":{"tool_calls":[...],"reasoning_opaque":"opaque-second"}}]}
   

   The second reasoning_opaque throws.
5. kaminski-dev commented on Oct 4, 2026

   

   kaminski-dev

   on Oct 4, 2026

   More actions

   Hey, I hit the same error on 1.18.34, macOS, Opus 5.5. This device is a company notebook. I OAuth login via github.com.

   Here is how I reached the chat-completions path: GET https://api.githubcopilot.com/models is blocked by our internal Zscaler (307 to a block page), while POST /chat/completions on the same host is allowed. opencode swallows the failed /models call in packages/opencode/src/plugin/github-copilot/copilot.ts and falls back to the static catalogue, which puts Claude on chat-completions instead of the Messages API (api.npm becomes @ai-sdk/github-copilot instead of @ai-sdk/anthropic, visible with opencode models github-copilot --verbose). With Opus 5.5 and thinking on, the stream then carries one reasoning_opaque per tool call and the parser throws. Any multi-tool turn does it.

   I tested the fix on my company notebook using the branch version. Calling multiple tools as well as subagents works, and across multiple threads I didn't get this issue at all.

   Fix PR: #53191. It would also help to log the failed /models call instead of swallowing it, so this fallback isn't silent.
6. kaminski-dev commented on Oct 5, 2026

   

   kaminski-dev

   on Oct 5, 2026

   More actions

   Hey @kitlangton, let me know if you find a second to check it.
7. rekram1-node commented on Oct 7, 2026

   

   rekram1-node

   on Oct 7, 2026

   Collaborator

   More actions

   will look.

   ❤️React with ❤️4Reacted by kaminski-dev, Mikael Valen-Sendstad, wzhkdi and Jean-Sébastien Eude
8. rekram1-node

   closed this as completedin #54207on Oct 9, 2026
9. rekram1-node commented on Oct 9, 2026

   

   rekram1-node

   on Oct 9, 2026

   Collaborator

   More actions

   updating:

   • models.dev to say the anthropic models can use anthropic sdk
   • chat parser to match vscode impl (on v1 and v2)
10. rekram1-node commented on Oct 9, 2026

    

    rekram1-node

    on Oct 9, 2026

    Collaborator

    More actions

    next release of v1 and v2 should fix.

    Note, recommend updating to v2:
    https://opencode.ai/v2/docs