feat(opencode): add copilot specific provider to properly handle copilot reasoning tokens

[SteffenDE]

What does this PR do?

This PR adds a copilot specific provider for the copilot completions API (there's already copilot specific code for the responses API). The code already states that this code is only meant for copilot, so I decided to rename the folder accordingly for clarity. While it would be great to not need this, I currently don't see a way to not have the copilot specifics (apart from extracting it into a separate repository).

It is similar to #5346, but it changes the completions code to properly store the reasoning_opaque field and send it back to the copilot API.

This PR is based on code I wrote for tidewave. I used Claude to implement the same changes based on a fresh copy of the upstream openai-compatible provider: https://github.com/vercel/ai/tree/%40ai-sdk/openai-compatible%401.0.30/packages/openai-compatible/src/chat

There are multiple small commits to make reviewing this easier and I also added tests for the important cases I encountered when handling the reasoning fields.

In the past, the Copilot API failed if the reasoning signature (reasoning_opaque) was not sent back for the Gemini 3 models. At some point GitHub seems to have changed this. Note though that the models still behave differently if the reasoning tokens are not passed back. For example, in Tidewave we've often seen Copilot's Gemini 2.5 spiral into a loop when omitting the reasoning, while it doesn't do that when including those.

How did you verify your code works?

You can chat with Gemini 2.5 Pro (3 Pro Preview is currently broken because of #8829, but it works in the Tidewave version of the code and all the fields are the same).

Closes #6864.

[github-actions]

Hey! Your PR title Add copilot specific provider to properly handle copilot reasoning tokens doesn't follow conventional commit format.

Please update it to start with one of:

• feat: or feat(scope): new feature
• fix: or fix(scope): bug fix
• docs: or docs(scope): documentation changes
• chore: or chore(scope): maintenance tasks
• refactor: or refactor(scope): code refactoring
• test: or test(scope): adding or updating tests

Where scope is the package name (e.g., app, desktop, opencode).

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

Related PRs Found

PR #5346: feat: reasoning text for Gemini 3 Pro in GH Copilot
#5346

Why it's related: This PR is explicitly mentioned in the description as similar work. PR #8900 builds on the approach from #5346 but extends it to properly handle reasoning_opaque field storage and transmission for the completions API, whereas #5346 focused on reasoning text for the responses API.

---

PR #5877: fix(github-copilot): auto-route GPT-5+ models to Responses API
#5877

Why it's related: Deals with GitHub Copilot routing and API handling, which is in the same domain as the copilot-specific provider changes.

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[Coruscant11]

Hello, I can confirm the reasoning works with Gemini 2.5 Pro however it doesn't seem to be the case with GPT 5.2 Codex / GPT 5.2 and Opus 4.5

[SteffenDE]

@Coruscant11 thanks for trying! GPT-5 variants use the responses API version, which this PR does not affect. I need to check Opus - it's possible that reasoning needs to be explicitly enabled.

[aadishv]

Here's the debugging I did so far -- lmk if there's something wrong with my methodology!

I added basic logging for the opaque signatures 1) received by the API and 2) received by convertToOpenAICompatibleChatMessages. See aadishv@5776493

Then I ran a basic test conversation:

The logs indicated that convertToOpenAICompatibleChatMessages didn't receive any opaque signatures back (see the "Parsed back:" log)

The research I did earlier indicates that this because the AI SDK version currently being used (5.0.x?) doesn't automatically propagate providerMetadata from response -> next request. iirc, we can get this PR to work by upgrading to a version of the AI SDK that does do this, but I don't recall the specifics.

Hope this helps @SteffenDE :)

[aadishv]

I also remember that @​rekram1-node (not pinging rn) was quite knowledgeable on the Copilot API/AI SDK relationship so he might be able to give some insights

[SteffenDE]

@aadishv the place you log in convertToOpenAICompatibleChatMessages is wrong though, it’s just after initializing the variable so you’ll never see opaque not being undefined there. Try logging here https://github.com/aadishv/opencode/blob/5776493d4add243cf2bb155ed3989dac835328a9/packages/opencode/src/provider/sdk/copilot/chat/openai-compatible-chat-language-model.ts#L320, which is where I tested and did see it working. But as mentioned I’m only on my phone right now - will verify tomorrow :)

I was also using ai-sdk v5 in another project from which I adapted the code and it‘s definitely passing the metadata back.

[aadishv]

Hmm interesting, added this log:

But I still don't see the signature:

Conversation (the last message is the one where the logs are from):

Not sure if there's something else wrong with my local setup though. If it worked for you then no worries!

[SteffenDE]

@aadishv the reasoning_opaque is part of the previous assistant messages, not the user message. So your check would need to be args.messages.at(-2).reasoning_opaque

[caozhiyuan]

@SteffenDE Perhaps you can refer to this https://github.com/caozhiyuan/copilot-api/tree/all/src/routes/messages, which supports claude thinking & interleaved thinking (not native, by prompt), gemini thinking, and gpt responses API. The copilot models API will return model information, such as support for responses API / chat completions / message API (message API may be supported in the next version of claude or the next major version of vscode).

[{
	"billing": {
		"is_premium": true,
		"multiplier": 1,
		"restricted_to": [
			"pro",
			"pro_plus",
			"business",
			"enterprise"
		]
	},
	"capabilities": {
		"family": "gpt-5.2-codex",
		"limits": {
			"max_context_window_tokens": 400000,
			"max_output_tokens": 128000,
			"max_prompt_tokens": 272000,
			"vision": {
				"max_prompt_image_size": 3145728,
				"max_prompt_images": 1,
				"supported_media_types": [
					"image/jpeg",
					"image/png",
					"image/webp",
					"image/gif"
				]
			}
		},
		"object": "model_capabilities",
		"supports": {
			"parallel_tool_calls": true,
			"streaming": true,
			"structured_outputs": true,
			"tool_calls": true,
			"vision": true
		},
		"tokenizer": "o200k_base",
		"type": "chat"
	},
	"id": "gpt-5.2-codex",
	"is_chat_default": false,
	"is_chat_fallback": false,
	"model_picker_category": "powerful",
	"model_picker_enabled": true,
	"name": "GPT-5.2-Codex",
	"object": "model",
	"policy": {
		"state": "enabled",
		"terms": "Enable access to the latest GPT-5.2-Codex model from OpenAI. [Learn more about how GitHub Copilot serves GPT-5.2-Codex](https://gh.io/copilot-openai)."
	},
	"preview": false,
	"supported_endpoints": [
		"/responses"
	],
	"vendor": "OpenAI",
	"version": "gpt-5.2-codex"
},{
    "billing": {
      "is_premium": true,
      "multiplier": 3,
      "restricted_to": [
        "pro",
        "pro_plus",
        "max",
        "business",
        "enterprise"
      ]
    },
    "capabilities": {
      "family": "claude-opus-4.5",
      "limits": {
        "max_context_window_tokens": 160000,
        "max_output_tokens": 16000,
        "max_prompt_tokens": 128000,
        "vision": {
          "max_prompt_image_size": 3145728,
          "max_prompt_images": 5,
          "supported_media_types": [
            "image/jpeg",
            "image/png",
            "image/webp"
          ]
        }
      },
      "object": "model_capabilities",
      "supports": {
        "max_thinking_budget": 32000,
        "min_thinking_budget": 1024,
        "parallel_tool_calls": true,
        "streaming": true,
        "tool_calls": true,
        "vision": true
      },
      "tokenizer": "o200k_base",
      "type": "chat"
    },
    "id": "claude-opus-4.5",
    "is_chat_default": false,
    "is_chat_fallback": false,
    "model_picker_category": "powerful",
    "model_picker_enabled": true,
    "name": "Claude Opus 4.5",
    "object": "model",
    "policy": {
      "state": "disabled",
      "terms": "Enable access to the latest Claude Opus 4.5 model from Anthropic. [Learn more about how GitHub Copilot serves Claude Opus 4.5](https://gh.io/copilot-anthropic)."
    },
    "preview": false,
    "supported_endpoints": [
      "/chat/completions"
    ],
    "vendor": "Anthropic",
    "version": "claude-opus-4.5"
  }]

[caozhiyuan]

I discovered through copilot-api that it's possible to bypass the max_prompt_tokens limit, allowing even Claude models to exceed 200k, though I'm not sure if this is a bug in Copilot. However, I haven't abused this.

[rekram1-node]

okay I played w/ this a bit, I think it's good now

[caozhiyuan]

seems need change "Openai-Intent" to "conversation-agent" and variants logic change to if (model.id.includes("claude")) {
return {
high: { thinking_budget: Math.min(15_999, Math.floor(model.limit.output / 2 - 1)) },
max: { thinking_budget: Math.min(31_999, model.limit.output - 1) },
}
} for claude model thinking, but only can thinking in first turn (chat comletion not support interleaved thinking). @SteffenDE and The gemini flash 3 model missed the reasoning_opaque thinking signature processing.

[caozhiyuan]

@rekram1-node can add a Flag.OPENCODE_EXPERIMENTAL_xxx (default false) for claude message api ?

[rekram1-node]

@caozhiyuan sure, but note that it does have a little bit to it, did you see I actually was using it in prod for a while but we ran into rate limit issues for users of it

[rekram1-node]

seems need change "Openai-Intent" to "conversation-agent"

What's this for again?

[rekram1-node]

lol u know the api rlly well

[caozhiyuan]

seems need change "Openai-Intent" to "conversation-agent"

What's this for again?

@rekram1-node for chat completions api , claude thinking need conversation-agent.

[NateSmyth]

@caozhiyuan I'm pretty sure the opencode client is actually blocked on copilots messages proxy now. It just 404's on messages now for me (unless I spoof being vscode, with a token obtained with the vscode client id). Could just be me but the fact that it works with the vscode auth makes me think its actually blocked

And yeah this does need to handle claude vs. gemini opaque_reasoning, right now if you switch from gemini -> claude mid-session claude will get a gemini reasoning signature and 400

[rekram1-node]

#11569

working on a fix for all that

[rekram1-node]

for chat completions api , claude thinking need conversation-agent.

Does this actually make a meaningful difference or is this a consistency thing

I think copilot team was telling me this header may not be necessary but i forget

[caozhiyuan]

seems need change "Openai-Intent" to "conversation-agent" and variants logic change to if (model.id.includes("claude")) {
return {
high: { thinking_budget: Math.min(15_999, Math.floor(model.limit.output / 2 - 1)) },
max: { thinking_budget: Math.min(31_999, model.limit.output - 1) },
}
} for claude model thinking, but only can thinking in first turn (chat comletion not support interleaved thinking). @SteffenDE and The gemini flash 3 model missed the reasoning_opaque thinking signature processing.

@rekram1-node This morning I tried the dev branch code, and the Claude model wasn't outputting thinking information. After changing the edits in the header to agent and adding the high max variant, it worked.

[rekram1-node]

It won't output thinking unless u use thinking variant tho, hit ctrl+t and youll see it, we should default to it on tho prolly

[SteffenDE]

@caozhiyuan I explicitly tested thinking with both intent header values and it did not make a difference

[rekram1-node]

cool, ill go over my pr one more time make sure everything still works and then merge it and should fix some small issues

[caozhiyuan]

It's possible that Copilot has relaxed its restrictions. Previously, when thinking budget was first supported, only agent calls were supported. This morning, I installed the latest version of OpenCode via npm. Typing prompt didn't display any output (although it did return output). After restarting OpenCode, it worked normally, and switching to the previous session also displayed the information correctly. Pressing Ctrl+T didn't display variant information either, so I tried making some changes. @SteffenDE @rekram1-node

[caozhiyuan]

@rekram1-node It would be better if it could support different variations of the thinking budget.

[rekram1-node]

Yeah we cna do that im game