feature about chat completions reasoning, support claude model enable thinking & interleaved thinking , and support gemini-3 thinking #167

caozhiyuan · 2026-01-06T05:47:57Z

This pull request introduces significant updates to the translation layer between OpenAI and Anthropic message formats, especially around supporting "thinking" blocks for Claude models, and improves protocol compliance for interleaved thinking. It also updates API versioning and various headers, and brings in several bug fixes and refactors for message streaming and token counting.

Anthropic "thinking" protocol and translation improvements:

Added support for "thinking" blocks in Claude model translations, including handling of reasoning_text and reasoning_opaque fields, filtering by signature, and ensuring protocol compliance by injecting reminders and requirements into system prompts. [1] [2] [3] [4] [5] [6] [7] [8]
Enhanced translation from OpenAI to Anthropic by reconstructing "thinking" blocks from OpenAI's reasoning_text and reasoning_opaque fields, and ensuring proper block ordering in responses. [1] [2] [3]
Improved mapping of system prompts and user messages to enforce Claude's interleaved thinking protocol, including reminders and minimum reasoning requirements.

API and protocol versioning:

Updated Copilot and API version constants to use 0.35.0 and 2025-10-01 respectively, and changed the openai-intent header to "conversation-agent". [1] [2]

Streaming and state management enhancements:

Refactored streaming translation logic to modularize event handling, added support for thinkingBlockOpen state, and ensured correct block closing and event emission for thinking and tool blocks. [1] [2] [3] [4] [5]

Token counting and content mapping fixes:

Updated token counting logic to skip the reasoning_opaque field, ensuring accurate token usage reporting.
Refined content mapping to exclude thinking blocks from user-visible content except where required, and improved image block handling. [1] [2] [3]

Type and import updates:

Added and updated types for model capabilities and Anthropic message blocks, including signature and thinkingBlockOpen fields. [1] [2] [3]

These changes collectively improve Claude model support, ensure protocol compliance, and enhance translation reliability between OpenAI and Anthropic message formats.

…g order when stream=false and exclude reasoning_opaque from token calculation in calculateMessageTokens

…ool block state

…arity

…llback VSCode version

… 1.107.0

…alue

…inking budget integration

…reasoning_opaque in different deltas

When account type is not specified or set to 'individual', use the default api.githubcopilot.com URL instead of constructing a subdomain-based URL. This restores previous behavior where business users could work without explicitly specifying their account type, as the default URL works for both individual and business accounts. Only constructs account-type-specific URLs (api.business.githubcopilot.com, api.enterprise.githubcopilot.com) when those account types are explicitly specified.

fix: use default API URL when account type is individual

caozhiyuan · 2026-01-08T13:05:38Z

Note: Currently, please configure claude code settings.json according to the following link https://github.com/ericc-ch/copilot-api?tab=readme-ov-file#manual-configuration-with-settingsjson .

{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:4141",
    "ANTHROPIC_AUTH_TOKEN": "dummy",
    "ANTHROPIC_MODEL": "xxxxx",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "xxxxx",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "gpt-5-mini",
    "CLAUDE_CODE_SUBAGENT_MODEL": "gpt-5-mini",
    "DISABLE_NON_ESSENTIAL_MODEL_CALLS": "1",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "BASH_MAX_TIMEOUT_MS": "600000"
  },
  "permissions": {
    "allow": [
      "mcp__fetch__fetch",
      "Bash(gh pr:*)",
      "Bash(gh status:*)",
      "Bash(gh repo view:*)",
      "Bash(node -e:*)"
    ],
    "deny": [
      "WebSearch"
    ]
  },
  "alwaysThinkingEnabled": true
}

caozhiyuan · 2026-01-09T08:31:01Z

if run in opencode , .config opencode opencode.json samples （if not use github copilot provider , add a custom local provider）:

{
	"$schema": "https://opencode.ai/config.json",
	"model": "local/gpt-5.2",
	"small_model": "local/gpt-5-mini",
	"agent": {
		"build": {
			"model": "local/gpt-5.2"
		},
		"plan": {
			"model": "local/gpt-5.2"
		},
		"explore": {
			"model": "local/gpt-5-mini"
		}
	},
	"mcp": {
		"fetch": {
			"type": "local",
			"command": ["uvx", "mcp-server-fetch", "--ignore-robots-txt"],
			"enabled": true
		}
	},
	"provider": {
		"local": {
			"npm": "@ai-sdk/anthropic",
			"name": "My Local",
			"options": {
				"baseURL": "http://localhost:4141/v1",
				"apiKey": "dummy"
			},
			"models": {
				"gpt-5.2": {
					"name": "gpt-5.2",
					"limit": {
						"context": 200000,
						"output": 64000
					}
				},
				"gpt-5-mini": {
					"name": "gpt-5-mini",
					"limit": {
						"context": 128000,
						"output": 64000
					}
				},
				"claude-sonnet-4.5": {
					"id": "claude-sonnet-4.5",
					"name": "claude-sonnet-4.5",
					"limit": {
						"context": 128000,
						"output": 16000
					},
					"options": {
						"thinking": {
							"type": "enabled",
							"budgetTokens": 15999
						}
					}
				}
			}
		}
	}
}

Arminova · 2026-01-10T18:20:16Z

  		"claude-sonnet-4.5": {
  			"id": "claude-sonnet-4.5",
  			"name": "claude-sonnet-4.5",
  			"limit": {
  				"context": 200000,
  				"output": 16000
  			},
  			"options": {
  				"thinking": {
  					"type": "enabled",
  					"budgetTokens": 16000
  				}
  			}
  		}

why 200000 context ? it must be 128K, none of Copilot models support 200000

caozhiyuan · 2026-01-10T23:46:52Z

@Arminova copy mistake, i changed. gpt seems support max 200k context.

22Goose · 2026-01-11T14:22:42Z

Hi, have you tested the thinking mode on Opus 4.5?

 ERROR  Failed to create chat completions Response {}                                                                                                                                                                                     10:20:06 PM


 ERROR  Error occurred: Failed to create chat completions                                                                                                                                                                                 10:20:06 PM

    at <anonymous> (src/services/copilot/create-chat-completions.ts:39:15)
    at async <anonymous> (src/routes/messages/handler.ts:82:26)
    at async handleCompletion (src/routes/messages/handler.ts:67:16)
    at async <anonymous> (src/routes/messages/route.ts:12:18)
    at async dispatch (node_modules/hono/dist/compose.js:22:23)
    at async cors2 (node_modules/hono/dist/middleware/cors/index.js:84:11)
    at async dispatch (node_modules/hono/dist/compose.js:22:23)
    at async logger2 (node_modules/hono/dist/middleware/logger/index.js:38:11)
    at async dispatch (node_modules/hono/dist/compose.js:22:23)
    at async <anonymous> (node_modules/hono/dist/hono-base.js:201:31)
    at processTicksAndRejections (native:7:39)


 ERROR  HTTP error: { error:                                                                                                                                                                                                              10:20:06 PM
   { message: 'The requested model is not supported.',
     code: 'model_not_supported',
     param: 'model',
     type: 'invalid_request_error' } }

I got this error.

caozhiyuan · 2026-01-11T14:30:34Z

@22Goose use https://github.com/caozhiyuan/copilot-api/tree/all , run bun run start start -v, and see logs in user .local\share\copilot-api\logs

22Goose · 2026-01-11T14:47:54Z

.local\share\copilot-api\logs

Besides, I noticed some discussion on the copilot may not support the thinking model of Opus 4.5. I don't know whether we can use the copilot api in this way.

caozhiyuan · 2026-01-11T14:51:47Z

Note: Currently, please configure claude code settings.json according to the following link https://github.com/ericc-ch/copilot-api?tab=readme-ov-file#manual-configuration-with-settingsjson .

{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:4141",
    "ANTHROPIC_AUTH_TOKEN": "dummy",
    "ANTHROPIC_MODEL": "xxxxx",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "xxxxx",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "gpt-5-mini",
    "CLAUDE_CODE_SUBAGENT_MODEL": "gpt-5-mini",
    "DISABLE_NON_ESSENTIAL_MODEL_CALLS": "1",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1",
    "BASH_MAX_TIMEOUT_MS": "600000"
  },
  "permissions": {
    "allow": [
      "mcp__fetch__fetch",
      "Bash(gh pr:*)",
      "Bash(gh status:*)",
      "Bash(gh repo view:*)",
      "Bash(node -e:*)"
    ],
    "deny": [
      "WebSearch"
    ]
  },
  "alwaysThinkingEnabled": true
}

@22Goose change user .claude/settings.json like this. and start a new conversation

caozhiyuan · 2026-01-11T14:53:13Z

Currently, based on my usage, gpt-5.2 is stronger than other models; I mainly use it for backend complex logic development @22Goose in my usage, ANTHROPIC_MODEL ANTHROPIC_DEFAULT_SONNET_MODEL all set gpt-5.2 (use responeses api high think efforts)

22Goose · 2026-01-11T15:09:48Z

Currently, based on my usage, gpt-5.2 is stronger than other models; I mainly use it for backend complex logic development @22Goose in my usage, ANTHROPIC_MODEL ANTHROPIC_DEFAULT_SONNET_MODEL all set gpt-5.2 (use responeses api high think efforts)

Thanks for advise. TBH, I also mostly use gpt-5.2. I just want to test the Opus 4.5 due to Boris Cherny's interview. Just try to figure out which one is better.

22Goose · 2026-01-11T15:21:27Z

Note: Currently, please configure claude code settings.json according to the following link https://github.com/ericc-ch/copilot-api?tab=readme-ov-file#manual-configuration-with-settingsjson .
@22Goose change user .claude/settings.json like this. and start a new conversation

I change the setting, but the outcomes are quite the same.

caozhiyuan · 2026-01-11T23:25:11Z

@22Goose bun run start start -v can see model names,You entered the wrong model name

D GitHub Copilot Token fetched successfully!                                                                
i Available models:                                                                                         
- gpt-5-mini
- gpt-5
- gpt-4o-mini-2024-07-18
- gpt-4o-2024-11-20
- gpt-4o-2024-08-06
- grok-code-fast-1
- gpt-5.1
- gpt-5.1-codex
- gpt-5.1-codex-mini
- gpt-5.1-codex-max
- gpt-5-codex
- text-embedding-3-small
- text-embedding-3-small-inference
- claude-sonnet-4
- claude-sonnet-4.5
- claude-opus-4.5
- claude-haiku-4.5
- gemini-3-pro-preview
- gemini-3-flash-preview
- gemini-2.5-pro
- gpt-4.1-2025-04-14
- oswe-vscode-prime
- oswe-vscode-secondary
- gpt-5.2
- gpt-41-copilot
- gpt-3.5-turbo-0613
- gpt-4
- gpt-4-0613
- gpt-4-0125-preview
- gpt-4o-2024-05-13
- gpt-4-o-preview
- gpt-4.1
- gpt-3.5-turbo
- gpt-4o-mini
- gpt-4
- gpt-4o
- gpt-4-o-preview
- text-embedding-ada-002

22Goose · 2026-01-12T02:14:57Z

@caozhiyuan I may find out where goes wrong.
tl,dr: directly change the model by using /model claude-opus-4.5 in CC for VS Code extension may let this error

Here is my setting:


{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:4141",
    "ANTHROPIC_AUTH_TOKEN": "dummy",
    "ANTHROPIC_MODEL": "claude-opus-4.5",
    "ANTHROPIC_SMALL_FAST_MODEL": "claude-sonnet-4.5",
    "DISABLE_NON_ESSENTIAL_MODEL_CALLS": "1",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  },
  "permissions": {
    "allow": [
      "mcp__fetch__fetch",
      "Bash(gh pr:*)",
      "Bash(gh status:*)",
      "Bash(gh repo view:*)",
      "Bash(node -e:*)"
    ],
    "deny": [
      "WebSearch"
    ]
  },
  "alwaysThinkingEnabled": true
}

In CC Cli:
The "ANTHROPIC_MODEL": "claude-opus-4.5" will take effect. No error.
The /model claude-opus-4.5 will take effect. No error.
The /model opus 4.5 will bring the error.

In CC VS Code extension:
The "ANTHROPIC_MODEL": "claude-opus-4.5" will not take effect. Sonnet 4 is in use. No error.
The /model claude-opus-4.5 will not take effect, which also bring the error.
The /model opus 4.5 will bring the error.

I get upper error log basically when I working with CC VS Code extension.

caozhiyuan · 2026-01-12T02:48:22Z

@caozhiyuan I may find out where goes wrong. tl,dr: directly change the model by using /model claude-opus-4.5 in CC for VS Code extension may let this error

Here is my setting:
{
  "$schema": "https://json.schemastore.org/claude-code-settings.json",
  "env": {
    "ANTHROPIC_BASE_URL": "http://localhost:4141",
    "ANTHROPIC_AUTH_TOKEN": "dummy",
    "ANTHROPIC_MODEL": "claude-opus-4.5",
    "ANTHROPIC_SMALL_FAST_MODEL": "claude-sonnet-4.5",
    "DISABLE_NON_ESSENTIAL_MODEL_CALLS": "1",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  },
  "permissions": {
    "allow": [
      "mcp__fetch__fetch",
      "Bash(gh pr:*)",
      "Bash(gh status:*)",
      "Bash(gh repo view:*)",
      "Bash(node -e:*)"
    ],
    "deny": [
      "WebSearch"
    ]
  },
  "alwaysThinkingEnabled": true
}
In CC Cli: The "ANTHROPIC_MODEL": "claude-opus-4.5" will take effect. No error. The /model claude-opus-4.5 will take effect. No error. The /model opus 4.5 will bring the error.

In CC VS Code extension: The "ANTHROPIC_MODEL": "claude-opus-4.5" will not take effect. Sonnet 4 is in use. No error. The /model claude-opus-4.5 will not take effect, which also bring the error. The /model opus 4.5 will bring the error.

I get upper error log basically when I working with CC VS Code extension.

@22Goose this all need addd
"ANTHROPIC_MODEL": "xxxxx",
"ANTHROPIC_DEFAULT_SONNET_MODEL": "xxxxx",
"ANTHROPIC_DEFAULT_HAIKU_MODEL": "gpt-5-mini",
"CLAUDE_CODE_SUBAGENT_MODEL": "gpt-5-mini",
"DISABLE_NON_ESSENTIAL_MODEL_CALLS": "1",
"CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
and CLAUDE_CODE_SUBAGENT_MODEL / ANTHROPIC_DEFAULT_HAIKU_MODEL must gpt-5-mini, otherwise cloud code will waste you Premium requests a lot . you can config ANTHROPIC_MODEL ANTHROPIC_DEFAULT_SONNET_MODEL = claude-sonnet-4.5. ANTHROPIC_DEFAULT_OPUS_MODEL = claude-opus-4.5 please read https://code.claude.com/docs/en/model-config

caozhiyuan and others added 17 commits November 19, 2025 21:32

feat: support copilot reasoning_opaque and reasoning_text

29668ce

feat: add signature field to AnthropicThinkingBlock

a2467d3

feat: add idleTimeout configuration for bun server

58f7a45

feat: enhance reasoning handling in tool calls and change the thinkin…

3fa5519

…g order when stream=false and exclude reasoning_opaque from token calculation in calculateMessageTokens

feat: conditionally handle reasoningOpaque in handleFinish based on t…

dfb40d2

…ool block state

fix: handleReasoningOpaqueInToolCalls add isToolBlockOpen judge

7657d87

feat: support claude model thinking block

7f8187b

feat: enhance thinking budget calculation and rename variables for cl…

cbe12eb

…arity

feat: update Copilot version and API version in api-config; adjust fa…

ebcacb2

…llback VSCode version

feat: update Copilot version to 0.35.0 and fallback VSCode version to…

0d6f7aa

… 1.107.0

fix: simplify copilotBaseUrl logic and correct openai-intent header v…

dcafbe1

…alue

feat: interleaved thinking support

5175245

feat: enhance system prompt handling for interleaved thinking with th…

dd80c8d

…inking budget integration

feat: compatible with copilot API returning content->reasoning_text->…

e45c6db

…reasoning_opaque in different deltas

Merge pull request #65 from HyunggyuJang/fix-individual-account-url

8191930

fix: use default API URL when account type is individual

feat: enforce interleaved thinking protocol in message handling

67b357a

This was referenced Jan 10, 2026

gpt-5-codex model endpoint support + Codex CLI documentation #117

Open

WebSearch tool returns 'did 0 searches' when used with Claude Code #169

Open

This was referenced Jan 15, 2026

github copilot 将本该免费的 gpt-5-mini 也计算进 premium-request usage 里了 caozhiyuan/copilot-api#70

Closed

GitHub Copilot is also counting the supposed free gpt-5-mini in the premium-request usage. #173

Open

Uh oh!

feature about chat completions reasoning, support claude model enable thinking & interleaved thinking , and support gemini-3 thinking #167

Are you sure you want to change the base?

feature about chat completions reasoning, support claude model enable thinking & interleaved thinking , and support gemini-3 thinking #167

Uh oh!

Conversation

caozhiyuan commented Jan 6, 2026

Uh oh!

caozhiyuan commented Jan 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

caozhiyuan commented Jan 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Arminova commented Jan 10, 2026

Uh oh!

caozhiyuan commented Jan 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

22Goose commented Jan 11, 2026

Uh oh!

caozhiyuan commented Jan 11, 2026

Uh oh!

22Goose commented Jan 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

caozhiyuan commented Jan 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

caozhiyuan commented Jan 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

22Goose commented Jan 11, 2026

Uh oh!

22Goose commented Jan 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

caozhiyuan commented Jan 11, 2026

Uh oh!

22Goose commented Jan 12, 2026

Uh oh!

caozhiyuan commented Jan 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

caozhiyuan commented Jan 8, 2026 •

edited

Loading

caozhiyuan commented Jan 9, 2026 •

edited

Loading

caozhiyuan commented Jan 10, 2026 •

edited

Loading

22Goose commented Jan 11, 2026 •

edited

Loading

caozhiyuan commented Jan 11, 2026 •

edited

Loading

caozhiyuan commented Jan 11, 2026 •

edited

Loading

22Goose commented Jan 11, 2026 •

edited

Loading