support responses api , support native message-api, fix inconsistent credit consumption in chat #170

caozhiyuan · 2026-01-12T07:37:11Z

This pull request introduces a new configuration system, structured logging, and support for the /v1/responses endpoint, and support for the claude native message api, along with improvements to model selection and request handling. The most important changes are grouped below:

Responses API Integration:

Added full support for the /v1/responses endpoint, including a new handler (src/routes/responses/handler.ts) that validates model support, streams or returns results, and logs all activity.
Enhanced src/routes/messages/handler.ts to route requests to the Responses API when supported by the selected model, including translation logic for payloads and results.
Updated the API documentation in README.md to include the new /v1/responses endpoint and clarify its purpose.

Claude Native Message API:

support Claude Native Message API

Configuration Management:

Added a new src/lib/config.ts module to provide persistent application configuration, including support for model-specific prompts, reasoning effort levels, and default model selection. Configuration is stored in a new config.json file in the app data directory, with automatic creation and safe permissions. [1] [2]

Logging Improvements:

Implemented a new buffered, file-based logging utility in src/lib/logger.ts for handler-level logging, with log rotation, retention, and structured output. Integrated this logger into key request handlers for better diagnostics. [1] [2] [3] [4] [5]

Token Counting Logic:

Refactored token counting in src/lib/tokenizer.ts to more accurately account for tool calls, array parameters, and model-specific behaviors (including GPT and Anthropic/Grok models). Added support for excluding certain schema keys and improved calculation for nested parameters. [1] [2] [3] [4] [5] [6] [7] [8]

Fix Credit Consumption Inconsistency:

fix inconsistent credit consumption in chat and Merge tool_result and text blocks into tool_result to avoid consuming premium requests

…n utility

…nt function

…se output message type

…nd state management

…ure in translation tests

…andlers

…arsing and allign with vscode-copilot-chat extractThinkingData, otherwise it will cause miss cache occasionally

…ing signature check and update prompt

…ing small model if no tools are used 2.add bun idleTimeout = 0 3.feat: Compatible with Claude code JSONL file usage error scenarios, delay closeBlockIfOpen and map responses api to anthropic support tool_use and fix spelling errors 4.feat: add configuration management with extra prompt handling and ensure config file creation

…is incompatible with gpt-5-mini

…ssage translation

…okens limit

…just runServer to set verbose level correctly

…ponses-api # Conflicts: # src/start.ts

…ogic

…adjusting input token calculations and handling tool prompts

…e tab characters

Some clients, like RooCode may send `service_tier` to `/responses` endpoint, but Copilot do not support this field and returns error

… expanded reasoning options and add doc

…ndling in responses

… code skill tool_result

caozhiyuan · 2026-01-12T07:48:51Z

@ericc-ch also fix inconsistent credit consumption in chat and adapter claude code skill tool_result. opencode had fixed it.

caozhiyuan · 2026-01-12T07:51:26Z

Also supports the vscode extension, not sure if you need it: https://github.com/caozhiyuan/copilot-api/tree/feature/vscode-extension. Does not depend on bun.

…uming premium requests (caused by skill invocations, edit hooks or to do reminders)

…ent array handling

getaaron · 2026-01-18T00:10:52Z

@ericc-ch this looks like a great improvement, can you please merge?

GitHub Copilot's Responses API returns different IDs for the same item in 'added' vs 'done' events, which causes @ai-sdk/openai to throw errors: - 'activeReasoningPart.summaryParts' undefined - 'text part not found' This fix: - Tracks IDs from 'added' events and reuses them in 'done' events - Removes empty summary arrays from reasoning items that cause AI SDK parsing issues - Handles output_item, content_part, output_text, and response.completed events - Synchronizes item_id for message-type outputs across all related events

… API , simpler version * fix: sync stream IDs for @ai-sdk/openai compatibility with Responses API GitHub Copilot's Responses API returns different IDs for the same item in 'added' vs 'done' events, which causes @ai-sdk/openai to throw errors: - 'activeReasoningPart.summaryParts' undefined - 'text part not found' This fix: - Tracks IDs from 'added' events and reuses them in 'done' events - Removes empty summary arrays from reasoning items that cause AI SDK parsing issues - Handles output_item, content_part, output_text, and response.completed events - Synchronizes item_id for message-type outputs across all related events * simpler version of #72

… blocks

hgcode1130 · 2026-01-22T07:25:50Z

We need wire_api = "responses". Hope these willl be merge soon

FlorianBruniaux · 2026-01-23T19:22:48Z

✅ Successfully tested with Claude Code CLI %

Thanks @caozhiyuan for this excellent work on the /responses endpoint support!

We extensively tested your fork with our project cc-copilot-bridge - a multi-provider wrapper for Claude Code CLI that
switches between Anthropic, GitHub Copilot, and Ollama.

Test Results: 6/6 Passed ✅

Model	Test	Result
gpt-5.2-codex	Simple prompt	✅ Pass (Extended Thinking works!)
gpt-5.1-codex	Simple prompt	✅ Pass
gpt-5.1-codex-mini	Simple prompt	✅ Pass
gpt-5.1-codex-max	Simple prompt	✅ Pass
gpt-5 (regression)	Simple prompt	✅ Pass
claude-sonnet-4.5 (regression)	Simple prompt	✅ Pass

What we tested

All 5 Codex models work without the 400: not accessible via /chat/completions error
No regressions on existing models (Claude, GPT-5, GPT-4.1, Gemini)
Extended Thinking feature works on gpt-5.2-codex (premium feature)
Response times: 1-5 seconds (comparable to non-Codex models)

Our setup

We created a fork launcher script that:

Clones your branch automatically
Builds with bun install && bun run build
Runs the proxy on port 4141
Auto-detects when this PR is merged to switch back to official release

Script: launch-responses-fork.sh

Recommendation

Strongly recommend merging this PR. It unlocks all Codex models for Claude Code users via Copilot, which is a significant improvement.

We've documented our findings in detail here:

Test Results Report
All Model Commands (42 models documented)

Thanks again for the great work! 🚀

- CHANGELOG.md: Add v1.5.0 section documenting Codex models via fork - README.md: Add "GPT Codex Models" section with setup instructions - CLAUDE.md: Update Model Compatibility Matrix (Codex now supported) - scripts/VERSION: Bump to 1.5.0 PR tracking: ericc-ch/copilot-api#170 Co-Authored-By: Claude <noreply@anthropic.com>

zhujian0805 · 2026-01-27T07:22:54Z

nice feature, i have beening test it and work as expected:
i have dockenized it this is my repo: https://github.com/Chat2AnyLLM/copilot-api-nginx-proxy.git

caozhiyuan and others added 27 commits September 27, 2025 13:43

feature gpt-5-codex responses api

a57c238

feat: enhance output type for function call and add content conversio…

87899a1

…n utility

refactor: optimize content conversion logic in convertToolResultConte…

4fc0fa0

…nt function

refactor: remove unused function call output type and simplify respon…

2b9733b

…se output message type

feat: add signature and reasoning handling to responses translation a…

505f648

…nd state management

feat: add signature to thinking messages and enhance reasoning struct…

9477b45

…ure in translation tests

refactor: remove summaryIndex from ResponsesStreamState and related h…

44551f9

…andlers

feat: enhance streaming response handling with ping mechanism

708ae33

feat: responses translation add cache_read_input_tokens

47fb3e4

feat: enhance response event handling with event types and improved p…

2800ed3

…arsing and allign with vscode-copilot-chat extractThinkingData, otherwise it will cause miss cache occasionally

feat: improve event log and enhance reasoning content handling by add…

619d482

…ing signature check and update prompt

fix: the cluade code small model where max_tokens is only 512, which …

32cb10a

…is incompatible with gpt-5-mini

feat: add model reasoning efforts configuration and integrate into me…

9051a21

…ssage translation

fix: ensure application directory is created when config file is missing

eeeb820

feat: consola file logger for handler.ts

3f69f13

fix: copolit function call returning infinite line breaks until max_t…

4c0d775

…okens limit

feat: add verbose logging configuration to enhance log detail level

1ec12db

fix: update verbose property to be required in State interface and ad…

174e868

…just runServer to set verbose level correctly

Merge remote-tracking branch 'remotes/origin/master' into feature/res…

83cdfde

…ponses-api # Conflicts: # src/start.ts

fix: correct typo in warning message and refine whitespace handling l…

6f47926

…ogic

fix: update token counting logic for GPT and Claude and Grok models, …

01d4adb

…adjusting input token calculations and handling tool prompts

fix: extend whitespace handling in updateWhitespaceRunState to includ…

3cdc32c

…e tab characters

Remove incompatible with copilot responses service_tier field (#45)

f7835a4

Some clients, like RooCode may send `service_tier` to `/responses` endpoint, but Copilot do not support this field and returns error

feat(config): enhance model configuration with automatic defaults and…

318855e

… expanded reasoning options and add doc

feat(config): add useFunctionApplyPatch option and implement patch ha…

afb7a5c

…ndling in responses

fix: fix inconsistent credit consumption in chat , and adapter claude…

ee5df50

… code skill tool_result

caozhiyuan mentioned this pull request Jan 12, 2026

Inconsistent credit consumption compared to VS Code Copilot #171

Open

fix: Merge tool_result and text blocks into tool_result to avoid cons…

f2b8476

…uming premium requests (caused by skill invocations, edit hooks or to do reminders)

hzura mentioned this pull request Jan 16, 2026

[Feature Request] support responses api (openAI's new generation API supports model thinking) Joouis/agent-maestro#113

Closed

fix: improve merging of tool results and text blocks to optimize cont…

f3bef04

…ent array handling

cuipengfei and others added 4 commits January 18, 2026 18:33

fix: update mergeToolResultForClaude to handle opencode request

bc205a6

fix: add default thinking text for opencode compatibility in response…

6e93cfc

… blocks

feat: support messages-api

7e16a65

caozhiyuan changed the title ~~support responses api (openAI's new generation API supports model thinking) , support native message-api, fix inconsistent credit consumption in chat~~ support responses api , support native message-api, fix inconsistent credit consumption in chat Jan 25, 2026

caozhiyuan added 2 commits January 28, 2026 12:32

feat: add compact model usage configuration and detection

de424c1

feat: filter valid thinking blocks for Claude models in Messages API

de08ef3

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

support responses api , support native message-api, fix inconsistent credit consumption in chat #170

support responses api , support native message-api, fix inconsistent credit consumption in chat #170

Uh oh!

caozhiyuan commented Jan 12, 2026 •

edited

Loading

Uh oh!

caozhiyuan commented Jan 12, 2026

Uh oh!

caozhiyuan commented Jan 12, 2026

Uh oh!

getaaron commented Jan 18, 2026

Uh oh!

hgcode1130 commented Jan 22, 2026

Uh oh!

FlorianBruniaux commented Jan 23, 2026

Uh oh!

zhujian0805 commented Jan 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

Uh oh!

support responses api , support native message-api, fix inconsistent credit consumption in chat #170

Are you sure you want to change the base?

support responses api , support native message-api, fix inconsistent credit consumption in chat #170

Uh oh!

Conversation

caozhiyuan commented Jan 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

caozhiyuan commented Jan 12, 2026

Uh oh!

caozhiyuan commented Jan 12, 2026

Uh oh!

getaaron commented Jan 18, 2026

Uh oh!

hgcode1130 commented Jan 22, 2026

Uh oh!

FlorianBruniaux commented Jan 23, 2026

✅ Successfully tested with Claude Code CLI %

Test Results: 6/6 Passed ✅

What we tested

Our setup

Recommendation

Uh oh!

zhujian0805 commented Jan 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

caozhiyuan commented Jan 12, 2026 •

edited

Loading