Spaces:
Paused
Paused
Commit History
-
row-proxy inicial
7fa88ca
-
fix: update go.mod to 1.23 and add debug output
justincnn- Update go version from 1.21 to 1.23 in go.mod - Add -x flag and error handling to go mod download for debuggingfedb52d -
fix: add Node.js 24 support and fix Docker build
justincnn- Add FORCE_JAVASCRIPT_ACTIONS_TO_NODE24=true to workflow env - Add git and ca-certificates to builder stage for dependency resolution - Add ca-certificates to final stage for HTTPS supportb9e6e15 -
fix: remove GOPROXY and go mod verify from Dockerfile
justincnn- Remove domestic Go proxy setting (goproxy.cn) as build runs on GitHub/DockerHub - Remove go mod verify to simplify build processd3b54a1 -
fix: update GitHub Actions to v6 and Go to 1.23
justincnn- Update docker/build-push-action from v5 to v6 to fix Node.js 20 deprecation warning - Update Go version from 1.21 to 1.23 in Dockerfile - Add go mod verify after go mod download for better error handlinge03f5eb -
build: 添加 Go 代理配置并统一 Go 版本为 1.21
justincnn06edf9e
-
Add Docker build workflow and compose configurations
justincnn629a8be
-
feat: lowercase model ids, add glm-5, socks5 proxy pool, remove arm build
kao0312cbe30d3
-
chore:ci&readme updates
d41adbf -
test: expand tool calling test suite with assertions and new cases
Add automated pass/fail assertions (check/check_not helpers) and new test cases: -tools model suffix, tool_choice=required, tool_choice with specific function, and streaming regression. Print summary with pass/fail counts.701b56b -
feat: add Anthropic Messages API compatibility (/v1/messages)
Implement full Anthropic Messages API translation layer including: - Request/response model types (messages, content blocks, streaming events) - Claude model name resolution to GLM backend models - Streaming (SSE) and non-streaming response conversion - Thinking/reasoning block passthrough - Tool use (tool_call → tool_use) format translation - System prompt, tool_choice, and multi-turn conversation support4f59ab4 -
feat: add Claude-to-GLM model name mapping
Map Claude model names (opus, sonnet, haiku) to corresponding GLM models with automatic thinking and tools tag resolution. Default unknown models to glm-4.7 as fallback.3a315de -
fix: improve streaming tool_call block detection and add debug logging
Support multiple close markers (</tool_call>, </think>, next <tool_call>) when detecting tool call block boundaries in streaming mode. Add debug logging for stream payload inspection.85aaf7e -
fix: inject system and tool prompts as user+assistant pairs
z.ai ignores system role messages, so convert system messages and tool definitions into user+assistant conversation pairs at the start of the dialog. Add debug logging for upstream message inspection.ae13261 -
refactor: simplify tool system prompt for better compliance
Rewrite tool injection prompt to be more concise and authoritative, framing it as an integrated runtime capability rather than verbose rules. Strengthen tool_choice enforcement wording.fddf449 -
fix: use brace-counting to extract tool_call JSON from nested structures
Replace regex-based <tool_call> extraction with a brace-counting parser that correctly handles nested JSON objects and missing/malformed close tags. Also supports FuncName{args} shorthand format.9506c04 -
chore: update .gitignore to exclude generated files
ab5972f -
fix: improve tool calling reliability with multi-format parsing and Delta pointer fix
- Enhance system prompt with Chinese instructions and few-shot examples for more reliable <tool_call> output from GLM models - Add fallback parsing for [TOOL]...[/TOOL], [TOOL_CALL]...[/TOOL_CALL], and markdown JSON block formats in prompt injection mode - Change Choice.Delta to *Delta pointer so omitempty correctly omits the field in non-streaming responses (fixes extra "delta":{} in JSON) - Convert tool_calls/tool messages to plain text for upstream z.ai API which doesn't support native tools field - Add comprehensive tests for all new parsing formats and serialization8646505 -
feat: support tool/function calling (OpenAI compatible)
Add tools and tool_choice fields to chat requests, parse upstream tool_call responses, and return them in OpenAI-compatible format with finish_reason=tool_calls. Includes builtin tools, test script, and unit tests.5a55e77
-
refactor: split internal package into domain-based sub-packages
Split the flat internal/ package into auth, config, filter, handler, logger, model, upstream, and version sub-packages by responsibility.48d903a
-
feat: support GLM-4.7
NtGdi04438fd
-
feat: support GLM-4.6-V
NtGdi106464e
-
feat: add anonymous token support
NtGdic9a9c69
-
ci: add Docker image auto-build
NtGdi71280f6 -
fix: support multimodal image upload
kao03123dad4b0
-
feat: support search result references
NtGdi12aa997 -
first commit
NtGdi553c00b