Home
  • English

    • V1.2.6.1 - UECopilot Changelog

    V1.2.6.1

    New Feature: Universal LLM Provider Integration

    No code changes required — simply enter the Base URL and API key on the settings page to connect any LLM service.

    • One-click templates: Seven built-in presets — OpenRouter, Groq, Claude-compatible proxy, one-api/new-api relay, Ollama, vLLM, and LM Studio — ready to use.
    • Protocol compatibility: Supports OpenAI Chat Completions, Anthropic Messages, and Gemini GenerateContent.
    • Flexible configuration: Customize the authentication method (Bearer, request header, URL parameter, or no auth), request path, and extra parameters; automatic retry with backoff is supported.
    • Network adaptive: Direct connection on desktop and automatic proxy on web, with manual override available.
    • Model management: Configure the model list manually, or use one-click "Fetch Models" to sync automatically.
    • Improved usability: Toggle API key visibility; enable/disable switch; warning with the number of linked models before deletion; bilingual UI hints.

    The integration method and configuration of the existing 15 built-in platforms remain completely unaffected.

    Fixes & Improvements: Agent / MCP Stability

    • More accurate tool-call detection: Model responses are now classified into four categories — "tool call / final answer / format error / empty response" — preventing malformed output from being mistaken for a final answer and ending the conversation prematurely.
    • Automatic correction of malformed calls: When unclosed code blocks, dangling JSON, or other anomalies occur, the system automatically prompts the model to correct them, retrying up to 2 times and then reporting a clear error instead of silently hanging.
    • Avoids long runs without progress: When all tool calls are skipped and there is no progress for 2 consecutive rounds, execution terminates automatically to avoid wasted consumption.
    • Native XML tool-call compatibility: When the model outputs native formats such as <tool_calls>/<invoke>, they are parsed and executed correctly, parameter types (objects, arrays, numbers, booleans) are preserved, and incomplete calls are never executed.
    • History cleanup: Protocol residue is automatically removed from historical messages so old anomalous records no longer affect the model.
    • Prompt unification: MCP protocol descriptions are deduplicated and aligned, contradictory requirements are eliminated, and the model is explicitly forbidden from imitating the XML in examples.
    • Observability: Structured execution logs are output each round (step, classification, call count, repair, termination reason) for easier troubleshooting, without recording prompts or tool parameters.

    Upgrade note: Simply install over the existing version. No manual migration of sessions or configurations is required.