AI 程式代理
Open Interpreter 0.0.42 Adds Local Qwen Tool-Calling Diagnostics, Separating Text Responses from Actual Side Effects
The new release streamlines multi-provider model and transport settings and provides a reproducible Ollama/Qwen tool-calling smoke test. The diagnostic script independently verifies file side effects, reducing false positives when an agent merely claims in its response that it completed the work.
Open Interpreter has released version 0.0.42 for its Rust path, aligning its baseline with upstream Codex `rust-v0.153.4`. The main focus of this update is not new agent capabilities, but making it easier for users to identify the model provider, model ID, and transport protocol in use. Chat Completions settings have been added to the root command, `exec` documentation, and Bash, Zsh, Fish, and PowerShell autocompletion, while error messages and cloud settings now consistently use Open Interpreter branding.
The updated model guide lists recent model IDs from OpenAI, Google AI Studio, Anthropic, and Z.AI, while clearly distinguishing the standard Z.AI Chat Completions path from ZCode's Messages-compatible endpoint. This is useful for multi-provider agents because even a correct model name can still fail due to account permissions, region, endpoint, or wire format. The project also notes that the list is only a starting point for configuration; users should ultimately rely on `/model` and the models currently reported as available by each provider.
Of greater interest to engineering teams is the new `scripts/ollama_qwen_smoke.py`. It runs a Qwen tool-calling test against a local Ollama instance, caps the amount of output collected, cleans up the entire process group on timeout, records whether a tool-call signal actually appeared, and verifies the requested file side effect separately from the assistant's final statement. This can catch a common failure that text-based evaluations easily miss: the model says, “The file has been created,” even though it never executed the tool.
Deterministic Python tests can be incorporated into routine regression testing, but actual local model inference remains opt-in by default, and slower models may require longer timeouts. This release also provides no cross-model success-rate, latency, or resource benchmarks. It is therefore better understood as an engineering release that improves diagnosability and portability, rather than proof that Open Interpreter reliably supports every listed model. Future developments to watch include whether this smoke-test harness expands to cover more tools, side-effect types, and CI hardware environments.