Continue + Ollama (qwen2.5-coder) in agent mode: replies come back as raw JSON tool calls instead of running tools

发生场景

Source(s): https://github.com/continuedev/continue/issues/7085 Status: source reports fixed; not reproduced by us (issue closed as stale; fix confirmed by several commenters).

问题

In Continue's agent mode with a local Ollama model, the chat shows JSON blobs like {"name": ..., "arguments": ...} and no tool ever runs.

报错原文（已脱敏/规范化）

Chat prints {"name": ..., "arguments": ...} and no tool runs

适用环境

Continue VS Code extension, agent mode, local Ollama qwen2.5-coder 7b/14b/32b.

如何确认是这个问题

Agent-mode replies contain the tool-call JSON as plain text and nothing executes.

不是这个问题时 / 边界

One commenter said that with this setting on, tool calls work in Continue, but other agent CLIs (qwen-code, continue cli, opencode) with the same model still printed raw JSON - the setting only changes Continue. (our note: models with solid native tool calling don't need this.)

根因

A Continue contributor said the model doesn't do native tool calling reliably. The "system message tools" setting describes the tools in the system prompt and parses the calls out of the text instead.

修复步骤

1. In Continue's settings, turn on "Only use system message tools".
2. Restart VS Code (one confirmer restarted "all the things"). → expect: the same prompt runs the tool.

如何验证

The same prompt now runs the tool instead of printing JSON.

试过没用的办法

Switching quantization (q8_0). One user on Linux Mint got an empty Python file after turning the setting on.

风险提示

The continuedev/continue repository README says it is no longer actively maintained and is read-only.

验证范围

Continue + Ollama qwen2.5-coder, checked Oct 2026