[Bug]: Tool call still not working

[SyscallBrain]

Summary

2026-02-15T22:43:00.354808Z WARN zeroclaw::agent::loop_: Malformed <tool_call> JSON: expected value at line 1 column 1

Current behavior

2026-02-15T22:43:00.354808Z WARN zeroclaw::agent::loop_: Malformed <tool_call> JSON: expected value at line 1 column 1

Expected behavior

Execute the tool

Steps to reproduce

zeroclaw agent

Logs / stack traces

ZeroClaw version

0.1.0

Rust version

1.92.0

Operating system

ArchLinux

Regression?

No, first-time setup

[casidiablo]

I was going to open a related issue. I am going to add it here as to not add more noise.

In Telegram, the tools do not seem to be called but included in the response message instead:

Needless to say the files are never actually updated. This happens with the most common models out there (Sonnet 4.5, Opus 4.6, Gemini 3 Pro)

[jmlago]

Same issue here

[mai1015]

Right now, this is not that standardized as tools called are not specified in the OpenAI API (with tool definition), so some models cannot use any tool. They need tools like Claude or GPT to call tools properly.

[neuronbuzdes]

I am using ZeroClaw with a local setup (Ollama + Qwen2.5-Coder:32B) on an RTX 4090. The model correctly identifies the need to use a tool (e.g., file_write) and generates a valid JSON schema for the tool call. However, ZeroClaw fails to intercept this JSON and instead prints it as raw text in the chat interface.

This prevents any autonomous actions like file creation or shell execution, rendering the "autonomous" mode useless for local models. This seems closely related to issue #234.

Environment:
OS: Windows 11 (PowerShell)
GPU: RTX 4090
Model: Qwen2.5-Coder:32B (via Ollama)

ZeroClaw version: Latest build from source (Rust)
Steps to Reproduce:
Configure config.toml to use provider = "ollama" and level = "full".
Run cargo run -- agent.
Ask the agent: "Create a simple python script and save it to test.py".

Actual Behavior:
The agent responds with: "Sure, I will save the file..." followed by a raw JSON block like {"name": "file_write", "arguments": {...}} printed directly in the terminal. No file is created.

Expected Behavior:
ZeroClaw should parse the JSON output, execute the file_write tool, and confirm the file creation.

[neuronbuzdes]

The project states it is a "Sovereign AI stack", but in practice, there is a massive gap in the implementation of tool calling for local providers like Ollama. Even when the model correctly identifies the tool and generates valid JSON, the system fails to execute it, simply printing the output to the terminal.

An "autonomous agent" that requires the user to manually copy-paste JSON commands to perform a file write is not an agent—it's just a chat interface.

Are there any plans to implement a more robust parser that can handle the conversational "noise" typical of local models? Currently, the "Full Autonomy" level seems to be hardcoded/optimized for proprietary APIs only, which contradicts the "100% Agnostic" and "Sovereign" spirit of the project.

[chumyin]

Opened a fix PR for this: #252\n\nThis patch makes tool-call parsing robust against noisy outputs from local/open-source models (e.g. markdown fenced JSON, conversational text around JSON, or raw JSON tool-call objects outside tags).\n\nIt also adds regression tests covering these cases.\n\nIf reviewers agree, merging #252 should resolve the tool-call interception failures tracked here.