Hollow

What's new in v5.1.0

New capability: fs_edit

Agents can now make surgical edits to existing files — replacing a specific string in a file without rewriting the whole thing. This is exposed as a new API endpoint (POST /fs/edit) and registered as a tool agents can call directly. Previously agents could only create or fully overwrite files, making incremental code changes impractical.

Model upgrade: Qwen3.5:9b-gpu everywhere

All agent LLM calls now route to qwen3.5:9b-gpu. The previous setup used mistral-nemo:12b for low-complexity tasks, which was returning empty responses on simple prompts. Everything now goes through Qwen consistently.

Context window: 8,192 tokens

Both the /ollama/chat and /ollama/generate endpoints now pass num_ctx: 8192 to Ollama, overriding the default 4,096-token cap. Larger files and longer agent reasoning chains no longer get silently truncated.

Template leak fix

Fixed a bug in autonomy_loop.py where any fs_write step that followed a shell_exec or ollama_chat would have the previous step's full output appended to the file content — even when the file's content string didn't contain a {result} placeholder. Files written by agents now contain exactly what was planned.

think: false

Qwen's extended thinking mode is disabled. Thinking is off by default; it can be re-enabled by setting "think": True in both ollama payload lines in api/server.py.

One-click launcher: HollowAgentOS.bat

New bat file at the repo root. Double-click to build the Docker image from local source, start all three services (API, dashboard, store), check for Ollama, and pull qwen3.5:9b-gpu if not already present. Ends by launching the live monitor TUI.

Requirements

• Docker Desktop (running)
• Ollama with qwen3.5:9b-gpu pulled (ollama pull qwen3.5:9b-gpu)
• Python 3.12+ (for the live monitor only)

Quick start from zip

1. Extract HollowAgentOS.zip
2. Make sure Docker Desktop is running
3. Double-click HollowAgentOS.bat
4. Open http://localhost:7778 for the dashboard

The bat file builds the Docker image from source on first run (2–5 min), then starts everything automatically.