Hollow

What changed

This release fixes the agent loop problem — the single most critical blocker before anything else in the roadmap can work.

The problem

Agents were cycling on the same goal indefinitely. Scout had been stuck on one goal for 900+ daemon cycles (several hours). Every 6 seconds it tried, hit the same wall, and got scheduled again. Nothing was getting done.

Three fixes

1. Stall detection now abandons the goal (daemon.py)
When an agent makes no progress for 5 consecutive cycles, the daemon used to just cool it off and let it retry the same goal later. Now it abandons the goal. The agent gets a fresh, identity-driven task after cooling off.

2. Artifact validation failure counter (autonomy_loop.py)
When a goal reaches 100% progress but produces no verifiable artifact, it used to reset to 85% and retry forever. Now it tracks artifact_check_failures. After 3 failures, the goal is permanently abandoned with the reason logged to memory.

3. Timeouts actually enforced (execution_engine.py)
_call_with_timeout accepted a timeout_ms parameter but called the function directly with no timeout mechanism. A hanging Ollama call would block a daemon thread indefinitely. Now uses ThreadPoolExecutor.submit().result(timeout=...) to actually enforce the limit.

Also fixed: JSONL log writers were reading entire files then rewriting on every step (O(n) per write). Changed to append mode.

Confirmed

23:53:40 [daemon] WARNING  scout stalled on 'Identify commented-out code...' — goal abandoned
23:54:23 [daemon] INFO     scout → goal=goal-9aab1fca8d30 progress=0.30 steps=3
23:54:37 [daemon] INFO     scout → goal=goal-9aab1fca8d30 progress=0.60 steps=3

Also in this release

• ROADMAP.md — 6-phase roadmap from stabilization through standalone OS, with decision log
• AUTONOMY_LOG.md — research record distinguishing human-designed decisions from autonomous agent decisions
• monitor.py — workspace file viewer (f key), agent name resolution from names.json
• live_capabilities.py — fs_write append mode so agents can share logs without overwriting each other

Install

Windows — double-click:

1. Download ZIP and extract
2. Double-click install.bat
3. Done — Docker, Ollama, models, and the TUI are all handled automatically

Mac/Linux:
```bash
ollama pull mistral-nemo:12b && ollama pull nomic-embed-text
git clone https://github.com/ninjahawk/hollow-agentOS && cd hollow-agentOS
cp config.example.json config.json && docker compose up -d
pip install -r requirements-monitor.txt && python monitor.py
```

Pre-built Docker image (no build step):
```bash
docker pull ghcr.io/ninjahawk/hollow-agentos:latest
```