Agent Terminal & Filesystem Capabilities
Also available: Markdown · Plain text
Agent Terminal & Filesystem Capabilities
Status: Implemented in v0.2 agent runtime (Phases 0–3).
Summary
Gnomad is a command-and-filesystem agent, not a chatbot that pretends to run shell commands. Cloud chat uses LLM tool calls (shell_run, fs_*, workspace_info). Local Ollama falls back to <gnomad-run> tags.
Trust modes
| Mode | Behavior |
|---|---|
| Standard (default) | Read/write/search under workspace ($HOME until changed). Paths outside workspace need approval. Sudo Gate for risky shell. |
| YOLO! | Full-machine file access; fewer path prompts. Sudo Gate may still apply for destructive shell. |
Architecture
- Rust:
agent_runtime,agent_fs,agent_settings,shell_session(PTY + cwd tracking) - Frontend:
agentLoop.tsorchestrates up to 10 tool steps per message - Audit:
{app_data}/gnomad/agent-audit.jsonl
User actions
- Settings → Agent access: workspace folder, trust mode
- Stop while thinking: interrupts PTY
- Paste
$ commandor use terminal button for direct run
See plan history in .cursor/plans/ for full approach comparison.
Local command planner (optional)
Settings → Agent access → Local command planner:
- Checkbox to enable
- Planner model (Ollama) — small/fast tag (e.g.
llama3.2:1b), separate from main chat model - Optional Use chat local model when in Ollama mode
- GGUF path — stored for future direct inference; v1 uses Ollama only
Runs when the main model emits prose instead of a valid shell command.