-
Notifications
You must be signed in to change notification settings - Fork 3
Settings
Dominic edited this page Jul 10, 2026
·
1 revision
Open the settings drawer from the gear icon in the left rail. Every option persists to the on-disk SQLite router_state table — changes apply on the next agent turn.
| Setting | Where | Notes |
|---|---|---|
| Orchestrator | Settings → Model Configuration | The model that replies to you. Pick a strong cloud model — this is where most of the thinking happens. |
| Atlas / Artemis | Settings → Model Configuration | Sub-agent model for research and review. Cloud or strong local both work. |
| Toolcall model (Iris / Byte / Dexter) | Settings → Model Configuration | Fast tool-execution model. Use a small local model — latency matters more than reasoning here. |
| Council Seats × 3 | Settings → Model Configuration → Council | Skeptic, Pragmatist, Synthesist. Different models = genuinely different perspectives. Blank inherits Atlas. |
| Setting | Notes |
|---|---|
| num_ctx — orchestrator | Override the orchestrator's context window. Choose a common value from the dropdown, or leave blank for the model default. |
| num_ctx — Atlas | Override for Atlas and The Council. Independent from the orchestrator ctx. |
| num_ctx — toolcall | Override for Iris / Byte / Dexter. Small models usually need less context than the orchestrator. |
| Thinking mode | Global default: On (first turn + always-think models), Max (every iteration), or Off. No per-chat override. |
| Setting | Notes |
|---|---|
| Ollama URL | Endpoint for the local Ollama daemon (typically http://localhost:11434). Models from /api/tags populate all model dropdowns, including :cloud variants. |
| Mercury | Off / RAG / Summary / Full. Controls automatic context injection from conversation memory. |
| Refresh model list | The drawer fetches the current Ollama model list when opened. Reopen it after ollama pull. |
-
Location:
logs/dockbox.log(stdout) andlogs/dockbox.error.log(stderr) under the project directory. - Size: shown next to each log path; click Refresh size to update.
-
Access: use the Process Logs tab, or
tail -f logs/dockbox.login the project directory. - Delete: truncates both files to 0 bytes. Irreversible.
- Automation model — model used for scheduled task runs (Dexter's cron jobs).
- Friendly names — display name shown in the header. Defaults to "Warden".
- Restart server — restart the Warden parent process. Use after config changes that need a fresh process.
Every setting persists to router_state in store/messages.db. There's no separate config file. To reset a value, clear the field and save — the empty string is the "use default" sentinel.
Don't manually edit router_state while Warden is running — the in-memory cache will overwrite your change. Use the dashboard or restart the server first.