General
Limits per turn
Applies to every session, sub-agent and review agent.
Model calls per turn
How many times one turn may call the model before it stops. Switch the limit off to let a turn run until the model is done.
Context ceiling
How full the model's context window may get before older messages are compacted into a summary.
Prompt cache
Anthropic models only. A longer cache keeps sessions you type into cheaper when you pause for more than 5 minutes between prompts.
Provider errors
What a turn does when the provider fails before it answers.
Attempts per error
How many times a turn sends a request again after it fails with a rate limit, an overload, a server error or a dropped connection. Then the turn gives up. Waits grow from 2 seconds, or follow the wait that the provider names.
Recovery waits
When every attempt failed because the provider is down or overloaded, the turn waits 15 to 60 seconds and tries again. It does this up to this many times, then it gives up. 0 gives up at once.
Storage
What runs leave on disk.
Tool Result Retention
How long a saved tool result is kept before the next launch deletes it, in whole days. 0 keeps every saved result: the sweep never runs.
Close behaviour
What happens when you close Koda?
Work still running
Defaults and config keys
| Setting | Default | config.toml |
|---|---|---|
| Model calls per turn | 100, limit on (1–10000) | max_steps, limit_steps |
| Context ceiling | 50% of the window (20–95) | context_ceiling_percent |
| Prompt cache | 5 minutes (Auto, 5 minutes, 1 hour, Off) | prompt_cache = auto, 5m, 1h, off |
| Attempts per error | 3 (1–20) | [reliability] attempts |
| Recovery waits | 5 (0–20) | [reliability] recovery_cycles |
| Tool Result Retention | 30 days. 0 keeps everything | spill_retention_days |
| Work still running | Ask (Ask, Keep running, Pause) | [background] on_close = ask, keep_running, pause |
With Prompt cache on Auto, sessions you type into use the 1-hour cache.