Runtime & UI

These settings control the technical limits and the default behavior of the agents when interacting with tools and the long-term memory.

1. Tool Loop Timeout (Seconds)

Determines the maximum time an agent may spend in a single “loop” calling tools.

• Range: 60 to 3600 seconds.
• Default: 600 seconds (10 minutes).
• Purpose: Prevents agents from getting into infinite tool call loops or consuming excessive resources when they cannot find a solution.

2. Memory Context Size (Top K)

Defines how many relevant fragments from the vector memory are passed to the LLM per request.

• Range: 1 to 50 entries.
• Default: 5 entries.
• Note: Higher values provide more context but consume more tokens and can confuse the model (“Lost in the Middle”).

3. Automatic Memory Storage

Controls the agents’ default write access to the memory.

• Allow write access: If active, agents can automatically store important information from the conversation in the long-term memory.
• Effect: Applies as the default for all new agents/tasks, but can be overridden by specific policies (see Memory documentation).

4. Provider Requests per Minute

A global rate limiting for outgoing API calls to AI providers (OpenAI, Anthropic etc.).

• Range: 1 to 500 requests.
• Default: 10 requests per minute.
• Purpose: Protection against unexpected costs and avoidance of “429 Too Many Requests” errors at the providers.

5. System Timezone

Determines the local time for the entire Ontheia host.

• Format: IANA timezone string (e.g., Europe/Berlin, UTC).
• Default: Europe/Berlin (or value from APP_TIMEZONE).
• Effect:
  • Chat Titles: Automatically generated titles use this timezone for dates.
  • Logs (Trace): Events are converted to this local time for display.
  • Cron Jobs: Schedules are executed based on this timezone.
  • Agent Context: The “Current Time” injected into the agent follows this setting.

6. Response Streaming

A global switch that enables or disables token-by-token streaming of LLM responses into the chat.

• Default: enabled.
• Effect: When enabled, agent responses appear in the chat while the model generates them. When disabled, the full response appears as one block after generation completes.

Scope: The switch applies to the Anthropic API path and all OpenAI-compatible providers (OpenAI, xAI, Google, Ollama, etc.). CLI providers (e.g. Claude CLI) always deliver the response as a block by nature. Individual OpenAI-compatible providers whose endpoint does not support SSE can be excluded via provider or model metadata with "stream": false — they then keep responding as a block regardless of the global switch.

When to disable? Only when something misbehaves — for example if a provider rejects streaming requests, or a reverse proxy in front of Ontheia buffers SSE responses so streaming never reaches the browser anyway.

7. Prompt Caching (Anthropic API)

A global switch that enables or disables prompt caching on the Anthropic API path.

• Default: enabled.
• Effect: When enabled, Ontheia places cache_control markers on the stable prefix (tools + system prompt) and the growing chat history. Recurring requests then read that prefix at the heavily reduced cache price (~0.1× input).

Why Anthropic only? Anthropic is the provider where caching can cost more than it saves without any way for us to opt out per model: writing the cache (cache_creation) is billed at ~1.25× the normal input price. If the cached prefix is not re-read within the 5-minute TTL — e.g. for sporadic single-shot runs (cron tasks with a single LLM call, long thinking pauses in chat) — you pay the write premium without ever collecting the cheap read: +25% instead of savings. With Anthropic, we decide whether cache_control is sent — which is why an on/off switch works here technically.

OpenAI (except gpt-5.6), xAI and Mistral cache automatically and without a write premium (there is only “Input” and “Cached input”, no write column). Caching can never be more expensive there than no caching — this switch does not affect them.

Exception since gpt-5.6 (July 2026): the gpt-5.6 models (sol/terra/luna) now also bill a 1.25× cache-write premium — the exact same mechanic as Anthropic (pricing table example: gpt-5.6-sol $5.00 input → $6.25 cache write). Unlike Anthropic, though, this caching is implicit on OpenAI’s side (mode auto, the documented default) — Ontheia sends no parameter of its own that this switch could toggle. This switch therefore does not cover gpt-5.6. OpenAI instead offers an explicit mode where you mark exactly which prefixes to cache — that isn’t a simple on/off setting but a structural change to how the request is built, and it isn’t implemented yet (see roadmap: Responses API expansion).

When to disable? If your Anthropic usage consists mostly of sporadic single-shot runs and you see no cache savings (the ⚡ symbol with cache-read tokens) in the response token usage. For dense tool loops and fast chat sequences the switch should stay enabled — the cheap read clearly dominates there.

Observe gpt-5.6 instead of switching it off: the ⚡ badge in the trace panel now also surfaces OpenAI’s cache-write tokens (as cache creation, mirroring Anthropic). Use it to judge whether gpt-5.6 is actually costing you extra in your usage pattern (sporadic runs) or whether repeated reads make it a net saving (dense tool loops, fast chat sequences).