Configuration

Most day-to-day platform options live under AI ‣ Configuration. Global defaults that map to system parameters are on the Settings page (Settings users only).

Settings

Open AI ‣ Configuration ‣ Settings (or the AI section on the main Settings app).

Guided system setup

• Start setup assistant — opens chat with the guided setup skill for system administrators. Module install is never free-form: when setup plans are used, install goes through a confirmable plan (MVP may propose only).

Abuse guardrails

Setting	Meaning
Ai Max Strikes	Violations in the window before a temporary ban (default 3).
Ai Strike Window Hours	Hours over which strikes accumulate (default 24).
Ai Ban Minutes	How long a ban lasts (default 60).

Rate and usage limits

Setting	Meaning
Ai Rate Limit Rpm	Max requests per minute per user (default 20).
Ai Max Daily Calls	0 = unlimited.
Ai Max Daily Cost Usd	Soft cost cap using model unit prices; 0 = unlimited. Unpriced models may not enforce cost well.
Ai Max Conversations Per Day	0 = unlimited.
Ai Max Successive Calls	Bound on chained model turns (default 10).
Ai Max Tool Calls	Bound on tool invocations per turn loop (default 10).
Ai Write Mode	Interactive chat only — auto / confirm / hybrid. Agent runs use each task’s own Write Mode instead (default confirm). See Using the assistant and Task write mode (agent runs only).
Chat write-proposal lifetime (minutes)	Minutes until a chat proposal expires; 0 = never.
Agent write-proposal lifetime (minutes)	Minutes until an agent proposal expires (default 1440); 0 = never.

Logging

• Ai Log Level — Metadata only (default, recommended) or Full payload (stores message/tool content; high sensitivity).

• Run Step Retention (Days) — purge horizon for ai.agent.run.step journal rows (default 90; 0 disables). Does not auto-purge ai.log or run headers.

Model

• AI Model — the model used by the interactive assistant. Its provider is used automatically.

This setting covers the interactive assistant only. A per-agent completion cap is set on the agent itself (Agents) and depends on the model record — see Completion caps.

Access policy defaults

Applied when no more specific ai.access.rule decides:

• Default read: Allow

• Default create / write / delete: Deny

See Access policy for rules and groups.

Runtime introspection

• Publish Installed Modules — whether environment facts may include the installed module list (still subject to the user’s ir.module.module ACL). Denying makes the assistant more cautious, not magically more accurate. Version, edition, company and user facts remain.

Web access and files

Requires the Web access capability to be enabled for tools to be offered.

Setting	Meaning
Search Backend	ai.web.provider used by web_search. Empty = search off.
Fetch Backend	Provider for web_fetch; empty falls back to built-in Direct fetch.
Allowed Domains	Comma-separated allow-list; empty = any public host (SSRF protections still apply to private ranges).
Daily Limit	Max web calls per user per company; 0 = unlimited.
Max Page Text	Characters retained from a fetched page.
Max Query Length	Cap on search query size.
Allowed Image Types, Allowed Document Types	MIME allow-lists for ingest.
Max File Size	Bytes (default 10 MiB).

System parameters and scheduled actions

A few platform switches have no field on the Settings page. They live in the Technical menus, which require developer mode.

System parameters

Settings ‣ Technical ‣ Parameters ‣ System Parameters

• ai.triage_enabled — global off switch for stage-1 triage, covering both the inbound check and a task’s Triage On Dispatch (default True). Set it to False and every instruction is executed unassessed.

• ai.waiting_approval_ttl_minutes — how long a run may sit waiting for a supervisor’s answer before it is failed (default 1440; 0 means never expire by age).

• ai.run_reaper_grace_seconds — how long past a task’s own time limit the stuck-run reaper waits before it declares a run abandoned (default 600, i.e. 10 minutes). This is not a kill switch: to stop reaping altogether, deactivate the scheduled action instead.

Scheduled actions

Settings ‣ Technical ‣ Automation ‣ Scheduled Actions

The module ships seven scheduled actions, all active after install:

• AI: Expire user bans (hourly) — lifts the temporary bans whose end time has passed.

• AI: Expire pending writes (every 15 minutes) — expires write proposals older than their configured lifetime. It first releases the proposals left behind by a run that has already finished, so a lifetime of 0 does not also disable that sweep.

• AI: dispatch scheduled agent tasks (every 5 minutes) — starts the agent tasks that are due. This is only how often the dispatcher itself wakes; each task’s own interval decides its cadence, so the cron has to wake at least as often as the shortest interval configured on a task.

• AI: drain inbound addressing requests (every minute) — starts the runs queued when somebody addresses an agent through a Discuss message, a chatter mention, an activity or an assignment. It is deliberately more frequent than the dispatcher, because somebody is waiting for the reply.

• AI: expire waiting-approval agent runs (every 15 minutes) — fails the runs that have waited for a supervisor’s answer longer than ai.waiting_approval_ttl_minutes, and expires the proposals they were waiting on.

• AI: reap stuck agent runs (every 15 minutes) — closes runs whose worker died, and also frees runs parked for approval with nothing left to approve. Both jobs, their cadence and their batch limit are described in Runs whose worker died.

• AI: memory maintenance (every 6 hours) — archives working memories past their expiry date, lets the salience of untouched entries decay, and flags high-salience personal entries as promotion candidates for an administrator.

LLM providers and models

AI ‣ Configuration ‣ Providers ‣ LLM Providers

Fields of interest:

• Provider type and API key (system parameter storage).

• Base Url — only override when you understand the trust boundary (a writable base URL can be an SSRF risk if pointed at internal services).

• Timeout, Max Retries, Call Deadline — wall-clock budget for one provider call including retries.

AI ‣ Configuration ‣ Providers ‣ Models

• Model — exact vendor id.

• Context Window, Max Output Tokens, tool and temperature support flags.

• Prompt / completion unit prices and price source (provider API, curated catalog, or manual — manual is not overwritten by refresh).

Completion caps

Max Output Tokens on the model is what the assistant caps its completions at, and it is also the precondition for a per-agent cap. Only providers that publish a completion limit fill it in; 0 means unknown.

An agent’s own Max Tokens (0 = no override) (Agents) reaches every provider — OpenAI, OpenRouter, xAI, Google Gemini and Anthropic — each under the parameter name its API expects. It is put on the wire only for a model that publishes a Max Output Tokens of its own. Point an agent at a model that still reads 0 and the agent’s value is dropped: the model writes to its own limit, and nothing on the agent form says why. Fill in Max Output Tokens on the model if you want the per-agent cap to bite.

0 on the agent — the default — means no override: the model’s own output cap applies, exactly as for a turn with no agent at all. Set a positive value only where you deliberately want that agent to be terser or cheaper, and remember that on reasoning models the cap counts thinking tokens too, so a low value truncates answers.

Note

Anthropic is the exception, in the other direction: its API always requires the parameter. An Anthropic model whose Max Output Tokens is still 0 therefore falls back to a conservative built-in value of 1024 tokens, not to the model’s real limit.

Warning

Max Output Tokens on a catalogue row is often almost the whole Context Window (for example 450 000 of 500 000). Providers such as OpenRouter reject the call when prompt + tools + max_tokens exceeds the window, even if the prompt itself still fits. The platform therefore clamps the completion cap on every call to whatever still fits after an overestimate of the prompt. A small first turn still gets the published output limit; a large tool round gets a smaller leftover. This cannot help when the prompt alone already exceeds the window — split the work (see Per-record work items).

Warning

Upgrading from an earlier version changes how agents that are already configured behave.

• A cap an administrator had already typed on an agent used to reach Anthropic only. It now applies on every provider, so a low value that was inert on OpenAI, OpenRouter, xAI or Gemini starts truncating answers there.

• The upgrade resets the old shipped default of 2048 to 0 on every agent still carrying it, so that no database keeps a cap nobody chose. On an Anthropic model that publishes a cap of its own, that 2048 was genuinely in force: those agents may now write up to the model’s full Max Output Tokens — more tokens, more cost and more time, and an answer long enough to exhaust the provider’s Call Deadline fails with an error instead of stopping at 2048 tokens.

Any other value you typed is kept. Re-enter a cap on every agent you meant to limit, including one you had deliberately set to 2048.

Web providers

AI ‣ Configuration ‣ Providers ‣ Web Providers

Configure search/fetch backends (factory rows may ship for common engines). Assign them in Settings as search/fetch backends.

MCP servers

AI ‣ Configuration ‣ Providers ‣ MCP Servers

Model Context Protocol servers expose external tools to the assistant when the MCP tools capability is enabled. Factory MCP definitions may ship inactive or empty of secrets.

Per server, typically configure:

• endpoint / transport;

• authentication secrets (never commit them to docs or git);

• which users/groups may use the server’s tools;

• enable/disable.

Every MCP call still re-checks capability and server audience. Treat MCP like installing untrusted plugins: least privilege, no production secrets in tool responses you do not need.

Prompt layers

AI ‣ Configuration ‣ Behavior ‣ Prompt Layers

Layered system prompts (core, global, module, …) shape behaviour. Core safety layers are not meant for casual editing. Global operational layers may be adjusted by AI administrators; migrations may refresh factory content only when still pristine.

Important

A layer’s Body is English only and not translatable. These are the rails the model is steered by, not user-facing text: rails that switched language with the interface would leave the instructions in one language and the tool names in another. The assistant still answers in the language of the conversation — that is decided by the user’s messages, never by this text.

On a database upgraded from an earlier version, each layer keeps its English body and any translated bodies are discarded. Re-check every global and per-module layer whose body you had edited in a translated interface.

Blocked patterns

AI ‣ Configuration ‣ Security ‣ Blocked Patterns

Regex (or pattern) rules on input/output that raise violations when matched (e.g. classic “ignore previous instructions” probes). Extend carefully — over-broad patterns create false positives and strike noise.

Allowed actions

AI ‣ Configuration ‣ Security ‣ Allowed Actions

Allow-list of business methods the Action capability may invoke via call_action (e.g. mark activity done). Methods not listed cannot be called through AI even if Write is on. Prefer allow-listed actions over raw field writes for lifecycle transitions.

Customizations ledger

AI ‣ Configuration ‣ Customizations

When Customize UI & schema is used, changes are tracked in a ledger with import/export for recovery. Primary UX is still chat tools for admins; the ledger is the admin recovery path.

Setup plans

AI ‣ Configuration ‣ Setup Plans (Settings users)

Plans proposed by the guided setup flow for module installation / configuration steps that require explicit confirmation.