Agent

Language preferences, active provider + model, reasoning level, the experimental codemode switch, and the user-editable rules the agent follows on every turn.

URL: /settings?tab=agent

Personal agent settings (regular users)

Regular users open Settings and see the personal Agent, Heartbeat, Email and Secrets tabs. The Agent tab holds their own:

• Chat model — any enabled model of any provider.
• Thinking level — off to max.
• Language — the agent's reply language and the language speech-to-text transcribes in.
• Timezone — the timezone of the agent's "now", your daily notes and your cronjobs and reminders (see Timezone).

Each field offers Use default (admin: …). On default, the user follows the admin's global settings below, including later changes by the admin. An override applies from that user's next message and affects nobody else.

The thinking-level button next to the chat input is available to everyone: it changes the user's own level, and the global level for admins. The same applies to /model and /thinking in the web chat and Telegram — see Chat → Slash commands. Regular users can also run /model default and /thinking default to go back to the default.

The remaining settings on this page (rules, provider, codemode, uploads, resilience, …) and all other tabs are admin-only. Admin settings are the global defaults for everyone; admins have no personal layer.

Agent Rules

A card at the top of the panel links to /data/config/AGENTS.md — the file the agent reads on every conversation as its "user-editable behavior rules" block.

Click Open editor to jump to the Agent Instructions page, where the file opens in a full-screen Markdown editor with a Restore default button.

See Agent Instructions for what belongs in this file, the default template, and editing tips.

Agent language

Forces the agent's reply language. Mapped to the <language> block in the system prompt.

Value	Behavior
match	Mirror the user's language on each turn.
English, German, French, Spanish, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean	Reply in this language regardless of the user's input.

Applies immediately on the next turn.

{ "language": "German" }

Timezone

The current date in this timezone is injected into the system prompt, and the minute-level time is appended to each user message, so the agent always knows "now" (the time is kept out of the system prompt to avoid invalidating provider prompt caches every minute — see System Prompt → layer 16). Also used for cron evaluation (Tasks & Heartbeat) and the naming of memory/daily/<date>.md files.

Default: UTC. Mirrors the container's TZ env var if set.

On a multi-user instance this is the global default (the admin's timezone). Every other user can pick their own timezone under Settings → Agent or keep Use default (admin: …). A user's timezone drives:

• the current date and time the agent sees in that user's chats,
• the file date and HH:MM headings of that user's daily notes,
• when that user's cronjobs and reminders fire.

When a user changes their timezone, their enabled cronjobs and reminders are rescheduled immediately. Admins can set a user's language and timezone when creating the account on the Users page.

{ "timezone": "Europe/Vienna" }

Provider

The active provider + model used for all normal chat conversations. The dropdown shows every enabled model across every configured provider (e.g. ChatGPT Plus (gpt-5.4-mini), Anthropic (claude-sonnet-4), Ollama (qwen2.5-coder)).

Changing this value activates the chosen combination immediately — in-flight sessions will use the new provider on their next turn.

Internally stored as two keys:

{
  "activeProviderId": "openai-chatgpt-plus",
  "activeModelId": "gpt-5.4-mini"
}

Configure providers themselves (add new ones, enable/disable models, set API keys) on the Providers page, not here.

Thinking level

How hard the main chat agent reasons before replying. Higher levels are slower and more expensive.

The dropdown only offers the levels the active model supports (from the model catalog or the model's Edit Model settings). A model without reasoning only offers off. If the stored level isn't supported by the current model — e.g. after switching models — it is marked not supported and the model runs with the nearest supported level instead (the next higher one first, otherwise the next lower one).

Value	Use for
off	Plain chat, no reasoning tokens.
minimal	Tiny amount of reasoning — default for most people.
low	Quick internal planning.
medium	Multi-step problems.
high	Hard reasoning, tool-heavy flows.
xhigh	Extra effort — only models that explicitly support it.
max	Maximum effort — only models that explicitly support it.

This only applies to the interactive chat agent. Background jobs (tasks, heartbeat) have their own setting in Tasks → Background thinking level.

{ "thinkingLevel": "minimal" }

Codemode (experimental)

Adds the codemode tool to the interactive chat agent on every channel (web UI and Telegram) — a JavaScript script that batches, chains and filters tool calls, so only the script's output reaches the model. Off by default, and the section is marked Experimental in the UI: the feature may change between releases, and script quality depends on the model. Applies from the next turn — no restart needed.

Background task agents have their own switch in Tasks → Codemode.

{ "codemode": { "mainAgent": true } }

Upload retention

How many days uploaded files in /data/uploads/ are kept before the cleanup job removes them. Applies to images, audio, and any other files users attach from the web UI or Telegram. Default: 30. Set to 0 to have the next cleanup run delete all uploads.

{
  "uploads": {
    "retentionDays": 30
  }
}

The cleanup job also prunes the referencing rows in the database so stale upload metadata doesn't linger after the files are gone.

Resilience

How the agent reacts when a provider misbehaves mid-turn. All five fields are read at the start of every turn, so a save applies to the next message — no restart needed.

Automatic retry

When a turn fails with a transient provider error (429, 5xx, timeout, dropped stream, or a watchdog stall abort), Axiom discards the failed attempt and re-runs the turn from the existing transcript — the user message is never sent twice. Errors the provider won't recover from on its own (invalid API key, quota, billing) fail immediately, and a turn you stop yourself is never retried.

Field	Default	Range	Effect
Automatic retry	on	—	Master switch. Off means every provider error ends the turn immediately.
Maximum retries	3	0 – 10	Retry budget per turn. 0 behaves like the switch being off.
Base delay	2000 ms	100 – 60000 ms	Backoff base; attempt n waits base × 2^(n-1) — with the defaults 2s / 4s / 8s.

While a retry is pending the chat shows a Retrying (n/max)… status. Once the budget is exhausted, the turn ends with a persisted error message containing the provider's error text.

{
  "retry": { "enabled": true, "maxRetries": 3, "baseDelayMs": 2000 }
}

Stall thresholds

The watchdog measures how long a turn goes without a single chunk from the provider. At the warn threshold a provider_stall message is written to the chat (it survives a reload and is updated in place when the provider recovers); at the abort threshold the stream is hard-aborted, which counts as a retryable error. Time spent executing tool calls (shell commands, image generation, …) does not count — the clock only runs while Axiom is waiting for the provider.

Field	Default	Range
Stall warning threshold	30000 ms	1000 – 600000 ms
Stall abort threshold	90000 ms	1000 – 3600000 ms, must be ≥ the warning threshold

{
  "watchdog": { "stallWarnMs": 30000, "stallAbortMs": 90000 }
}

Stall frequency and average duration (split by recovered vs. aborted) are aggregated on the Token Usage page. Telegram delivery of the warning is opt-in — see Telegram → Send stall warnings.

Context compaction

When a chat approaches the model's context window, older messages are replaced by a summary and the most recent messages are kept verbatim. All thresholds derive from the active model's context window — there are no budgets to set. How this works — triggers, the summary format, the per-window caps — is explained in Context Compaction. The switch is read fresh before every check, so a save applies without a restart.

Field	Default	Effect
Automatic compaction	on	Master switch for automatic compaction (main chat and tasks). /compact works even when off.

Whether background tasks are compacted too is switched under Tasks → Context compaction.

{ "compaction": { "enabled": true } }