Tasks
Defaults and safety rails for background tasks and cronjobs — the long-running jobs Axiom kicks off without a user sitting in front of a chat window. For what tasks are and how to create them, see Tasks & Cronjobs.
URL: /settings?tab=tasks
General Task Settings
Default provider
Provider + model used for task execution when the task itself doesn't specify one. Defaults to the currently active chat provider; override if you want background jobs to use a different model than your interactive chat.
Tasks are built for hard work — long-running, tool-heavy jobs without a human waiting on the other end. Feel free to pick a stronger, slower model here than you'd tolerate in chat (e.g. a top-tier reasoning model). Latency barely matters; quality of the end result does.
{ "tasks": { "defaultProvider": "openai:gpt-5.4" } }Max duration
Hard upper bound on a single task run, in minutes. Any task hitting this limit is killed. Default: 30. Range: 1 – 1440.
Use this to protect your wallet against runaway agents in an infinite tool loop.
Axiom has no built-in hard spend cap. Max duration and loop detection bound how long a task runs and catch runaway loops, but they do not limit absolute cost. The hard cost ceiling is set at the provider: for subscription/OAuth providers it is your plan allowance (shown as the subscriber usage quota on the Providers page), and for pay-per-token API keys it is a spend limit you configure with the provider.
{ "tasks": { "maxDurationMinutes": 30 } }Telegram delivery
How task results are pushed to the user who owns them (if Telegram is configured).
| Value | Behavior |
|---|---|
auto | Deliver via Telegram only when the user isn't actively using the web UI. |
always | Deliver via Telegram every time, even if they're online in the web UI. |
{ "tasks": { "telegramDelivery": "auto" } }Background thinking level
Reasoning level for tasks and other internal background jobs (heartbeat, consolidation-side work). Separate from the main chat thinking level so you can keep chat snappy and background jobs thoughtful (or vice versa).
Values: off, minimal, low, medium, high, xhigh, max. Like the chat setting, the dropdown only offers the levels supported by the selected Default provider model (or the active model when tasks use the active provider); unsupported levels are rounded to the nearest supported one.
{ "tasks": { "backgroundThinkingLevel": "minimal" } }Codemode (experimental)
Adds the codemode tool — a JavaScript script that batches, chains and filters tool calls, so only the script's output reaches the model — to task agents: user tasks, agent-created tasks, cronjob runs and heartbeat runs. 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 to tasks started after the change; already-running tasks are unaffected. Memory consolidation never uses codemode. Individual cronjobs can override this switch with the three-way control in the cronjob form (inherit / on / off).
{ "codemode": { "tasks": true } }The interactive chat agent has its own switch in Agent → Codemode.
When to use it, what scripts can do, the limits and where the switches' effects show up: Codemode.
Context compaction
Compact background tasks (default: on) summarizes older turns of long-running tasks so they stay inside the model's context window. It needs the master switch Agent → Context compaction → Automatic compaction; while that is off, this switch is disabled. Thresholds derive from the task model's context window. How it works: Context Compaction.
{ "compaction": { "tasks": { "enabled": true } } }Loop detection
Tasks can get stuck — repeating the same failing tool call over and over, or going in circles without making progress. Loop detection catches this and terminates the task. Which of the two it catches depends on the detection method below.
Enable loop detection
Master toggle. Default: true. Leave it on unless you're debugging an agent that actively needs to retry the same tool hundreds of times.
{ "tasks": { "loopDetection": { "enabled": true } } }Detection method
| Value | How it works |
|---|---|
systematic | Pure rule-based. Fires when the same tool is called with the same arguments and returns the same error maxConsecutiveFailures times in a row. A call counts as failed when the tool reports an error, for example edit_file not finding the text it should replace. Fast, zero extra tokens. |
smart | Periodically asks a small LLM "is this agent making progress?". Slower, costs tokens, catches subtle loops. |
auto | Start with systematic; escalate to smart if the rule-based signal is ambiguous. A good all-round choice, though the built-in default is systematic. |
{ "tasks": { "loopDetection": { "method": "auto" } } }Max consecutive failures
How many identical failing tool calls in a row count as a loop. Default: 3. Range: 1 – 20.
{ "tasks": { "loopDetection": { "maxConsecutiveFailures": 5 } } }Smart provider
Only shown for smart / auto. Which LLM judges "is the agent making progress?". Pick something small (gpt-5.4-mini, claude-haiku, etc.).
{ "tasks": { "loopDetection": { "smartProvider": "openai:gpt-5.4-mini" } } }Smart check interval
Only shown for smart / auto. How often (every N tool calls) the smart check runs. Default: 5. Range: 1 – 50. Lower = more sensitive, more tokens; higher = cheaper, slower to react.
{ "tasks": { "loopDetection": { "smartCheckInterval": 10 } } }Status updates
While a task runs, Axiom can push periodic progress signals into the parent chat (the session that triggered the task) and, when Telegram delivery is configured, into that user's Telegram DM. Each signal is a short <task_status type="periodic_update"> line with the task's runtime, tool-call count and token estimate plus the /kill_task <id> hint.
Enable periodic status updates
Master toggle. Default: false — opt-in to avoid noisy chats out of the box. Turn it on if you want a visible heartbeat for long-running background tasks.
{ "tasks": { "statusUpdates": { "enabled": true } } }Status update interval
How often to emit a status update, in minutes. Default: 10. Range: 1 – 120. Set higher for long, quiet tasks; lower if you want a more responsive heartbeat.
{ "tasks": { "statusUpdates": { "intervalMinutes": 10 } } }