Skip to Content
RiusPlatformAlertsFailure modes

Failure modes

Rius files every problem it finds in a run under a failure mode, named by a catalog id such as CF-15. The id is the same wherever you meet it: on a finding in the trace view, in an alert, and in the failure_mode argument of the MCP tools. This page lists every id in the catalog and what each one means.

The ids

IdNameWhat it means
CF-01Tool call repetitionThe agent called the same tool with the same arguments again and again in one run. Reserved for the repetition detector. No finding carries this id yet.
CF-03Lost historyA call’s input shrank sharply from the previous call of the same conversation, because history was dropped or the context was compacted.
CF-15Context rotThe context became a problem: a call ran close to its model’s context window, one tool’s results took over the input, or the output was cut off at its limit.
CF-16AmnesiaReserved for the amnesia detector. No finding carries this id yet.

The catalog grows, so treat an id you do not see here as new rather than invalid.

CF-01: tool call repetition

Reserved for the repetition detector, which finds an agent stuck calling the same tool with the same arguments. No finding carries this id yet. Until it does, the Tool Loop alert is the way to catch an agent in a loop.

CF-03: lost history

Context rot reports this mode in two cases:

  • A call’s input fell by at least half and nothing reported a compaction, or a compaction was marked and explains the drop. The Context rot page describes this case.
  • Automatic compactions followed each other closely, so the agent keeps compacting instead of working.

CF-15: context rot

Context rot reports this mode in three cases:

  • Context too full: a call used at least 70% of its model’s context window.
  • One tool takes over: one tool’s results made up a large share of a call’s input and stayed that large.
  • Output cut off: a call stopped at its output token limit.

The Context rot page has the thresholds and what to do about each.

Use the ids

  • In the MCP server, pass a catalog id as failure_mode to narrow traces or findings to one mode. See MCP tools.
  • In the console, a firing or finding names its failure mode by id. Look the id up here.
  • Failure modes are separate from alert kinds. Context rot is both an alert and the source of two ids, and an alert can carry more than one id.
Last updated on