Summary
Goal mode fires a continuation turn on a fixed, short cadence (observed: roughly every 40 seconds) while the goal is active. There is no configuration surface for this interval. Each continuation turn is a full model call — it consumes tokens/quota even when the agent has nothing to do but wait for an external event (e.g. a long-running background build, a CI run, or a dispatched subagent that will notify on completion).
Motivation
Long-running autonomous goals frequently have idle waiting phases: the agent has dispatched background work and is correctly waiting for the completion notification. During these phases the fixed tick produces dozens of no-op turns per hour, each costing tokens and (for users with notifications enabled) producing noise. The only current escape hatch is /goal pause, which also stops legitimate event-driven progress checks and must be remembered to resume.
Proposed solution
Make the continuation-turn interval configurable, e.g.:
[goal]
continuation_interval_s = 300 # default: current behavior
A reasonable addition would be an idle-aware backoff: when the previous N turns produced no tool calls and no state change, lengthen the interval exponentially up to a cap, resetting on any real progress (tool call, background-task completion, user message).
Alternatives considered
/goal pause during waits: works, but is manual, easy to forget, and disables the goal lifecycle entirely (no automatic continuation when the waited-on event fires — the agent only resumes on the user's manual action).- Ending turns with minimal output: still burns the model call and its input tokens every tick.
Additional context
Observed on Windows with kimi-code CLI, goal mode active with background tasks pending. Happy to provide session logs if useful.
Summary
Goal mode fires a continuation turn on a fixed, short cadence (observed: roughly every 40 seconds) while the goal is active. There is no configuration surface for this interval. Each continuation turn is a full model call — it consumes tokens/quota even when the agent has nothing to do but wait for an external event (e.g. a long-running background build, a CI run, or a dispatched subagent that will notify on completion).
Motivation
Long-running autonomous goals frequently have idle waiting phases: the agent has dispatched background work and is correctly waiting for the completion notification. During these phases the fixed tick produces dozens of no-op turns per hour, each costing tokens and (for users with notifications enabled) producing noise. The only current escape hatch is
/goal pause, which also stops legitimate event-driven progress checks and must be remembered to resume.Proposed solution
Make the continuation-turn interval configurable, e.g.:
A reasonable addition would be an idle-aware backoff: when the previous N turns produced no tool calls and no state change, lengthen the interval exponentially up to a cap, resetting on any real progress (tool call, background-task completion, user message).
Alternatives considered
/goal pauseduring waits: works, but is manual, easy to forget, and disables the goal lifecycle entirely (no automatic continuation when the waited-on event fires — the agent only resumes on the user's manual action).Additional context
Observed on Windows with kimi-code CLI, goal mode active with background tasks pending. Happy to provide session logs if useful.