agent.default_watchdog¶
agent.default_watchdog(
) -> {
max_consecutive_errors: number?,
max_rounds_without_success: number?,
refractory_rounds: number?,
refractory_seconds: number?,
repeat_sentence_window: number?,
repeat_sentence_threshold: number?,
repeat_sentence_threshold_loose: number?,
min_sentence_chars: number?,
min_armed_sentences: number?,
repeat_call_window: number?,
repeat_call_threshold: number?,
repeat_call_max_len: number?,
max_think_nudges: number?,
interrupt_on_cap: boolean?,
nudge_text: string?,
break_loop_text: string?,
}?
Returns the watchdog policy set by agent.set_default_watchdog, or nil
when none was set. A personality can adopt it explicitly by writing
watchdog = agent.default_watchdog(); when unset that reads as nil,
leaving the personality's watchdog off.
Returns
{max_consecutive_errors: number?, max_rounds_without_success: number?, refractory_rounds: number?, refractory_seconds: number?, repeat_sentence_window: number?, repeat_sentence_threshold: number?, repeat_sentence_threshold_loose: number?, min_sentence_chars: number?, min_armed_sentences: number?, repeat_call_window: number?, repeat_call_threshold: number?, repeat_call_max_len: number?, max_think_nudges: number?, interrupt_on_cap: boolean?, nudge_text: string?, break_loop_text: string?}?max_consecutive_errors(defaultdefaults :: MAX_CONSECUTIVE_ERRORS) — The number of consecutive real tool failures before the error-streak trigger arms.max_rounds_without_success(defaultdefaults :: MAX_ROUNDS_WITHOUT_SUCCESS) — The number of consecutive rounds ending without a successful tool result before the no-progress trigger arms.refractory_rounds(defaultdefaults :: REFRACTORY_ROUNDS) — The number of rounds that must complete after a nudge before another may fire. Both this andrefractory_secondsmust elapse.refractory_seconds(defaultdefaults :: REFRACTORY_SECONDS) — The wall-clock delay that must elapse after a nudge before another may fire, in seconds. Both this andrefractory_roundsmust elapse.repeat_sentence_window(defaultdefaults :: REPEAT_SENTENCE_WINDOW) — The number of sentences the repetition detector keeps in its sliding window.repeat_sentence_threshold(defaultdefaults :: REPEAT_SENTENCE_THRESHOLD) — The number of occurrences of one normalized sentence before the strict tier fires.repeat_sentence_threshold_loose(defaultdefaults :: REPEAT_SENTENCE_THRESHOLD_LOOSE) — Occurrences of one content-word set before the loose tier fires.min_sentence_chars(defaultdefaults :: MIN_SENTENCE_CHARS) — Sentences shorter than this are not counted.min_armed_sentences(defaultdefaults :: MIN_ARMED_SENTENCES) — Sentences that must have been observed before the detector arms.repeat_call_window(defaultdefaults :: REPEAT_CALL_WINDOW) — Tool invocations the call-repetition detector keeps in its window.repeat_call_threshold(defaultdefaults :: REPEAT_CALL_THRESHOLD) — The number of occurrences of one call sequence before the call-repetition trigger arms.repeat_call_max_len(defaultdefaults :: REPEAT_CALL_MAX_LEN) — The maximum length of the call sequence the detector counts.max_think_nudges(defaultdefaults :: MAX_THINK_NUDGES) — The per-turn cap on break-the-loop nudges.interrupt_on_cap(defaulttrue) — Whether the turn is interrupted when the break-the-loop nudge cap is exceeded. When false the watchdog does nothing at the cap.nudge_text— The escalation nudge text, used verbatim.break_loop_text— The break-the-loop nudge text, used verbatim.