runtime/

watchdog.rs

// This Source Code Form is subject to the terms of the Mozilla Public
// License, v. 2.0. If a copy of the MPL was not distributed with this
// file, You can obtain one at https://mozilla.org/MPL/2.0/.
// Copyright (c) 2026 2youg1 and the sprawling contributors

//! The disposal surface. The stall verdict is
//! `kernel::stall`'s alone; this module never re-derives it and never
//! forwards its internals — it answers one question: what happens next.
//! Disposal is graded on purpose: a corrective steer first (name the
//! repetition to the model), freezing only when correction failed.
//! Terminal-only watchdogs kill recoverable sessions; that lesson is the
//! reason this type exists.

use std::num::NonZeroU64;

use kernel::event::record::{FiredAction, WatchdogFired};
use kernel::{AxCode, AxError, Payload, Retries, Retry, RunId, StallVerdict, TimeMs};

/// One watchdog per run: it holds the correction history and the ceiling
/// the person set on retries, nothing else.
#[derive(Debug, Default)]
pub struct Watchdog {
    corrections: u32,
    provider_failures: u32,
    /// Failures since the provider last answered. The backoff and the
    /// person's ceiling both read this rather than the run's total: a
    /// provider that recovered owes the next outage no minute-long
    /// first wait, and a ceiling on retries is a ceiling on asking the
    /// same call again.
    streak: u32,
    retries: Retries,
    /// Folded from the run's id once, so the run's jitter is the same
    /// on every replay and different from its neighbours'.
    jitter_seed: u64,
}

/// Deliberately exhaustive: a new disposal must force every run loop to
/// decide, not fall through a catch-all.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum Disposal {
    Proceed,
    CorrectiveSteer {
        text: String,
    },
    /// Try the same call again, not before this moment. The moment
    /// comes from the backoff schedule [`Watchdog`] owns, and the
    /// failure travels with it: "backed off" without what it backed
    /// off from is a line nobody can act on, and a second writer of
    /// that half of the fact could drift from this one.
    BackOff {
        until: TimeMs,
        code: AxCode,
        subject: String,
    },
    Freeze {
        reason: FreezeReason,
    },
}

impl Disposal {
    /// The verdict that asks the same call again, holding what failed.
    fn backing_off(until: TimeMs, failure: &AxError) -> Disposal {
        Disposal::BackOff {
            until,
            code: *failure.code(),
            subject: failure.subject().to_owned(),
        }
    }
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum FreezeReason {
    Stall,
    ProviderRefused,
}

impl FreezeReason {
    fn as_str(self) -> &'static str {
        match self {
            FreezeReason::Stall => "stall",
            FreezeReason::ProviderRefused => "provider_refused",
        }
    }
}

impl Watchdog {
    #[must_use]
    pub fn new(retries: Retries, run: RunId) -> Watchdog {
        Watchdog {
            retries,
            jitter_seed: seed_of(run),
            ..Watchdog::default()
        }
    }

    /// Consumes the stall verdict verbatim. First hit: a corrective
    /// steer naming the repetition. Second hit: freeze.
    pub fn on_stall(&mut self, verdict: &StallVerdict) -> Disposal {
        match verdict {
            StallVerdict::Ok => Disposal::Proceed,
            StallVerdict::Stall { repeats } => {
                if self.corrections == 0 {
                    self.corrections = 1;
                    Disposal::CorrectiveSteer {
                        text: format!(
                            "You have repeated the same call {repeats} times in a row. \
                             Change the approach or report why the goal cannot be met."
                        ),
                    }
                } else {
                    Disposal::Freeze {
                        reason: FreezeReason::Stall,
                    }
                }
            }
        }
    }

    /// Classifies one provider failure. The producer of the error
    /// already decided whether the same call is worth making again, so
    /// this reads `AxError::retry` rather than guessing a second time.
    ///
    /// `Retry::No` freezes on the first one: repeating a request the
    /// provider rejected on its shape buys the same rejection again.
    /// `Yes` and `Unknown` back off (a model call's only effect is an
    /// answer the city never received) from `now` by the
    /// schedule [`Watchdog::backoff_ms`] owns, or by the provider's own
    /// `retry_after_ms` when that is longer: asking before the time it
    /// named buys one more refusal.
    ///
    /// **A retriable failure freezes only against a ceiling the person
    /// set.** Under [`Retries::UntilHalted`] what stops a run that keeps
    /// failing is `Halt`, the city's one brake; under
    /// [`Retries::AtMost`] it is the number they entered, because a
    /// number entered on a form that nothing reads is worse than no
    /// field at all.
    pub fn on_provider_failure(&mut self, failure: &AxError, now: TimeMs) -> Disposal {
        self.provider_failures = self.provider_failures.saturating_add(1);
        self.streak = self.streak.saturating_add(1);
        let own = self.backoff_ms();
        let wait = failure.retry_after_ms().map_or(own, |told| own.max(told));
        let not_before = TimeMs::new(now.value().saturating_add(wait));
        let refused = Disposal::Freeze {
            reason: FreezeReason::ProviderRefused,
        };
        if failure.retry() == Retry::No {
            return refused;
        }
        match self.retries {
            Retries::UntilHalted => Disposal::backing_off(not_before, failure),
            // Counted from the first failure, so `AtMost(0)` spends its
            // one attempt and reports.
            Retries::AtMost(ceiling) if self.streak <= ceiling => {
                Disposal::backing_off(not_before, failure)
            }
            Retries::AtMost(_) => refused,
        }
    }

    /// The provider answered, which ends the streak of failures.
    pub fn on_provider_answered(&mut self) {
        self.streak = 0;
    }

    /// How long the failure just counted waits: a base of 500 ms doubled
    /// for each failure in a row before it, capped at a minute, plus up
    /// to half the base again. Both numbers are the provider's scale,
    /// the time an overloaded service takes to recover, so no reading of
    /// this machine moves them. The jitter only adds, so the base stays
    /// the floor; it is drawn from the run's seed and the streak, so
    /// runs cut by one outage spread out and each replays its own.
    fn backoff_ms(&self) -> u64 {
        const FIRST_MS: u64 = 500;
        const CEILING_MS: u64 = 60_000;
        let doublings = self.streak.saturating_sub(1).min(16);
        let base = FIRST_MS
            .checked_shl(doublings)
            .map_or(CEILING_MS, |wait| wait.min(CEILING_MS));
        let spread = NonZeroU64::MIN.saturating_add(base >> 1);
        base.saturating_add(mixed(self.jitter_seed ^ u64::from(self.streak)) % spread)
    }

    /// The `watchdog_fired` payload (E_LOOP_SUSPECTED's carrier when the
    /// reason is a stall). Proceed never fires.
    ///
    /// # Errors
    /// `E_INVALID_ARGS` for [`Disposal::Proceed`], which records
    /// nothing, and whatever `Payload::of` says about the encoding.
    pub fn fired_payload(&self, disposal: &Disposal) -> Result<Payload, AxError> {
        let action = match disposal {
            Disposal::Proceed => {
                return Err(AxError::failure(
                    AxCode::InvalidArgs,
                    "encode watchdog_fired",
                    "Proceed does not fire",
                )
                .with_recovery(
                    "report this against runtime::watchdog: `Proceed` is the verdict                      that records nothing, and the caller asked it for a payload",
                ));
            }
            Disposal::CorrectiveSteer { text } => FiredAction::Steer { text: text.clone() },
            Disposal::BackOff {
                until,
                code,
                subject,
            } => FiredAction::BackOff {
                until_ms: until.value(),
                code: code.as_str().to_owned(),
                subject: subject.clone(),
            },
            Disposal::Freeze { reason } => FiredAction::Freeze {
                reason: reason.as_str().to_owned(),
            },
        };
        Payload::of(&WatchdogFired {
            action,
            corrections: self.corrections,
            provider_failures: self.provider_failures,
        })
    }
}

/// FNV-1a over the id's sixteen bytes. A uuid v7 keeps its random bits
/// at the end, which is where runs started in one millisecond differ,
/// and FNV-1a lets the last byte move the whole word.
fn seed_of(run: RunId) -> u64 {
    run.as_bytes()
        .iter()
        .fold(0xcbf2_9ce4_8422_2325, |hash, byte| {
            (hash ^ u64::from(*byte)).wrapping_mul(0x0000_0100_0000_01b3)
        })
}

/// splitmix64's finishing mix, so neighbouring streaks and seeds land
/// far apart before the remainder takes the low bits.
fn mixed(word: u64) -> u64 {
    let word = (word ^ (word >> 30)).wrapping_mul(0xbf58_476d_1ce4_e5b9);
    let word = (word ^ (word >> 27)).wrapping_mul(0x94d0_49bb_1331_11eb);
    word ^ (word >> 31)
}

#[cfg(test)]
#[allow(
    clippy::unwrap_used,
    clippy::expect_used,
    clippy::panic,
    clippy::indexing_slicing,
    clippy::wildcard_enum_match_arm,
    clippy::let_underscore_must_use,
    clippy::let_underscore_untyped,
    clippy::arithmetic_side_effects,
    reason = "test code"
)]
mod tests {
    use super::*;
    use kernel::ActionFingerprint;
    use kernel::stall::observe;

    /// The schedule's floor for each failure in a row, before jitter.
    const BASES: [u64; 9] = [
        500, 1_000, 2_000, 4_000, 8_000, 16_000, 32_000, 60_000, 60_000,
    ];

    /// A uuid v7 run id; runs that differ only in `tail` started in the
    /// same millisecond.
    fn run(tail: u8) -> RunId {
        let mut bytes = [
            0x01, 0x93, 0x2a, 0x7c, 0x10, 0x00, 0x70, 0x00, 0x80, 0, 0, 0, 0, 0, 0, 0,
        ];
        bytes[15] = tail;
        RunId::from_bytes(bytes)
    }

    /// The waits one watchdog hands out for `count` retriable failures
    /// in a row, measured from zero.
    fn waits(dog: &mut Watchdog, count: usize) -> Vec<u64> {
        (0..count)
            .map(
                |_| match dog.on_provider_failure(&provider_error(true), TimeMs::new(0)) {
                    Disposal::BackOff { until, .. } => until.value(),
                    other => panic!("a retriable failure under its ceiling backs off: {other:?}"),
                },
            )
            .collect()
    }

    fn within_schedule(waits: &[u64]) -> bool {
        waits
            .iter()
            .zip(BASES)
            .all(|(wait, base)| (base..=base + base / 2).contains(wait))
    }

    /// Runs cut by the same outage in the same millisecond spread their
    /// next calls instead of arriving together, and one run replays its
    /// own waits byte for byte.
    #[test]
    fn each_run_waits_its_own_jittered_schedule() {
        let schedule = |tail: u8| waits(&mut Watchdog::new(Retries::UntilHalted, run(tail)), 9);
        let (first, second) = (schedule(1), schedule(2));
        assert!(
            within_schedule(&first) && within_schedule(&second),
            "{first:?} {second:?}"
        );
        assert_eq!(first, schedule(1), "the same run waits the same schedule");
        assert_ne!(first, second, "two runs do not ask again in step");
    }

    #[test]
    fn disposal_is_graded_steer_first_freeze_second() {
        let mut dog = Watchdog::new(Retries::UntilHalted, run(1));
        let same = ActionFingerprint::derive(b"exec identical");
        let sample = vec![same, same, same];
        let verdict = observe(&sample);
        let first = dog.on_stall(&verdict);
        match &first {
            Disposal::CorrectiveSteer { text } => {
                assert!(text.contains("repeated the same call 3 times"))
            }
            other => panic!("first hit must steer, got {other:?}"),
        }
        let second = dog.on_stall(&verdict);
        assert_eq!(
            second,
            Disposal::Freeze {
                reason: FreezeReason::Stall
            }
        );
        // The payload names the action; Proceed refuses to fire.
        let payload = serde_json::to_value(dog.fired_payload(&second).unwrap()).unwrap();
        assert_eq!(payload["action"], "freeze");
        assert_eq!(payload["reason"], "stall");
        assert!(dog.fired_payload(&Disposal::Proceed).is_err());
    }

    #[test]
    fn ok_verdicts_never_dispose() {
        let mut dog = Watchdog::new(Retries::UntilHalted, run(1));
        assert_eq!(dog.on_stall(&StallVerdict::Ok), Disposal::Proceed);
        assert_eq!(dog.on_stall(&StallVerdict::Ok), Disposal::Proceed);
    }

    fn provider_error(retriable: bool) -> AxError {
        let draft = AxError::failure(AxCode::Provider, "call the model", "the provider said no");
        let draft = if retriable { draft.retriable() } else { draft };
        draft.with_recovery("dispatch again, or attach a second endpoint")
    }

    #[test]
    fn a_failure_the_provider_will_repeat_stops_after_one() {
        let mut dog = Watchdog::new(Retries::UntilHalted, run(1));
        assert_eq!(
            dog.on_provider_failure(&provider_error(false), TimeMs::new(9_000)),
            Disposal::Freeze {
                reason: FreezeReason::ProviderRefused
            },
            "a request the provider rejected on its shape buys the same rejection again"
        );
    }

    #[test]
    fn an_answer_ends_the_streak_the_backoff_counts() {
        let mut dog = Watchdog::new(Retries::AtMost(2), run(1));
        let before = waits(&mut dog, 2);
        assert!(within_schedule(&before), "{before:?}");
        dog.on_provider_answered();
        assert_eq!(
            waits(&mut dog, 2),
            before,
            "an outage after a recovery starts from the first wait, under a fresh ceiling"
        );
    }

    #[test]
    fn a_retriable_failure_backs_off_and_never_freezes_by_itself() {
        let mut dog = Watchdog::new(Retries::UntilHalted, run(1));
        let waits: Vec<u64> = (0..64u64)
            .map(
                |_| match dog.on_provider_failure(&provider_error(true), TimeMs::new(1_000)) {
                    Disposal::BackOff { until, .. } => until.value().saturating_sub(1_000),
                    other => {
                        panic!("only Halt stops a run that is waiting out a provider: {other:?}")
                    }
                },
            )
            .collect();
        assert!(
            within_schedule(&waits),
            "each failure in a row doubles the wait, up to a minute: {waits:?}"
        );
        assert!(
            waits[9..]
                .iter()
                .all(|wait| (60_000..=90_000).contains(wait))
        );
        let payload = serde_json::to_value(
            dog.fired_payload(&Disposal::BackOff {
                until: TimeMs::new(1_000),
                code: AxCode::Provider,
                subject: "the provider said no".to_owned(),
            })
            .unwrap(),
        )
        .unwrap();
        assert_eq!(payload["action"], "back_off");
        assert_eq!(payload["until_ms"], 1_000);
        assert_eq!(payload["provider_failures"], 64);
        assert_eq!(
            payload["subject"], "the provider said no",
            "a line saying it backed off says what it backed off from"
        );
    }

    fn waits_of_fresh_run() -> Vec<u64> {
        waits(&mut Watchdog::new(Retries::UntilHalted, run(1)), 2)
    }

    #[test]
    fn a_provider_that_names_its_wait_is_not_asked_sooner() {
        let mut dog = Watchdog::new(Retries::UntilHalted, run(1));
        let told = |wait_ms: u64| {
            AxError::failure(AxCode::Provider, "call the model", "answered 429")
                .retriable_after(wait_ms)
                .with_recovery("wait, then ask again")
        };
        let waits: Vec<u64> = [30_000, 10]
            .map(
                |wait_ms| match dog.on_provider_failure(&told(wait_ms), TimeMs::new(1_000)) {
                    Disposal::BackOff { until, .. } => until.value().saturating_sub(1_000),
                    other => {
                        panic!("a provider that says when to come back is asked again: {other:?}")
                    }
                },
            )
            .to_vec();
        let own = waits_of_fresh_run();
        assert_eq!(
            waits,
            [30_000, own[1]],
            "the longer of the provider's word and the schedule's own wait"
        );
    }
}