orch/engine/

cooldown.rs

//! Per-agent and per-model cooldown tracking.
//!
//! Shared by `engine::runner::response` (recording failures) and
//! `engine::router::config` (avoiding cooled models during pool selection).
//!
//! # Persistence
//!
//! All cooldowns are persisted to the KV store in SQLite so they survive
//! service restarts. Call [`init_cooldown_store`] once at startup (after
//! the `TaskStore` is open) to register the store and pre-load unexpired
//! cooldowns.
//!
//! When a rate limit includes a "try again at {date}" message, the cooldown
//! is set to that specific timestamp instead of the default duration.
//! Relative usage windows (e.g. "5-hour usage limit") are also detected and
//! used to set a model-specific cooldown matching the advertised window.

use chrono::TimeZone;
use regex::Regex;
use std::collections::HashMap;
use std::collections::HashSet;
use std::sync::LazyLock;
use std::sync::MutexGuard;
use std::sync::{Arc, Mutex, OnceLock};
use tokio::sync;

use crate::engine::router::config;
#[cfg(test)]
use crate::engine::runner::agents::opencode;
use crate::store::TaskStore;

/// Default fallback cooldown for agents when no store is available: 30 minutes.
///
/// Used in contexts that need a concrete duration and cannot compute exponential
/// backoff (e.g. computing how long to wait before retrying when a cooldown
/// expiry timestamp is unknown). New code should use [`BACKOFF_BASE_SECS`] instead.
pub const AGENT_COOLDOWN_SECS: i64 = 30 * 60;

/// Base backoff for generic agent/model failures: 5 minutes.
pub const BACKOFF_BASE_SECS: i64 = 5 * 60;

/// Maximum backoff for generic agent/model failures: 4 hours.
pub const BACKOFF_MAX_SECS: i64 = 4 * 60 * 60;

/// Base backoff for persistent model failures (silent exit/provider errors): 4 hours.
pub const PERSISTENT_MODEL_BACKOFF_BASE_SECS: i64 = 4 * 60 * 60;

/// Maximum backoff for persistent model failures: 7 days.
pub const PERSISTENT_MODEL_BACKOFF_MAX_SECS: i64 = 7 * 24 * 60 * 60;

/// Base backoff for a one-off model failure that does not mean the model is gone: 5 minutes.
pub const TRANSIENT_MODEL_BACKOFF_BASE_SECS: i64 = 5 * 60;

/// Maximum backoff for a one-off model failure: 15 minutes.
pub const TRANSIENT_MODEL_BACKOFF_MAX_SECS: i64 = 15 * 60;

/// Base backoff for credit exhaustion (out_of_credits): 1 hour.
pub const CREDIT_BACKOFF_BASE_SECS: i64 = 60 * 60;

/// Base backoff for org-level disabling: 2 hours.
pub const ORG_BACKOFF_BASE_SECS: i64 = 2 * 60 * 60;

/// Standard cap for credit exhaustion and org-level disabling: 8 hours.
/// Repeated failures at the cap move to the extended tiers (48h, then 96h)
pub const CREDIT_BACKOFF_MAX_SECS: i64 = 8 * 60 * 60;

/// Base cooldown for billing cycle exhaustion: 24 hours.
///
/// Unlike per-request rate limits, billing cycle exhaustion is a calendar event
/// that won't self-resolve for days or weeks. The first failure starts at 24h;
/// subsequent failures escalate via `compute_backoff()` up to
/// `BILLING_CYCLE_MAX_SECS` (7 days). This prevents daily retry-and-fail cycles
/// when a monthly billing quota is exhausted.
pub const BILLING_CYCLE_COOLDOWN_SECS: i64 = 24 * 60 * 60;

/// Maximum cooldown for billing cycle exhaustion: 7 days.
///
/// Monthly billing cycles typically reset on a fixed date. 7 days is a
/// reasonable cap — the cooldown will expire before the next monthly reset
/// in most cases, allowing a single probe to re-check availability.
pub const BILLING_CYCLE_MAX_SECS: i64 = 7 * 24 * 60 * 60;

/// Credit exhaustion reason detected from error message.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum CreditExhaustionReason {
    /// Per-model credit exhaustion (out_of_credits).
    OutOfCredits,
    /// Organization-level disabling (org_level_disabled).
    OrgLevelDisabled,
    /// Monthly billing cycle exhaustion — escalating cooldown (24h base, 7d cap).
    BillingCycleExhausted,
}

impl CreditExhaustionReason {
    /// Stable string for `task_runs.outcome`, distinct from the cooldown reason
    /// string used by `record_credit_exhaustion`. Lets callers persist the actual
    /// failure mode instead of collapsing it into a generic "rate_limit"/"failed".
    pub fn outcome_str(self) -> &'static str {
        match self {
            CreditExhaustionReason::OutOfCredits => "credit_exhausted",
            CreditExhaustionReason::OrgLevelDisabled => "org_level_disabled",
            CreditExhaustionReason::BillingCycleExhausted => "billing_cycle_exhausted",
        }
    }
}

/// Short agent cooldown applied on silence detection (120 seconds).
///
/// Forces the router to pick a different agent immediately on re-route,
/// without long-term blocking. The model cooldown (exponential, starting at
/// 5 min) keeps the dead model out while this short cooldown just breaks
/// the same-agent loop.
pub const SILENCE_AGENT_COOLDOWN_SECS: u64 = 120;

/// Silence detections before applying an extended cooldown (rolling window).
pub const SILENCE_COUNT_THRESHOLD: usize = 5;

/// Rolling window for silence detections (24 hours).
pub const SILENCE_COUNT_WINDOW_SECS: i64 = 24 * 60 * 60;

/// Extended cooldown applied after repeated silence detections (4 hours).
pub const SILENCE_EXTENDED_COOLDOWN_SECS: u64 = 4 * 60 * 60;

/// KV key prefix for persisted cooldowns (both agent and model).
const KV_PREFIX: &str = "cooldown:";

/// KV key prefix for rolling silence counters.
const SILENCE_COUNT_PREFIX: &str = "silence_count:";

/// KV key prefix for per-agent and per-agent:model failure counts (drives exponential backoff).
const FAILURE_COUNT_PREFIX: &str = "failure_count:";

/// KV key prefix for credit-exhaustion failure counts (separate from generic failures to prevent
/// cross-contamination of backoff escalation — credit exhaustion uses different base/max durations).
const CREDIT_FAILURE_COUNT_PREFIX: &str = "credit_failure_count:";

/// KV key prefix for persistent model failure counts (separate from [`FAILURE_COUNT_PREFIX`] so
/// `record_agent_success()` resetting the generic counter on an intervening success doesn't erase
/// escalation progress for structurally unreliable models — see `record_persistent_model_failure`).
const PERSISTENT_FAILURE_COUNT_PREFIX: &str = "persistent_failure_count:";

/// KV key prefix for the unix timestamp after which `rate_limits` rows count toward degraded
const RATE_LIMITS_CLEARED_AT_PREFIX: &str = "rate_limits_cleared_at:";

/// Agent name from a cooldown key (`claude:sonnet` -> `claude`)
fn agent_from_key(key: &str) -> &str {
    key.split_once(':').map_or(key, |(agent, _)| agent)
}

struct CooldownEntry {
    /// Unix timestamp when the cooldown expires.
    cooldown_until: i64,
    reason: String,
}

#[derive(Debug)]
pub struct SilenceCountResult {
    pub count: usize,
    pub extended_cooldown_applied: bool,
}

/// Global in-memory cooldown map, protected by a Mutex.
///
/// Uses `std::sync::Mutex` because this map is also read from sync helper
/// functions (`is_in_cooldown`, `is_model_in_cooldown`, etc.).  All lock
/// holders perform only in-memory HashMap operations — **no `.await` is ever
/// called while a guard is held**.  This invariant must be preserved: if a
/// future caller needs to await while holding this lock, switch to
/// `tokio::sync::Mutex` and make the helpers async.
fn cooldowns() -> &'static Mutex<HashMap<String, CooldownEntry>> {
    static COOLDOWNS: OnceLock<Mutex<HashMap<String, CooldownEntry>>> = OnceLock::new();
    COOLDOWNS.get_or_init(|| Mutex::new(HashMap::new()))
}

/// Global store reference used for background KV writes.
///
/// Uses `tokio::sync::Mutex` because this value is only ever accessed from
/// async functions. A `tokio::sync::Mutex` yields the Tokio task instead of
/// blocking the worker thread if the lock is contended.
fn cooldown_store() -> &'static sync::Mutex<Option<Arc<TaskStore>>> {
    static STORE: OnceLock<sync::Mutex<Option<Arc<TaskStore>>>> = OnceLock::new();
    STORE.get_or_init(|| sync::Mutex::new(None))
}

/// Initialise persistent cooldowns.
///
/// Must be called once at engine startup, after the `TaskStore` is opened.
/// Loads unexpired cooldowns from KV and registers the store for future writes.
/* Parses the `{"until": N, "reason": "..."}` cooldown KV format, falling back
to the legacy bare-integer format (reason unknown) so existing rows keep
working across the upgrade without a migration. */
fn parse_cooldown_value(value: &str) -> Option<(i64, String)> {
    if let Ok(serde_json::Value::Object(map)) = serde_json::from_str(value) {
        let until = map.get("until")?.as_i64()?;
        let reason = map
            .get("reason")
            .and_then(|r| r.as_str())
            .unwrap_or("persisted")
            .to_string();
        return Some((until, reason));
    }
    value
        .parse::<i64>()
        .ok()
        .map(|until| (until, "persisted".to_string()))
}

pub async fn init_cooldown_store(store: Arc<TaskStore>) {
    match store.kv_list_prefix(KV_PREFIX).await {
        Ok(rows) => {
            let now = chrono::Utc::now().timestamp();
            let mut map = cooldown_lock();
            let mut loaded = 0;
            for (key, value) in rows {
                let cooldown_key = key.trim_start_matches(KV_PREFIX);
                if let Some((cooldown_until, reason)) = parse_cooldown_value(&value)
                    && now < cooldown_until
                {
                    map.insert(
                        cooldown_key.to_string(),
                        CooldownEntry {
                            cooldown_until,
                            reason,
                        },
                    );
                    loaded += 1;
                }
                // Expired entries are simply not loaded — they'll be overwritten
                // on the next failure or cleaned up by future writes.
            }
            if loaded > 0 {
                tracing::info!(loaded, "loaded persisted cooldowns from KV store");
            }
        }
        Err(e) => {
            tracing::warn!(err = %e, "failed to load cooldowns from KV store");
        }
    }

    // Load GitHub 5xx circuit breaker state if it was persisted before restart.
    if let Ok(Some(raw)) = store.kv_get("cooldown:github:5xx").await
        && let Ok(cooldown_until) = raw.parse::<i64>()
    {
        let now = chrono::Utc::now().timestamp();
        if now < cooldown_until {
            let mut open = github_5xx_circuit_open_lock();
            *open = Some(cooldown_until);
            let remaining = cooldown_until - now;
            tracing::warn!(
                remaining_secs = remaining,
                "GitHub 5xx circuit breaker restored from KV (still open)"
            );
        }
    }

    *cooldown_store().lock().await = Some(store);
}

/// Reconcile the in-memory cooldown map with the persisted KV state.
///
/// KV is the source of truth. This is called once per tick by
/// `router.refresh_health` so the running service notices state mutated by a
/// separate process (e.g. `orch cooldown clear`). Without this sync, the CLI
/// clears KV but the engine's in-memory map keeps the stale entries until they
/// expire on their own, producing the "all agents cooled" + "No active
/// cooldowns" inconsistency.
///
/// Reconciliation rules (KV authoritative for every key it has):
/// - KV row missing for an in-memory key: leave memory alone — likely a
///   `set_cooldown_async` race where memory was written but the KV row isn't
///   yet visible. The next tick will catch up.
/// - KV row present with `cooldown_until <= now` (including the "0" tombstone
///   written by `clear_cooldown`): drop from memory.
/// - KV row present with `cooldown_until > now`: insert/update memory to match.
pub async fn sync_from_kv(store: &Arc<TaskStore>) {
    let rows = match store.kv_list_prefix(KV_PREFIX).await {
        Ok(rows) => rows,
        Err(e) => {
            tracing::warn!(err = %e, "cooldown sync_from_kv: KV list failed");
            return;
        }
    };
    let now = chrono::Utc::now().timestamp();

    // Some((ts, reason)) for active cooldowns, None for cleared/expired rows.
    // Missing key means no KV row at all, left alone below.
    let mut explicit_kv: HashMap<String, Option<(i64, String)>> =
        HashMap::with_capacity(rows.len());
    for (kv_key, value) in &rows {
        let cooldown_key = kv_key.trim_start_matches(KV_PREFIX);
        if let Some((ts, reason)) = parse_cooldown_value(value) {
            let normalized = if ts > now { Some((ts, reason)) } else { None };
            explicit_kv.insert(cooldown_key.to_string(), normalized);
        }
    }

    // The 5xx circuit breaker lives in its own statics, so mirror its KV row there
    match explicit_kv.get("github:5xx") {
        Some(None) => {
            let mut open = github_5xx_circuit_open_lock();
            if open.take().is_some() {
                tracing::info!("GitHub 5xx circuit breaker closed by KV clear");
                github_5xx_timestamps_lock().clear();
            }
        }
        Some(Some((ts, _))) => {
            *github_5xx_circuit_open_lock() = Some(*ts);
        }
        None => {}
    }

    let mut map = cooldown_lock();
    let mut dropped = 0usize;
    let mut updated = 0usize;

    // Drop in-memory entries that KV explicitly marks cleared/expired.
    map.retain(|key, _| {
        if matches!(explicit_kv.get(key), Some(None)) {
            dropped += 1;
            false
        } else {
            true
        }
    });

    // Adopt KV's authoritative timestamp for active cooldowns.
    for (key, normalized) in explicit_kv {
        if let Some((ts, reason)) = normalized {
            if let Some(entry) = map.get_mut(&key) {
                if entry.cooldown_until != ts {
                    entry.cooldown_until = ts;
                    updated += 1;
                }
            } else {
                map.insert(
                    key,
                    CooldownEntry {
                        cooldown_until: ts,
                        reason,
                    },
                );
                updated += 1;
            }
        }
    }

    if dropped > 0 || updated > 0 {
        tracing::info!(
            dropped,
            updated,
            "cooldown sync_from_kv: reconciled in-memory map with KV"
        );
    }
}

/// Failure count threshold (times at max backoff) for first extended tier (24h).
///
/// After a model has been at the max cap this many times, apply a 6x multiplier.
const EXTENDED_TIER_1_THRESHOLD: u64 = 2;

/// Multiplier for first extended tier (6x = 24h when max=4h).
const EXTENDED_TIER_1_MULTIPLIER: i64 = 6;

/// Failure count threshold (times at max backoff) for second extended tier (48h).
///
/// After a model has been at the max cap this many times, apply a 12x multiplier.
const EXTENDED_TIER_2_THRESHOLD: u64 = 5;

/// Multiplier for second extended tier (12x = 48h when max=4h).
const EXTENDED_TIER_2_MULTIPLIER: i64 = 12;

/// Compute exponential backoff duration using base-3 growth.
///
/// Returns `min(base * 3^(count-1), max)` for count >= 1.
/// For count == 0 or 1 (first failure), returns `base`.
///
/// | count | base=300s | result |
/// |-------|-----------|--------|
/// | 0     | 300       | 300    |
/// | 1     | 300       | 300    |
/// | 2     | 300       | 900    |
/// | 3     | 300       | 2700   |
/// | 4     | 300       | 8100   |
/// | 5-6   | 14400     | 14400 (max, not yet extended) |
/// | 7-9   | 86400     | 86400 (24h extended tier) |
/// | 10+   | 172800    | 172800 (48h extended tier) |
///
/// The extended tier applies when the model has hit the max cap multiple times,
/// indicating a persistently broken model that won't self-resolve quickly.
pub fn compute_backoff(count: u64, base: i64, max: i64, allow_extended: bool) -> i64 {
    if count <= 1 {
        return base;
    }
    let exp = u32::try_from(count.saturating_sub(1)).unwrap_or(u32::MAX);
    let factor = 3_i64.saturating_pow(exp);
    let standard = base.saturating_mul(factor).min(max);

    // Extended tier: after N failures at max backoff, apply longer window
    // to reduce noise from persistently broken models.
    //
    // First find the count at which exponential growth first hits max:
    // solve: base * 3^(first_max-1) >= max
    // => 3^(first_max-1) >= max / base
    // => first_max-1 >= log_3(max / base)
    // => first_max >= 1 + log_3(max / base)
    let first_max = if base <= 0 {
        0
    } else {
        // Use integer approximation: count where base * 3^(count-1) would first exceed max
        let mut accumulated = base;
        let mut first_max_count = 1u64;
        while accumulated < max && first_max_count < count {
            accumulated = accumulated.saturating_mul(3);
            first_max_count += 1;
        }
        first_max_count
    };

    // Times we've been at the max cap (excluding the first time we hit it)
    let times_at_max = count.saturating_sub(first_max);

    if !allow_extended {
        return standard;
    }

    if standard == max && times_at_max >= EXTENDED_TIER_2_THRESHOLD {
        max.saturating_mul(EXTENDED_TIER_2_MULTIPLIER)
    } else if standard == max && times_at_max >= EXTENDED_TIER_1_THRESHOLD {
        max.saturating_mul(EXTENDED_TIER_1_MULTIPLIER)
    } else {
        standard
    }
}

/// Read the failure count for a key from KV, increment it, write it back, and return the new count.
///
/// Always increments — the backoff formula's `max` parameter prevents the
/// cooldown duration from growing beyond the cap, so unbounded counts are safe.
///
/// Returns 1 when no store is configured (unit-test contexts without a store).
/// Returns `u64::MAX` on store errors (e.g. lock contention) so backoff applies the cap duration
/// instead of resetting to the base, preventing rapid re-dispatch during store outages.
async fn read_and_increment_failure_count(store_opt: Option<&Arc<TaskStore>>, key: &str) -> u64 {
    read_and_increment_failure_count_with_prefix(store_opt, FAILURE_COUNT_PREFIX, key).await
}

/// Like [`read_and_increment_failure_count`] but with an explicit KV prefix.
///
/// Credit exhaustion uses [`CREDIT_FAILURE_COUNT_PREFIX`] so its backoff escalation
/// is independent of generic agent failures (which use [`FAILURE_COUNT_PREFIX`]).
async fn read_and_increment_failure_count_with_prefix(
    store_opt: Option<&Arc<TaskStore>>,
    prefix: &str,
    key: &str,
) -> u64 {
    let kv_key = format!("{prefix}{key}");
    if let Some(store) = store_opt {
        match store.kv_increment(&kv_key).await {
            Ok(n) => n,
            Err(e) => {
                tracing::warn!(key = %kv_key, err = %e, "failed to increment failure count — treating as high count for safe backoff");
                u64::MAX
            }
        }
    } else {
        1
    }
}

/// Reset failure counts for an agent and agent:model combo after a successful run.
///
/// Call this from the runner's success path so that the next failure starts
/// backoff from the base duration again, not from wherever it left off.
/// Resets both generic and credit-exhaustion failure counters.
pub async fn record_agent_success(agent_name: &str, model: &str) {
    let store_opt = cooldown_store().lock().await.clone();
    if let Some(store) = store_opt {
        let agent_key = format!("{FAILURE_COUNT_PREFIX}{agent_name}");
        let credit_key = format!("{CREDIT_FAILURE_COUNT_PREFIX}{agent_name}");
        if let Err(e) = store.kv_set(&agent_key, "0").await {
            tracing::warn!(key = agent_key, err = %e, "failed to reset failure count");
        }
        if !model.is_empty() {
            let model_key = format!("{FAILURE_COUNT_PREFIX}{agent_name}:{model}");
            if let Err(e) = store.kv_set(&model_key, "0").await {
                tracing::warn!(key = model_key, err = %e, "failed to reset failure count");
            }
        }
        if let Err(e) = store.kv_set(&credit_key, "0").await {
            tracing::warn!(key = credit_key, err = %e, "failed to reset failure count");
        }
    }
}

/// Record that an agent has failed and should be temporarily avoided.
///
/// Applies exponential backoff based on the agent's failure count in KV.
/// If `error_message` contains "try again at {date}", that vendor date is used
/// instead of the computed backoff (vendor dates are always authoritative).
pub async fn record_agent_failure_with_message(agent_name: &str, error_message: &str) {
    // Vendor-specified retry date takes priority over backoff — but only
    // if it is in the present or future. If the parsed timestamp is in the
    // past (stale vendor message), fall through to the exponential backoff
    // path instead of returning early and bypassing backoff entirely.
    if let Some(cooldown_until) = parse_retry_at(error_message) {
        let now = chrono::Utc::now().timestamp();
        if cooldown_until > now {
            set_cooldown_async(agent_name, cooldown_until, "agent_error").await;
            return;
        }
        tracing::debug!(
            agent = agent_name,
            cooldown_until,
            "vendor retry_at is in the past — using backoff instead"
        );
    }

    let store_opt = cooldown_store().lock().await.clone();
    let count = read_and_increment_failure_count(store_opt.as_ref(), agent_name).await;
    let base = config::get_agent_backoff_base(agent_name);
    let cooldown_until =
        chrono::Utc::now().timestamp() + compute_backoff(count, base, BACKOFF_MAX_SECS, true);
    set_cooldown_async(agent_name, cooldown_until, "agent_error").await;
}

/// Detect credit exhaustion from an error message and return the stable
/// `task_runs.outcome` string for it (see `CreditExhaustionReason::outcome_str`).
///
/// Callers that classify `AgentError::RateLimit`/`AgentError::Auth` into an
/// outcome string should consult this first — generic rate-limit patterns like
/// "usage limit" also match billing-cycle exhaustion messages (e.g. Kimi's
/// "You've reached your usage limit for this billing cycle"), which would
/// otherwise be stored as a plain "rate_limit" outcome.
pub fn detect_credit_exhaustion_outcome(error_message: &str) -> Option<&'static str> {
    detect_credit_exhaustion(error_message).map(CreditExhaustionReason::outcome_str)
}

/// Detect credit exhaustion reasons from error message.
///
/// Returns `Some(CreditExhaustionReason)` if the error indicates
/// out_of_credits or org_level_disabled, which require longer agent-wide cooldowns.
pub fn detect_credit_exhaustion(error_message: &str) -> Option<CreditExhaustionReason> {
    let lower = error_message.to_lowercase();

    let overage_patterns = [
        "overagedisabledreason",
        "overage_disabled_reason",
        "overage disabled reason",
    ];

    if overage_patterns.iter().any(|p| lower.contains(p)) {
        if lower.contains("out_of_credits") || lower.contains("outofcredits") {
            return Some(CreditExhaustionReason::OutOfCredits);
        }
        if lower.contains("org_level_disabled") || lower.contains("orgdisabled") {
            return Some(CreditExhaustionReason::OrgLevelDisabled);
        }
    }

    let billing_patterns = [
        "credit balance too low",
        "credit balance insufficient",
        "insufficient credit balance",
        "billing quota exceeded",
        "billing limit exceeded",
        "organization has been disabled",
        "org disabled",
    ];

    if billing_patterns.iter().any(|p| lower.contains(p)) {
        return Some(CreditExhaustionReason::OrgLevelDisabled);
    }

    // Billing cycle exhaustion: monthly quota reset — check before generic quota
    // patterns because "monthly quota exceeded" would otherwise match "quota exceeded".
    let billing_cycle_patterns = [
        "billing cycle",
        "quota refreshed next cycle",
        "monthly quota",
        "quota will be refreshed",
        "next billing cycle",
        "refreshed in the next cycle",
        "weekly/monthly limit exhausted",
        "limit will reset at",
        "upgrade to pro", // codex/chatgpt plan usage cap: requires billing upgrade
        "upgrade your token plan", // minimax: Token Plan usage limit reached
        "token plan usage limit", // minimax: Token Plan usage limit reached
        "purchase credits for more", // minimax: ...purchase Credits for more usage
    ];

    if billing_cycle_patterns.iter().any(|p| lower.contains(p)) {
        return Some(CreditExhaustionReason::BillingCycleExhausted);
    }

    if lower.contains("out of credits")
        || lower.contains("outofcredits")
        || lower.contains("insufficient funds")
        || lower.contains("no credits remaining")
        || lower.contains("insufficient_quota")
        || lower.contains("quota exceeded")
        || lower.contains("insufficient balance")
        || lower.contains("please recharge")
        || lower.contains("no resource package")
        || lower.contains("balance exhausted")
        || lower.contains("402 payment required")
    {
        return Some(CreditExhaustionReason::OutOfCredits);
    }

    None
}

/// Record an agent-level cooldown for credit exhaustion.
///
/// Applies exponential backoff based on the agent's failure count.
/// All variants use escalating backoff:
/// - `OutOfCredits`: 1h, 3h, then 8h (credits can be refilled any time), 48h from the 5th failure, 96h from the 8th
/// - `OrgLevelDisabled`: 2h, 6h, then 8h, 48h from the 5th failure, 96h from the 8th
/// - `BillingCycleExhausted`: starts at 24h, caps at 7 days (monthly event)
///
/// Billing cycle exhaustion was previously a flat 24h, but this caused daily
/// retry-and-fail cycles when a monthly quota was exhausted (the 24h cooldown
/// would expire, kimi would be retried, fail immediately, and get another 24h).
/// Now each recurrence escalates: 24h → 72h → 7d (capped).
pub async fn record_credit_exhaustion(agent_name: &str, reason: CreditExhaustionReason) {
    let reason_str = match reason {
        CreditExhaustionReason::OutOfCredits => "credit_exhaustion_out_of_credits",
        CreditExhaustionReason::OrgLevelDisabled => "credit_exhaustion_org_level_disabled",
        CreditExhaustionReason::BillingCycleExhausted => "billing_cycle_exhausted",
    };

    // A failure from a run that started before the cooldown is the same event
    if is_in_cooldown_with_reason(agent_name, reason_str) {
        tracing::debug!(
            agent = agent_name,
            reason = reason_str,
            "credit exhaustion already cooled, not escalating"
        );
        return;
    }

    let store_opt = cooldown_store().lock().await.clone();
    // Use a distinct failure counter (CREDIT_FAILURE_COUNT_PREFIX) so credit-
    // exhaustion backoff escalation is independent of generic agent failures.
    // Previously both paths shared FAILURE_COUNT_PREFIX, causing generic failures
    // to inflate credit-exhaustion backoff (and vice versa).
    let count = read_and_increment_failure_count_with_prefix(
        store_opt.as_ref(),
        CREDIT_FAILURE_COUNT_PREFIX,
        agent_name,
    )
    .await;

    let (base, max) = match reason {
        CreditExhaustionReason::OutOfCredits => (CREDIT_BACKOFF_BASE_SECS, CREDIT_BACKOFF_MAX_SECS),
        CreditExhaustionReason::OrgLevelDisabled => {
            (ORG_BACKOFF_BASE_SECS, CREDIT_BACKOFF_MAX_SECS)
        }
        CreditExhaustionReason::BillingCycleExhausted => {
            (BILLING_CYCLE_COOLDOWN_SECS, BILLING_CYCLE_MAX_SECS)
        }
    };
    let cooldown_secs = compute_backoff(
        count,
        base,
        max,
        !matches!(reason, CreditExhaustionReason::BillingCycleExhausted),
    );
    let cooldown_until = chrono::Utc::now().timestamp() + cooldown_secs;
    set_cooldown_async(agent_name, cooldown_until, reason_str).await;
    tracing::warn!(
        agent = agent_name,
        reason = reason_str,
        attempt = count,
        cooldown_secs,
        "credit exhaustion detected: applying exponential agent-wide cooldown"
    );
}

// Returns false when the message is no credit exhaustion. Model-scoped billing cycle cools the model only
pub async fn record_credit_exhaustion_from_message(
    agent_name: &str,
    model: Option<&str>,
    message: &str,
) -> bool {
    let Some(reason) = detect_credit_exhaustion(message) else {
        return false;
    };
    match (reason, model) {
        (CreditExhaustionReason::BillingCycleExhausted, Some(model)) => {
            record_persistent_model_failure(agent_name, model).await;
        }
        _ => record_credit_exhaustion(agent_name, reason).await,
    }
    true
}

/// Record that a specific agent+model combo has failed.
///
/// Applies exponential backoff based on the model's failure count in KV.
pub async fn record_model_failure(agent_name: &str, model: &str) {
    let key = format!("{agent_name}:{model}");
    let store_opt = cooldown_store().lock().await.clone();
    let count = read_and_increment_failure_count(store_opt.as_ref(), &key).await;
    let base = config::get_agent_backoff_base(agent_name);
    let cooldown_until =
        chrono::Utc::now().timestamp() + compute_backoff(count, base, BACKOFF_MAX_SECS, true);
    set_cooldown_async(&key, cooldown_until, "model_error").await;
}

/// Record a rate-limit event for an agent+model.
///
/// If the error message advertises a relative usage window (e.g.
/// "5-hour usage limit"), the cooldown is applied to the specific model for
/// that window length. The quota is model-scoped, so the whole agent is not
/// cooled. If no window is detected, fall back to the generic agent+model
/// exponential backoff path.
pub async fn record_rate_limit(agent_name: &str, model: Option<&str>, error_message: &str) {
    if let Some(window_secs) = parse_relative_usage_window(error_message)
        && let Some(m) = model
    {
        let cooldown_until = chrono::Utc::now()
            .timestamp()
            .saturating_add(i64::try_from(window_secs).unwrap_or(i64::MAX));
        set_model_cooldown(agent_name, m, window_secs, "rate_limit_usage_window").await;
        tracing::info!(
            agent = agent_name,
            model = m,
            window_secs,
            cooldown_until,
            "relative usage window detected: applying model-specific cooldown"
        );
        return;
    }

    record_agent_failure_with_message(agent_name, error_message).await;
    if let Some(m) = model {
        record_model_failure(agent_name, m).await;
    }
}

/// Record a persistent model-specific failure and apply longer exponential backoff.
///
/// Used for failure modes that are commonly broken for long periods (for example
/// silent exit-0 or opaque provider errors). Starts at 4h and escalates with
/// the generic backoff formula up to 7 days.
///
/// Uses [`PERSISTENT_FAILURE_COUNT_PREFIX`], a counter independent of the generic
/// `failure_count:` key that `record_agent_success()` resets on every successful
/// run. Otherwise a single intervening success between persistent failures would
/// reset escalation back to the 4h base, letting structurally unreliable models
/// (intermittent parse errors, truncated runs) stay in rotation indefinitely.
pub async fn record_persistent_model_failure(agent_name: &str, model: &str) {
    let key = format!("{agent_name}:{model}");
    let store_opt = cooldown_store().lock().await.clone();
    let count = read_and_increment_failure_count_with_prefix(
        store_opt.as_ref(),
        PERSISTENT_FAILURE_COUNT_PREFIX,
        &key,
    )
    .await;
    let cooldown_secs = compute_backoff(
        count,
        PERSISTENT_MODEL_BACKOFF_BASE_SECS,
        PERSISTENT_MODEL_BACKOFF_MAX_SECS,
        false,
    );
    let cooldown_until = chrono::Utc::now().timestamp() + cooldown_secs;
    set_cooldown_async(&key, cooldown_until, "persistent_model_error").await;
}

/// Record a `ModelUnavailable` failure for an agent+model.
///
/// "Not found", "not supported" and region blocks never lift on their own, so they get the
/// persistent backoff (4h base, 7d max). Any other message gets [`record_model_failure`]
pub async fn record_model_unavailable(agent_name: &str, model: &str, message: &str) {
    if is_model_permanently_gone(message) {
        record_persistent_model_failure(agent_name, model).await;
    } else {
        record_model_failure(agent_name, model).await;
    }
}

fn is_model_permanently_gone(message: &str) -> bool {
    let lower = message.to_lowercase();
    lower.contains("not found")
        || lower.contains("not supported")
        || lower.contains("not available in your country")
        || lower.contains("not available in your region")
}

/// Record a run that failed without showing the model is gone: empty or cut off output, an
/// unparseable review, an opaque provider error.
///
/// A stopped session or one odd answer causes these, so the backoff is 5 then 15 minutes. Billing
/// exhaustion and "model not found" keep [`record_persistent_model_failure`]
pub async fn record_transient_model_failure(agent_name: &str, model: &str) {
    let key = format!("{agent_name}:{model}");
    let store_opt = cooldown_store().lock().await.clone();
    let count = read_and_increment_failure_count(store_opt.as_ref(), &key).await;
    let cooldown_secs = compute_backoff(
        count,
        TRANSIENT_MODEL_BACKOFF_BASE_SECS,
        TRANSIENT_MODEL_BACKOFF_MAX_SECS,
        false,
    );
    let cooldown_until = chrono::Utc::now().timestamp() + cooldown_secs;
    set_cooldown_async(&key, cooldown_until, "transient_model_error").await;
}

/// Set a model cooldown with a custom duration (in seconds) and reason.
///
/// Generic model-cooldown setter shared by silence detection and the
/// rate-limit relative-usage-window path. The reason is persisted with
/// the cooldown (KV + in-memory entry) so operators can tell the two
/// apart in `orch cooldown list`.
pub async fn set_model_cooldown(agent_name: &str, model: &str, duration_secs: u64, reason: &str) {
    let key = format!("{agent_name}:{model}");
    let cooldown_until = chrono::Utc::now()
        .timestamp()
        .saturating_add(i64::try_from(duration_secs).unwrap_or(i64::MAX));
    set_cooldown_async(&key, cooldown_until, reason).await;
}

/// Set a model cooldown that ends at a vendor-provided Unix timestamp.
pub async fn set_model_cooldown_until(agent_name: &str, model: &str, until: i64) {
    let key = format!("{agent_name}:{model}");
    set_cooldown_async(&key, until, "vendor_reset_at").await;
}

/// Set a short agent-level cooldown (in seconds).
///
/// Used by silence detection to temporarily block the whole agent so the
/// router picks a different one on re-route. Unlike `record_agent_failure`
/// (exponential backoff), this uses a short duration (typically 120s) —
/// just enough to force one re-route cycle to a different agent.
pub async fn set_agent_cooldown(agent_name: &str, duration_secs: u64) {
    let cooldown_until = chrono::Utc::now()
        .timestamp()
        .saturating_add(i64::try_from(duration_secs).unwrap_or(i64::MAX));
    set_cooldown_async(agent_name, cooldown_until, "silence_agent_cooldown").await;
}

/// Record a silence detection for an agent+model and apply extended cooldowns
/// when repeated silences exceed the threshold within the rolling window.
pub async fn record_silence_detection(agent_name: &str, model: &str) -> Option<SilenceCountResult> {
    let store_opt = cooldown_store().lock().await.clone();
    let Some(store) = store_opt else {
        tracing::debug!(
            agent = agent_name,
            model,
            "skipping silence count record (no KV store)"
        );
        return None;
    };

    let key = format!("{SILENCE_COUNT_PREFIX}{agent_name}:{model}");
    let now = chrono::Utc::now().timestamp();
    let window_start = now - SILENCE_COUNT_WINDOW_SECS;

    let mut timestamps = match store.kv_get(&key).await {
        Ok(Some(raw)) => match serde_json::from_str::<Vec<i64>>(&raw) {
            Ok(ts) => ts,
            Err(e) => {
                tracing::warn!(
                    kv_key = %key,
                    err = %e,
                    "failed to parse silence timestamps from KV — resetting window"
                );
                Vec::new()
            }
        },
        Ok(None) => Vec::new(),
        Err(err) => {
            tracing::warn!(
                kv_key = %key,
                err = %err,
                "failed to load silence count from KV"
            );
            Vec::new()
        }
    };

    timestamps.retain(|ts| *ts >= window_start);
    timestamps.push(now);

    let count = timestamps.len();
    let mut extended_cooldown_applied = false;
    if count >= SILENCE_COUNT_THRESHOLD {
        set_model_cooldown(
            agent_name,
            model,
            SILENCE_EXTENDED_COOLDOWN_SECS,
            "silence_detected",
        )
        .await;
        extended_cooldown_applied = true;
        timestamps.clear();
    }

    let value = serde_json::to_string(&timestamps).unwrap_or_else(|_| "[]".to_string());
    if let Err(err) = store.kv_set(&key, &value).await {
        tracing::warn!(kv_key = key, err = %err, "failed to persist silence count");
    }

    Some(SilenceCountResult {
        count,
        extended_cooldown_applied,
    })
}

/// Check if a specific agent+model combo is in cooldown.
pub fn is_model_in_cooldown(agent_name: &str, model: &str) -> bool {
    let key = format!("{agent_name}:{model}");
    is_in_cooldown(&key)
}

/// Check if an agent is currently in cooldown period.
pub fn is_agent_in_cooldown(agent_name: &str) -> bool {
    is_in_cooldown(agent_name)
}

/// Return the cooldown expiry timestamp for a key (agent or agent:model), if active.
pub fn cooldown_until(key: &str) -> Option<i64> {
    let map = cooldown_lock();
    map.get(key).and_then(|entry| {
        let now = chrono::Utc::now().timestamp();
        if now < entry.cooldown_until {
            Some(entry.cooldown_until)
        } else {
            None
        }
    })
}

/// Return the reason string recorded when a cooldown was set, if the cooldown is still active.
pub fn cooldown_reason(key: &str) -> Option<String> {
    let map = cooldown_lock();
    map.get(key).and_then(|entry| {
        let now = chrono::Utc::now().timestamp();
        if now < entry.cooldown_until {
            Some(entry.reason.clone())
        } else {
            None
        }
    })
}

/// Return all currently active cooldowns as `(key, cooldown_until_unix, reason)`.
///
/// Reads from the in-memory map only — always fast, no async needed.
/// Expired entries are filtered out. Used by `orch cooldown list`.
pub fn list_all_cooldowns() -> Vec<(String, i64, String)> {
    let now = chrono::Utc::now().timestamp();
    let map = cooldown_lock();
    let mut result: Vec<(String, i64, String)> = map
        .iter()
        .filter(|(_, entry)| now < entry.cooldown_until)
        .map(|(key, entry)| (key.clone(), entry.cooldown_until, entry.reason.clone()))
        .collect();
    result.sort_by(|a, b| a.0.cmp(&b.0));
    result
}

/// Write `rate_limits_cleared_at:{agent}` = now so the next health check
/// ignores older `rate_limits` rows. Rows stay in the table for metrics
pub async fn mark_rate_limits_cleared(store: &Arc<TaskStore>, agents: &[String]) {
    let now = chrono::Utc::now().timestamp().to_string();
    let mut seen = HashSet::new();
    for key in agents {
        let agent = agent_from_key(key);
        if agent.is_empty() || !seen.insert(agent.to_string()) {
            continue;
        }
        let kv_key = format!("{RATE_LIMITS_CLEARED_AT_PREFIX}{agent}");
        if let Err(e) = store.kv_set(&kv_key, &now).await {
            tracing::warn!(agent, err = %e, "failed to mark rate limits cleared");
        }
    }
}

/// Clear a specific cooldown by key, or all cooldowns when `key == "*"`.
///
/// Removes from the in-memory map, writes a past timestamp to KV, and resets
/// the corresponding `failure_count:` entries so the next failure starts
/// backoff from the base duration. Also writes `rate_limits_cleared_at:{agent}`
/// so a degraded agent held only by rate rows recovers on the next health check
pub async fn clear_cooldown(key: &str, store: &Arc<TaskStore>) {
    let mut agents_to_mark: Vec<String> = Vec::new();
    if key == "*" {
        let keys: Vec<String> = {
            let mut map = cooldown_lock();
            let keys: Vec<String> = map.keys().cloned().collect();
            map.clear();
            keys
        };
        for k in &keys {
            let kv_key = format!("{KV_PREFIX}{k}");
            if let Err(e) = store.kv_set(&kv_key, "0").await {
                tracing::warn!(key = kv_key, err = %e, "failed to clear cooldown");
            }
            // Reset failure count so backoff restarts from base.
            let fc_key = format!("{FAILURE_COUNT_PREFIX}{k}");
            if let Err(e) = store.kv_set(&fc_key, "0").await {
                tracing::warn!(key = fc_key, err = %e, "failed to reset failure count");
            }
            // Also reset credit-specific failure count.
            let credit_fc_key = format!("{CREDIT_FAILURE_COUNT_PREFIX}{k}");
            if let Err(e) = store.kv_set(&credit_fc_key, "0").await {
                tracing::warn!(key = credit_fc_key, err = %e, "failed to reset credit failure count");
            }
            // Also reset the persistent-model-failure counter.
            let persistent_fc_key = format!("{PERSISTENT_FAILURE_COUNT_PREFIX}{k}");
            if let Err(e) = store.kv_set(&persistent_fc_key, "0").await {
                tracing::warn!(key = persistent_fc_key, err = %e, "failed to reset persistent failure count");
            }
        }
        // Also reset any persisted failure_count keys that are not in the
        // in-memory map (e.g. survived a restart or were set without a
        // corresponding in-memory cooldown entry).
        if let Ok(fc_entries) = store.kv_list_prefix(FAILURE_COUNT_PREFIX).await {
            for (fc_key, _) in fc_entries {
                if let Err(e) = store.kv_set(&fc_key, "0").await {
                    tracing::warn!(key = fc_key, err = %e, "failed to reset failure count");
                }
            }
        }
        if let Ok(cfc_entries) = store.kv_list_prefix(CREDIT_FAILURE_COUNT_PREFIX).await {
            for (cfc_key, _) in cfc_entries {
                if let Err(e) = store.kv_set(&cfc_key, "0").await {
                    tracing::warn!(key = cfc_key, err = %e, "failed to reset credit failure count");
                }
            }
        }
        if let Ok(pfc_entries) = store.kv_list_prefix(PERSISTENT_FAILURE_COUNT_PREFIX).await {
            for (pfc_key, _) in pfc_entries {
                if let Err(e) = store.kv_set(&pfc_key, "0").await {
                    tracing::warn!(key = pfc_key, err = %e, "failed to reset persistent failure count");
                }
            }
        }
        tracing::info!(
            count = keys.len(),
            "cleared all cooldowns and failure counts"
        );
        agents_to_mark.extend(keys.iter().map(|k| agent_from_key(k).to_string()));
        match store.rate_limit_agents().await {
            Ok(rl_agents) => agents_to_mark.extend(rl_agents),
            Err(e) => tracing::warn!(err = %e, "failed to list rate_limits agents while clearing"),
        }
        match store.kv_list_prefix(RATE_LIMITS_CLEARED_AT_PREFIX).await {
            Ok(marks) => {
                for (k, _) in marks {
                    let agent = k.trim_start_matches(RATE_LIMITS_CLEARED_AT_PREFIX);
                    if !agent.is_empty() {
                        agents_to_mark.push(agent.to_string());
                    }
                }
            }
            Err(e) => tracing::warn!(err = %e, "failed to list rate_limits_cleared_at keys"),
        }
    } else {
        {
            let mut map = cooldown_lock();
            map.remove(key);
        }
        let kv_key = format!("{KV_PREFIX}{key}");
        if let Err(e) = store.kv_set(&kv_key, "0").await {
            tracing::warn!(key = kv_key, err = %e, "failed to clear cooldown");
        }
        // Reset failure count so backoff restarts from base.
        let fc_key = format!("{FAILURE_COUNT_PREFIX}{key}");
        if let Err(e) = store.kv_set(&fc_key, "0").await {
            tracing::warn!(key = fc_key, err = %e, "failed to reset failure count");
        }
        // Also reset credit-specific failure count.
        let credit_fc_key = format!("{CREDIT_FAILURE_COUNT_PREFIX}{key}");
        if let Err(e) = store.kv_set(&credit_fc_key, "0").await {
            tracing::warn!(key = credit_fc_key, err = %e, "failed to reset credit failure count");
        }
        // Also reset the persistent-model-failure counter.
        let persistent_fc_key = format!("{PERSISTENT_FAILURE_COUNT_PREFIX}{key}");
        if let Err(e) = store.kv_set(&persistent_fc_key, "0").await {
            tracing::warn!(key = persistent_fc_key, err = %e, "failed to reset persistent failure count");
        }
        tracing::info!(key, "cleared cooldown and failure count");
        agents_to_mark.push(agent_from_key(key).to_string());
    }
    mark_rate_limits_cleared(store, &agents_to_mark).await;
}

/// Set a cooldown with a specific expiry timestamp. Updates in-memory state only.
///
/// Returns `true` if the cooldown was applied, `false` if an existing longer
/// cooldown prevented the update (never-shorten rule).
///
/// **Callers are responsible for persisting to KV.** Use [`set_cooldown_async`]
/// for critical cooldowns that must survive a crash, or this function combined
/// with a fire-and-forget `tokio::spawn` for non-critical short cooldowns.
fn set_cooldown_in_memory(key: &str, cooldown_until: i64, reason: &str) -> bool {
    let mut map = cooldown_lock();
    if let Some(existing) = map.get(key) {
        // Never shorten an existing cooldown. This prevents a short
        // retry window (e.g., generic rate limit) from overriding a longer
        // billing-cycle cooldown.
        if existing.cooldown_until >= cooldown_until {
            return false;
        }
    }
    map.insert(
        key.to_string(),
        CooldownEntry {
            cooldown_until,
            reason: reason.to_string(),
        },
    );
    true
}

/// Set a cooldown with a specific expiry timestamp and persist to KV inline.
///
/// This is the preferred function for critical cooldowns (credit exhaustion,
/// billing cycle, agent/model failures) that must survive a service crash.
/// The KV write is awaited before returning, so the caller can be confident
/// the cooldown is durable once this function returns `true`.
///
/// Returns `true` if the cooldown was applied, `false` if an existing longer
/// cooldown prevented the update (never-shorten rule).
async fn set_cooldown_async(key: &str, cooldown_until: i64, reason: &str) -> bool {
    if !set_cooldown_in_memory(key, cooldown_until, reason) {
        return false;
    }
    let store_opt = cooldown_store().lock().await.clone();
    if let Some(store) = store_opt {
        let kv_key = format!("{KV_PREFIX}{key}");
        let value = serde_json::json!({"until": cooldown_until, "reason": reason}).to_string();
        if let Err(e) = store.kv_set(&kv_key, &value).await {
            tracing::warn!(
                kv_key,
                err = %e,
                "failed to persist cooldown to KV store — rolling back in-memory entry"
            );
            let mut map = cooldown_lock();
            // Only roll back if nothing has superseded our write.
            // A racing thread may have set a longer cooldown in the meantime;
            // removing unconditionally would erase that valid entry.
            if map.get(key).map(|e| e.cooldown_until) == Some(cooldown_until) {
                map.remove(key);
            }
            return false;
        }
    }
    true
}

/// Helper to acquire the cooldown map lock, logging and recovering on poison.
fn cooldown_lock() -> MutexGuard<'static, HashMap<String, CooldownEntry>> {
    cooldowns().lock().unwrap_or_else(|e| {
        tracing::error!("cooldown map lock poisoned — recovering with partial state; backoff and degraded-agent decisions may be incorrect");
        e.into_inner()
    })
}

/// Helper to acquire the degraded agents lock, logging and recovering on poison.
fn degraded_agents_lock() -> MutexGuard<'static, HashMap<String, String>> {
    degraded_agents().lock().unwrap_or_else(|e| {
        tracing::error!("degraded agents lock poisoned — recovering with partial state; health check decisions may be incorrect");
        e.into_inner()
    })
}

/// Helper to acquire the GitHub 5xx timestamps lock, logging and recovering on poison.
fn github_5xx_timestamps_lock() -> MutexGuard<'static, Vec<i64>> {
    github_5xx_timestamps().lock().unwrap_or_else(|e| {
        tracing::error!("github 5xx timestamps lock poisoned — recovering with partial state; circuit breaker may operate incorrectly");
        e.into_inner()
    })
}

/// Helper to acquire the GitHub 5xx circuit open lock, logging and recovering on poison.
fn github_5xx_circuit_open_lock() -> MutexGuard<'static, Option<i64>> {
    github_5xx_circuit_open().lock().unwrap_or_else(|e| {
        tracing::error!("github 5xx circuit open lock poisoned — recovering with partial state; circuit breaker may operate incorrectly");
        e.into_inner()
    })
}

/// Check if a key is in cooldown with the given reason.
fn is_in_cooldown_with_reason(key: &str, reason: &str) -> bool {
    let map = cooldown_lock();
    map.get(key).is_some_and(|entry| {
        entry.reason == reason && chrono::Utc::now().timestamp() < entry.cooldown_until
    })
}

/// Check if a key is currently in cooldown.
fn is_in_cooldown(key: &str) -> bool {
    let map = cooldown_lock();
    if let Some(entry) = map.get(key) {
        let now = chrono::Utc::now().timestamp();
        return now < entry.cooldown_until;
    }
    false
}

// ---------------------------------------------------------------------------
// Degraded-agent tracking (pre-emptive health check)
// ---------------------------------------------------------------------------

/// Default health check window for recent rate-limit events (hours).
pub const DEFAULT_HEALTH_CHECK_WINDOW_HOURS: u32 = 6;

/// Default threshold: minimum rate-limit events within the window to consider
/// an agent degraded.  A single transient 429 shouldn't mark an agent; we need
/// a pattern of repeated failures.
pub const DEFAULT_DEGRADED_RATE_LIMIT_THRESHOLD: i64 = 3;

/// In-memory map of agents currently flagged as degraded, keyed by agent
/// with the reason that holds now
fn degraded_agents() -> &'static Mutex<HashMap<String, String>> {
    static DEGRADED: OnceLock<Mutex<HashMap<String, String>>> = OnceLock::new();
    DEGRADED.get_or_init(|| Mutex::new(HashMap::new()))
}

/// Check if an agent is currently flagged as degraded by the health check.
///
/// Degraded means: all configured models are cooled **or** the agent has
/// exceeded the rate-limit threshold within the configured lookback window.
pub fn is_agent_degraded(agent: &str) -> bool {
    let set = degraded_agents_lock();
    set.contains_key(agent)
}

/// Reason the agent is currently degraded, if it is
pub fn agent_degraded_reason(agent: &str) -> Option<String> {
    degraded_agents_lock().get(agent).cloned()
}

/// Active degraded agents as `(agent, reason)`, sorted by agent name
pub fn list_degraded_agents() -> Vec<(String, String)> {
    let mut entries: Vec<(String, String)> = degraded_agents_lock()
        .iter()
        .map(|(agent, reason)| (agent.clone(), reason.clone()))
        .collect();
    entries.sort_by(|a, b| a.0.cmp(&b.0));
    entries
}

/// Mark an agent as degraded (called by [`refresh_degraded_agents`]).
pub fn mark_agent_degraded(agent: &str) {
    mark_agent_degraded_with_reason(agent, "degraded");
}

fn mark_agent_degraded_with_reason(agent: &str, reason: &str) {
    let mut set = degraded_agents_lock();
    set.insert(agent.to_string(), reason.to_string());
}

/// Clear the degraded flag for an agent.
pub fn clear_agent_degraded(agent: &str) {
    let mut set = degraded_agents_lock();
    set.remove(agent);
}

/// Refresh the degraded-agent set from the rate_limits table and cooldown state.
///
/// An agent is marked degraded when:
/// 1. It is in agent-level cooldown (explicit cooldown KV key exists and hasn't expired), **or**
/// 2. All its configured models are in cooldown, **or**
/// 3. It has >= `threshold` `limit_type = 'rate'` events within `window_hours`.
///
/// Failures keep their own cooldowns and do not count toward the threshold
/// Agents that no longer meet these criteria are cleared.
///
/// Note: This function calls `clear_expired_cooldowns()` first to ensure the
/// in-memory cooldown map reflects current state. An agent should NOT be marked
/// degraded simply because an expired cooldown entry hasn't been cleaned up yet.
pub async fn refresh_degraded_agents(
    store: &Arc<TaskStore>,
    available_agents: &[String],
    model_checker: &(dyn Fn(&str) -> bool + Send + Sync),
    window_hours: u32,
    threshold: i64,
) {
    // Clear expired cooldowns from in-memory map before evaluating agent state.
    // This ensures we don't incorrectly mark agents as degraded due to stale
    // in-memory entries that should have expired but weren't cleaned up.
    clear_expired_cooldowns();

    // Time the DB query to surface slow health-checks that can dominate the
    // engine tick latency. If the query fails, log and return early.
    let start = chrono::Utc::now();
    let counts = match store.recent_rate_limit_counts(window_hours).await {
        Ok(c) => {
            let dur_ms = (chrono::Utc::now() - start).num_milliseconds();
            tracing::info!(
                duration_ms = dur_ms,
                window_hours,
                "recent_rate_limit_counts query completed"
            );
            c
        }
        Err(e) => {
            let dur_ms = (chrono::Utc::now() - start).num_milliseconds();
            tracing::warn!(
                err = %e,
                duration_ms = dur_ms,
                "failed to query recent rate limit counts for health check"
            );
            return;
        }
    };

    for agent in available_agents {
        let in_cooldown = is_agent_in_cooldown(agent);
        let no_models = !model_checker(agent);
        let rate_limit_count = counts.get(agent.as_str()).copied().unwrap_or(0);
        let over_threshold = rate_limit_count >= threshold;

        let degraded = in_cooldown || no_models || over_threshold;

        if degraded {
            let reason = if in_cooldown {
                "agent in cooldown"
            } else if no_models {
                "all models cooled"
            } else {
                "rate limit threshold exceeded"
            };
            let previous = agent_degraded_reason(agent);
            match previous.as_deref() {
                None => tracing::warn!(
                    agent,
                    reason,
                    rate_limit_count,
                    window_hours,
                    threshold,
                    "pre-emptive health check: marking agent as degraded"
                ),
                Some(prev) if prev != reason => tracing::warn!(
                    agent,
                    reason,
                    previous = prev,
                    rate_limit_count,
                    window_hours,
                    threshold,
                    "pre-emptive health check: degraded reason changed"
                ),
                Some(_) => {}
            }
            mark_agent_degraded_with_reason(agent, reason);
        } else {
            if is_agent_degraded(agent) {
                tracing::info!(
                    agent,
                    rate_limit_count,
                    "pre-emptive health check: agent recovered, clearing degraded flag"
                );
            }
            clear_agent_degraded(agent);
        }
    }
}

/// Clear expired cooldowns from the in-memory map.
pub fn clear_expired_cooldowns() {
    let mut map = cooldown_lock();
    let now = chrono::Utc::now().timestamp();
    map.retain(|_key, entry| now < entry.cooldown_until);
}

// ---------------------------------------------------------------------------
// GitHub 5xx circuit breaker
// ---------------------------------------------------------------------------

/// Sliding window for counting GitHub 5xx errors (seconds).
const GITHUB_5XX_WINDOW_SECS: i64 = 120;

/// Number of 5xx errors within the window to trip the circuit breaker.
const GITHUB_5XX_THRESHOLD: usize = 5;

/// Duration the circuit breaker stays open after tripping (seconds).
const GITHUB_5XX_COOLDOWN_SECS: i64 = 180;

/// Global sliding window of 5xx error timestamps.
fn github_5xx_timestamps() -> &'static Mutex<Vec<i64>> {
    static TS: OnceLock<Mutex<Vec<i64>>> = OnceLock::new();
    TS.get_or_init(|| Mutex::new(Vec::new()))
}

/// Whether the GitHub 5xx circuit breaker is currently open (tripped).
fn github_5xx_circuit_open() -> &'static Mutex<Option<i64>> {
    static OPEN: OnceLock<Mutex<Option<i64>>> = OnceLock::new();
    OPEN.get_or_init(|| Mutex::new(None))
}

/// Record a GitHub 5xx server error and trip the circuit breaker if the
/// sliding window threshold is exceeded.
///
/// Emits a single high-level log entry when the circuit transitions to open.
pub async fn record_github_5xx() {
    let now = chrono::Utc::now().timestamp();
    let window_start = now - GITHUB_5XX_WINDOW_SECS;

    let should_trip = {
        let mut ts = github_5xx_timestamps_lock();
        ts.retain(|t| *t >= window_start);
        ts.push(now);
        ts.len() >= GITHUB_5XX_THRESHOLD
    };

    if should_trip {
        let cooldown_until = {
            let mut open = github_5xx_circuit_open_lock();
            if open.is_none() {
                let cd = now + GITHUB_5XX_COOLDOWN_SECS;
                *open = Some(cd);
                tracing::warn!(
                    errors_in_window = GITHUB_5XX_THRESHOLD,
                    window_secs = GITHUB_5XX_WINDOW_SECS,
                    cooldown_secs = GITHUB_5XX_COOLDOWN_SECS,
                    "GitHub 5xx circuit breaker OPEN — throttling non-critical work"
                );
                Some(cd)
            } else {
                None
            }
        };
        // Persist to KV (outside the mutex lock so the future is Send).
        if let Some(cd) = cooldown_until {
            let store_opt = cooldown_store().lock().await.clone();
            if let Some(store) = store_opt {
                let _ = store.kv_set("cooldown:github:5xx", &cd.to_string()).await;
            }
            // Also set the generic cooldown so is_agent_in_cooldown("github:5xx")
            // returns true immediately, avoiding a race window where concurrent
            // requests bypass the circuit breaker.
            set_agent_cooldown("github:5xx", GITHUB_5XX_COOLDOWN_SECS as u64).await;
        }
        // Clear the sliding window after tripping so we don't re-trip immediately
        let mut ts = github_5xx_timestamps_lock();
        ts.clear();
    }
}

/// Check if the GitHub 5xx circuit breaker is currently open.
///
/// When open, non-critical background work (cleanup, polling, mention scans)
/// should be skipped. Critical operations (PR creation) should still attempt
/// with their existing exponential backoff.
///
/// Automatically recovers (logs a recovery message) when the cooldown expires.
pub fn is_github_circuit_open() -> bool {
    // Check the dedicated in-memory circuit flag first. If present, honour it
    // and perform auto-recovery when it expires (logging once on close).
    {
        let mut open = github_5xx_circuit_open_lock();
        if let Some(until) = *open {
            let now = chrono::Utc::now().timestamp();
            if now >= until {
                *open = None;
                tracing::info!("GitHub 5xx circuit breaker CLOSED — resuming normal operations");
                return false;
            }
            return true;
        }
    }

    // If the dedicated flag is not set, consult the generic cooldown map so
    // callers see a consistent view regardless of which code path set the
    // persisted cooldown (e.g. send_with_retries -> set_agent_cooldown).
    if let Some(until) = cooldown_until("github:5xx") {
        // Populate the dedicated in-memory flag so future checks and the
        // auto-recovery path go through the same code and logging.
        let mut open = github_5xx_circuit_open_lock();
        if open.is_none() {
            *open = Some(until);
            let now = chrono::Utc::now().timestamp();
            if now < until {
                let remaining = until - now;
                tracing::warn!(
                    remaining_secs = remaining,
                    "GitHub 5xx circuit breaker OPEN (discovered via generic cooldown) — throttling non-critical work"
                );
            }
        }
        return true;
    }

    false
}

/// Returns the remaining seconds of the GitHub 5xx circuit breaker cooldown,
/// or 0 if the circuit is closed.
pub fn github_circuit_remaining_secs() -> u64 {
    // Prefer the dedicated in-memory value when present.
    {
        let open = github_5xx_circuit_open_lock();
        if let Some(until) = *open {
            let now = chrono::Utc::now().timestamp();
            if now < until {
                return u64::try_from(until - now).unwrap_or(0);
            }
            return 0;
        }
    }

    // Fall back to the generic cooldown map if the dedicated flag isn't set.
    if let Some(until) = cooldown_until("github:5xx") {
        let now = chrono::Utc::now().timestamp();
        if now < until {
            return u64::try_from(until - now).unwrap_or(0);
        }
    }
    0
}

/// Parse a "try again at {date}" or "reset at {date}" from an error message.
///
/// Returns a Unix timestamp if a retry-at date is found.
pub(crate) fn parse_retry_at(error_message: &str) -> Option<i64> {
    if error_message.is_empty() {
        return None;
    }

    let lower = error_message.to_lowercase();
    // Accept a wider set of vendor-provided markers. Claude commonly uses
    // variations like "resets 10:30am" (no explicit "at"), so include
    // both "resets" and the older "reset at"/"try again at" markers.
    for retry_marker in [
        "try again at ",
        "reset at ",
        "resets at ",
        "resets ",
        "reset ",
    ] {
        if let Some(idx) = lower.find(retry_marker) {
            let date_str = &lower[idx + retry_marker.len()..];
            // Take until period, newline, or end
            let date_str = date_str
                .split(['.', '\n'])
                .next()
                .unwrap_or(date_str)
                .trim();

            // Try common date formats
            // "Mar 26th, 2026 5:55 AM" → strip ordinal suffixes
            let cleaned = date_str
                .replace("st,", ",")
                .replace("nd,", ",")
                .replace("rd,", ",")
                .replace("th,", ",");

            for fmt in &[
                "%b %d, %Y %I:%M %p", // Mar 26, 2026 5:55 AM
                "%b %d, %Y %I:%M%p",  // Mar 26, 2026 5:55AM
                "%B %d, %Y %I:%M %p", // March 26, 2026 5:55 AM
                "%Y-%m-%dT%H:%M:%S",  // 2026-03-26T05:55:00
                "%Y-%m-%d %H:%M:%S",  // 2026-03-26 05:55:00
                "%Y-%m-%d %H:%M",     // 2026-03-26 05:55
            ] {
                if let Ok(dt) = chrono::NaiveDateTime::parse_from_str(&cleaned, fmt) {
                    // Vendor messages (e.g. Codex/GLM) emit times in the machine's
                    // local timezone, not UTC. Interpret naive datetime as local time.
                    let utc = match chrono::Local.from_local_datetime(&dt) {
                        chrono::LocalResult::Single(local_dt) => {
                            local_dt.with_timezone(&chrono::Utc)
                        }
                        chrono::LocalResult::Ambiguous(earliest, _) => {
                            // DST fold: pick the earlier (conservative) interpretation
                            earliest.with_timezone(&chrono::Utc)
                        }
                        chrono::LocalResult::None => {
                            // DST gap: fall back to treating as UTC to avoid panic
                            dt.and_utc()
                        }
                    };
                    tracing::info!(
                        retry_at = %utc,
                        raw = date_str,
                        marker = retry_marker,
                        "parsed rate limit retry-at date"
                    );
                    return Some(utc.timestamp());
                }
            }
            tracing::debug!(
                raw = date_str,
                marker = retry_marker,
                "could not parse retry-at date"
            );

            // If the vendor provided only a time (e.g. "resets 10:30am"),
            // attempt to parse a time-only and interpret it as the next
            // occurrence of that time in the local timezone.
            // Strip parenthesised timezone hints like "(America/Sao_Paulo)".
            let time_only = date_str.split_once('(').map_or(date_str, |(t, _)| t).trim();
            // Hour-only values like "8am" have no minute field, which chrono needs, so insert ":00"
            let mut compact = time_only.replace(" am", "am").replace(" pm", "pm");
            let split = compact.rfind(' ').map_or(0, |i| i + 1);
            let last = &compact[split..];
            if !last.contains(':')
                && (last.ends_with("am") || last.ends_with("pm"))
                && last.len() > 2
            {
                let at = split + last.len() - 2;
                compact.insert_str(at, ":00");
            }
            // "Jun 9 at 1am" has no year, so take the next future occurrence
            let md = compact.replace(" at ", " ");
            let now_local = chrono::Local::now();
            let this_year = chrono::Datelike::year(&now_local);
            for year in [this_year, this_year + 1] {
                let with_year = format!("{year} {md}");
                if let Ok(dt) =
                    chrono::NaiveDateTime::parse_from_str(&with_year, "%Y %b %d %I:%M%p")
                    && let Some(local_dt) = chrono::Local.from_local_datetime(&dt).earliest()
                    && local_dt > now_local
                {
                    return Some(local_dt.timestamp());
                }
            }
            let time_only = compact.as_str();
            for tf in &["%I:%M %p", "%I:%M%p", "%I %p", "%H:%M"] {
                if let Ok(t) = chrono::NaiveTime::parse_from_str(time_only, tf) {
                    // Build a NaiveDateTime for today at that time, then choose
                    // today or tomorrow depending on whether the time has passed.
                    let local_now = chrono::Local::now();
                    let today = local_now.date_naive();
                    let candidate = chrono::NaiveDateTime::new(today, t);
                    // Use the TimeZone trait methods; annotate types so the
                    // compiler can infer the concrete DateTime types.
                    let candidate_local: chrono::DateTime<chrono::Local> =
                        match chrono::Local.from_local_datetime(&candidate) {
                            chrono::LocalResult::Single(dt) => dt,
                            chrono::LocalResult::Ambiguous(earliest, _) => earliest,
                            chrono::LocalResult::None => {
                                // DST gap — fall back to UTC interpretation
                                let utc_dt = candidate.and_utc();
                                return Some(utc_dt.timestamp());
                            }
                        };
                    let chosen = if candidate_local <= local_now {
                        // Use tomorrow
                        let tomorrow = today.succ_opt().unwrap_or(today);
                        let ndt = chrono::NaiveDateTime::new(tomorrow, t);
                        match chrono::Local.from_local_datetime(&ndt) {
                            chrono::LocalResult::Single(dt) => dt.with_timezone(&chrono::Utc),
                            chrono::LocalResult::Ambiguous(earliest, _) => {
                                earliest.with_timezone(&chrono::Utc)
                            }
                            chrono::LocalResult::None => ndt.and_utc(),
                        }
                    } else {
                        candidate_local.with_timezone(&chrono::Utc)
                    };
                    tracing::info!(
                        retry_at = %chosen,
                        raw = date_str,
                        marker = retry_marker,
                        "parsed rate limit retry-at time-only; interpreted as next occurrence"
                    );
                    return Some(chosen.timestamp());
                }
            }
        }
    }

    // Without a specific retry/reset date, return None and let the caller use
    // its default cooldown (exponential backoff). A blanket long fallback here
    // would cause false billing-cycle cooldowns on agents without those limits.
    None
}

/// Parse a relative usage window (e.g. "5-hour usage limit") from an error
/// message.
///
/// Returns the window length in seconds. Only matches in usage/quota/limit
/// contexts ("usage limit", "quota", "window", "limit") to avoid interpreting
/// unrelated numbers as cooldown durations.
///
/// `pub(crate)` so `engine::runner::fallback::handle_error()` can apply the
/// same window-detection the `needs_review` call site
/// (`record_rate_limit()`, below) already uses — see issue #3580.
pub(crate) fn parse_relative_usage_window(error_message: &str) -> Option<u64> {
    static WINDOW_RE: LazyLock<Regex> = LazyLock::new(|| {
        Regex::new(
            r"(?i)(?:^|\s|[\(\)\.,;:!?])(?:(\d+)\s*-?\s*(hour|hours|hr|hrs|minute|minutes|min|mins|day|days)|(hourly|daily|weekly|monthly))(?:$|\s|[\(\)\.,;:!?])"
        )
        .expect("usage window regex is valid")
    });

    if error_message.is_empty() {
        return None;
    }

    let lower = error_message.to_lowercase();
    let has_context = lower.contains("usage limit")
        || lower.contains("quota")
        || lower.contains("window")
        || lower.contains("limit");
    if !has_context {
        return None;
    }

    let mut max_secs: Option<u64> = None;
    for caps in WINDOW_RE.captures_iter(&lower) {
        let secs = if let Some(num) = caps.get(1) {
            let n: u64 = num.as_str().parse().ok()?;
            let unit = caps.get(2)?.as_str().to_lowercase();
            let multiplier = match unit.as_str() {
                "hour" | "hours" | "hr" | "hrs" => 3600,
                "minute" | "minutes" | "min" | "mins" => 60,
                "day" | "days" => 86400,
                _ => return None,
            };
            n.checked_mul(multiplier)?
        } else if let Some(word) = caps.get(3) {
            match word.as_str().to_lowercase().as_str() {
                "hourly" => 3600,
                "daily" => 86400,
                "weekly" => 604_800,
                "monthly" => 2_592_000,
                _ => return None,
            }
        } else {
            continue;
        };
        max_secs = Some(max_secs.map_or(secs, |m| m.max(secs)));
    }

    max_secs
}

#[cfg(test)]
pub(crate) async fn reset_global_state() {
    cooldowns().lock().unwrap().clear();
    *cooldown_store().lock().await = None;
    github_5xx_timestamps().lock().unwrap().clear();
    *github_5xx_circuit_open().lock().unwrap() = None;
    degraded_agents().lock().unwrap().clear();
    opencode::reset_model_caches_for_test();
}

#[cfg(test)]
mod tests {
    use std::time::Duration;

    use super::*;
    use chrono::Datelike;
    use chrono::TimeZone;
    use serial_test::serial;
    use tokio::time;

    async fn test_store() -> Arc<TaskStore> {
        Arc::new(TaskStore::open_memory().await.expect("in-memory store"))
    }

    fn sqlite_datetime_offset(secs: i64) -> String {
        (chrono::Utc::now() + chrono::Duration::seconds(secs))
            .format("%Y-%m-%d %H:%M:%S")
            .to_string()
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn init_loads_unexpired_cooldowns_from_kv() {
        super::reset_global_state().await;
        let store = test_store().await;
        let cooldown_until = chrono::Utc::now().timestamp() + 3600; // 1 hour from now

        store
            .kv_set(
                &format!("{KV_PREFIX}kimi:k2p5"),
                &cooldown_until.to_string(),
            )
            .await
            .unwrap();

        // Also add an already-expired entry
        let expired = chrono::Utc::now().timestamp() - 1;
        store
            .kv_set(
                &format!("{KV_PREFIX}opencode:old-free"),
                &expired.to_string(),
            )
            .await
            .unwrap();

        {
            let mut map = cooldown_lock();
            map.clear();
        }

        init_cooldown_store(store).await;

        assert!(
            is_model_in_cooldown("kimi", "k2p5"),
            "unexpired cooldown should be loaded"
        );
        assert!(
            !is_model_in_cooldown("opencode", "old-free"),
            "expired cooldown should not be loaded"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn credit_exhaustion_inside_active_cooldown_keeps_counter_at_one() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let msg =
            "API error (status 402 Payment Required): usage balance exhausted, out of credits";
        assert!(record_credit_exhaustion_from_message("credit_dup_agent", Some("m"), msg).await);
        assert!(record_credit_exhaustion_from_message("credit_dup_agent", Some("m"), msg).await);

        assert!(is_agent_in_cooldown("credit_dup_agent"));
        let count = store
            .kv_get(&format!("{CREDIT_FAILURE_COUNT_PREFIX}credit_dup_agent"))
            .await
            .unwrap();
        assert_eq!(count.as_deref(), Some("1"));
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn non_credit_message_is_not_recorded_as_credit_exhaustion() {
        super::reset_global_state().await;
        assert!(!record_credit_exhaustion_from_message("x_agent", Some("m"), "boom").await);
        assert!(!is_agent_in_cooldown("x_agent"));
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_model_failure_writes_to_kv() {
        super::reset_global_state().await;
        let store = test_store().await;

        *cooldown_store().lock().await = Some(Arc::clone(&store));

        record_model_failure("testagent_persist", "testmodel_persist").await;

        // Give the async KV write a moment to settle before querying.
        time::sleep(Duration::from_millis(50)).await;

        let kv_key = format!("{KV_PREFIX}testagent_persist:testmodel_persist");
        let stored = store.kv_get(&kv_key).await.unwrap();
        assert!(stored.is_some(), "model failure should be persisted to KV");
        let (ts, reason) = parse_cooldown_value(&stored.unwrap()).expect("cooldown value");
        let now = chrono::Utc::now().timestamp();
        // Should be in the future (cooldown_until, not failed_at)
        assert!(ts > now, "persisted timestamp should be in the future");
        assert_eq!(reason, "model_error");
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn cooldown_reason_survives_restart_via_kv() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_reason_restart";
        let future = chrono::Utc::now().timestamp() + 600;

        init_cooldown_store(Arc::clone(&store)).await;
        set_cooldown_async(agent, future, "billing_cycle_exhausted").await;

        // Simulate a service restart: drop the in-memory map, reload from KV.
        {
            let mut map = cooldown_lock();
            map.clear();
        }
        init_cooldown_store(store).await;

        let entries = list_all_cooldowns();
        let entry = entries.iter().find(|(key, ..)| key == agent);
        assert_eq!(
            entry.map(|(_, _, reason)| reason.as_str()),
            Some("billing_cycle_exhausted"),
            "cooldown reason must survive a restart, not collapse to 'persisted'"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_persistent_model_failure_escalates_and_increments_failure_count() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "testagent_persistent";
        let model = "testmodel_persistent";
        let key = format!("{agent}:{model}");

        // First failure -> base 4h.
        record_persistent_model_failure(agent, model).await;
        let now = chrono::Utc::now().timestamp();
        let first_until = cooldown_until(&key).expect("first cooldown should exist");
        let first_remaining = first_until.saturating_sub(now);
        assert!(
            first_remaining >= PERSISTENT_MODEL_BACKOFF_BASE_SECS - 5,
            "first persistent model cooldown should be ~4h, got {first_remaining}s"
        );

        // Second failure -> 12h (3x growth).
        record_persistent_model_failure(agent, model).await;
        let now2 = chrono::Utc::now().timestamp();
        let second_until = cooldown_until(&key).expect("second cooldown should exist");
        let second_remaining = second_until.saturating_sub(now2);
        assert!(
            second_remaining >= (PERSISTENT_MODEL_BACKOFF_BASE_SECS * 3) - 5,
            "second persistent model cooldown should be ~12h, got {second_remaining}s"
        );

        let count_key = format!("{PERSISTENT_FAILURE_COUNT_PREFIX}{agent}:{model}");
        let stored_count = store
            .kv_get(&count_key)
            .await
            .expect("kv read should succeed")
            .expect("failure count should be written");
        assert_eq!(stored_count, "2");
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_agent_success_does_not_reset_persistent_failure_count() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "testagent_persistent_success";
        let model = "testmodel_persistent_success";
        let key = format!("{agent}:{model}");

        // Two persistent failures escalate the counter to 2 (~12h cooldown).
        record_persistent_model_failure(agent, model).await;
        record_persistent_model_failure(agent, model).await;

        let count_key = format!("{PERSISTENT_FAILURE_COUNT_PREFIX}{agent}:{model}");
        let stored_count = store
            .kv_get(&count_key)
            .await
            .expect("kv read should succeed")
            .expect("failure count should be written");
        assert_eq!(stored_count, "2");

        // A successful run resets the generic failure_count but must NOT touch
        // the persistent counter — otherwise an intervening success would erase
        // escalation progress for a structurally unreliable model.
        record_agent_success(agent, model).await;

        let stored_count_after_success = store
            .kv_get(&count_key)
            .await
            .expect("kv read should succeed")
            .expect("persistent failure count should still exist");
        assert_eq!(
            stored_count_after_success, "2",
            "persistent failure count must survive an intervening success"
        );

        // The next persistent failure should escalate from 3 (~36h), not reset to base (4h).
        record_persistent_model_failure(agent, model).await;
        let now = chrono::Utc::now().timestamp();
        let third_until = cooldown_until(&key).expect("third cooldown should exist");
        let third_remaining = third_until.saturating_sub(now);
        assert!(
            third_remaining > PERSISTENT_MODEL_BACKOFF_BASE_SECS,
            "third persistent model failure should escalate past the 4h base, got {third_remaining}s"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_persistent_model_failure_stays_capped_at_7_days() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "testagent_persistent_capped";
        let model = "testmodel_persistent_capped";
        let key = format!("{agent}:{model}");

        for _ in 0..9 {
            record_persistent_model_failure(agent, model).await;
        }

        let remaining = cooldown_until(&key)
            .expect("cooldown should exist")
            .saturating_sub(chrono::Utc::now().timestamp());
        assert!(
            remaining >= PERSISTENT_MODEL_BACKOFF_MAX_SECS - 5,
            "persistent model cooldown should stay capped at 7 days, got {remaining}s"
        );
        assert!(
            remaining <= PERSISTENT_MODEL_BACKOFF_MAX_SECS + 5,
            "persistent model cooldown should not exceed 7 days, got {remaining}s"
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn parse_retry_at_codex_format() {
        let msg = "You've hit your usage limit. Upgrade to Pro, visit https://chatgpt.com/codex/settings/usage to purchase more credits or try again at Mar 26th, 2026 5:55 AM.";
        let ts = parse_retry_at(msg);
        assert!(ts.is_some(), "should parse codex retry-at date");

        // The parsed timestamp must equal "Mar 26 2026 05:55 local" in UTC.
        // We reconstruct what the function should produce and compare directly.
        let naive = chrono::NaiveDateTime::new(
            chrono::NaiveDate::from_ymd_opt(2026, 3, 26).unwrap(),
            chrono::NaiveTime::from_hms_opt(5, 55, 0).unwrap(),
        );
        let expected_utc = match chrono::Local.from_local_datetime(&naive) {
            chrono::LocalResult::Single(local_dt) => local_dt.with_timezone(&chrono::Utc),
            chrono::LocalResult::Ambiguous(earliest, _) => earliest.with_timezone(&chrono::Utc),
            chrono::LocalResult::None => naive.and_utc(),
        };
        assert_eq!(
            ts.unwrap(),
            expected_utc.timestamp(),
            "timestamp should reflect local-timezone interpretation of the vendor date"
        );

        // Sanity: the UTC date must still be March 26 (true for all timezones within ±14h)
        let dt = chrono::DateTime::from_timestamp(ts.unwrap(), 0).unwrap();
        assert_eq!(dt.month(), 3);
        assert_eq!(dt.day(), 26);
    }

    #[serial(cooldown_state)]
    #[test]
    fn parse_retry_at_hour_only_and_month_day() {
        let now = chrono::Utc::now().timestamp();
        for msg in [
            "You've hit your session limit · resets 8am (America/Sao_Paulo)",
            "You've hit your session limit · resets 6pm (America/Sao_Paulo)",
            "You've hit your session limit · resets 12am (America/Sao_Paulo)",
            "You've hit your weekly limit · resets Jun 9 at 1am (America/Sao_Paulo)",
        ] {
            let ts = parse_retry_at(msg).unwrap_or_else(|| panic!("should parse: {msg}"));
            assert!(ts > now, "{msg}");
        }
    }

    #[serial(cooldown_state)]
    #[test]
    fn parse_retry_at_weekly_resets_without_year() {
        let t = chrono::Local::now() + chrono::Duration::days(2);
        let msg = format!(
            "You've hit your weekly limit · resets {} {} at 1am (America/Sao_Paulo)",
            t.format("%b"),
            t.day()
        );
        let ts = parse_retry_at(&msg).expect("should parse year-less reset date");
        let remaining = ts - chrono::Utc::now().timestamp();
        assert!(
            remaining > 3600 && remaining < 3 * 86400,
            "expected ~2d retry-at, got {remaining}s"
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn parse_retry_at_no_date() {
        assert!(parse_retry_at("generic rate limit error").is_none());
        assert!(parse_retry_at("").is_none());
    }

    #[serial(cooldown_state)]
    #[test]
    fn parse_retry_at_glm_reset_format_with_seconds() {
        let msg = "Weekly/Monthly Limit Exhausted. Your limit will reset at 2026-04-23 18:33:48";
        let ts = parse_retry_at(msg).expect("should parse GLM reset-at date");

        let naive = chrono::NaiveDateTime::new(
            chrono::NaiveDate::from_ymd_opt(2026, 4, 23).expect("valid date"),
            chrono::NaiveTime::from_hms_opt(18, 33, 48).expect("valid time"),
        );
        let expected_utc = match chrono::Local.from_local_datetime(&naive) {
            chrono::LocalResult::Single(local_dt) => local_dt.with_timezone(&chrono::Utc),
            chrono::LocalResult::Ambiguous(earliest, _) => earliest.with_timezone(&chrono::Utc),
            chrono::LocalResult::None => naive.and_utc(),
        };

        assert_eq!(
            ts,
            expected_utc.timestamp(),
            "timestamp should reflect local-timezone interpretation of vendor reset date"
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn parse_retry_at_billing_cycle_without_date_returns_none() {
        // Without a specific "try again at" date, parse_retry_at should return
        // None so callers use their default cooldown (30m agent / 1h model).
        let msg = "You've reached your usage limit for this billing cycle. Your quota will be refreshed in the next cycle. Upgrade to get more.";
        assert!(
            parse_retry_at(msg).is_none(),
            "billing cycle without date should return None"
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn parse_relative_usage_window_extracts_kimi_five_hour_window() {
        let msg = "Failed to authenticate. API Error: 403 You've reached your 5-hour usage limit. Your quota will reset when the current 5-hour window expires.";
        assert_eq!(
            parse_relative_usage_window(msg),
            Some(5 * 3600),
            "should extract the 5-hour window as seconds"
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn parse_relative_usage_window_handles_common_durations() {
        assert_eq!(
            parse_relative_usage_window("You've reached your hourly usage limit."),
            Some(3600)
        );
        assert_eq!(
            parse_relative_usage_window("You've reached your daily usage limit."),
            Some(86400)
        );
        assert_eq!(
            parse_relative_usage_window("You've reached your weekly limit."),
            Some(604_800)
        );
        assert_eq!(
            parse_relative_usage_window("You've reached your 30-minute usage limit."),
            Some(1800)
        );
        assert_eq!(
            parse_relative_usage_window("You've reached your 2-day quota."),
            Some(172_800)
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn parse_relative_usage_window_returns_none_without_usage_context() {
        // A bare rate-limit message with a number should not be interpreted as a
        // relative usage window unless usage/quota/window context is present.
        assert!(parse_relative_usage_window("429 Too Many Requests").is_none());
        assert!(parse_relative_usage_window("rate limit exceeded").is_none());
        assert!(parse_relative_usage_window("").is_none());
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_rate_limit_uses_model_specific_window_cooldown() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_record_rl_window_agent";
        let model = "test_record_rl_window_model";
        let msg = "You've reached your 5-hour usage limit. Your quota will reset when the current 5-hour window expires.";

        record_rate_limit(agent, Some(model), msg).await;

        assert!(
            is_model_in_cooldown(agent, model),
            "model should be cooled for the relative usage window"
        );
        assert!(
            !is_agent_in_cooldown(agent),
            "agent should not be cooled for a model-scoped quota"
        );

        let key = format!("{agent}:{model}");
        let remaining = cooldown_until(&key)
            .expect("model cooldown should exist")
            .saturating_sub(chrono::Utc::now().timestamp());
        assert!(
            remaining >= 5 * 3600 - 5,
            "model cooldown should be at least ~5 hours, got {remaining}s"
        );
        assert_eq!(
            cooldown_reason(&key).as_deref(),
            Some("rate_limit_usage_window"),
            "relative usage-window cooldown must be labeled as a rate limit, not silence"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_rate_limit_without_window_falls_back_to_agent_and_model() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_record_rl_fallback_agent";
        let model = "test_record_rl_fallback_model";
        let msg = "429 Too Many Requests";

        record_rate_limit(agent, Some(model), msg).await;

        assert!(
            is_agent_in_cooldown(agent),
            "agent should be cooled when no relative window is detected"
        );
        assert!(
            is_model_in_cooldown(agent, model),
            "model should be cooled when no relative window is detected"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn agent_cooldown_persists_to_kv() {
        super::reset_global_state().await;
        // Agent-level cooldowns persist to KV via set_cooldown_async
        let agent = "test_agent_persist_check";
        record_agent_failure_with_message(agent, "").await;
        assert!(is_agent_in_cooldown(agent));
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn cooldown_never_shortens() {
        super::reset_global_state().await;
        let agent = "test_agent_no_shorten";
        set_agent_cooldown(agent, 24 * 60 * 60).await;
        let initial = {
            let map = cooldowns().lock().unwrap();
            map.get(agent)
                .expect("cooldown entry should exist")
                .cooldown_until
        };

        // Shorter cooldown attempt should not override the existing one.
        record_agent_failure_with_message(agent, "rate limit").await;
        let after = {
            let map = cooldowns().lock().unwrap();
            map.get(agent)
                .expect("cooldown entry should exist")
                .cooldown_until
        };
        assert_eq!(
            initial, after,
            "cooldown should not be shortened by later failures"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn silence_agent_cooldown_is_short_lived() {
        super::reset_global_state().await;
        let agent = "test_silence_agent_cd";
        assert!(!is_agent_in_cooldown(agent));

        set_agent_cooldown(agent, SILENCE_AGENT_COOLDOWN_SECS).await;
        assert!(is_agent_in_cooldown(agent));

        // Verify it's a short cooldown (120s), not the exponential backoff one
        let map = cooldowns().lock().unwrap();
        let entry = map.get(agent).expect("should have cooldown entry");
        let now = chrono::Utc::now().timestamp();
        let remaining = entry.cooldown_until - now;
        assert!(
            remaining <= i64::try_from(SILENCE_AGENT_COOLDOWN_SECS).unwrap(),
            "silence agent cooldown should be <= {SILENCE_AGENT_COOLDOWN_SECS}s, got {remaining}s"
        );
        assert!(remaining > 0, "cooldown should still be active");
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_silence_detection_applies_extended_cooldown() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_silence_count_agent";
        let model = "test_silence_count_model";
        assert!(!is_model_in_cooldown(agent, model));

        for _ in 0..SILENCE_COUNT_THRESHOLD {
            let result = record_silence_detection(agent, model).await;
            assert!(result.is_some());
        }

        assert!(is_model_in_cooldown(agent, model));
        let key = format!("{agent}:{model}");
        let remaining = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(&key).expect("cooldown entry should exist");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };
        assert!(
            remaining >= i64::try_from(SILENCE_EXTENDED_COOLDOWN_SECS).unwrap() - 5,
            "extended cooldown should be applied, got {remaining}s"
        );

        let kv_key = format!("{SILENCE_COUNT_PREFIX}{agent}:{model}");
        let stored = store.kv_get(&kv_key).await.unwrap().unwrap();
        assert_eq!(stored, "[]");
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_out_of_credits() {
        let msg = "Error: API error - {\"error\":{\"message\":\"You exceeded your current quota\",\"type\":\"insufficient_quota\",\"param\":null,\"code\":\"insufficient_quota\"}}";
        assert_eq!(
            detect_credit_exhaustion(msg),
            Some(CreditExhaustionReason::OutOfCredits)
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_out_of_credits_variations() {
        assert_eq!(
            detect_credit_exhaustion("out of credits"),
            Some(CreditExhaustionReason::OutOfCredits)
        );
        assert_eq!(
            detect_credit_exhaustion("outofcredits"),
            Some(CreditExhaustionReason::OutOfCredits)
        );
        assert_eq!(
            detect_credit_exhaustion("no credits remaining"),
            Some(CreditExhaustionReason::OutOfCredits)
        );
        assert_eq!(
            detect_credit_exhaustion("insufficient funds"),
            Some(CreditExhaustionReason::OutOfCredits)
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_org_level_disabled() {
        let msg = "Error: API error - {\"error\":{\"message\":\"Organization has been disabled\",\"type\":\"org_disabled\",\"code\":\"overageDisabledReason:org_level_disabled\"}}";
        assert_eq!(
            detect_credit_exhaustion(msg),
            Some(CreditExhaustionReason::OrgLevelDisabled)
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_org_level_disabled_variations() {
        assert_eq!(
            detect_credit_exhaustion("organization has been disabled"),
            Some(CreditExhaustionReason::OrgLevelDisabled)
        );
        assert_eq!(
            detect_credit_exhaustion("org disabled"),
            Some(CreditExhaustionReason::OrgLevelDisabled)
        );
        assert_eq!(
            detect_credit_exhaustion("billing quota exceeded"),
            Some(CreditExhaustionReason::OrgLevelDisabled)
        );
        assert_eq!(
            detect_credit_exhaustion("credit balance too low"),
            Some(CreditExhaustionReason::OrgLevelDisabled)
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_overage_disabled_reason_parsing() {
        let msg = "overageDisabledReason: out_of_credits";
        assert_eq!(
            detect_credit_exhaustion(msg),
            Some(CreditExhaustionReason::OutOfCredits)
        );

        let msg = "overageDisabledReason: org_level_disabled";
        assert_eq!(
            detect_credit_exhaustion(msg),
            Some(CreditExhaustionReason::OrgLevelDisabled)
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_glm_insufficient_balance() {
        // GLM/ZhipuAI 429: "Insufficient balance or no resource package. Please recharge."
        assert_eq!(
            detect_credit_exhaustion(
                "API Error: Request rejected (429) · Insufficient balance or no resource package. Please recharge."
            ),
            Some(CreditExhaustionReason::OutOfCredits)
        );
        assert_eq!(
            detect_credit_exhaustion("Insufficient balance"),
            Some(CreditExhaustionReason::OutOfCredits)
        );
        assert_eq!(
            detect_credit_exhaustion("please recharge your account"),
            Some(CreditExhaustionReason::OutOfCredits)
        );
        assert_eq!(
            detect_credit_exhaustion("no resource package available"),
            Some(CreditExhaustionReason::OutOfCredits)
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_none_for_regular_rate_limit() {
        assert!(detect_credit_exhaustion("rate limit exceeded").is_none());
        assert!(detect_credit_exhaustion("too many requests").is_none());
        assert!(detect_credit_exhaustion("429 Too Many Requests").is_none());
        assert!(detect_credit_exhaustion("you've hit your usage limit").is_none());
        assert!(
            detect_credit_exhaustion("You've hit your usage limit, try again in a few minutes.")
                .is_none()
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_credit_exhaustion_applies_agent_cooldown() {
        super::reset_global_state().await;
        let agent = "test_credit_exhaust_agent";
        assert!(!is_agent_in_cooldown(agent));

        record_credit_exhaustion(agent, CreditExhaustionReason::OutOfCredits).await;
        assert!(is_agent_in_cooldown(agent));

        let remaining = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(agent).expect("cooldown entry should exist");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };
        // First failure: base = 1h
        assert!(
            remaining >= CREDIT_BACKOFF_BASE_SECS - 5,
            "first out_of_credits cooldown should be ~1 hour, got {remaining}s"
        );
        assert!(
            remaining <= CREDIT_BACKOFF_MAX_SECS,
            "cooldown should not exceed cap of 8h, got {remaining}s"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_credit_exhaustion_org_level_applies_longer_cooldown() {
        super::reset_global_state().await;
        let agent = "test_org_disabled_agent";
        assert!(!is_agent_in_cooldown(agent));

        record_credit_exhaustion(agent, CreditExhaustionReason::OrgLevelDisabled).await;
        assert!(is_agent_in_cooldown(agent));

        let remaining = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(agent).expect("cooldown entry should exist");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };
        // First failure: base = 2h (ORG_BACKOFF_BASE_SECS)
        assert!(
            remaining >= ORG_BACKOFF_BASE_SECS - 5,
            "first org-level disabled cooldown should be ~2 hours, got {remaining}s"
        );
        assert!(
            remaining <= CREDIT_BACKOFF_MAX_SECS,
            "cooldown should not exceed cap of 8h, got {remaining}s"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_credit_exhaustion_billing_cycle_applies_24h() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_billing_cycle_agent";
        assert!(!is_agent_in_cooldown(agent));

        record_credit_exhaustion(agent, CreditExhaustionReason::BillingCycleExhausted).await;
        assert!(is_agent_in_cooldown(agent));

        let remaining = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(agent).expect("cooldown entry should exist");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };
        // First failure: base = 24h
        assert!(
            remaining >= BILLING_CYCLE_COOLDOWN_SECS - 5,
            "first billing cycle cooldown should be ~24 hours, got {remaining}s"
        );
        assert!(
            remaining <= BILLING_CYCLE_MAX_SECS,
            "cooldown should not exceed cap of 7 days, got {remaining}s"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_credit_exhaustion_billing_cycle_escalates() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_billing_cycle_escalation";

        // First failure: 24h
        record_credit_exhaustion(agent, CreditExhaustionReason::BillingCycleExhausted).await;
        let remaining_1 = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(agent).expect("cooldown entry should exist");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };
        assert!(
            (BILLING_CYCLE_COOLDOWN_SECS - 5..=BILLING_CYCLE_COOLDOWN_SECS + 5)
                .contains(&remaining_1),
            "first billing cycle cooldown should be ~24h, got {remaining_1}s"
        );

        // Clear in-memory cooldown (but NOT failure count) to allow next set_cooldown_async
        {
            let mut map = cooldowns().lock().unwrap();
            map.remove(agent);
        }

        // Second failure: 24h * 3 = 72h (3 days)
        record_credit_exhaustion(agent, CreditExhaustionReason::BillingCycleExhausted).await;
        let remaining_2 = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(agent).expect("cooldown entry should exist");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };
        let expected_2 = BILLING_CYCLE_COOLDOWN_SECS * 3; // 72h
        assert!(
            remaining_2 >= expected_2 - 5,
            "second billing cycle cooldown should be ~72h, got {remaining_2}s"
        );

        // Clear in-memory cooldown again
        {
            let mut map = cooldowns().lock().unwrap();
            map.remove(agent);
        }

        // Third failure: 24h * 9 = 216h → capped at 7 days (168h)
        record_credit_exhaustion(agent, CreditExhaustionReason::BillingCycleExhausted).await;
        let remaining_3 = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(agent).expect("cooldown entry should exist");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };
        assert!(
            remaining_3 >= BILLING_CYCLE_MAX_SECS - 5,
            "third billing cycle cooldown should be capped at 7 days, got {remaining_3}s"
        );
        assert!(
            remaining_3 <= BILLING_CYCLE_MAX_SECS + 5,
            "third billing cycle cooldown should not exceed 7 days, got {remaining_3}s"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_credit_exhaustion_billing_cycle_stays_capped_after_repeats() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_billing_cycle_agent_capped";

        for i in 0..5 {
            if i > 0 {
                // Each repeat is a new exhaustion after the previous cooldown ended
                cooldowns().lock().unwrap().remove(agent);
            }
            record_credit_exhaustion(agent, CreditExhaustionReason::BillingCycleExhausted).await;
        }

        let remaining = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(agent).expect("cooldown entry should exist");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };
        assert!(
            remaining >= BILLING_CYCLE_MAX_SECS - 5,
            "billing cycle cooldown should stay capped at 7 days, got {remaining}s"
        );
        assert!(
            remaining <= BILLING_CYCLE_MAX_SECS + 5,
            "billing cycle cooldown should not exceed 7 days, got {remaining}s"
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn credit_exhaustion_backoff_constants() {
        // OutOfCredits first failure: 1h base
        assert_eq!(
            compute_backoff(1, CREDIT_BACKOFF_BASE_SECS, CREDIT_BACKOFF_MAX_SECS, true),
            CREDIT_BACKOFF_BASE_SECS
        );
        // OrgLevelDisabled first failure: 2h base
        assert_eq!(
            compute_backoff(1, ORG_BACKOFF_BASE_SECS, CREDIT_BACKOFF_MAX_SECS, true),
            ORG_BACKOFF_BASE_SECS
        );
        // BillingCycleExhausted: 24h base, 7d cap
        assert_eq!(BILLING_CYCLE_COOLDOWN_SECS, 24 * 60 * 60);
        assert_eq!(BILLING_CYCLE_MAX_SECS, 7 * 24 * 60 * 60);
    }

    #[serial(cooldown_state)]
    #[test]
    fn compute_backoff_long_cap_without_extended_stays_capped() {
        assert_eq!(
            compute_backoff(
                3,
                BILLING_CYCLE_COOLDOWN_SECS,
                BILLING_CYCLE_MAX_SECS,
                false
            ),
            BILLING_CYCLE_MAX_SECS,
            "7-day capped paths should stop at the documented max"
        );
        assert_eq!(
            compute_backoff(
                9,
                PERSISTENT_MODEL_BACKOFF_BASE_SECS,
                PERSISTENT_MODEL_BACKOFF_MAX_SECS,
                false,
            ),
            PERSISTENT_MODEL_BACKOFF_MAX_SECS,
            "persistent model failures should not extend past 7 days"
        );
    }

    // ---- Degraded-agent tracking tests ----

    #[serial(cooldown_state)]
    #[test]
    fn mark_and_check_agent_degraded() {
        let agent = "test_degraded_mark";
        assert!(!is_agent_degraded(agent));
        mark_agent_degraded(agent);
        assert!(is_agent_degraded(agent));
        clear_agent_degraded(agent);
        assert!(!is_agent_degraded(agent));
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn sync_from_kv_drops_in_memory_entry_cleared_externally() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_sync_external_clear";
        let future = chrono::Utc::now().timestamp() + 600;

        // Register the store and put a cooldown in both memory and KV.
        init_cooldown_store(Arc::clone(&store)).await;
        set_cooldown_async(agent, future, "test").await;
        assert!(is_agent_in_cooldown(agent));

        // Simulate `orch cooldown clear --all` from a separate CLI process: it
        // can only write KV (to "0"); the running engine's in-memory map is
        // untouched.
        store
            .kv_set(&format!("{KV_PREFIX}{agent}"), "0")
            .await
            .expect("kv_set tombstone");
        assert!(
            is_agent_in_cooldown(agent),
            "in-memory cooldown should still be present before sync"
        );

        sync_from_kv(&store).await;

        assert!(
            !is_agent_in_cooldown(agent),
            "sync_from_kv must drop in-memory entries the KV tombstoned"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn sync_from_kv_closes_github_circuit_when_row_cleared() {
        super::reset_global_state().await;
        let store = test_store().await;
        {
            *github_5xx_circuit_open_lock() = Some(chrono::Utc::now().timestamp() + 600);
            github_5xx_timestamps_lock().push(chrono::Utc::now().timestamp());
        }
        assert!(is_github_circuit_open());

        store
            .kv_set("cooldown:github:5xx", "0")
            .await
            .expect("kv_set tombstone");
        sync_from_kv(&store).await;

        assert!(!is_github_circuit_open());
        assert!(github_5xx_timestamps_lock().is_empty());
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn sync_from_kv_keeps_5xx_window_when_expired_row_and_breaker_closed() {
        super::reset_global_state().await;
        let store = test_store().await;
        let now = chrono::Utc::now().timestamp();
        {
            github_5xx_timestamps_lock().push(now);
            github_5xx_timestamps_lock().push(now);
        }

        store
            .kv_set("cooldown:github:5xx", &(now - 600).to_string())
            .await
            .expect("kv_set past");
        sync_from_kv(&store).await;

        assert!(!is_github_circuit_open());
        assert_eq!(github_5xx_timestamps_lock().len(), 2);

        github_5xx_timestamps_lock().clear();
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn sync_from_kv_keeps_github_circuit_open_for_future_row() {
        super::reset_global_state().await;
        let store = test_store().await;
        let future = chrono::Utc::now().timestamp() + 600;
        *github_5xx_circuit_open_lock() = Some(future);

        store
            .kv_set("cooldown:github:5xx", &future.to_string())
            .await
            .expect("kv_set future");
        sync_from_kv(&store).await;

        assert!(is_github_circuit_open());

        *github_5xx_circuit_open_lock() = None;
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn sync_from_kv_keeps_in_memory_entry_when_kv_row_missing() {
        // Race window: set_cooldown_async writes memory first, then KV. If a
        // sync runs between those two writes, the KV row is missing — we must
        // leave the in-memory entry alone, not drop it.
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_sync_race_window";
        let future = chrono::Utc::now().timestamp() + 600;

        // Simulate the race: memory has the entry, KV does not.
        set_cooldown_in_memory(agent, future, "raced");
        assert!(is_agent_in_cooldown(agent));

        sync_from_kv(&store).await;

        assert!(
            is_agent_in_cooldown(agent),
            "sync_from_kv must NOT drop entries whose KV row is missing — that is the set_cooldown_async race window"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn sync_from_kv_adopts_kv_timestamp() {
        // KV is the source of truth: if KV has a different active timestamp,
        // memory adopts it (covers external manual KV edits).
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_sync_adopt_kv";
        let original = chrono::Utc::now().timestamp() + 600;
        let updated = chrono::Utc::now().timestamp() + 1200;

        init_cooldown_store(Arc::clone(&store)).await;
        set_cooldown_async(agent, original, "initial").await;

        store
            .kv_set(&format!("{KV_PREFIX}{agent}"), &updated.to_string())
            .await
            .expect("kv_set updated");

        sync_from_kv(&store).await;

        let until = cooldown_until(agent).expect("cooldown still active");
        assert_eq!(
            until, updated,
            "sync_from_kv must adopt KV's authoritative timestamp"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn refresh_degraded_agents_marks_cooled_agent() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_refresh_cooled";

        // Put agent in cooldown
        record_agent_failure_with_message(agent, "").await;
        assert!(is_agent_in_cooldown(agent));

        // model_checker always returns true (agent has models)
        let agents = vec![agent.to_string()];
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;

        assert!(
            is_agent_degraded(agent),
            "agent in cooldown should be marked degraded"
        );

        // Cleanup
        clear_agent_degraded(agent);
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn refresh_degraded_agents_marks_no_models_agent() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_refresh_no_models";

        // model_checker returns false (no available models)
        let agents = vec![agent.to_string()];
        refresh_degraded_agents(&store, &agents, &|_| false, 6, 3).await;

        assert!(
            is_agent_degraded(agent),
            "agent with no available models should be marked degraded"
        );

        // Cleanup
        clear_agent_degraded(agent);
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn refresh_degraded_agents_marks_rate_limited_agent() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_refresh_rate_limited";

        // Insert rate limit events exceeding threshold
        for _ in 0..4 {
            store.record_rate_limit(agent, "rate", None).await.unwrap();
        }

        let agents = vec![agent.to_string()];
        // threshold=3, window=6h — 4 events should trigger degraded
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;

        assert!(
            is_agent_degraded(agent),
            "agent with rate limit count >= threshold should be marked degraded"
        );

        // Cleanup
        clear_agent_degraded(agent);
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn refresh_degraded_agents_clears_healthy_agent() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_refresh_healthy";

        // Pre-mark as degraded
        mark_agent_degraded(agent);
        assert!(is_agent_degraded(agent));

        // No cooldown, has models, no rate limit events
        let agents = vec![agent.to_string()];
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;

        assert!(
            !is_agent_degraded(agent),
            "healthy agent should have degraded flag cleared"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn refresh_degraded_agents_ignores_expired_cooldowns() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_refresh_expired_cooldown";

        // Manually insert an already-expired cooldown entry directly into the map
        // (simulating a stale entry that wasn't cleaned up)
        let expired_ts = chrono::Utc::now().timestamp() - 100; // 100 seconds ago
        {
            let mut map = cooldowns().lock().unwrap();
            map.insert(
                agent.to_string(),
                CooldownEntry {
                    cooldown_until: expired_ts,
                    reason: "expired_test".to_string(),
                },
            );
        }

        // Verify the expired entry is in the map but should not cause degraded status
        assert!(
            !is_agent_in_cooldown(agent),
            "expired cooldown should not be active"
        );

        // Pre-mark as degraded to verify it gets cleared
        mark_agent_degraded(agent);
        assert!(is_agent_degraded(agent));

        // Run refresh — should clear degraded because cooldown is expired
        let agents = vec![agent.to_string()];
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;

        assert!(
            !is_agent_degraded(agent),
            "agent with only expired cooldown should NOT be marked degraded"
        );

        // Cleanup
        clear_agent_degraded(agent);
        {
            let mut map = cooldowns().lock().unwrap();
            map.remove(agent);
        }
    }

    #[serial(cooldown_state)]
    #[test]
    fn compute_backoff_grows_exponentially() {
        // count=0 or 1: base
        assert_eq!(compute_backoff(0, 300, 14400, true), 300);
        assert_eq!(compute_backoff(1, 300, 14400, true), 300);
        // count=2: base * 3
        assert_eq!(compute_backoff(2, 300, 14400, true), 900);
        // count=3: base * 9
        assert_eq!(compute_backoff(3, 300, 14400, true), 2700);
        // count=4: base * 27 = 8100 < 14400
        assert_eq!(compute_backoff(4, 300, 14400, true), 8100);
        // count=5: base * 81 = 24300 → capped at 14400
        assert_eq!(compute_backoff(5, 300, 14400, true), 14400);
    }

    #[serial(cooldown_state)]
    #[test]
    fn compute_backoff_extended_tier_1_applies_24h() {
        // With base=300, max=14400:
        // count=4: 8100s (not yet at max)
        // count=5: 14400s (first time at max, times_at_max=0, no extension)
        // count=6: 14400s (times_at_max=1, no extension yet, need >=2)
        // count=7: 86400s (times_at_max=2, extended tier 1: 6x = 24h)
        assert_eq!(
            compute_backoff(5, 300, 14400, true),
            14400,
            "first at max, not extended"
        );
        assert_eq!(
            compute_backoff(6, 300, 14400, true),
            14400,
            "times_at_max=1, not extended"
        );
        assert_eq!(
            compute_backoff(7, 300, 14400, true),
            86400,
            "times_at_max=2, extended 6x"
        );
        assert_eq!(
            compute_backoff(8, 300, 14400, true),
            86400,
            "times_at_max=3, still extended 6x"
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn compute_backoff_extended_tier_2_applies_48h() {
        // With base=300, max=14400:
        // count=10: times_at_max=5 (first_max=5, count=10), extended tier 2: 12x = 48h
        assert_eq!(
            compute_backoff(9, 300, 14400, true),
            86400,
            "times_at_max=4, still tier 1"
        );
        assert_eq!(
            compute_backoff(10, 300, 14400, true),
            172_800,
            "times_at_max=5, extended 12x"
        );
        assert_eq!(
            compute_backoff(15, 300, 14400, true),
            172_800,
            "times_at_max=10, still tier 2"
        );
        assert_eq!(
            compute_backoff(20, 300, 14400, true),
            172_800,
            "times_at_max=15, still tier 2"
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn compute_backoff_extended_tier_with_different_base_max() {
        // With base=60, max=3600:
        // count=3: 60*9=540 < 3600
        // count=4: 60*27=1620 < 3600
        // count=5: 60*81=4860 → capped at 3600
        // count=6: times_at_max=1, still 3600
        // count=7: times_at_max=2, extended 6x = 21600 (6h)
        assert_eq!(compute_backoff(5, 60, 3600, true), 3600, "first at max");
        assert_eq!(compute_backoff(6, 60, 3600, true), 3600, "times_at_max=1");
        assert_eq!(
            compute_backoff(7, 60, 3600, true),
            21600,
            "times_at_max=2, extended 6x"
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn compute_backoff_extended_tier_persistent_fails_example() {
        // Real-world example from issue: glm:haiku with 21 failures
        // With base=300, max=14400:
        // count=21 → times_at_max=16 (first_max=5), extended tier 2: 12x = 48h
        assert_eq!(
            compute_backoff(21, 300, 14400, true),
            172_800,
            "21 failures should get 48h extended cooldown"
        );

        // kimi:haiku with 36 failures
        assert_eq!(
            compute_backoff(36, 300, 14400, true),
            172_800,
            "36 failures should also get 48h extended cooldown"
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn compute_backoff_extended_tier_zero_base() {
        // Zero base should still work without panicking
        assert_eq!(compute_backoff(1, 0, 14400, true), 0);
        assert_eq!(compute_backoff(5, 0, 14400, true), 0);
        assert_eq!(compute_backoff(10, 0, 14400, true), 0);
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_billing_cycle() {
        assert_eq!(
            detect_credit_exhaustion("billing cycle"),
            Some(CreditExhaustionReason::BillingCycleExhausted)
        );
        assert_eq!(
            detect_credit_exhaustion("quota refreshed next cycle"),
            Some(CreditExhaustionReason::BillingCycleExhausted)
        );
        assert_eq!(
            detect_credit_exhaustion(
                "You've reached your usage limit for this billing cycle. Your quota will be refreshed in the next cycle."
            ),
            Some(CreditExhaustionReason::BillingCycleExhausted)
        );
        assert_eq!(
            detect_credit_exhaustion("monthly quota exceeded"),
            Some(CreditExhaustionReason::BillingCycleExhausted)
        );
        assert_eq!(
            detect_credit_exhaustion(
                "Weekly/Monthly Limit Exhausted. Your limit will reset at 2026-04-23 18:33:48"
            ),
            Some(CreditExhaustionReason::BillingCycleExhausted)
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_billing_cycle_minimax() {
        assert_eq!(
            detect_credit_exhaustion(
                "minimax rate limit: API Error: Request rejected (429) · Token Plan usage limit reached: Upgrade your Token Plan or purchase Credits for more usage."
            ),
            Some(CreditExhaustionReason::BillingCycleExhausted)
        );
        assert_eq!(
            detect_credit_exhaustion("Upgrade your Token Plan"),
            Some(CreditExhaustionReason::BillingCycleExhausted)
        );
        assert_eq!(
            detect_credit_exhaustion("Token Plan usage limit reached"),
            Some(CreditExhaustionReason::BillingCycleExhausted)
        );
        assert_eq!(
            detect_credit_exhaustion("purchase Credits for more usage"),
            Some(CreditExhaustionReason::BillingCycleExhausted)
        );
        assert_eq!(
            detect_credit_exhaustion(
                "You've hit your usage limit. Upgrade to Pro (https://chatgpt.com/explore/pro), visit https://chatgpt.com/codex/settings/usage to purchase more credits"
            ),
            Some(CreditExhaustionReason::BillingCycleExhausted)
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_outcome_kimi_billing_cycle() {
        // Regression test for issue #3529: the exact Kimi billing-cycle exhaustion
        // message from task_runs (tasks 156734/156678) must resolve to the
        // "billing_cycle_exhausted" outcome string, not fall through to a generic
        // rate-limit/auth classification.
        let msg = "Failed to authenticate. API Error: 403 You've reached your usage limit for \
                    this billing cycle. Your quota will be refreshed soon.";
        assert_eq!(
            detect_credit_exhaustion_outcome(msg),
            Some("billing_cycle_exhausted")
        );
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_outcome_maps_all_reasons() {
        assert_eq!(
            detect_credit_exhaustion_outcome("out of credits"),
            Some("credit_exhausted")
        );
        assert_eq!(
            detect_credit_exhaustion_outcome("organization has been disabled"),
            Some("org_level_disabled")
        );
        assert_eq!(
            detect_credit_exhaustion_outcome("monthly quota exceeded"),
            Some("billing_cycle_exhausted")
        );
    }

    #[test]
    fn detect_credit_exhaustion_outcome_returns_none_for_unrelated_errors() {
        assert_eq!(
            detect_credit_exhaustion_outcome("429 Too Many Requests"),
            None
        );
        assert_eq!(detect_credit_exhaustion_outcome("connection reset"), None);
    }

    #[serial(cooldown_state)]
    #[test]
    fn detect_credit_exhaustion_overage_takes_priority_over_billing_cycle() {
        // A message containing both overage and billing cycle language should resolve
        // to OrgLevelDisabled/OutOfCredits (detected first) not BillingCycleExhausted.
        let msg = "overageDisabledReason: org_level_disabled for billing cycle";
        assert_eq!(
            detect_credit_exhaustion(msg),
            Some(CreditExhaustionReason::OrgLevelDisabled)
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn refresh_degraded_agents_below_threshold_not_degraded() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_refresh_below_threshold";

        // Insert rate limit events below threshold
        for _ in 0..2 {
            store.record_rate_limit(agent, "rate", None).await.unwrap();
        }

        let agents = vec![agent.to_string()];
        // threshold=3, only 2 events — should NOT be degraded
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;

        assert!(
            !is_agent_degraded(agent),
            "agent with rate limit count < threshold should not be degraded"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn refresh_degraded_agents_ignores_error_limit_type() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_refresh_error_rows";

        for _ in 0..3 {
            store.record_rate_limit(agent, "error", None).await.unwrap();
        }

        let agents = vec![agent.to_string()];
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;

        assert!(
            !is_agent_degraded(agent),
            "three error rows in 6 hours must not mark an agent degraded"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn refresh_degraded_agents_marks_three_rate_rows() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_refresh_three_rate_rows";

        for _ in 0..3 {
            store.record_rate_limit(agent, "rate", None).await.unwrap();
        }

        let agents = vec![agent.to_string()];
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;

        assert!(
            is_agent_degraded(agent),
            "three rate rows in 6 hours must mark an agent degraded"
        );
        assert_eq!(
            agent_degraded_reason(agent).as_deref(),
            Some("rate limit threshold exceeded")
        );
        let listed = list_degraded_agents();
        assert!(
            listed
                .iter()
                .any(|(a, r)| a == agent && r == "rate limit threshold exceeded"),
            "orch cooldown list must show the degraded reason, got {listed:?}"
        );

        clear_agent_degraded(agent);
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn refresh_degraded_agents_logs_reason_change() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_refresh_reason_change";
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        set_agent_cooldown(agent, 600).await;
        let agents = vec![agent.to_string()];
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;
        assert_eq!(
            agent_degraded_reason(agent).as_deref(),
            Some("agent in cooldown")
        );

        clear_cooldown(agent, &store).await;
        let after_clear = sqlite_datetime_offset(60);
        for _ in 0..3 {
            store
                .record_rate_limit_at(agent, "rate", None, &after_clear)
                .await
                .unwrap();
        }
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;
        assert_eq!(
            agent_degraded_reason(agent).as_deref(),
            Some("rate limit threshold exceeded"),
            "degraded reason must update when the hold changes"
        );

        clear_agent_degraded(agent);
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn clear_un_degrades_on_next_health_check() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_clear_un_degrades";
        let before = sqlite_datetime_offset(-30);
        let after = sqlite_datetime_offset(60);

        for _ in 0..3 {
            store
                .record_rate_limit_at(agent, "rate", None, &before)
                .await
                .unwrap();
        }

        let agents = vec![agent.to_string()];
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;
        assert!(
            is_agent_degraded(agent),
            "three rate rows must mark an agent degraded"
        );

        clear_cooldown(agent, &store).await;
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;
        assert!(
            !is_agent_degraded(agent),
            "clear must un-degrade on the next health check"
        );

        store
            .record_rate_limit_at(agent, "rate", None, &after)
            .await
            .unwrap();
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;
        assert!(
            !is_agent_degraded(agent),
            "one rate row after the mark must not degrade again"
        );

        for _ in 0..2 {
            store
                .record_rate_limit_at(agent, "rate", None, &after)
                .await
                .unwrap();
        }
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;
        assert!(
            is_agent_degraded(agent),
            "three rate rows after the mark must degrade again"
        );

        clear_agent_degraded(agent);
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn clear_all_un_degrades_rate_limited_agent() {
        super::reset_global_state().await;
        let store = test_store().await;
        let agent = "test_clear_all_un_degrades";
        let before = sqlite_datetime_offset(-30);

        for _ in 0..3 {
            store
                .record_rate_limit_at(agent, "rate", None, &before)
                .await
                .unwrap();
        }

        let agents = vec![agent.to_string()];
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;
        assert!(is_agent_degraded(agent));

        clear_cooldown("*", &store).await;
        refresh_degraded_agents(&store, &agents, &|_| true, 6, 3).await;
        assert!(
            !is_agent_degraded(agent),
            "clear --all must un-degrade an agent held only by rate rows"
        );
    }

    // ---- Exponential backoff integration tests ----

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn repeated_agent_failures_escalate_backoff() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_escalation_agent";

        // First failure: should be ~5 min (BACKOFF_BASE_SECS)
        record_agent_failure_with_message(agent, "").await;
        let remaining_1 = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(agent).expect("should have cooldown");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };
        assert!(
            (BACKOFF_BASE_SECS - 5..=BACKOFF_BASE_SECS + 5).contains(&remaining_1),
            "first failure should be ~5 min, got {remaining_1}s"
        );

        // Clear the cooldown (but NOT failure count) to allow the next set_cooldown_async to apply
        {
            let mut map = cooldowns().lock().unwrap();
            map.remove(agent);
        }

        // Second failure: should be ~15 min (5 * 3)
        record_agent_failure_with_message(agent, "").await;
        let remaining_2 = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(agent).expect("should have cooldown");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };
        assert!(
            remaining_2 >= BACKOFF_BASE_SECS * 3 - 5,
            "second failure should be ~15 min, got {remaining_2}s"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn vendor_past_retry_falls_back_to_backoff() {
        super::reset_global_state().await;
        // If a vendor-provided "try again at" date is in the past, the
        // code must NOT early-return and must instead apply the normal
        // exponential backoff (increment failure count and set a future cooldown).
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_vendor_past_retry";

        // Ensure no pre-existing cooldown
        {
            let mut map = cooldowns().lock().unwrap();
            map.remove(agent);
        }

        // A clearly past date that parse_retry_at can parse
        let past_msg = "You've hit the limit; try again at Jan 01, 2000 01:00 AM.";

        record_agent_failure_with_message(agent, past_msg).await;

        // After handling a past vendor timestamp, we should have applied
        // exponential backoff (base = BACKOFF_BASE_SECS) and a persisted
        // failure count of 1.
        assert!(
            is_agent_in_cooldown(agent),
            "agent should be in cooldown after failure"
        );

        let remaining = {
            let map = cooldowns().lock().unwrap();
            let entry = map.get(agent).expect("cooldown entry should exist");
            entry.cooldown_until - chrono::Utc::now().timestamp()
        };

        assert!(
            remaining >= BACKOFF_BASE_SECS - 5,
            "expected backoff of ~{BACKOFF_BASE_SECS}s when vendor timestamp is past, got {remaining}s"
        );

        // Verify the failure count was incremented and persisted to KV
        let kv_key = format!("{FAILURE_COUNT_PREFIX}{agent}");
        let stored = store
            .kv_get(&kv_key)
            .await
            .unwrap()
            .expect("failure count should be stored");
        assert_eq!(stored, "1");
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_agent_success_resets_backoff() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_success_reset_agent";
        let model = "test_model";

        // Accumulate 3 failures to escalate the backoff
        for _ in 0..3 {
            let _ = read_and_increment_failure_count(Some(&store), agent).await;
        }
        // Verify count is 3
        let kv_key = format!("{FAILURE_COUNT_PREFIX}{agent}");
        let count: u32 = store
            .kv_get(&kv_key)
            .await
            .unwrap()
            .unwrap()
            .parse()
            .unwrap();
        assert_eq!(count, 3);

        // Success should reset to 0
        record_agent_success(agent, model).await;

        let count_after: u32 = store
            .kv_get(&kv_key)
            .await
            .unwrap()
            .unwrap()
            .parse()
            .unwrap();
        assert_eq!(count_after, 0, "success should reset failure count to 0");

        // Model-specific count should also be reset
        let model_kv_key = format!("{FAILURE_COUNT_PREFIX}{agent}:{model}");
        let model_count: u32 = store
            .kv_get(&model_kv_key)
            .await
            .unwrap()
            .unwrap()
            .parse()
            .unwrap();
        assert_eq!(
            model_count, 0,
            "success should reset model failure count to 0"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn record_agent_success_without_model_skips_empty_model_key() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_success_without_model";

        record_agent_success(agent, "").await;

        let agent_kv_key = format!("{FAILURE_COUNT_PREFIX}{agent}");
        let agent_count = store
            .kv_get(&agent_kv_key)
            .await
            .unwrap()
            .expect("agent failure count should be stored");
        assert_eq!(agent_count, "0");

        let empty_model_kv_key = format!("{FAILURE_COUNT_PREFIX}{agent}:");
        let empty_model_count = store.kv_get(&empty_model_kv_key).await.unwrap();
        assert!(
            empty_model_count.is_none(),
            "success without model should not create an empty model failure count key"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn clear_cooldown_resets_failure_counts() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agent = "test_clear_fc_agent";

        // Accumulate failures
        for _ in 0..5 {
            let _ = read_and_increment_failure_count(Some(&store), agent).await;
        }
        // Set a cooldown so clear_cooldown has something to clear
        set_agent_cooldown(agent, 3600).await;

        // Verify failure count is 5
        let kv_key = format!("{FAILURE_COUNT_PREFIX}{agent}");
        let count: u32 = store
            .kv_get(&kv_key)
            .await
            .unwrap()
            .unwrap()
            .parse()
            .unwrap();
        assert_eq!(count, 5);

        // Clear the cooldown
        clear_cooldown(agent, &store).await;

        // Failure count should be reset to 0
        let count_after: u32 = store
            .kv_get(&kv_key)
            .await
            .unwrap()
            .unwrap()
            .parse()
            .unwrap();
        assert_eq!(
            count_after, 0,
            "clear_cooldown should reset failure count to 0"
        );

        // Cooldown should be removed
        assert!(
            !is_agent_in_cooldown(agent),
            "agent should not be in cooldown after clear"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn clear_all_cooldowns_resets_all_failure_counts() {
        super::reset_global_state().await;
        let store = test_store().await;
        *cooldown_store().lock().await = Some(Arc::clone(&store));

        let agents = ["test_clear_all_a", "test_clear_all_b"];
        for agent in &agents {
            let _ = read_and_increment_failure_count(Some(&store), agent).await;
            set_agent_cooldown(agent, 3600).await;
        }

        clear_cooldown("*", &store).await;

        for agent in &agents {
            assert!(!is_agent_in_cooldown(agent));
            let kv_key = format!("{FAILURE_COUNT_PREFIX}{agent}");
            let count: u32 = store
                .kv_get(&kv_key)
                .await
                .unwrap()
                .unwrap()
                .parse()
                .unwrap();
            assert_eq!(
                count, 0,
                "clear --all should reset failure count for {agent}"
            );
        }
    }

    #[serial(cooldown_state)]
    #[test]
    fn compute_backoff_saturates_on_large_counts() {
        // Very large count should not overflow — extended tier applies 48h
        assert_eq!(
            compute_backoff(100, BACKOFF_BASE_SECS, BACKOFF_MAX_SECS, true),
            BACKOFF_MAX_SECS * 12
        );
        assert_eq!(
            compute_backoff(u64::from(u32::MAX), 300, 14400, true),
            172_800
        );
        assert_eq!(compute_backoff(u64::MAX, 300, 14400, true), 172_800);
    }

    #[serial(cooldown_state)]
    #[test]
    fn compute_backoff_zero_base_returns_zero() {
        assert_eq!(compute_backoff(1, 0, 14400, true), 0);
        assert_eq!(compute_backoff(5, 0, 14400, true), 0);
    }

    // ---- GitHub 5xx circuit breaker tests ----

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn github_circuit_stays_closed_below_threshold() {
        super::reset_global_state().await;
        // Reset circuit breaker state
        {
            let mut open = github_5xx_circuit_open().lock().unwrap();
            *open = None;
            let mut ts = github_5xx_timestamps().lock().unwrap();
            ts.clear();
            // Also clear any generic cooldown map entry for the github 5xx
            // circuit so tests are isolated from other tests that may have
            // set the generic cooldown via set_agent_cooldown("github:5xx", ...).
            let mut map = cooldown_lock();
            map.remove("github:5xx");
        }

        // Record fewer errors than threshold
        for _ in 0..GITHUB_5XX_THRESHOLD - 1 {
            record_github_5xx().await;
        }

        assert!(
            !is_github_circuit_open(),
            "circuit should stay closed below threshold"
        );
        assert_eq!(github_circuit_remaining_secs(), 0);
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn github_circuit_trips_at_threshold() {
        super::reset_global_state().await;
        // Reset circuit breaker state
        {
            let mut open = github_5xx_circuit_open_lock();
            *open = None;
            let mut ts = github_5xx_timestamps_lock();
            ts.clear();
            let mut map = cooldown_lock();
            map.remove("github:5xx");
        }

        // Record exactly threshold errors
        for _ in 0..GITHUB_5XX_THRESHOLD {
            record_github_5xx().await;
        }

        assert!(is_github_circuit_open(), "circuit should trip at threshold");
        assert!(
            github_circuit_remaining_secs() > 0,
            "remaining should be > 0 when open"
        );
        assert!(
            github_circuit_remaining_secs() <= GITHUB_5XX_COOLDOWN_SECS as u64,
            "remaining should not exceed cooldown duration"
        );

        // Clean up
        {
            let mut open = github_5xx_circuit_open().lock().unwrap();
            *open = None;
        }
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn github_circuit_cannot_double_trip() {
        super::reset_global_state().await;
        // Reset circuit breaker state
        {
            let mut open = github_5xx_circuit_open_lock();
            *open = None;
            let mut ts = github_5xx_timestamps_lock();
            ts.clear();
            let mut map = cooldown_lock();
            map.remove("github:5xx");
        }

        // Trip the circuit
        for _ in 0..GITHUB_5XX_THRESHOLD {
            record_github_5xx().await;
        }
        assert!(is_github_circuit_open());

        let first_remaining = github_circuit_remaining_secs();

        // More 5xx errors should NOT reset the cooldown (never-shorten)
        for _ in 0..GITHUB_5XX_THRESHOLD {
            record_github_5xx().await;
        }

        let second_remaining = github_circuit_remaining_secs();
        assert!(
            second_remaining <= first_remaining,
            "additional errors should not extend cooldown"
        );

        // Clean up
        {
            let mut open = github_5xx_circuit_open().lock().unwrap();
            *open = None;
        }
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn github_circuit_recovers_after_cooldown() {
        super::reset_global_state().await;
        // Reset circuit breaker state
        {
            let mut open = github_5xx_circuit_open_lock();
            *open = None;
            let mut ts = github_5xx_timestamps_lock();
            ts.clear();
            let mut map = cooldown_lock();
            map.remove("github:5xx");
        }

        // Set the circuit to an expired cooldown
        {
            let mut open = github_5xx_circuit_open().lock().unwrap();
            *open = Some(chrono::Utc::now().timestamp() - 1);
        }

        // Should be closed now (auto-recovery)
        assert!(
            !is_github_circuit_open(),
            "circuit should auto-recover after cooldown"
        );
    }

    #[serial(cooldown_state)]
    #[tokio::test]
    async fn github_circuit_sliding_window_ages_out() {
        super::reset_global_state().await;
        // Reset circuit breaker state
        {
            let mut open = github_5xx_circuit_open_lock();
            *open = None;
            let mut ts = github_5xx_timestamps_lock();
            ts.clear();
            let mut map = cooldown_lock();
            map.remove("github:5xx");
        }

        // Record errors that are outside the window
        {
            let mut ts = github_5xx_timestamps().lock().unwrap();
            let old = chrono::Utc::now().timestamp() - GITHUB_5XX_WINDOW_SECS - 10;
            for _ in 0..GITHUB_5XX_THRESHOLD {
                ts.push(old);
            }
        }

        // A new error should clear old ones and not trip the circuit
        record_github_5xx().await;

        assert!(
            !is_github_circuit_open(),
            "old errors outside window should not trip circuit"
        );
    }
}