feat(sse): wire the PROVIDER_PROFILES window gate into the global provider cooldown

providerFailureThreshold / providerFailureWindowMs / providerCooldownMs shipped
in PROVIDER_PROFILES with no runtime consumer (2026-08-31 docs audit, P0.1).
Provider-level entries in providerCooldownTracker now honor them: the whole
provider only counts as cooling after providerFailureThreshold failures inside
providerFailureWindowMs, then cools for providerCooldownMs. Connection-level
entries keep the pre-existing exponential backoff, and the layer stays opt-in
(PROVIDER_COOLDOWN_ENABLED, default off) — default behavior is unchanged.

TDD: tests/unit/provider-cooldown-window-gate.test.ts written first (4 red on
the old behavior), then the wiring; legacy tracker suite aligned to the new
contract (23/23 green). Docs: AGENTS.md breaker section + RESILIENCE_GUIDE
opt-in layer subsection; executors soft-drift refresh (104 -> 106).

AGENTS.md

@@ -131,10 +131,14 @@ breaker runs on `circuitBreakerThreshold` / `circuitBreakerReset`:
 | API key |         `7` |                                 `12` |                         `30s` |
 | Local   |   (derived) |                                  `2` |                         `15s` |
 
-`PROVIDER_PROFILES` also defines `providerFailureThreshold` (10/15/2) and `providerCooldownMs`
-(5min/10min/1min); those fields are loaded into the profile but have **no runtime consumer
-today** — do not tune or document them as the live breaker. Every default is overridable
-through the `OMNIROUTE_PROVIDER_BREAKER_*` and `OMNIROUTE_CIRCUIT_BREAKER_*` env vars; the
+`PROVIDER_PROFILES` also defines `providerFailureThreshold` (10/15/2),
+`providerFailureWindowMs` (15/30/5 min) and `providerCooldownMs` (5/10/1 min): these power the
+**window gate of the opt-in global Provider Cooldown** (`PROVIDER_COOLDOWN_ENABLED`, default
+off) — a provider-level entry in `open-sse/services/providerCooldownTracker.ts` only counts as
+cooling after `providerFailureThreshold` failures inside `providerFailureWindowMs`, and then
+cools for `providerCooldownMs`. They are NOT the live breaker's thresholds — do not tune them
+expecting breaker behavior. Every default is overridable through the
+`OMNIROUTE_PROVIDER_BREAKER_*` and `OMNIROUTE_CIRCUIT_BREAKER_*` env vars; the
 runtime-accurate reference table lives in `docs/architecture/RESILIENCE_GUIDE.md`.
 
 Only provider-level failure statuses should trip the provider breaker:

docs/architecture/ARCHITECTURE.md

@@ -17,7 +17,7 @@ It provides a single OpenAI-compatible endpoint (`/v1/*`) and routes traffic acr
 
 Core capabilities:
 
-- OpenAI-compatible API surface for CLI/tools (352 providers, 104 executors)
+- OpenAI-compatible API surface for CLI/tools (352 providers, 106 executors)
 - Request/response translation across provider formats
 - Model combo fallback (multi-model sequence)
 - Structured combo steps (`provider + model + connection`) with runtime ordering by `compositeTiers`

docs/architecture/CODEBASE_DOCUMENTATION.md

@@ -450,7 +450,7 @@ open-sse/
 ├── types.d.ts
 ├── config/                 Provider registries, header profiles, identity, …
 ├── handlers/               Request handlers (chat, embeddings, audio, image, …)
-├── executors/              104 provider-specific HTTP executors
+├── executors/              106 provider-specific HTTP executors
 ├── translator/             Format conversion (OpenAI ↔ Claude ↔ Gemini ↔ Cursor ↔ Kiro)
 ├── transformer/            Responses API ↔ Chat Completions stream transformer
 ├── services/               80+ service modules (combos, fallback, quotas, identity, …)
@@ -480,7 +480,7 @@ open-sse/
 
 ### 4.2 `open-sse/executors/`
 
-104 provider executors, each extending `BaseExecutor` (`base.ts`):
+106 provider executors, each extending `BaseExecutor` (`base.ts`):
 
 `antigravity`, `azure-openai`, `blackbox-web`, `cliproxyapi`,
 `chatgpt-web-codex`, `cloudflare-ai`, `codex`, `commandCode`, `cursor`, `default`, `devin-cli`,

docs/architecture/RESILIENCE_GUIDE.md

@@ -50,6 +50,25 @@ OmniRoute has three distinct but related resilience mechanisms. Each has a diffe
 
 ---
 
+### Opt-in global Provider Cooldown (window gate)
+
+A fourth, **opt-in** layer (`PROVIDER_COOLDOWN_ENABLED`, default **off**) keeps a
+cross-request memory of failing providers in
+`open-sse/services/providerCooldownTracker.ts`, consulted by combo target
+resolution so consecutive combo requests stop re-walking a provider that just
+failed. Provider-level entries honor the `PROVIDER_PROFILES` window gate:
+
+| Profile | trips after (`providerFailureThreshold`) | inside (`providerFailureWindowMs`) | cools for (`providerCooldownMs`) |
+| ------- | ---------------------------------------: | ---------------------------------: | -------------------------------: |
+| OAuth   |                                     `10` |                            `15min` |                           `5min` |
+| API key |                                     `15` |                            `30min` |                          `10min` |
+
+Below the threshold the provider is **not** considered cooling; a success clears
+the window. Connection-level entries (`provider:connectionId`) keep the
+exponential `minRetryCooldownMs → maxRetryCooldownMs` backoff instead. Overrides:
+`OMNIROUTE_PROVIDER_BREAKER_{OAUTH,API_KEY}_{FAILURE_THRESHOLD,FAILURE_WINDOW_MS,COOLDOWN_MS}`.
+Regression guard: `tests/unit/provider-cooldown-window-gate.test.ts`.
+
 ## 2. Connection Cooldown
 
 **Scope:** single provider connection/account/key.

open-sse/services/providerCooldownTracker.ts

@@ -11,6 +11,8 @@ import {
   DEFAULT_RESILIENCE_SETTINGS,
   type ResilienceSettings,
 } from "../../src/lib/resilience/settings";
+import { PROVIDER_PROFILES } from "../config/constants.ts";
+import { getProviderCategory } from "../config/providerRegistry.ts";
 
 interface CooldownEntry {
   /** Timestamp of last recorded failure (ms since epoch) */
@@ -19,6 +21,47 @@ interface CooldownEntry {
   failureCount: number;
   /** How long this entry must be retained for cleanup purposes */
   retentionMs: number;
+  /**
+   * Provider-level entries only: timestamps of recent failures, pruned to the
+   * profile's `providerFailureWindowMs`. Powers the PROVIDER_PROFILES window
+   * gate (`providerFailureThreshold` failures inside the window trip a
+   * `providerCooldownMs` cooldown for the whole provider).
+   */
+  failureTimestamps?: number[];
+}
+
+// ── PROVIDER_PROFILES window gate (whole-provider scope) ─────────────────────
+// `providerFailureThreshold` / `providerFailureWindowMs` / `providerCooldownMs`
+// shipped in PROVIDER_PROFILES with no runtime consumer (2026-08-31 docs
+// audit, P0.1). Provider-level entries (no connectionId) now honor them: the
+// provider only counts as cooling after `providerFailureThreshold` failures
+// inside `providerFailureWindowMs`, and then cools for `providerCooldownMs`.
+// Connection-level entries keep the pre-existing exponential backoff.
+function providerWindowProfile(provider: string) {
+  const category = getProviderCategory(provider);
+  const profile = PROVIDER_PROFILES[category] ?? PROVIDER_PROFILES.apikey;
+  return {
+    failureThreshold: profile.providerFailureThreshold,
+    failureWindowMs: profile.providerFailureWindowMs,
+    cooldownMs: profile.providerCooldownMs,
+  };
+}
+
+function pruneWindow(timestamps: number[], windowMs: number, now: number): number[] {
+  const cutoff = now - windowMs;
+  const pruned = timestamps.filter((t) => t >= cutoff);
+  // Memory bound: the gate only ever needs `failureThreshold` recent samples;
+  // keep a small multiple so bursts cannot grow the array unbounded.
+  return pruned.length > 200 ? pruned.slice(-200) : pruned;
+}
+
+function providerWindowCooldownMs(provider: string, entry: CooldownEntry, now: number): number {
+  const { failureThreshold, failureWindowMs, cooldownMs } = providerWindowProfile(provider);
+  const inWindow = pruneWindow(entry.failureTimestamps ?? [], failureWindowMs, now);
+  if (inWindow.length < failureThreshold) return 0;
+  const elapsed = now - entry.lastFailureAt;
+  const remaining = cooldownMs - elapsed;
+  return remaining > 0 ? remaining : 0;
 }
 
 // Global cooldown state: keyed by "provider:connectionId" or "provider"
@@ -90,8 +133,21 @@ export function recordProviderCooldown(
     existing.lastFailureAt = now;
     existing.failureCount++;
     existing.retentionMs = Math.max(existing.retentionMs, retentionMs);
+    if (!connectionId) {
+      const { failureWindowMs } = providerWindowProfile(provider);
+      existing.failureTimestamps = pruneWindow(
+        [...(existing.failureTimestamps ?? []), now],
+        failureWindowMs,
+        now
+      );
+    }
   } else {
-    cooldownMap.set(key, { lastFailureAt: now, failureCount: 1, retentionMs });
+    cooldownMap.set(key, {
+      lastFailureAt: now,
+      failureCount: 1,
+      retentionMs,
+      ...(connectionId ? {} : { failureTimestamps: [now] }),
+    });
   }
 
   startCleanupIfNeeded();
@@ -119,6 +175,11 @@ export function isProviderInCooldown(
   if (entry.failureCount === 0) return false;
 
   const now = Date.now();
+
+  if (!connectionId) {
+    return providerWindowCooldownMs(provider, entry, now) > 0;
+  }
+
   const elapsed = now - entry.lastFailureAt;
 
   const minCooldownMs =
@@ -151,6 +212,12 @@ export function getRemainingCooldownMs(
   if (!entry) return 0;
 
   const now = Date.now();
+
+  if (!connectionId) {
+    if (entry.failureCount === 0) return 0;
+    return providerWindowCooldownMs(provider, entry, now);
+  }
+
   const elapsed = now - entry.lastFailureAt;
 
   const minCooldownMs =
@@ -183,8 +250,9 @@ export function recordProviderSuccess(provider: string, connectionId: string | u
   const key = cooldownKey(provider, connectionId);
   const entry = cooldownMap.get(key);
   if (entry) {
-    // Reset failure count but keep the entry
+    // Reset failure count and the provider-level failure window, keep the entry
     entry.failureCount = 0;
+    entry.failureTimestamps = [];
   }
 }
 

tests/unit/provider-cooldown-window-gate.test.ts

@@ -0,0 +1,121 @@
+import { test, beforeEach, mock } from "node:test";
+import assert from "node:assert";
+import {
+  recordProviderCooldown,
+  isProviderInCooldown,
+  getRemainingCooldownMs,
+  recordProviderSuccess,
+  clearCooldownState,
+} from "../../open-sse/services/providerCooldownTracker.ts";
+import { PROVIDER_PROFILES } from "../../open-sse/config/constants.ts";
+import { DEFAULT_RESILIENCE_SETTINGS } from "../../src/lib/resilience/settings.ts";
+
+// Provider-level entries (no connectionId) must honor the PROVIDER_PROFILES
+// window gate: `providerFailureThreshold` failures inside
+// `providerFailureWindowMs` put the whole provider in a `providerCooldownMs`
+// cooldown — below the threshold the provider must NOT be considered cooling.
+// These fields shipped in PROVIDER_PROFILES with no runtime consumer (2026-08-31
+// docs audit, P0.1); this suite is the regression guard for wiring them in.
+// Connection-level entries keep the pre-existing exponential-backoff behavior.
+
+const settings = DEFAULT_RESILIENCE_SETTINGS;
+// "openai" resolves to the apikey category in the provider registry.
+const APIKEY = PROVIDER_PROFILES.apikey;
+
+beforeEach(() => {
+  clearCooldownState();
+});
+
+test("provider-level: below providerFailureThreshold the provider is NOT in cooldown", () => {
+  for (let i = 0; i < APIKEY.providerFailureThreshold - 1; i++) {
+    recordProviderCooldown("openai", undefined, settings);
+  }
+  assert.equal(
+    isProviderInCooldown("openai", undefined, settings),
+    false,
+    `expected no provider-level cooldown below the ${APIKEY.providerFailureThreshold}-failure threshold`
+  );
+  assert.equal(getRemainingCooldownMs("openai", undefined, settings), 0);
+});
+
+test("provider-level: reaching providerFailureThreshold trips a providerCooldownMs cooldown", () => {
+  for (let i = 0; i < APIKEY.providerFailureThreshold; i++) {
+    recordProviderCooldown("openai", undefined, settings);
+  }
+  assert.equal(isProviderInCooldown("openai", undefined, settings), true);
+  const remaining = getRemainingCooldownMs("openai", undefined, settings);
+  assert.ok(
+    remaining > 0 && remaining <= APIKEY.providerCooldownMs,
+    `remaining ${remaining}ms should be within (0, providerCooldownMs=${APIKEY.providerCooldownMs}]`
+  );
+  assert.ok(
+    remaining > APIKEY.providerCooldownMs - 5_000,
+    `a freshly tripped cooldown should last ~providerCooldownMs (got ${remaining}ms)`
+  );
+});
+
+test("provider-level: failures outside providerFailureWindowMs do not count toward the threshold", () => {
+  mock.timers.enable({ apis: ["Date"], now: 1_000_000 });
+  try {
+    // threshold-1 failures, then jump past the window before the next one
+    for (let i = 0; i < APIKEY.providerFailureThreshold - 1; i++) {
+      recordProviderCooldown("openai", undefined, settings);
+    }
+    mock.timers.setTime(1_000_000 + APIKEY.providerFailureWindowMs + 60_000);
+    recordProviderCooldown("openai", undefined, settings);
+    assert.equal(
+      isProviderInCooldown("openai", undefined, settings),
+      false,
+      "stale failures beyond the window must not trip the provider cooldown"
+    );
+  } finally {
+    mock.timers.reset();
+  }
+});
+
+test("provider-level: the cooldown expires providerCooldownMs after the tripping failure", () => {
+  mock.timers.enable({ apis: ["Date"], now: 2_000_000 });
+  try {
+    for (let i = 0; i < APIKEY.providerFailureThreshold; i++) {
+      recordProviderCooldown("openai", undefined, settings);
+    }
+    assert.equal(isProviderInCooldown("openai", undefined, settings), true);
+    mock.timers.setTime(2_000_000 + APIKEY.providerCooldownMs + 1_000);
+    assert.equal(
+      isProviderInCooldown("openai", undefined, settings),
+      false,
+      "provider cooldown must expire after providerCooldownMs"
+    );
+  } finally {
+    mock.timers.reset();
+  }
+});
+
+test("provider-level: recordProviderSuccess clears the failure window", () => {
+  for (let i = 0; i < APIKEY.providerFailureThreshold; i++) {
+    recordProviderCooldown("openai", undefined, settings);
+  }
+  assert.equal(isProviderInCooldown("openai", undefined, settings), true);
+  recordProviderSuccess("openai", undefined);
+  assert.equal(isProviderInCooldown("openai", undefined, settings), false);
+  recordProviderCooldown("openai", undefined, settings);
+  assert.equal(
+    isProviderInCooldown("openai", undefined, settings),
+    false,
+    "one failure after a success must not re-trip the threshold gate"
+  );
+});
+
+test("connection-level entries keep the pre-existing backoff behavior (no window gate)", () => {
+  recordProviderCooldown("openai", "conn-1", settings);
+  assert.equal(
+    isProviderInCooldown("openai", "conn-1", settings),
+    true,
+    "a single connection-level failure still starts the legacy min-cooldown backoff"
+  );
+  const remaining = getRemainingCooldownMs("openai", "conn-1", settings);
+  assert.ok(
+    remaining > 0 && remaining <= settings.providerCooldown.minRetryCooldownMs,
+    `connection-level cooldown should follow minRetryCooldownMs (got ${remaining}ms)`
+  );
+});

tests/unit/services/providerCooldownTracker.test.ts

@@ -9,6 +9,7 @@ import {
   getCooldownEntryCount,
   cleanupExpiredCooldownEntries,
 } from "../../../open-sse/services/providerCooldownTracker.ts";
+import { PROVIDER_PROFILES } from "../../../open-sse/config/constants.ts";
 import {
   resolveResilienceSettings,
   DEFAULT_RESILIENCE_SETTINGS,
@@ -183,8 +184,16 @@ test("different connections have independent cooldowns", () => {
 
 test("provider-only key works without connectionId", () => {
   const settings = makeSettings();
+  // Provider-level entries honor the PROVIDER_PROFILES window gate: a single
+  // failure no longer cools the whole provider (2026-08-31 audit, P0.1 wiring).
   recordProviderCooldown("openai", undefined, settings);
+  assert.equal(isProviderInCooldown("openai", undefined, settings), false);
 
+  // Reaching the profile threshold trips the whole-provider cooldown, still
+  // independent from any connection-level key.
+  for (let i = 1; i < PROVIDER_PROFILES.apikey.providerFailureThreshold; i++) {
+    recordProviderCooldown("openai", undefined, settings);
+  }
   assert.ok(isProviderInCooldown("openai", undefined, settings));
   assert.equal(isProviderInCooldown("openai", "conn-1", settings), false);
 });