fix(host-cloudflare): una caída breve de D1 ya no hace fallar search ni invoke

.changeset/d1-retry-outage-budget.md

@@ -0,0 +1,5 @@
+---
+"executor": patch
+---
+
+Reads and unreached writes to D1 are retried for up to 20 seconds instead of five attempts in under four seconds, so a D1 outage of about 15 seconds no longer fails `search` and `invoke` with `Internal tool error`.

apps/host-cloudflare/src/db/d1-retry.test.ts

@@ -114,17 +114,39 @@ describe("retryTransientD1Call", () => {
     }),
   );
 
-  it.effect("gives up after five attempts and surfaces the original error", () =>
+  it.effect("rides out a 16 second outage of a read", () =>
     Effect.gen(function* () {
-      const { call, calls } = failingCall(
-        [INCIDENT_ERROR, INCIDENT_ERROR, INCIDENT_ERROR, INCIDENT_ERROR, INCIDENT_ERROR],
-        "rows",
-      );
+      const outage = { down: true };
+      const call = (): Promise<string> =>
+        outage.down
+          ? // oxlint-disable-next-line executor/no-promise-reject -- test double: a D1 call that rejects like the binding does
+            Promise.reject(bindingRejection(INCIDENT_ERROR))
+          : Promise.resolve("rows");
+
+      const fiber = yield* Effect.forkChild(Effect.exit(retryTransientD1Call("all", true, call)));
+      yield* TestClock.adjust(Duration.seconds(16));
+      outage.down = false;
+      yield* TestClock.adjust(Duration.seconds(4));
+      const exit = yield* Fiber.join(fiber);
+
+      expect(exit).toEqual(Exit.succeed("rows"));
+    }),
+  );
+
+  it.effect("gives up once the retry budget is spent and surfaces the original error", () =>
+    Effect.gen(function* () {
+      const calls = { count: 0 };
+      const call = (): Promise<string> => {
+        calls.count += 1;
+        // oxlint-disable-next-line executor/no-promise-reject -- test double: a D1 call that rejects like the binding does
+        return Promise.reject(bindingRejection(INCIDENT_ERROR));
+      };
 
       const exit = yield* runWithVirtualTime(retryTransientD1Call("all", true, call));
 
       expect(exit).toEqual(Exit.fail(new D1CallError({ cause: bindingRejection(INCIDENT_ERROR) })));
-      expect(calls.count).toBe(5);
+      expect(calls.count).toBeGreaterThan(5);
+      expect(calls.count).toBeLessThan(25);
     }),
   );
 

apps/host-cloudflare/src/db/d1-retry.ts

@@ -80,16 +80,18 @@ export const isRetryableD1Failure = (failure: unknown, readOnly: boolean): boole
   return failureClass === "maybe-executed" && readOnly;
 };
 
-export const D1_RETRY_MAX_ATTEMPTS = 5;
+export const D1_RETRY_BUDGET = Duration.seconds(20);
 const D1_RETRY_BASE_DELAY = Duration.millis(200);
 const D1_RETRY_MAX_DELAY = Duration.seconds(2);
 
-/** 200ms, 400ms, 800ms, 1600ms (each jittered to 80-120%, capped at 2s):
- *  five attempts within about 3.6 seconds, so an MCP call still answers. */
+/** 200ms, 400ms, 800ms, 1600ms, then 2s (each jittered to 80-120%), for as
+ *  long as the 20 second budget lasts. D1 outages seen in production last
+ *  about 15 seconds, and most failed attempts return in milliseconds, so a
+ *  count-based limit gave up after under 4 seconds, long before D1 recovered. */
 export const d1RetrySchedule = Schedule.exponential(D1_RETRY_BASE_DELAY).pipe(
   Schedule.modifyDelay((_, delay) => Effect.succeed(Duration.min(delay, D1_RETRY_MAX_DELAY))),
   Schedule.jittered,
-  Schedule.both(Schedule.recurs(D1_RETRY_MAX_ATTEMPTS - 1)),
+  Schedule.both(Schedule.during(D1_RETRY_BUDGET)),
 );
 
 /** A rejection from the D1 binding, kept as-is in `cause`. */
@@ -108,13 +110,12 @@ export const retryTransientD1Call = <A>(
     const current = attempt;
     return Effect.tryPromise({ try: call, catch: (cause) => new D1CallError({ cause }) }).pipe(
       Effect.tapError(({ cause: failure }) =>
-        current < D1_RETRY_MAX_ATTEMPTS && isRetryableD1Failure(failure, readOnly)
+        isRetryableD1Failure(failure, readOnly)
           ? Effect.logWarning("D1 transient failure, retrying").pipe(
               Effect.annotateLogs({
                 "d1.operation": operation,
                 "d1.read_only": readOnly,
                 "d1.attempt": current,
-                "d1.max_attempts": D1_RETRY_MAX_ATTEMPTS,
                 "d1.failure_class": classifyD1Failure(failure),
                 "d1.failure": failureMessages(failure)[0] ?? "unknown",
               }),