From fb2515649a60ef1e7db214928779147b46f1b5c3 Mon Sep 17 00:00:00 2001
From: Patrick Erichsen <patrick.a.erichsen@gmail.com>
Date: Tue, 11 Aug 2026 11:30:43 -0700
Subject: [PATCH] fix(search): bound rolling usage query batches (#3456)

---
 convex/lib/canonicalSkillSearchBounds.ts |  2 +-
 convex/search.test.ts                    | 54 +++++++++++++++++++++++-
 convex/search.ts                         |  4 +-
 e2e/clawhub.e2e.test.ts                  |  5 ++-
 4 files changed, 58 insertions(+), 7 deletions(-)

diff --git a/convex/lib/canonicalSkillSearchBounds.ts b/convex/lib/canonicalSkillSearchBounds.ts
index dea490c6..95c35e5c 100644
--- a/convex/lib/canonicalSkillSearchBounds.ts
+++ b/convex/lib/canonicalSkillSearchBounds.ts
@@ -5,5 +5,5 @@ export const CANONICAL_SKILL_SEARCH_BOUNDS = {
   externalCandidateLimitPerIndex: 50,
   externalIndexedReadCount: 6,
   rollingAdoptionDays: 60,
-  rollingUsageBatchSize: 40,
+  rollingUsageBatchSize: 20,
 } as const;
diff --git a/convex/search.test.ts b/convex/search.test.ts
index 126d21b8..5e0959b6 100644
--- a/convex/search.test.ts
+++ b/convex/search.test.ts
@@ -1295,12 +1295,12 @@ describe("search helpers", () => {
       getRollingSkillSearchUsageHandler(
         { db: { query: vi.fn() } },
         {
-          skillIds: Array.from({ length: 41 }, (_, index) => `skills:${index}`),
+          skillIds: Array.from({ length: 21 }, (_, index) => `skills:${index}`),
           startDay: 100,
           endDay: 159,
         },
       ),
-    ).rejects.toThrow("skillIds exceeds 40");
+    ).rejects.toThrow("skillIds exceeds 20");
   });
 
   it("returns one ordered native and external contract with canonical routes and install refs", async () => {
@@ -1370,6 +1370,56 @@ describe("search helpers", () => {
     });
   });
 
+  it("preserves rolling adoption ranking when usage reads require smaller transactions", async () => {
+    generateEmbeddingMock.mockRejectedValueOnce(new Error("embedding unavailable"));
+    const native = Array.from({ length: 100 }, (_, index) => ({
+      skill: makePublicSkill({
+        id: `skills:calendar-${index}`,
+        slug: `calendar-${index}`,
+        displayName: `Calendar ${index}`,
+      }),
+      version: null,
+      ownerHandle: "openclaw",
+      owner: null,
+    }));
+    const runQuery = vi.fn(
+      async (ref: Parameters<typeof getFunctionName>[0], args?: { skillIds?: string[] }) => {
+        switch (getFunctionName(ref)) {
+          case "search:getExactSkillSlugMatch":
+          case "search:lexicalFallbackSkills":
+          case "search:getExternalSkillSearchCandidates":
+            return [];
+          case "search:directPrefixSkillMatches":
+            return native;
+          case "search:getRollingSkillSearchUsage": {
+            const skillIds = args?.skillIds ?? [];
+            if (skillIds.length > 20) {
+              throw new Error("Function execution timed out (maximum duration: 1s)");
+            }
+            return skillIds.map((skillId) => ({
+              skillId,
+              installs: skillId === "skills:calendar-99" ? 10_000 : 1,
+              bookmarks: 0,
+            }));
+          }
+          default:
+            throw new Error(`Unexpected query ${getFunctionName(ref)}`);
+        }
+      },
+    );
+
+    const result = await canonicalSearchSkillsHandler(
+      { runQuery, vectorSearch: vi.fn() },
+      { query: "calendar", limit: 100 },
+    );
+
+    expect(result).toHaveLength(100);
+    expect(result[0]).toMatchObject({
+      id: "clawhub:skills:calendar-99",
+      metrics: { rolling60DayInstalls: 10_000 },
+    });
+  });
+
   it("keeps exact mode deterministic across canonical sources", async () => {
     const native = {
       skill: makePublicSkill({
diff --git a/convex/search.ts b/convex/search.ts
index 2d4fbc63..150cc936 100644
--- a/convex/search.ts
+++ b/convex/search.ts
@@ -664,8 +664,8 @@ type CanonicalSkillSearchResult = {
 const CANONICAL_NATIVE_CANDIDATE_LIMIT = CANONICAL_SKILL_SEARCH_BOUNDS.nativeCandidateLimit;
 const CANONICAL_RESULT_LIMIT_MAX = CANONICAL_SKILL_SEARCH_BOUNDS.resultLimit;
 const ROLLING_ADOPTION_DAYS = CANONICAL_SKILL_SEARCH_BOUNDS.rollingAdoptionDays;
-// Forty candidates can read at most 2,400 daily rows, leaving headroom below
-// Convex's per-transaction document/byte limits for imported production-shaped data.
+// Twenty candidates read at most 1,200 daily rows. Keep this well below the
+// query CPU ceiling: production-shaped 40-skill batches had intermittent 1s timeouts.
 const ROLLING_USAGE_QUERY_BATCH_SIZE = CANONICAL_SKILL_SEARCH_BOUNDS.rollingUsageBatchSize;
 
 function chunkValues<T>(values: T[], size: number) {
diff --git a/e2e/clawhub.e2e.test.ts b/e2e/clawhub.e2e.test.ts
index 6670b4df..3948f443 100644
--- a/e2e/clawhub.e2e.test.ts
+++ b/e2e/clawhub.e2e.test.ts
@@ -342,8 +342,9 @@ describe("clawhub e2e", () => {
     const response = await fetchWithTimeout(url.toString(), {
       headers: { Accept: "application/json" },
     });
-    expect(response.ok).toBe(true);
-    const json = (await response.json()) as unknown;
+    const body = await response.text();
+    expect(response.ok, `search failed with ${response.status}: ${body}`).toBe(true);
+    const json = JSON.parse(body) as unknown;
     const parsed = parseArk(ApiV1SearchResponseSchema, json, "API response");
     expect(Array.isArray(parsed.results)).toBe(true);
   });