Skip to content

Supercharge your AI agents with data from the web and beyond. Building the library for superintelligence. πŸ”₯

Update run.ts

Developers Digestc6f920a8f07d
diff --git a/apps/api/src/services/monitoring/search/run.ts b/apps/api/src/services/monitoring/search/run.ts
index 0f16e3430..2aa56a26c 100644
--- a/apps/api/src/services/monitoring/search/run.ts
+++ b/apps/api/src/services/monitoring/search/run.ts
@@ -39,8 +39,7 @@ import {
} from "./verify";
import { hasGeminiKey } from "./tuning";
-// Flat credits per result the judge evaluates (scraped + judged in deep mode).
-// Deterministic and known at check time β€” no token/at-cost math.
+// Flat, deterministic credits per result the judge evaluates.
export const JUDGE_CREDITS_PER_RESULT = 5;
function windowToTbs(window: string): string {
@@ -49,9 +48,8 @@ function windowToTbs(window: string): string {
return "qdr:w";
}
-// The provider rate-limits near-simultaneous identical queries and may answer
-// with a clean HTTP 200 empty body. Retry empty responses with backoff; a
-// genuinely-empty query just exhausts attempts and returns empty cleanly.
+// Retry empty responses with backoff (the provider rate-limits near-identical
+// queries with a 200 + empty body); a truly-empty query just exhausts attempts.
const SEARCH_RETRY_BACKOFF_MS = [400, 1200] as const;
async function searchWithRetry(
@@ -125,8 +123,7 @@ export type KnownPage = {
goalVersion: string;
lastCheckedAt?: string;
lastStatus?: string;
- // Carried through so unchanged-page upserts don't wipe event/verdict fields
- // persisted by earlier runs.
+ // Carried through so unchanged-page upserts don't wipe earlier event/verdict fields.
metadata?: Record<string, unknown>;
};
@@ -151,19 +148,12 @@ type SearchTargetRunResult = {
summary: string;
// Reserved provider-level signal; always false (see assembly below).
searchDegraded: boolean;
- // True when the deep path was expected to judge results but judged ~none
- // because every per-result scrape failed β€” otherwise indistinguishable from
- // a clean "nothing new".
+ // Deep path found candidates but judged none (every scrape failed).
judgeDegraded: boolean;
- // Explanation surfaced onto the check when judgeDegraded is set.
degradedReason: string | null;
sources: SearchSource[];
- // Search() call credits (β‰ˆ2 per 10 results), matching executeSearch's math.
searchCredits: number;
- // Flat judge billing: JUDGE_CREDITS_PER_RESULT per result the judge evaluated
- // (covers both the scrape and the judge). Zero in raw mode.
judgeCredits: number;
- // Denominator for judgeCredits; carried for observability / estimate parity.
resultsJudged: number;
pageUpserts: Array<{
url: string;
@@ -186,8 +176,7 @@ export async function runSearchTarget(params: {
teamId: string;
goal: string | null;
subject: string;
- // When false, every LLM judge stage is skipped and the target behaves like
- // depth:"raw" (deterministic URLs + dedup, no LLM credits).
+ // When false, every LLM judge stage is skipped (behaves like depth:"raw").
judgeEnabled: boolean;
};
target: SearchTargetInput;
@@ -217,15 +206,11 @@ export async function runSearchTarget(params: {
let pagesChecked = 0;
let skipped = 0;
let matches = 0;
- // For observability only β€” billing does not read from it, but the judge
- // helpers require a CostTracking argument.
+ // Required by the judge helpers; billing does not read from it.
const costTracking = new CostTracking();
- // Search calls bill β‰ˆ2 credits / 10 results, mirroring executeSearch().
let searchResultsBilled = 0;
- // Denominator for flat judge billing: judgeCredits = 5 * this.
let resultsJudged = 0;
- // Deep-path per-result failures (scrape threw / !success / unparseable verdict)
- // that `continue` before resultsJudged increments β€” drives judgeDegraded.
+ // Per-result failures that `continue` before resultsJudged increments.
let judgeScrapeFailures = 0;
const seenThisRun = new Map<string, number>();
@@ -868,24 +853,10 @@ export async function runSearchTarget(params: {
}
}
- // The public search() helper swallows provider errors and returns an empty
- // result, so the monitor can't distinguish a real provider failure from a
- // genuine no-results without reaching into core search. We deliberately don't:
- // empty-after-retries is treated as a clean no-results. The field is kept for
- // the check shape (and the deep-path judge-degraded signal below still fires).
const searchDegraded = false;
- // Deep-path degraded signal: judging WAS expected (depth deep + search
- // returned candidates) but produced ~nothing (resultsJudged === 0) BECAUSE the
- // per-result scrapes failed (judgeScrapeFailures > 0). This is the conservative
- // trigger β€” it is deliberately NOT set when:
- // - search legitimately returned 0 results (judgeExpected is false, no
- // candidates), or
- // - the judge legitimately evaluated results and ignored them all
- // (resultsJudged > 0), or
- // - only SOME scrapes failed but at least one was judged (resultsJudged > 0).
- // Without this, a deep check whose every scrape timed out completes "clean"
- // with 0 pages / 0 judged / 0 alerts β€” indistinguishable from "nothing new".
+ // A deep check that found candidates but judged none (every scrape failed)
+ // would otherwise look identical to a clean "nothing new".
const judgeExpected = depth === "deep" && candidates.length > 0;
const judgeDegraded =
judgeExpected && resultsJudged === 0 && judgeScrapeFailures > 0;
@@ -906,8 +877,6 @@ export async function runSearchTarget(params: {
}
}
- // Single reason surfaced onto the check when the deep path could not evaluate
- // the results it found.
const degradedReason: string | null = judgeDegraded
? "deep-path scrapes failed for every candidate so nothing was judged; results may be incomplete"
: null;