@agent-native/core-corpus 0.207.0-nightly-20261011124220 → 0.207.0

package/corpus/README.md

@@ -32,4 +32,4 @@ rg -n "defineAction|useActionQuery" node_modules/@agent-native/core-corpus/corpu
 
 ## Generated Counts
 
-- template files: 11961
+- template files: 11993

package/corpus/templates/analytics/.agents/skills/secrets/SKILL.md

@@ -25,8 +25,12 @@ never copy a value into app code or client-visible state.
 3. If the app truly needs its own API key, register it with the shared secrets
    surface. Do not build a parallel key form or storage table.
 4. Use deployment configuration only for a secret owned by the deployed app,
-   such as a webhook verification secret. A deployment key must not stand in
-   for a user's or customer's provider credential.
+   such as a webhook verification secret. LLM inference keys are an app-owned
+   exception: self-hosted apps may use provider keys from deployment env as a
+   shared fallback, while scoped user or organization keys take precedence.
+   Hosted-workspace deployments backed by a shared database must not share
+   those env keys across apps or tenants. This exception does not authorize
+   using env credentials to access user-owned provider data.
 
 ## Register an app-local API key
 

package/corpus/templates/analytics/CHANGELOG.md

@@ -3,16 +3,73 @@
 All notable user-facing changes to Agent-Native Analytics are documented here. Open it any
 time from the command menu (Cmd+K → "What's new") or from Settings.
 
-## 2026-10-09
+## 2026-10-11
+
+### Improved
+
+- Analytics finds data definitions faster and identifies user-organization relationships more clearly.
+
+## 2026-10-10
 
 ### Improved
 
+- Long session recordings start playback while remaining chunks load.
+- Read only the event fields needed for onboarding journeys.
+- The source index reports tracking calls with event names it cannot resolve, making event coverage gaps easier to spot.
+- The data index preserves dbt's declared primary entity separately from its row grain.
+
+### Fixed
+
+- Analytics discovery preserves source and freshness details when saved dictionary entries contain blank fields.
+- Analytics now shows BigQuery recommendations only when they apply and marks unavailable index counts clearly.
+- Onboarding journey queries reuse a single event read while preserving exact output attribution
+- Reduce duplicate reads when loading organization onboarding journeys
+- Replay friction scores now count rage clicks that span recording uploads.
+- Session summaries count replay errors and failed requests across uploads without double-counting retries
+- Storyboard export failures now report bounded handoff diagnostics without exposing raw exceptions.
+
+### Changed
+
+- Definitions and the source index now live on one Semantic layer page, which replaces the Source Status and Data Dictionary pages.
+
+## 2026-10-09
+
+### Added
+
+- Analytics can query owner-defined dbt metrics through a workspace connection
+- Compare later onboarding activity across first-party sessions and apps, with unknown identities and incomplete observation windows called out.
+- Dashboard folders can now sync with a GitHub repo folder: pull only the panels that changed in GitHub, or export app changes as a pull request, without overwriting edits made elsewhere in the app.
+- Dashboards can filter by several values at once with multi-select filters.
+- Journey capture can render recorded media in Chromium and save visible prompt provenance locally.
+- Review source connections and export the data dictionary from one status page
+
+### Improved
+
+- Analytics prioritizes reviewed definitions over broad generated matches, keeps exact runnable panels discoverable, reports source-index quality accurately, and continues BigQuery searches without repeating metadata reads.
+- Ask chats now open on their own shareable URL at /ask/<id>, and /ask always starts a blank chat.
+- Dashboard filters support searching, selecting all values, or keeping only one value.
+- Data Sources now lists integrations in the same rows as Settings, with connected sources first and setup in a side sheet.
+- Approved metric definitions rank ahead of dashboard examples for exact name searches.
+- Generated source entries now appear in data dictionary browsing, and new entries remain unapproved until reviewed.
+- Journey captures can check replay media and optionally save a private, bounded copy of visible user messages.
+- Onboarding journey event reads avoid an unused identity scan, cap query cost, and mark follow-up coverage incomplete when results are too large
+- Onboarding journeys now honor the selected date range when reading events.
+- Onboarding summaries return journey counts without waiting for replay metadata.
 - Onboarding journeys include Builder connection and custom-key outcomes through deeper paths
 - Onboarding journeys now show later activity within a chosen window, mark no-later results as right-censored, and identify when follow-up counts are incomplete.
 - You can add app origins to a public key's replay allowlist without replacing its existing origins.
 
 ### Fixed
 
+- First-party BigQuery dashboards use date-bounded weekly activity queries on both saved dashboard versions
+- GitHub source connections use dedicated OAuth app credentials, separate from sign-in.
+- Journey screenshots account for legacy clipping when checking embedded replay content.
+- Journey trees remain available when a matching recording is not visible
+- Onboarding journey queries avoid redundant event reads
+- Onboarding journeys can find recordings from exact replay links even when session IDs differ
+- Onboarding journey summaries load reliably for organizations using Builder Analytics.
+- Replay screenshots finish loading fonts before capture.
+- Replay screenshots support modern CSS colors.
 - Existing chat threads keep the composer anchored at the bottom while their messages load.
 - Journey captures now flag visible previews missing from the recording instead of saving blank screenshots
 - Journey recordings flag iframe content when clipping, masks, or filters make its visibility uncertain.
@@ -20,6 +77,10 @@ time from the command menu (Cmd+K → "What's new") or from Settings.
 - Session recordings keep their full history and performance insights across browser session changes and large replays.
 - Large onboarding journeys keep their step counts when follow-up coverage exceeds query limits
 
+### Removed
+
+- Removed the default HTTP 5xx spike alert, since server responses are no longer recorded as analytics events
+
 ## 2026-10-08
 
 ### Added
@@ -85,6 +146,7 @@ time from the command menu (Cmd+K → "What's new") or from Settings.
 
 ### Improved
 
+- Session replays now show the date and time the session started, and the Sessions list shows start times in a shorter, easier-to-scan format.
 - Agent dashboard edits are now checked against the live chart before they save, so the agent no longer reports a chart change that did not happen.
 
 ### Fixed

package/corpus/templates/analytics/actions/find-data.ts

@@ -37,13 +37,24 @@ export default defineAction({
       orgId: getRequestOrgId() || null,
     });
     const { candidates, ...paginationAndCoverage } = result;
+    const sourceIndexGuidance =
+      result.sourceIndexStatus === "not-configured"
+        ? "The generated source index is not configured, so absence from these references is not evidence that a warehouse relation does not exist. Use search-bigquery-schema for connected BigQuery schemas."
+        : result.sourceIndexStatus === "invalid" ||
+            result.sourceIndexStatus === "unavailable"
+          ? "The generated source index could not be read, so treat these references as incomplete and verify with search-bigquery-schema."
+          : "";
     return {
       question,
       results: candidates,
       ...paginationAndCoverage,
       resultType: "reference-metadata",
-      guidance:
+      guidance: [
         "Treat saved SQL and source-index entries as examples or unapproved metadata. Verify source scope, live schema, partitions, and current values before answering. If truncated is true and nextPage is present, continue with that cursor; if truncated is true and nextPage is null, refine the search or use an explicit dashboard/schema search because a source cap was reached.",
+        sourceIndexGuidance,
+      ]
+        .filter(Boolean)
+        .join(" "),
     };
   },
 });

package/corpus/templates/analytics/actions/get-data-dictionary-index-status.ts

@@ -39,7 +39,7 @@ export default defineAction({
       generatedAt: result.bundle.generatedAt,
       entryCount: result.bundle.entries.length,
       unresolvedTrackingCallSites:
-        result.bundle.scanSummary?.unresolvedTrackingCallSites ?? 0,
+        result.bundle.scanSummary?.unresolvedTrackingCallSites ?? null,
       sources: result.bundle.sources,
       sourceCounts: result.bundle.sources.map(({ id }) => ({
         source: id,

package/corpus/templates/analytics/actions/get-first-party-analytics-health.ts

@@ -74,7 +74,7 @@ const healthSchema = z.object({
 
 export default defineAction({
   description:
-    "Check whether the built-in first-party Analytics source is still a good fit for Neon/Postgres. Use this before a large or historical query. It reads compact daily rollups and the small slow-query pressure ledger, never the raw event table. For BigQuery cutovers, it also reports pending delivery count, age, and the last delivery error so a warehouse outage is visible before dashboards go quiet. A pending delivery-queue migration is reported as degraded delivery health instead of hiding the existing health result. `healthy` means Neon is fine, `monitor` means keep Neon but watch growth or delivery lag, and `recommend_bigquery` means use a configured external analytics backend for high-volume or historical queries. The result reports the two supported external options with safe setup links: BigQuery for warehouse SQL and history, and Amplitude for product analytics, funnels, and retention. A recommendation is based on 1M+ observed events, three or more slow queries in 24 hours, or any query timeout/30-second query. Connecting a backend does not silently reroute `/track` events or backfill existing Neon events.",
+    "Check whether the built-in first-party Analytics source is still a good fit for Neon/Postgres. Use this before a large or historical query. It reads compact daily rollups and the small slow-query pressure ledger, never the raw event table. For BigQuery cutovers, it also reports pending delivery count, age, and the last delivery error so a warehouse outage is visible before dashboards go quiet. A pending delivery-queue migration is reported as degraded delivery health instead of hiding the existing health result. `healthy` means Neon is fine, `monitor` means keep Neon but watch query pressure or delivery lag, and `recommend_bigquery` means move high-volume or historical queries to a configured external analytics backend. Event volume, three or more slow queries in 24 hours, or any query timeout/30-second query can trigger that recommendation while the first-party sink is Neon/Postgres. When BigQuery is already the first-party sink, event volume is not a migration reason; query pressure and delivery lag can still produce `monitor`. The result reports the two supported external options with safe setup links: BigQuery for warehouse SQL and history, and Amplitude for product analytics, funnels, and retention. Connecting a backend does not silently reroute `/track` events or backfill existing Neon events.",
   schema: z.object({}),
   outputSchema: healthSchema,
   http: { method: "GET" },

package/corpus/templates/analytics/actions/import-data-dictionary-index.ts

@@ -29,22 +29,17 @@ export default defineAction({
     const index = bundle as SourceIndexBundle;
     await putOrgSetting(admin.orgId, SOURCE_INDEX_SETTING_KEY, index);
     invalidateSourceIndexCache(admin.orgId);
-    const scanSummary = index.scanSummary
-      ? {
-          ...index.scanSummary,
-          unresolvedTrackingCallSites:
-            index.scanSummary.unresolvedTrackingCallSites ?? 0,
-        }
-      : null;
+    const scanSummary = index.scanSummary ?? null;
+    const unresolvedTrackingCallSites =
+      scanSummary?.unresolvedTrackingCallSites ?? null;
     return {
       entryCount: index.entries.length,
       generatedAt: index.generatedAt,
       sources: index.sources.map((source) => source.id),
       scanSummary,
-      unresolvedTrackingCallSites:
-        scanSummary?.unresolvedTrackingCallSites ?? 0,
+      unresolvedTrackingCallSites,
       message: scanSummary
-        ? `Replaced the source index with ${index.entries.length} unapproved entries; omitted ${scanSummary.unsafeEntriesOmitted} unsafe entries and ${scanSummary.unsafeFieldsOmitted} unsafe fields, marked ${scanSummary.truncatedFields} truncated fields, and found ${scanSummary.unresolvedTrackingCallSites} tracking call sites with unresolved event names.`
+        ? `Replaced the source index with ${index.entries.length} unapproved entries; omitted ${scanSummary.unsafeEntriesOmitted} unsafe entries and ${scanSummary.unsafeFieldsOmitted} unsafe fields, marked ${scanSummary.truncatedFields} truncated fields, and ${unresolvedTrackingCallSites === null ? "could not determine the unresolved tracking call count" : `found ${unresolvedTrackingCallSites} tracking call sites with unresolved event names`}.`
         : `Replaced the source index with ${index.entries.length} unapproved entries. Scan quality counts are unavailable for this bundle.`,
     };
   },

package/corpus/templates/analytics/app/components/layout/Sidebar.tsx

@@ -2635,12 +2635,12 @@ export function Sidebar({ mobile }: { mobile?: boolean } = {}) {
               </div>
             </nav>
 
-            <div className="mt-3 shrink-0 min-w-0 space-y-1 border-t border-border/70">
+            <div className="mt-3 shrink-0 min-w-0">
               <AppSidebarFooter
                 collapsed={false}
                 collapsible={false}
                 feedback={footerFeedback}
-                className="space-y-1 border-t-0 px-2"
+                className="border-border/70 px-2"
                 orgSwitcher={
                   <OrgSwitcher className="min-w-0 flex-1 !px-2 !bg-transparent !text-primary hover:!bg-accent/60 hover:!text-primary" />
                 }

package/corpus/templates/analytics/app/components/semantic-layer/IndexPanel.tsx

@@ -116,7 +116,7 @@ type IndexStatusAvailable = {
   status: "available";
   generatedAt: string;
   entryCount: number;
-  unresolvedTrackingCallSites: number;
+  unresolvedTrackingCallSites: number | null;
   sources: Array<{ id: string; revision?: string }>;
   sourceCounts?: Array<{ source: string; entryCount: number }>;
   ageDays: number;
@@ -393,7 +393,11 @@ function StatusSection() {
         {available ? (
           <SettingRow label={t("indexPanel.unresolvedTrackingCallSites")}>
             <span className="text-sm tabular-nums text-muted-foreground">
-              {formatters.formatNumber(available.unresolvedTrackingCallSites)}
+              {available.unresolvedTrackingCallSites === null
+                ? t("indexPanel.unresolvedTrackingCallSitesUnavailable")
+                : formatters.formatNumber(
+                    available.unresolvedTrackingCallSites,
+                  )}
             </span>
           </SettingRow>
         ) : null}

package/corpus/templates/analytics/app/i18n/en-US.ts

@@ -1869,6 +1869,7 @@ export default {
     presetWeekly: "Weekly on Monday at 06:00",
     status: "Status",
     unresolvedTrackingCallSites: "Unresolved tracking calls",
+    unresolvedTrackingCallSitesUnavailable: "Unavailable",
     dynamicTrackingCoverageCaveat:
       "Calls with runtime-computed event names are excluded; aliases or other tracking APIs may be missed.",
     overviewFailed: "Build state could not be loaded.",

package/corpus/templates/analytics/app/pages/DataSources.tsx

@@ -1588,9 +1588,7 @@ function useFirstPartyAnalyticsData() {
     | undefined;
   const healthStatus = healthQuery.isError ? "unavailable" : health?.status;
   const externalBackends = health?.externalBackends ?? [];
-  const externalBackendConfigured = externalBackends.some(
-    (backend) => backend.configured === true,
-  );
+  const externalBackendConfigured = health?.bigQuery.configured === true;
   const recommendsExternalBackend = healthStatus === "recommend_bigquery";
   const healthTitleKey = recommendsExternalBackend
     ? externalBackendConfigured

package/corpus/templates/analytics/changelog/2026-10-10-analytics-now-shows-bigquery-recommendations-only-when-they-.md

@@ -0,0 +1,6 @@
+---
+type: fixed
+date: 2026-10-10
+---
+
+Analytics now shows BigQuery recommendations only when they apply and marks unavailable index counts clearly.

package/corpus/templates/analytics/changelog/2026-10-11-analytics-finds-data-definitions-faster-and-identifies-user-.md

@@ -0,0 +1,6 @@
+---
+type: improved
+date: 2026-10-11
+---
+
+Analytics finds data definitions faster and identifies user-organization relationships more clearly.

package/corpus/templates/analytics/evals/README.md

@@ -10,8 +10,10 @@ The Analytics adapter assembles the production framework prompt, Analytics
 instructions, initial tools, read-only action registry, and request-time catalog
 prefetch. It invokes the shared production `runAgentLoop` with the real Analytics
 final-response guard, caller email, organization id, abort signal, and usage
-capture. The runner checks a bounded receipt for those conditions and rejects
-failed or timed-out prefetch, missing guard/usage, skipped cases, or cancellation.
+capture. The runner checks a bounded receipt for those conditions. A failed or
+timed-out catalog prefetch is reported to the model as a context note, as in
+production; the agent can still search with its tools. Aborted runs, missing
+guard/usage, skipped cases, and incorrect answers fail the eval.
 
 This exercises the shared production agent-loop boundary; it does not dispatch
 through the mounted HTTP chat handler or its thread/run persistence path. The
@@ -22,22 +24,41 @@ read from environment variables or app configuration.
 If either identity value or the production adapter is missing, the command
 exits with an error before running or scoring any case.
 
+## Production path before and after
+
+`results/analytics-smartness-before-after.csv` contains two runs per phase for
+the same redacted synthetic case based on the reported Builder.io user and
+organization question. It is not a replay of Brent's historical chat. The CSV
+contains only run timing, token and tool counts, credits, and scorer outcomes;
+it has no prompts, transcripts, emails, identifiers, row values, or secrets.
+
+Both phases completed through the production agent loop, but both failed the
+source-contract scorer because the production organization had no generated
+source index configured at evaluation time. Mean latency fell from 101.4s to
+83.7s (17.4%), mean LLM calls from 14.5 to 12 (17.2%), and mean input tokens
+from 908k to 855k (5.8%). This establishes a faster lookup path, not an
+accuracy improvement. The generated index now includes the missing membership
+grains; rerun the same cases after the index is imported to measure the
+accuracy gain. No production index or data rows were written for this eval.
+
 ## Synthetic source and grain cases
 
-`production-source-cases.eval.ts` adds six synthetic questions. They ask for
-model selection and declared grain only; they do not use captured prompts,
-people, identifiers, event rows, or production result values. Their scorer
-rejects failed or aborted runs, email-shaped output, and calls outside the
-metadata-only tool set. The CLI still runs each case through this directory's
-production adapter and the shared production agent loop.
+`production-source-cases.eval.ts` adds seven synthetic questions. They ask for
+source selection, declared grain, and relation or field attribution; they do
+not use captured prompts, people, identifiers, event rows, or production result
+values. Their scorer rejects failed or aborted runs, email-shaped output, and
+calls outside the metadata-only tool set. The CLI still runs each case through
+this directory's production adapter and the shared production agent loop.
 
 The expected source contracts are:
 
 - Builder.io user roster: `dbt_mart.dim_users_core` at user grain,
   `dbt_mart.dim_organizations` at organization grain, joined through
   `dbt_intermediate.user_organization_role` at user-organization membership
-  grain. The nonexistent `dbt_mapping.user_id_to_org_id` is not an accepted
-  alternative.
+  grain. `dbt_mapping.user_id_to_org_id` is also a real one-row-per-membership
+  map with `user_id`, `organization_id`, and `is_root_org`; it does not carry
+  `organization_role`, so it cannot replace the role bridge for role-aware
+  rosters.
 - Builder.io current and historical counts: `dbt_mart.organization_user_count`
   is one row per organization per date; `dbt_mart.aggregate_monthly_users_per_org`
   is one row per organization per month.

package/corpus/templates/analytics/evals/production-context.ts

@@ -51,20 +51,32 @@ function toMessages(
   return messages;
 }
 
-function appendReferences(
+export function buildAnalyticsPromptContext(
   systemPrompt: string,
-  candidates: Awaited<
-    ReturnType<typeof retrieveAnalyticsPromptReferences>
-  >["jevPromptCandidates"],
+  references: Awaited<ReturnType<typeof retrieveAnalyticsPromptReferences>>,
 ): string {
-  if (candidates.length === 0) return systemPrompt;
-  const references = candidates
-    .map(
-      (candidate) =>
-        `### ${candidate.name}\n${candidate.description}\n${candidate.content}`,
-    )
-    .join("\n\n");
-  return `${systemPrompt}\n\n<resource scope="analytics-catalog">\n${references}\n</resource>`;
+  const candidates = references.jevPromptCandidates;
+  let prompt = systemPrompt;
+  if (candidates.length > 0) {
+    const resource = candidates
+      .map(
+        (candidate) =>
+          `### ${candidate.name}\n${candidate.description}\n${candidate.content}`,
+      )
+      .join("\n\n");
+    prompt += `\n\n<resource scope="analytics-catalog">\n${resource}\n</resource>`;
+  }
+  if (
+    references.prefetchStatus === "timed_out" ||
+    references.prefetchStatus === "failed"
+  ) {
+    const reason =
+      references.prefetchStatus === "timed_out" ? "timed out" : "failed";
+    const subject =
+      candidates.length > 0 ? "Some preloaded context" : "Preloaded context";
+    prompt += `\n\n<context-note>${subject} was unavailable this turn (${reason}); search for anything relevant with the available tools before answering.</context-note>`;
+  }
+  return prompt;
 }
 
 export function filterAnalyticsEvalActions(
@@ -123,10 +135,7 @@ async function runAnalyticsProductionPath(args: {
     deadlineAt: Date.now() + 1_300,
   });
   const runId = `eval:${crypto.randomUUID()}`;
-  const prompt = appendReferences(
-    args.systemPrompt,
-    references.jevPromptCandidates,
-  );
+  const prompt = buildAnalyticsPromptContext(args.systemPrompt, references);
   const availableTools = actionsToEngineTools(args.actions);
   const initialNames = new Set(args.initialToolNames);
   const tools = availableTools.filter((tool) => initialNames.has(tool.name));
@@ -195,23 +204,6 @@ async function runAnalyticsProductionPath(args: {
     availableActionNames,
     readOnlyActionNames,
   };
-  if (
-    references.prefetchStatus !== "ok" &&
-    references.prefetchStatus !== "empty"
-  ) {
-    return {
-      output: {
-        text: "",
-        toolCalls: [],
-        ok: false,
-        error: `Analytics request preparation ${references.prefetchStatus}.`,
-        runId,
-        durationMs: 0,
-      },
-      receipt,
-    };
-  }
-
   const started = Date.now();
   let ok = true;
   try {