Merge remote-tracking branch 'origin/release/v3.8.50' into fix/release-v3.8.50-basereds

`assertValidEngine()` valida id, apply, compress, getConfigSchema e
validateConfig — não exige `metadata`. Uma engine registrada sem esse campo é,
portanto, um registro legal. Mas `canRunAtCompressionStage` lia
`engine.metadata.executionStages` sem guarda, então essa engine legal derrubava o
pipeline inteiro com `TypeError: Cannot read properties of undefined` em vez de
falhar aberto, que é o contrato da compressão.

Metadata ausente é o mesmo caso de "não declarou executionStages" e passa a cair
no mesmo fallback documentado: só pre-translation.

Isso destravava também `tests/unit/compression/pipeline-circuit-breaker.test.ts`,
que registra uma engine de teste sem metadata e vinha 8/9 na base — agora 9/9. O
teste novo torna o contrato explícito, em vez de deixá-lo dependendo de uma
reprodução incidental noutro arquivo.

changelog.d/fixes/10313-catalog-cache-key-hash.md

@@ -1 +0,0 @@
-- fix(api): hash the API key before using it as the model-catalog cache Map key (no raw credentials in process heap) (#10313)

changelog.d/fixes/9147-catalog-eventloop-yield.md

@@ -1 +0,0 @@
-- fix(api): yield the event loop during catalog builds and bulk-load override/hidden-model tables (#9147)

open-sse/services/compression/strategySelector.ts

@@ -800,7 +800,11 @@ function canRunAtCompressionStage(
   stage: CompressionStage | undefined
 ): boolean {
   const effectiveStage = stage ?? "pre-translation";
-  const stages = engine.metadata.executionStages;
+  // `assertValidEngine` não exige `metadata`, então uma engine registrada sem
+  // esse campo é legal — e sem a guarda derrubava o pipeline inteiro com
+  // TypeError em vez de falhar aberto. Metadata ausente é o mesmo caso de "não
+  // declarou estágio" e cai no mesmo fallback: só pre-translation.
+  const stages = engine.metadata?.executionStages;
   return stages ? stages.includes(effectiveStage) : effectiveStage === "pre-translation";
 }
 

src/app/api/v1/models/catalog.ts

@@ -6,9 +6,9 @@ import {
   getAllCustomModels,
   getSettings,
   getCachedProviderNodes,
+  getModelIsHidden,
   getModelAliases,
   getDatabaseSettings,
-  getHiddenModelsByProvider,
 } from "@/lib/localDb";
 import { createLazyConnectionView } from "@/lib/db/providers/lazyConnectionView";
 import { extractAliasBackedModels } from "./aliasBackedModels";
@@ -133,7 +133,7 @@ export {
 } from "./catalogCache";
 export type { CachedCatalog } from "./catalogCache";
 
-const BUILTIN_AUTO_YIELD_INTERVAL = 2;
+const BUILTIN_AUTO_YIELD_INTERVAL = 8;
 
 function yieldCatalogBuildTurn(): Promise<void> {
   return new Promise((resolve) => setImmediate(resolve));
@@ -157,8 +157,6 @@ export async function getUnifiedModelsResponse(
     try {
       settingsForAuth = await getSettings();
     } catch {}
-    // #9147: yield before auth check to allow event loop tick
-    await yieldCatalogBuildTurn();
     const authRejection = await getModelCatalogAuthRejection(request, settingsForAuth, {
       ...corsHeaders,
       ...diagnosticHeaders,
@@ -229,34 +227,7 @@ async function buildUnifiedModelsResponseCore(
   corsHeaders: Record<string, string> = {}
 ) {
   const diagnosticHeaders = getCatalogDiagnosticsHeaders({ request });
-  // #9147: this builder walks connections + model registries at catalog scale with no
-  // event-loop yield, so a large deployment pins the single Node.js thread for the
-  // whole build (reporter: 183 connections / 2000+ models → 10.1s stall that blocks the
-  // dashboard WS heartbeat). Yield every `catYIELD_EVERY` items across the hot loops.
-  const catYIELD_EVERY = 20;
-  let catYieldCount = 0;
-  const maybeYieldCatalogBuild = async (): Promise<void> => {
-    catYieldCount++;
-    if (catYieldCount % catYIELD_EVERY === 0) {
-      await yieldCatalogBuildTurn();
-    }
-  };
   try {
-    // #9147: `getModelIsHidden()` is a SQLite read per call (custom row + compat list)
-    // and the build consults it ~16× per entry. Bulk-load the hidden-model map once
-    // (one query — `getHiddenModelsByProvider`) and resolve from memory for the whole
-    // build. A provider absent from the map has no hidden models at all — `false`,
-    // no on-demand fallback (that would reintroduce the per-call SQLite reads).
-    // Deliberately kept INSIDE this try block (not hoisted above it): the builder's
-    // own catch below is what converts a build-time failure into a sanitized 500
-    // Response instead of a rejected promise — hoisting this bulk read above the
-    // try would let a crash here propagate as an unhandled rejection instead
-    // (catalogCache.ts's in-flight coalescing does not fully consume rejections).
-    const hiddenModelsByProvider = getHiddenModelsByProvider();
-    const isModelHiddenBulk = (providerId: string, modelId: string): boolean => {
-      const hiddenSet = hiddenModelsByProvider.get(providerId);
-      return hiddenSet ? hiddenSet.has(modelId) : false;
-    };
     let settings: Record<string, any> = {};
     try {
       settings = await getSettings();
@@ -267,10 +238,6 @@ async function buildUnifiedModelsResponseCore(
       ...diagnosticHeaders,
     });
     if (authRejection) return authRejection;
-
-    // #9147: yield after auth check before DB initialization prologue
-    await yieldCatalogBuildTurn();
-
     const { aliasToProviderId, providerIdToAlias } = buildAliasMaps();
     const _qp = new URL(request.url).searchParams.get("prefix");
     const prefixMode =
@@ -355,7 +322,6 @@ async function buildUnifiedModelsResponseCore(
 
     // Get combos
     let combos = [];
-    await yieldCatalogBuildTurn();
     try {
       combos = await getCombos();
     } catch (e) {
@@ -389,16 +355,7 @@ async function buildUnifiedModelsResponseCore(
       if ("alias" in p && typeof p.alias === "string") activeAliases.add(p.alias);
     }
 
-    // #9147 follow-up: this is called ~1-3x per model at catalog scale (providerSupportsModel,
-    // isExcludedByProviderConnections). Connections do not change mid-build, so memoize per
-    // unique (unordered) key-set instead of rescanning connectionsByProvider on every call —
-    // otherwise the O(models) hot loop regains an O(connections) cost per model and blows the
-    // single-stretch event-loop budget this file's own yield mechanism is meant to protect.
-    const connectionsForProviderCache = new Map<string, typeof connections>();
     const getConnectionsForProvider = (...keys: Array<string | null | undefined>) => {
-      const cacheKey = keys.filter((k): k is string => Boolean(k)).sort().join("");
-      const cached = connectionsForProviderCache.get(cacheKey);
-      if (cached) return cached;
       const seen = new Set<string>();
       const collected: typeof connections = [];
       for (const key of keys) {
@@ -409,7 +366,6 @@ async function buildUnifiedModelsResponseCore(
           collected.push(connection);
         }
       }
-      connectionsForProviderCache.set(cacheKey, collected);
       return collected;
     };
 
@@ -630,7 +586,7 @@ async function buildUnifiedModelsResponseCore(
           timestamp,
           (c) => buildComboCatalogMetadata(c, combos)
         );
-        const quotaFinal = await applyCatalogPostFilters(request, quotaModels, {
+        const quotaFinal = applyCatalogPostFilters(request, quotaModels, {
           connections,
           prefixMode,
           aliasToProviderId,
@@ -725,7 +681,7 @@ async function buildUnifiedModelsResponseCore(
       ) as ComboCatalogTarget[];
       const visibleTargets = comboTargets.filter((target) => {
         const resolved = getComboTargetModelId(target);
-        return resolved ? !isModelHiddenBulk(resolved.providerId, resolved.modelId) : true;
+        return resolved ? !getModelIsHidden(resolved.providerId, resolved.modelId) : true;
       });
       if (visibleTargets.length === 0) continue;
 
@@ -742,16 +698,11 @@ async function buildUnifiedModelsResponseCore(
         parent: null,
         ...comboMetadata,
       });
-
-      // #9147: combos can number hundreds at catalog scale — yield periodically.
-      await maybeYieldCatalogBuild();
     }
 
     let syncedModelsByProvider: Record<string, SyncedAvailableModel[]> = {};
     try {
-      await yieldCatalogBuildTurn();
       syncedModelsByProvider = await getAllActiveSyncedModels();
-      await yieldCatalogBuildTurn();
     } catch (e) {
       // DB unavailable — log and fall through; static models remain as defaults.
       console.log("[catalog] Could not fetch synced available models:", e);
@@ -839,7 +790,7 @@ async function buildUnifiedModelsResponseCore(
         if (!isModelSelectable(canonicalProviderId, model.id)) continue;
         if (!providerSupportsModel(canonicalProviderId, model.id)) continue;
         const aliasId = `${alias}/${model.id}`;
-        if (isModelHiddenBulk(canonicalProviderId, model.id)) continue;
+        if (getModelIsHidden(canonicalProviderId, model.id)) continue;
         if (isExcludedByProviderConnections(canonicalProviderId, model.id)) continue;
         if (shouldHidePaid(canonicalProviderId, model.id, (model as { pricing?: unknown }).pricing))
           continue;
@@ -894,15 +845,12 @@ async function buildUnifiedModelsResponseCore(
             ...thinkingCapabilities,
           });
         }
-
-        // #9147: static model walk is the densest loop — yield periodically.
-        await maybeYieldCatalogBuild();
       }
     }
 
     for (const modelId of CODEX_NATIVE_UNPREFIXED_MODELS) {
       if (!providerSupportsModel("codex", modelId)) continue;
-      if (isModelHiddenBulk("codex", modelId)) continue;
+      if (getModelIsHidden("codex", modelId)) continue;
 
       const alias = providerIdToAlias.codex || "cx";
       const aliasId = `${alias}/${modelId}`;
@@ -963,7 +911,7 @@ async function buildUnifiedModelsResponseCore(
           if (canonicalProviderId === "codex" && isCodexDiscoveryModelExcluded(sm)) {
             continue;
           }
-          if (isModelHiddenBulk(providerId, sm.id)) continue;
+          if (getModelIsHidden(providerId, sm.id)) continue;
           if (isExcludedByProviderConnections(canonicalProviderId, sm.id)) continue;
           // #6457: some upstream discovery catalogs (e.g. HuggingFace's live
           // `/v1/models`) return image/diffusion models with no modality info,
@@ -1076,9 +1024,6 @@ async function buildUnifiedModelsResponseCore(
               });
             }
           }
-
-          // #9147: synced-model union is usually the largest walk — yield periodically.
-          await maybeYieldCatalogBuild();
         }
       }
     } catch (err) {
@@ -1108,7 +1053,7 @@ async function buildUnifiedModelsResponseCore(
           if (hidePaid && !isFree) continue;
           // #9293: respect per-model hidden flags (e.g. operator hid google/chirp-3
           // from the OpenRouter provider, so it should not appear in the live catalog).
-          if (isModelHiddenBulk("openrouter", openRouterModel.id)) continue;
+          if (getModelIsHidden("openrouter", openRouterModel.id)) continue;
           const supportedParameters = Array.isArray(openRouterModel.supported_parameters)
             ? openRouterModel.supported_parameters
             : [];
@@ -1149,9 +1094,6 @@ async function buildUnifiedModelsResponseCore(
             ...(outputModalities.length > 0 ? { output_modalities: outputModalities } : {}),
             ...(Object.keys(capabilities).length > 0 ? { capabilities } : {}),
           });
-
-          // #9147: OpenRouter catalog can be large — yield periodically.
-          await maybeYieldCatalogBuild();
         }
       } catch (err) {
         console.error("[catalog] Error loading OpenRouter catalog:", err);
@@ -1196,7 +1138,7 @@ async function buildUnifiedModelsResponseCore(
 
     // Helper: strip the provider prefix from a specialty model ID to get the
     // provider-relative path (e.g. "openrouter/google/chirp-3" -> "google/chirp-3").
-    // This is the correct key used by the hidden-model lookup — using .split("/").pop()
+    // This is the correct key used by getModelIsHidden() — using .split("/").pop()
     // here would discard all but the last segment and miss stored flags for
     // providers whose model IDs carry a sub-path (e.g. OpenRouter scoped models).
     const getSpecialtyModelRelativeId = (modelId: string, provider: string): string =>
@@ -1207,7 +1149,7 @@ async function buildUnifiedModelsResponseCore(
       if (!isProviderActive(embModel.provider)) continue;
       const rawModelId = getSpecialtyModelRelativeId(embModel.id, embModel.provider);
       if (!providerSupportsModel(embModel.provider, rawModelId)) continue;
-      if (isModelHiddenBulk(embModel.provider, rawModelId)) continue;
+      if (getModelIsHidden(embModel.provider, rawModelId)) continue;
       if (hasEquivalentSpecialtyModel(embModel.provider, rawModelId, "embedding", embModel.id)) {
         continue;
       }
@@ -1235,7 +1177,7 @@ async function buildUnifiedModelsResponseCore(
       if (!isProviderActive(imgModel.provider)) continue;
       const rawModelId = getSpecialtyModelRelativeId(imgModel.id, imgModel.provider);
       if (!providerSupportsModel(imgModel.provider, rawModelId)) continue;
-      if (isModelHiddenBulk(imgModel.provider, rawModelId)) continue;
+      if (getModelIsHidden(imgModel.provider, rawModelId)) continue;
       models.push({
         id: imgModel.id,
         object: "model",
@@ -1255,7 +1197,7 @@ async function buildUnifiedModelsResponseCore(
       if (!isProviderActive(rerankModel.provider)) continue;
       const rawModelId = getSpecialtyModelRelativeId(rerankModel.id, rerankModel.provider);
       if (!providerSupportsModel(rerankModel.provider, rawModelId)) continue;
-      if (isModelHiddenBulk(rerankModel.provider, rawModelId)) continue;
+      if (getModelIsHidden(rerankModel.provider, rawModelId)) continue;
       if (hasEquivalentSpecialtyModel(rerankModel.provider, rawModelId, "rerank", rerankModel.id)) {
         continue;
       }
@@ -1274,7 +1216,7 @@ async function buildUnifiedModelsResponseCore(
       if (!isProviderActive(audioModel.provider)) continue;
       const rawModelId = getSpecialtyModelRelativeId(audioModel.id, audioModel.provider);
       if (!providerSupportsModel(audioModel.provider, rawModelId)) continue;
-      if (isModelHiddenBulk(audioModel.provider, rawModelId)) continue;
+      if (getModelIsHidden(audioModel.provider, rawModelId)) continue;
       models.push({
         id: audioModel.id,
         object: "model",
@@ -1290,7 +1232,7 @@ async function buildUnifiedModelsResponseCore(
       if (!isProviderActive(modModel.provider)) continue;
       const rawModelId = getSpecialtyModelRelativeId(modModel.id, modModel.provider);
       if (!providerSupportsModel(modModel.provider, rawModelId)) continue;
-      if (isModelHiddenBulk(modModel.provider, rawModelId)) continue;
+      if (getModelIsHidden(modModel.provider, rawModelId)) continue;
       models.push({
         id: modModel.id,
         object: "model",
@@ -1305,7 +1247,7 @@ async function buildUnifiedModelsResponseCore(
       if (!isProviderActive(videoModel.provider)) continue;
       const rawModelId = getSpecialtyModelRelativeId(videoModel.id, videoModel.provider);
       if (!providerSupportsModel(videoModel.provider, rawModelId)) continue;
-      if (isModelHiddenBulk(videoModel.provider, rawModelId)) continue;
+      if (getModelIsHidden(videoModel.provider, rawModelId)) continue;
       models.push({
         id: videoModel.id,
         object: "model",
@@ -1326,7 +1268,7 @@ async function buildUnifiedModelsResponseCore(
       if (!isProviderActive(musicModel.provider)) continue;
       const rawModelId = getSpecialtyModelRelativeId(musicModel.id, musicModel.provider);
       if (!providerSupportsModel(musicModel.provider, rawModelId)) continue;
-      if (isModelHiddenBulk(musicModel.provider, rawModelId)) continue;
+      if (getModelIsHidden(musicModel.provider, rawModelId)) continue;
       models.push({
         id: musicModel.id,
         object: "model",
@@ -1372,7 +1314,7 @@ async function buildUnifiedModelsResponseCore(
           if (!isUnifiedChatSourceModelSelectable(canonicalProviderId, { ...model, id: modelId }))
             continue;
           if (model.isHidden === true) continue;
-          if (isModelHiddenBulk(canonicalProviderId, modelId)) continue;
+          if (getModelIsHidden(canonicalProviderId, modelId)) continue;
           if (isExcludedByProviderConnections(canonicalProviderId, modelId)) continue;
           // #6328: apply hidePaidModels to user-defined custom rows too.
           // Custom entries do not carry pricing, so shouldHidePaid() decides
@@ -1508,9 +1450,6 @@ async function buildUnifiedModelsResponseCore(
               ...(providerVisionFields || {}),
             });
           }
-
-          // #9147: custom-model walk — yield periodically.
-          await maybeYieldCatalogBuild();
         }
       }
     } catch (e) {
@@ -1556,7 +1495,7 @@ async function buildUnifiedModelsResponseCore(
           continue;
         }
 
-        if (isModelHiddenBulk(canonicalProviderId, modelId)) continue;
+        if (getModelIsHidden(canonicalProviderId, modelId)) continue;
         if (isExcludedByProviderConnections(canonicalProviderId, modelId)) continue;
         // #6328: apply hidePaidModels to alias-backed rows too. Alias mappings
         // point at providerKey/modelId with no pricing, so shouldHidePaid()
@@ -1630,7 +1569,7 @@ async function buildUnifiedModelsResponseCore(
       for (const model of fallbackModels) {
         const modelId = typeof model.id === "string" ? model.id : null;
         if (!modelId) continue;
-        if (isModelHiddenBulk(canonicalProviderId, modelId)) continue;
+        if (getModelIsHidden(canonicalProviderId, modelId)) continue;
         if (isExcludedByProviderConnections(canonicalProviderId, modelId)) continue;
         // #6328: apply hidePaidModels to managed-fallback rows too. Compatible
         // provider fallbacks lack pricing; shouldHidePaid() decides via the
@@ -1658,9 +1597,6 @@ async function buildUnifiedModelsResponseCore(
           ...(contextLength ? { context_length: contextLength } : {}),
           ...(visionFields || {}),
         });
-
-        // #9147: per-connection fallback walk — yield periodically.
-        await maybeYieldCatalogBuild();
       }
     }
 
@@ -1706,7 +1642,7 @@ async function buildUnifiedModelsResponseCore(
       }
     }
     // ?configuredOnly — hide models that have no eligible DB connection.
-    finalModels = await applyCatalogPostFilters(request, finalModels, {
+    finalModels = applyCatalogPostFilters(request, finalModels, {
       connections,
       prefixMode,
       aliasToProviderId,

src/app/api/v1/models/catalogResponse.ts

@@ -32,7 +32,6 @@ import {
   enrichCatalogModelEntry,
   type CatalogEnrichmentSnapshot,
 } from "@/lib/modelMetadataRegistry";
-import { createModelCapabilityResolutionSnapshot } from "@/lib/modelCapabilityResolutionSnapshot";
 import { isModelCatalogNamesEnabled } from "@/shared/utils/featureFlags";
 import { extractApiKey } from "@/sse/services/auth";
 import { maybeOmitCatalogModelName } from "./catalogHelpers";
@@ -46,7 +45,7 @@ import { isCodexModelCatalogClient } from "./catalogRequest";
  * returns early, but it still owes the caller these steps — the discovery mirrors in
  * particular are what let Claude Code see a quota pool's models at all.
  */
-export async function applyCatalogPostFilters(
+export function applyCatalogPostFilters(
   request: Request,
   models: Array<Record<string, any>>,
   ctx: {
@@ -55,8 +54,7 @@ export async function applyCatalogPostFilters(
     aliasToProviderId: Record<string, string>;
     hideNoThinkVariants?: boolean;
   }
-): Promise<Array<Record<string, any>>> {
-  const yieldTurn = (): Promise<void> => new Promise((resolve) => setImmediate(resolve));
+): Array<Record<string, any>> {
   let finalModels = models;
 
   // variants are only generated for surviving models.
@@ -67,11 +65,6 @@ export async function applyCatalogPostFilters(
     });
   }
 
-  // #9147: the variant-append passes each walk the full model list (O(n) per pass),
-  // so a catalog-scale build must not run all of them in one synchronous stretch.
-  // Yield once between the expensive passes to let the event loop breathe.
-  await yieldTurn();
-
   // Advertise Claude reasoning-effort variants (claude/<model>-{low,medium,high[,xhigh]}).
   // Derived from the already key-filtered list so a variant only appears when its real
   // model is permitted. Runs before the no-thinking pass: the gateway already routes these
@@ -146,15 +139,11 @@ export async function applyCatalogPostFilters(
     );
   }
 
-  await yieldTurn();
-
   // #7694: advertise `<provider>/<model>-<tier>` variants for synced models that
   // captured `reasoning.supported_efforts` at sync time (capabilities.effort_tiers).
   // Derived from the already key-filtered list; skips codex/kimi (own suffix mechanism).
   finalModels = appendSyncedEffortVariants(finalModels);
 
-  await yieldTurn();
-
   // #4424 follow-up — drop exact-duplicate ids that slip through the per-source push
   // guards (e.g. `codex/gpt-5.5`, `veo-free/seedance` listed twice). Keyed by listing
   // identity (id, type, subtype) so the intentional same-id audio transcription/speech
@@ -218,45 +207,25 @@ export async function finalizeCatalogResponse(
   }
 
   const includeModelNames = isModelCatalogNamesEnabled();
-  // #9147: enrichment is the most expensive single stage of the catalog build —
-  // per-entry provider/model resolution plus pricing + token/context override
-  // lookups. Two fixes so a large catalog cannot pin the Node.js thread here:
-  //  (1) bulk-load the synced-capability + override tables ONCE into an in-memory
-  //      snapshot (#9199 machinery) so per-entry enrichment never hits SQLite;
-  //  (2) yield to the event loop every `YIELD_EVERY` entries so even the remaining
-  //      per-entry work is interleaved with other callers / the dashboard WS.
-  const yieldTurn = (): Promise<void> => new Promise((resolve) => setImmediate(resolve));
-  await yieldTurn();
-  const capabilityResolutionSnapshot = createModelCapabilityResolutionSnapshot();
-  const enriched: Array<Record<string, unknown>> = [];
-  const catYIELD_EVERY = 5;
-  let catEnrichCount = 0;
-  for (const model of finalModels) {
-    let listedModel: Record<string, unknown>;
-    if (model.owned_by === "combo") {
-      listedModel = maybeOmitCatalogModelName(model, includeModelNames);
-    } else {
-      const entry = enrichCatalogModelEntry(model, undefined, {
-        ...enrichmentSnapshot,
-        capabilityResolutionSnapshot,
-      });
-      const fallbackContextLength = getContextFallback(entry);
-      listedModel = fallbackContextLength
-        ? { ...entry, context_length: fallbackContextLength }
-        : entry;
-      listedModel = maybeOmitCatalogModelName(listedModel, includeModelNames);
-    }
-    enriched.push(listedModel);
-    catEnrichCount++;
-    if (catEnrichCount % catYIELD_EVERY === 0) {
-      await yieldTurn();
-    }
-  }
-  await yieldTurn();
-  const enrichedModels = disambiguateCatalogModelNames(enriched);
-  await yieldTurn();
+  const enrichedModels = disambiguateCatalogModelNames(
+    finalModels.map((model) => {
+      if (model.owned_by === "combo") {
+        return maybeOmitCatalogModelName(model, includeModelNames);
+      }
+      const enriched = enrichCatalogModelEntry(model, undefined, enrichmentSnapshot);
+      const fallbackContextLength = getContextFallback(enriched);
+      const listedModel = fallbackContextLength
+        ? { ...enriched, context_length: fallbackContextLength }
+        : enriched;
+      return maybeOmitCatalogModelName(listedModel, includeModelNames);
+    })
+  );
+  // Canonical provider-grouped publication: one contiguous block per provider,
+  // combos pinned first. Stable — preserves combo sort_order, connection priority,
+  // and equal-id audio twins. Grouped by owned_by (canonical identity), not the
+  // routing alias prefix. Applied after enrichment/disambiguation so the final
+  // serialized order is what every consumer sees; cached as part of the body.
   const orderedModels = sortCatalogModelsProviderGrouped(enrichedModels);
-  await yieldTurn();
   // Codex CLI compatibility: its model-catalog refresh (codex_models_manager) does
   // GET /v1/models?client_version=<v> and decodes a JSON object with a TOP-LEVEL
   // `models` array, so the OpenAI-standard `{object,data}` shape makes it fail with

src/lib/modelCapabilities.ts

@@ -567,12 +567,9 @@ function getContextOverride(
 /**
  * Resolve a persisted context override by canonical id, then by the exact raw
  * alias supplied by the caller. Neither lookup inherits to related models.
- *
- * `snapshot` is the #9147 build-local bulk load; when supplied the on-demand
- * SQLite read is skipped and the preloaded nested map is used instead.
  */
-export function getResolvedModelContextOverride(input: CapabilityInput, snapshot?: ModelCapabilityResolutionSnapshot | null): number | null {
-  return getContextOverride(resolveCapabilityInput(input), snapshot);
+export function getResolvedModelContextOverride(input: CapabilityInput): number | null {
+  return getContextOverride(resolveCapabilityInput(input));
 }
 
 function getInputTokenCapabilityOverride(resolved: {

src/lib/modelMetadataRegistry.ts

@@ -28,7 +28,6 @@ import {
   CANONICAL_EFFORT_VALUES,
   extendCodexGpt56EffortValues,
 } from "@/shared/reasoning/effortStandardization";
-import type { ModelCapabilityResolutionSnapshot } from "@/lib/modelCapabilityResolutionSnapshot";
 
 const MODEL_METADATA_SCHEMA_VERSION = "model-metadata-v1";
 
@@ -41,9 +40,6 @@ type JsonRecord = Record<string, unknown>;
 export interface CatalogEnrichmentSnapshot {
   modelsDevPricing: PricingByProvider | null;
   providerNodeIdsByPrefix?: Readonly<Record<string, string>>;
-  /** #9147: build-local bulk load of synced capabilities + token/context overrides
-   * so per-entry enrichment never hits SQLite again (see catalogResponse.ts). */
-  capabilityResolutionSnapshot?: ModelCapabilityResolutionSnapshot | null;
 }
 
 interface CatalogDiagnosticsOptions {
@@ -204,27 +200,20 @@ export function getCatalogDiagnosticsHeaders(
 export function getCanonicalModelMetadata(input: {
   provider?: string | null;
   model?: string | null;
-  snapshot?: ModelCapabilityResolutionSnapshot | null;
 }): CanonicalModelMetadata | null {
   const modelId = asNonEmptyString(input.model);
   if (!modelId) return null;
 
-  const resolved = getResolvedModelCapabilities(
-    {
-      provider: input.provider || null,
-      model: modelId,
-    },
-    undefined,
-    input.snapshot || null
-  );
+  const resolved = getResolvedModelCapabilities({
+    provider: input.provider || null,
+    model: modelId,
+  });
   const provider = resolved.provider;
   const providerAlias = provider ? PROVIDER_ID_TO_ALIAS[provider] || provider : null;
   const registryModel = getRegistryModel(providerAlias || provider, resolved.model || modelId);
   const staticSpec = getModelSpec(resolved.model || modelId);
   const syncedCapability =
-    provider && resolved.model
-      ? getSyncedCapability(provider, resolved.model, input.snapshot?.synced ?? null)
-      : null;
+    provider && resolved.model ? getSyncedCapability(provider, resolved.model) : null;
   const canonicalStaticAlias = resolveStaticModelAlias(resolved.model || modelId);
   const modalities = buildModalities(
     resolved.modalitiesInput,
@@ -431,11 +420,7 @@ export function enrichCatalogModelEntry<T extends JsonRecord>(
       return id;
     })();
 
-  const metadata = getCanonicalModelMetadata({
-    provider,
-    model,
-    snapshot: snapshot?.capabilityResolutionSnapshot ?? null,
-  });
+  const metadata = getCanonicalModelMetadata({ provider, model });
   if (!metadata) return entry;
   const registryModel = getRegistryModel(
     metadata.providerAlias || metadata.provider,
@@ -451,11 +436,7 @@ export function enrichCatalogModelEntry<T extends JsonRecord>(
     getAuthoritativeContextWindow(metadata.model) ??
     getAuthoritativeContextWindow(model);
   const specialtySurface = isNonChatCatalogSurface(entry.type);
-  const capabilitySnapshot = snapshot?.capabilityResolutionSnapshot ?? null;
-  const persistedContextWindow = getResolvedModelContextOverride(
-    { provider, model },
-    capabilitySnapshot
-  );
+  const persistedContextWindow = getResolvedModelContextOverride({ provider, model });
   const capabilityFields = {
     ...(typeof metadata.capabilities.vision === "boolean"
       ? { vision: metadata.capabilities.vision }
@@ -547,15 +528,10 @@ export function enrichCatalogModelEntry<T extends JsonRecord>(
   }
 
   const persistedOutputLimit =
-    getModelCapabilityOverride(provider, model, "max_output_tokens", capabilitySnapshot?.maxTokenOverrides) ??
-    getModelCapabilityOverride(provider, model, "max_token", capabilitySnapshot?.maxTokenOverrides) ??
-    getModelCapabilityOverride(
-      publicProvider,
-      model,
-      "max_output_tokens",
-      capabilitySnapshot?.maxTokenOverrides
-    ) ??
-    getModelCapabilityOverride(publicProvider, model, "max_token", capabilitySnapshot?.maxTokenOverrides);
+    getModelCapabilityOverride(provider, model, "max_output_tokens") ??
+    getModelCapabilityOverride(provider, model, "max_token") ??
+    getModelCapabilityOverride(publicProvider, model, "max_output_tokens") ??
+    getModelCapabilityOverride(publicProvider, model, "max_token");
   if (persistedOutputLimit !== null) {
     nextEntry.max_output_tokens = persistedOutputLimit;
   } else if (

tests/unit/10313-catalog-cache-key-hashing.test.ts

@@ -1,131 +0,0 @@
-import test from "node:test";
-import assert from "node:assert/strict";
-import fs from "node:fs";
-import os from "node:os";
-import path from "node:path";
-
-const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-catalog-keyleak-"));
-process.env.DATA_DIR = TEST_DATA_DIR;
-process.env.API_KEY_SECRET = process.env.API_KEY_SECRET || "catalog-keyleak-test-secret";
-
-const core = await import("../../src/lib/db/core.ts");
-const apiKeysDb = await import("../../src/lib/db/apiKeys.ts");
-const v1ModelsCatalog = await import("../../src/app/api/v1/models/catalog.ts");
-const catalogCacheMod = await import("../../src/app/api/v1/models/catalogCache.ts");
-
-const SECRET = "sk-live-PROBE-10313-SUPER-SECRET-TOKEN";
-
-test.beforeEach(() => {
-  core.resetDbInstance();
-  apiKeysDb.resetApiKeyState();
-  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
-  fs.mkdirSync(TEST_DATA_DIR, { recursive: true });
-  v1ModelsCatalog.__resetCatalogBuilderRunsForTest();
-});
-
-test.after(() => {
-  core.resetDbInstance();
-  apiKeysDb.resetApiKeyState();
-  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
-});
-
-function captureMapKeys(): { keys: string[]; restore: () => void } {
-  const capturedKeys: string[] = [];
-  const originalSet = Map.prototype.set;
-  // eslint-disable-next-line @typescript-eslint/no-explicit-any
-  (Map.prototype as any).set = function (key: unknown, value: unknown) {
-    if (typeof key === "string") capturedKeys.push(key);
-    return originalSet.call(this, key, value);
-  };
-  return {
-    keys: capturedKeys,
-    release: () => {
-      Map.prototype.set = originalSet;
-    },
-  };
-}
-
-// buildCatalogCacheKey emits `prefix|isCodex|apiKeyFingerprint|configuredOnly|hideAuto|hideNoThink`
-// (6 pipe-delimited fields). Other in-flight keys (e.g. `x-request-id`) don't match.
-function isCatalogCacheKey(k: string): boolean {
-  return k.split("|").length === 6;
-}
-
-test("catalog cache Map keys must not contain the raw bearer API key (#10313)", async () => {
-  const probe = captureMapKeys();
-
-  try {
-    const request = new Request("http://localhost/v1/models", {
-      headers: { Authorization: `Bearer ${SECRET}` },
-    });
-    const res = await v1ModelsCatalog.getUnifiedModelsResponse(request);
-    assert.ok(res.status === 200 || res.status === 401 || res.status === 403);
-  } finally {
-    probe.release();
-  }
-
-  const catalogKeys = probe.keys.filter(isCatalogCacheKey);
-  assert.ok(
-    catalogKeys.length > 0,
-    "no catalog cache Map.set() calls observed — the probe did not exercise catalogCache/catalogInFlight"
-  );
-
-  const leaked = catalogKeys.filter((k) => k.includes(SECRET));
-  assert.deepEqual(
-    leaked,
-    [],
-    `catalog cache Map key retained the raw API key verbatim: ${JSON.stringify(leaked)} — ` +
-      `buildCatalogCacheKey() must hash the secret (e.g. sha256) before using it as a Map key`
-  );
-
-  assert.ok(catalogCacheMod.CATALOG_CACHE_TTL_MS_DEFAULT > 0);
-});
-
-test("cache keys embed the sha256 digest of the secret, never the raw secret (#10313)", async () => {
-  // Capture ALL Map keys emitted across BOTH requests (a 2nd identical-secret request
-  // may be a cache hit and emit no new catalog key — irrelevant here: we assert on the
-  // digest that did appear).
-  const probe = captureMapKeys();
-  try {
-    const resA = await v1ModelsCatalog.getUnifiedModelsResponse(
-      new Request("http://localhost/v1/models", {
-        headers: { Authorization: "Bearer sk-10313-DIGEST-A" },
-      })
-    );
-    const resB = await v1ModelsCatalog.getUnifiedModelsResponse(
-      new Request("http://localhost/v1/models", {
-        headers: { Authorization: "Bearer sk-10313-DIGEST-B" },
-      })
-    );
-    assert.ok(resA.status === 200 || resA.status === 401 || resA.status === 403);
-    assert.ok(resB.status === 200 || resB.status === 401 || resB.status === 403);
-  } finally {
-    probe.release();
-  }
-
-  const catalogKeys = probe.keys.filter(isCatalogCacheKey);
-  assert.ok(catalogKeys.length > 0, "expected catalog cache Map.set() calls");
-
-  // #10538 sync note: buildCatalogCacheKey() delegates to the canonical
-  // fingerprintCatalogAuthKey() (landed on release/v3.8.50 independently of #10313),
-  // which truncates the sha256 hex digest to 16 chars for a shorter Map key. Derive
-  // the expected fingerprint the same way rather than re-hardcoding the full digest.
-  const digestA = catalogCacheMod.fingerprintCatalogAuthKey("sk-10313-DIGEST-A");
-  const digestB = catalogCacheMod.fingerprintCatalogAuthKey("sk-10313-DIGEST-B");
-  const rawA = "sk-10313-DIGEST-A";
-  const rawB = "sk-10313-DIGEST-B";
-
-  // The hashed fingerprint, not the raw secret, rides in the cache keys.
-  const keysWithDigestA = catalogKeys.filter((k) => k.includes(digestA));
-  const keysWithDigestB = catalogKeys.filter((k) => k.includes(digestB));
-  assert.ok(keysWithDigestA.length > 0, `expected a cache key embedding the fingerprint of A: ${catalogKeys.join(",")}`);
-  assert.ok(keysWithDigestB.length > 0, `expected a cache key embedding the fingerprint of B: ${catalogKeys.join(",")}`);
-
-  // Raw secrets must never appear (issue #10313 root cause).
-  assert.ok(!catalogKeys.some((k) => k.includes(rawA) || k.includes(rawB)));
-
-  // Identical secrets ⇒ identical key (memoized reuse); different ⇒ distinct.
-  assert.ok(keysWithDigestA.every((k) => k === keysWithDigestA[0]), "all A keys must be identical");
-  assert.ok(keysWithDigestB.every((k) => k === keysWithDigestB[0]), "all B keys must be identical");
-  assert.notEqual(keysWithDigestA[0], keysWithDigestB[0]);
-});

tests/unit/9147-catalog-eventloop-yield.test.ts

@@ -1,88 +0,0 @@
-import test from "node:test";
-import assert from "node:assert/strict";
-import fs from "node:fs";
-import os from "node:os";
-import path from "node:path";
-
-const TEST_DATA_DIR = fs.mkdtempSync(path.join(os.tmpdir(), "omniroute-9147-"));
-process.env.DATA_DIR = TEST_DATA_DIR;
-process.env.API_KEY_SECRET = process.env.API_KEY_SECRET || "catalog-9147-test-secret";
-
-const core = await import("../../src/lib/db/core.ts");
-const apiKeysDb = await import("../../src/lib/db/apiKeys.ts");
-const v1ModelsCatalog = await import("../../src/app/api/v1/models/catalog.ts");
-
-const CONNECTION_COUNT = 60;
-const MODELS_PER_CONNECTION = 12; // ~720 synced models total
-
-async function resetStorage() {
-  core.resetDbInstance();
-  apiKeysDb.resetApiKeyState();
-  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
-  fs.mkdirSync(TEST_DATA_DIR, { recursive: true });
-  v1ModelsCatalog.__resetCatalogBuilderRunsForTest();
-}
-
-async function seedCatalogScaleDataset() {
-  const db = core.getDbInstance();
-  const now = new Date().toISOString();
-  const insertConn = db.prepare(
-    `INSERT INTO provider_connections (id, provider, auth_type, name, priority, is_active, api_key, created_at, updated_at)
-     VALUES (?, 'openai-compatible', 'apikey', ?, ?, 1, ?, ?, ?)`
-  );
-  const insertModels = db.prepare(
-    `INSERT INTO key_value (namespace, key, value) VALUES ('syncedAvailableModels', ?, ?)`
-  );
-  const seedTx = db.transaction(() => {
-    for (let i = 0; i < CONNECTION_COUNT; i++) {
-      const id = `probe-conn-${i}`;
-      insertConn.run(id, `probe-connection-${i}`, i, `sk-probe-${i}`, now, now);
-      const models = Array.from({ length: MODELS_PER_CONNECTION }, (_, m) => ({
-        id: `probe-model-${i}-${m}`,
-        name: `Probe Model ${i}-${m}`,
-        contextLength: 128000,
-      }));
-      insertModels.run(`openai-compatible:${id}`, JSON.stringify(models));
-    }
-  });
-  seedTx();
-}
-
-test.beforeEach(async () => {
-  await resetStorage();
-});
-
-test.after(async () => {
-  core.resetDbInstance();
-  apiKeysDb.resetApiKeyState();
-  fs.rmSync(TEST_DATA_DIR, { recursive: true, force: true });
-});
-
-test("#9147 — catalog build at catalog-scale must not pin the event loop for a long stretch", async () => {
-  await seedCatalogScaleDataset();
-  const req = new Request("http://localhost/v1/models");
-  let settled = false;
-  const buildPromise = v1ModelsCatalog.getUnifiedModelsResponse(req).then((res) => {
-    settled = true;
-    return res;
-  });
-  let lastTick = performance.now();
-  let maxGapMs = 0;
-  let ticks = 0;
-  while (!settled) {
-    await new Promise((resolve) => setTimeout(resolve, 0));
-    const now = performance.now();
-    maxGapMs = Math.max(maxGapMs, now - lastTick);
-    lastTick = now;
-    ticks++;
-    if (ticks > 20000) break;
-  }
-  const res = await buildPromise;
-  assert.equal(res.status, 200);
-  assert.ok(
-    maxGapMs < 150,
-    `event loop was blocked for ${maxGapMs.toFixed(1)}ms in a single stretch while building the ` +
-      `catalog for ${CONNECTION_COUNT} connections / ${CONNECTION_COUNT * MODELS_PER_CONNECTION} models ` +
-      `(${ticks} interleaved ticks observed) — the builder is not yielding to the event loop`
-  );
-});

tests/unit/compression/engine-stage-gate-metadata.test.ts

@@ -0,0 +1,72 @@
+/**
+ * Gate de estágio × engine sem `metadata`.
+ *
+ * `assertValidEngine()` (engines/registry.ts) valida `id`, `apply`, `compress`,
+ * `getConfigSchema` e `validateConfig` — NÃO exige `metadata`. Uma engine sem
+ * esse campo é, portanto, um registro legal. Mas `canRunAtCompressionStage`
+ * lia `engine.metadata.executionStages` sem guarda, então essa engine legal
+ * derrubava o pipeline inteiro com `TypeError: Cannot read properties of
+ * undefined` — em vez de falhar aberto, que é o contrato da compressão.
+ *
+ * O fallback documentado para quem não declara `executionStages` é "só
+ * pre-translation". Metadata ausente é o mesmo caso de "não declarou", e tem de
+ * cair no mesmo fallback.
+ */
+import { describe, it, beforeEach, afterEach } from "node:test";
+import assert from "node:assert/strict";
+
+import {
+  registerCompressionEngine,
+  unregisterCompressionEngine,
+} from "../../../open-sse/services/compression/engines/registry.ts";
+import { applyStackedCompression } from "../../../open-sse/services/compression/strategySelector.ts";
+import type { CompressionEngine } from "../../../open-sse/services/compression/engines/types.ts";
+
+const ENGINE_ID = "metadata-less-test-engine";
+
+const body = () => ({ messages: [{ role: "user", content: "hello world" }] });
+
+describe("stage gate — engine sem metadata", () => {
+  beforeEach(() => {
+    // Registro deliberadamente sem `metadata`: é o que assertValidEngine aceita.
+    registerCompressionEngine({
+      id: ENGINE_ID,
+      name: "metadata-less test engine",
+      targets: ["messages"],
+      stackable: true,
+      apply(input: Record<string, unknown>) {
+        return { body: input, compressed: false, stats: null };
+      },
+      compress() {
+        return { text: "", stats: null };
+      },
+      getConfigSchema() {
+        return [];
+      },
+      validateConfig() {
+        return { valid: true, errors: [] };
+      },
+    } as unknown as CompressionEngine);
+  });
+
+  afterEach(() => unregisterCompressionEngine(ENGINE_ID));
+
+  it("não derruba o pipeline no estágio pré-tradução (default)", () => {
+    const r = applyStackedCompression(body(), [{ engine: ENGINE_ID }]);
+    assert.equal(r.compressed, false);
+    assert.deepEqual(r.body, body());
+  });
+
+  it("não derruba o pipeline no estágio pós-tradução — pula pelo fallback", () => {
+    const r = applyStackedCompression(body(), [{ engine: ENGINE_ID }], {
+      compressionStage: "post-translation",
+    });
+    assert.equal(r.compressed, false);
+    assert.deepEqual(r.body, body());
+    // Fallback documentado: quem não declara estágio só roda pre-translation.
+    assert.ok(
+      r.stats?.validationWarnings?.some((w) => w.includes(ENGINE_ID) && w.includes("skipped")),
+      `esperado aviso de skip por estágio, veio ${JSON.stringify(r.stats?.validationWarnings)}`
+    );
+  });
+});

tests/unit/models-catalog-functional-gateway.test.ts

@@ -24,11 +24,11 @@ function makeRequest(query = ""): Request {
   return new Request(`http://localhost/v1/models${query}`);
 }
 
-test("catalog post-filters do not add mirrors when gate off (default)", async () => {
+test("catalog post-filters do not add mirrors when gate off (default)", () => {
   const models = [
     { id: "deepseek/deepseek-v4-flash", owned_by: "deepseek", root: "deepseek-v4-flash" },
   ];
-  const out = await applyCatalogPostFilters(makeRequest(), models, {
+  const out = applyCatalogPostFilters(makeRequest(), models, {
     connections: [],
     prefixMode: "dual",
     aliasToProviderId: {},
@@ -41,7 +41,7 @@ test("final catalog permission filtering does not let a mirror inherit base acce
   setFunctionalGatewayProviderSetting("agentrouter", "on");
 
   const models = [{ id: "kmc/k3", owned_by: "kimi-coding", root: "k3" }];
-  const withMirror = await applyCatalogPostFilters(makeRequest(), models, {
+  const withMirror = applyCatalogPostFilters(makeRequest(), models, {
     connections: [
       {
         id: "conn-1",
@@ -77,14 +77,14 @@ test("final catalog permission filtering does not let a mirror inherit base acce
   );
 });
 
-test("catalog post-filters synthesize a gateway mirror when gate on and gateway has a connection", async () => {
+test("catalog post-filters synthesize a gateway mirror when gate on and gateway has a connection", () => {
   setFeatureFlagOverride(FLAG_KEY, "true");
   setFunctionalGatewayProviderSetting("agentrouter", "on");
 
   const models = [
     { id: "deepseek/deepseek-v4-flash", owned_by: "deepseek", root: "deepseek-v4-flash" },
   ];
-  const out = await applyCatalogPostFilters(makeRequest(), models, {
+  const out = applyCatalogPostFilters(makeRequest(), models, {
     connections: [
       {
         id: "conn-1",

tests/unit/models-catalog-route.test.ts

@@ -1340,24 +1340,18 @@ test("v1 models catalog returns 500 when model compatibility lookup crashes", as
 
   db.prepare = (sql) => {
     const statement = originalPrepare(sql);
-    // #9147: the catalog builder now resolves hidden models via a single bulk
-    // read (`getHiddenModelsByProvider()`, src/lib/db/models.ts) instead of the
-    // old per-provider `SELECT value FROM key_value WHERE namespace = ? AND
-    // key = ?` / readCompatList() lookup — intercept the bulk query's `.all()`
-    // call so this test still exercises "DB read for model visibility crashes
-    // -> catalog endpoint surfaces 500" against the current implementation.
-    if (
-      String(sql) !==
-      "SELECT namespace, key, value FROM key_value WHERE namespace IN ('modelCompatOverrides', 'customModels')"
-    ) {
+    if (String(sql) !== "SELECT value FROM key_value WHERE namespace = ? AND key = ?") {
       return statement;
     }
 
     return new Proxy(statement, {
       get(target, prop, receiver) {
-        if (prop === "all") {
+        if (prop === "get") {
           return (...args) => {
-            throw new Error("compat lookup boom");
+            if (args[0] === "modelCompatOverrides") {
+              throw new Error("compat lookup boom");
+            }
+            return target.get(...args);
           };
         }
         return Reflect.get(target, prop, receiver);