@vellumai/assistant 0.12.4-staging.2 → 0.12.4-staging.4

package/src/cli/lib/__tests__/plugins-install-offline.test.ts

@@ -172,6 +172,17 @@ describe("plugins install by name — disable-platform mode", () => {
     expect(message).toContain("not in the bundled marketplace catalog");
   });
 
+  test("does not install a feature-gated bundled integration", async () => {
+    process.env.VELLUM_DISABLE_PLATFORM = "true";
+    delete process.env.IS_PLATFORM;
+
+    await runInstall("gamma");
+
+    expect(installPluginCalls).toHaveLength(0);
+    expect(platformInstallCalls).toHaveLength(0);
+    expect(process.exitCode).toBe(1);
+  });
+
   test("uses the platform install endpoint when platform features are enabled", async () => {
     delete process.env.VELLUM_DISABLE_PLATFORM;
     delete process.env.IS_PLATFORM;
@@ -184,6 +195,21 @@ describe("plugins install by name — disable-platform mode", () => {
     expect(process.exitCode).not.toBe(1);
   });
 
+  test("does not install a name absent from a successful platform catalog", async () => {
+    delete process.env.VELLUM_DISABLE_PLATFORM;
+    delete process.env.IS_PLATFORM;
+    const errSpy = spyOn(console, "error").mockImplementation(() => {});
+
+    await runInstall("gamma");
+
+    expect(installPluginCalls).toHaveLength(0);
+    expect(platformInstallCalls).toHaveLength(0);
+    expect(process.exitCode).toBe(1);
+    expect(String(errSpy.mock.calls[0]?.[0])).toContain(
+      "not in the marketplace catalog",
+    );
+  });
+
   test("installs a platform-visible local package without the platform tar endpoint", async () => {
     delete process.env.VELLUM_DISABLE_PLATFORM;
     delete process.env.IS_PLATFORM;

package/src/cli/lib/inspect-plugin.ts

@@ -31,6 +31,7 @@ import {
   arePlatformFeaturesEnabled,
   readBundledLocalPluginCatalog,
 } from "./plugin-catalog-local.js";
+import { isPluginCatalogEntryVisible } from "./plugin-catalog-visibility.js";
 import { DEFAULT_PLUGIN_REF } from "./plugin-constants.js";
 import {
   compareFingerprint,
@@ -338,6 +339,10 @@ export async function inspectPlugin(
   const local = entry ? readLocal(entry, readInstallMeta(entry.target)) : null;
   const surfaces = entry ? detectPluginSurfaces(entry.target) : null;
 
+  if (!installed && !isPluginCatalogEntryVisible(name)) {
+    throw new PluginInspectNotFoundError(name);
+  }
+
   let remote: PluginRemoteInfo | null = null;
   let remoteError: string | null = null;
   const installedLocalSource = local?.source?.kind === "local";

package/src/cli/lib/plugin-catalog-cache.ts

@@ -14,6 +14,7 @@ import {
   readBundledPluginCatalog,
 } from "./plugin-catalog-local.js";
 import { fetchPluginCatalogFromPlatform } from "./plugin-catalog-platform.js";
+import { filterPluginCatalogByFeatureFlags } from "./plugin-catalog-visibility.js";
 import type { PluginCatalog, SearchPluginsDeps } from "./search-plugins.js";
 
 /** How long a fetched catalog is served before a refresh is attempted. */
@@ -24,6 +25,8 @@ interface CacheEntry {
   timestamp: number;
 }
 
+type PluginCatalogLoadDeps = Pick<SearchPluginsDeps, "fetch">;
+
 const cache = new Map<string, CacheEntry>();
 
 /** Add bundled local packages without overriding platform-authoritative rows. */
@@ -50,9 +53,9 @@ export function mergePlatformCatalogWithBundledLocals(
  * fetch failure propagates so the caller can surface it (e.g. map a rate-limit
  * to 503) — no stale catalog is ever served.
  */
-export async function getPluginCatalog(
+async function loadPluginCatalog(
   ref: string,
-  deps: SearchPluginsDeps,
+  deps: PluginCatalogLoadDeps,
 ): Promise<PluginCatalog> {
   if (!arePlatformFeaturesEnabled()) {
     return { ...readBundledPluginCatalog(), ref };
@@ -69,6 +72,25 @@ export async function getPluginCatalog(
   return merged;
 }
 
+/** Installable catalog with feature-flag visibility applied. */
+export async function getPluginCatalog(
+  ref: string,
+  deps: SearchPluginsDeps,
+): Promise<PluginCatalog> {
+  return filterPluginCatalogByFeatureFlags(
+    await loadPluginCatalog(ref, deps),
+    deps.featureFlagEnabled,
+  );
+}
+
+/** Category metadata for plugins already installed, including hidden entries. */
+export async function getPluginCatalogForInstalledMetadata(
+  ref: string,
+  deps: PluginCatalogLoadDeps,
+): Promise<PluginCatalog> {
+  return loadPluginCatalog(ref, deps);
+}
+
 /** Invalidate the cache (for testing or forced refresh). */
 export function invalidatePluginCatalogCache(): void {
   cache.clear();

package/src/cli/lib/plugin-catalog-local.ts

@@ -13,6 +13,10 @@
  */
 
 import bundledManifest from "./bundled-marketplace.json" with { type: "json" };
+import {
+  isPluginCatalogEntryVisible,
+  type PluginCatalogFeatureFlagResolver,
+} from "./plugin-catalog-visibility.js";
 import {
   type MarketplaceEntry,
   marketplaceManifestSchema,
@@ -27,6 +31,7 @@ import {
 // Re-exported so `local`-tagged CLI commands can gate on platform features
 // without importing `platform/` directly (cli/no-daemon-internals allows lib).
 export { arePlatformFeaturesEnabled } from "../../platform/feature-gate.js";
+export { isPluginCatalogEntryVisible };
 
 let memoizedEntries: readonly MarketplaceEntry[] | undefined;
 
@@ -80,6 +85,10 @@ export function readBundledLocalPluginCatalog(): PluginCatalog {
  */
 export function resolveBundledPluginSource(
   name: string,
+  isFeatureFlagEnabled?: PluginCatalogFeatureFlagResolver,
 ): ResolvedPluginSource | null {
+  if (!isPluginCatalogEntryVisible(name, isFeatureFlagEnabled)) {
+    return null;
+  }
   return resolveMarketplaceSource(name, bundledEntries());
 }

package/src/cli/lib/plugin-catalog-visibility.ts

@@ -0,0 +1,41 @@
+import { isAssistantFeatureFlagEnabled } from "../../config/assistant-feature-flags.js";
+import type { PluginCatalog } from "./search-plugins.js";
+
+export const MCP_CATALOG_QA_INTEGRATIONS_FLAG = "mcp-catalog-qa-integrations";
+
+const MCP_CATALOG_QA_INTEGRATION_NAMES = new Set([
+  "gamma",
+  "guru",
+  "intercom",
+  "ramp",
+  "semrush",
+  "typeform",
+]);
+
+export type PluginCatalogFeatureFlagResolver = (key: string) => boolean;
+
+export function isPluginCatalogEntryVisible(
+  name: string,
+  isFeatureFlagEnabled: PluginCatalogFeatureFlagResolver = isAssistantFeatureFlagEnabled,
+): boolean {
+  return (
+    !MCP_CATALOG_QA_INTEGRATION_NAMES.has(name) ||
+    isFeatureFlagEnabled(MCP_CATALOG_QA_INTEGRATIONS_FLAG)
+  );
+}
+
+/**
+ * Hide catalog entries whose OAuth flows are still being validated. Installed
+ * plugins are unaffected because this filter only applies to catalog reads.
+ */
+export function filterPluginCatalogByFeatureFlags(
+  catalog: PluginCatalog,
+  isFeatureFlagEnabled: PluginCatalogFeatureFlagResolver = isAssistantFeatureFlagEnabled,
+): PluginCatalog {
+  const matches = catalog.matches.filter((match) =>
+    isPluginCatalogEntryVisible(match.name, isFeatureFlagEnabled),
+  );
+  return matches.length === catalog.matches.length
+    ? catalog
+    : { ...catalog, matches };
+}

package/src/cli/lib/plugin-details.ts

@@ -46,6 +46,7 @@ import {
   findCatalogEntry,
   resolveSourceFromMatch,
 } from "./plugin-catalog-resolve.js";
+import { isPluginCatalogEntryVisible } from "./plugin-catalog-visibility.js";
 import { DEFAULT_PLUGIN_REF } from "./plugin-constants.js";
 import { readValidatedPluginIcon } from "./plugin-icon-file.js";
 import { fetchMarketplaceEntries } from "./plugin-marketplace.js";
@@ -156,8 +157,8 @@ export class PluginDetailsNotFoundError extends Error {
  * Resolve the detail view for {@link opts.name}.
  *
  * Throws {@link PluginDetailsNotFoundError} when the name is neither installed
- * locally nor present in the catalog (gated at the default ref, the GitHub
- * marketplace at an explicit historical ref). A gated-catalog outage
+ * locally nor present in the visibility-gated catalog at either the default
+ * or an explicit historical ref. A gated-catalog outage
  * ({@link PluginCatalogUnavailableError}) degrades to the on-disk fields only
  * when a local copy exists; with nothing installed to render it propagates so
  * the caller can map the transient failure to a retryable 503 rather than a
@@ -400,11 +401,10 @@ async function fetchRawFile(
 /**
  * Resolve the external catalog entry claiming {@link name}.
  *
- * The default ref reads the gated catalog (the same source search / install
- * use). An explicit historical {@link ref} reads the GitHub marketplace
- * manifest at that revision — the gated catalog is ref-agnostic, so honoring a
- * reviewed/rolled-back revision is inherently a git lookup, matching the pin
- * history / inspect carve-out.
+ * Both default and historical-ref lookups apply catalog visibility first. The
+ * default ref then reads the gated catalog used by search and install, while an
+ * explicit historical {@link ref} reads the GitHub marketplace manifest at
+ * that revision.
  *
  * A gated-catalog outage (fail-hard {@link PluginCatalogUnavailableError})
  * propagates when nothing is installed — there is nothing to render and the
@@ -420,6 +420,9 @@ async function resolveCatalogEntry(
   fetchFn: FetchLike,
   installed: boolean,
 ): Promise<PluginSearchMatch | null> {
+  if (!isPluginCatalogEntryVisible(name)) {
+    return null;
+  }
   try {
     const match =
       ref === DEFAULT_PLUGIN_REF

package/src/cli/lib/search-plugins.ts

@@ -19,6 +19,8 @@ import type { MarketplaceEntry } from "./plugin-marketplace.js";
 export interface SearchPluginsDeps {
   /** HTTP client. Production callers pass `globalThis.fetch.bind(globalThis)`. */
   readonly fetch: FetchLike;
+  /** Optional feature-flag resolver for deterministic tests. */
+  readonly featureFlagEnabled?: (key: string) => boolean;
 }
 
 /** Where a catalog match comes from. */

package/src/config/feature-flag-registry.json

@@ -338,6 +338,14 @@
       "description": "Show the Add Server action on the MCP settings page. The MCP page itself remains visible; this flag gates only creating new Model Context Protocol server connections.",
       "defaultEnabled": false
     },
+    {
+      "id": "mcp-catalog-qa-integrations",
+      "scope": "assistant",
+      "key": "mcp-catalog-qa-integrations",
+      "label": "MCP Catalog QA Integrations",
+      "description": "Show Gamma, Guru, Intercom, Ramp, Semrush, and Typeform in the installable plugin catalog while their OAuth flows are validated. When off, catalog search, detail, and install-by-name cannot resolve them. Existing installed connections remain available.",
+      "defaultEnabled": false
+    },
     {
       "id": "os-beta",
       "scope": "assistant",
@@ -557,7 +565,7 @@
       "scope": "both",
       "key": "sidebar-done",
       "label": "Sidebar Done",
-      "description": "Turns archiving a chat into marking it done: the sidebar row carries a Done check instead of the overflow menu, an Old chats page lists the full history, and done chats stay findable in search by default. Off keeps today's Archive wording, the overflow menu, and search that hides archived rows unless the query asks for them.",
+      "description": "Turns archiving a chat into marking it done: the sidebar row carries a Done check instead of the overflow menu, an All chats page lists the full history, and done chats stay findable in search by default. Off keeps today's Archive wording, the overflow menu, and search that hides archived rows unless the query asks for them.",
       "defaultEnabled": false
     }
   ]

package/src/plugins/defaults/memory/substrate/skill-content.ts

@@ -34,7 +34,7 @@ export const DEFAULT_CARD_CHARS = 500;
  * only, so a missing product is not evidence it is unsupported.
  */
 export const SKILLS_INJECTION_CATALOG_HINT =
-  "Injected skills are only ones currently in the workspace. Run `assistant plugins search <name>` and `assistant skills search <name>` before concluding a given integration or skill is unsupported";
+  "These skill cards describe availability, not tool instructions. When skill_load is available, use it to load a skill's current instructions and input schemas before calling its tools; follow those schemas over remembered arguments. Injected skills are only ones currently in the workspace. Run `assistant plugins search <name>` and `assistant skills search <name>` before concluding a given integration or skill is unsupported";
 
 /**
  * Render the prose-style capability statement embedded into the unified

package/src/plugins/defaults/memory/v3/__tests__/render-injection.test.ts

@@ -31,6 +31,9 @@ describe("renderInjectionBlockInner", () => {
     expect(inner).toContain("assistant plugins search <name>");
     expect(inner).toContain("assistant skills search <name>");
     expect(inner).toContain("currently in the workspace");
+    expect(inner).toContain("When skill_load is available, use it");
+    expect(inner).toContain("before calling its tools");
+    expect(inner).toContain("over remembered arguments");
     expect(inner.indexOf("# Skills")).toBeLessThan(inner.indexOf("# Skill:"));
     expect(inner.startsWith(`${V3_INJECTION_HEADER}\n\n# Skills\n`)).toBe(true);
   });

package/src/providers/__tests__/connection-model-compat.test.ts

@@ -57,6 +57,8 @@ describe("isConnectionCompatibleWithModel", () => {
   test("oauth_subscription connection is compatible with a Codex model", () => {
     const conn = { auth: oauthAuth };
     expect(isConnectionCompatibleWithModel(conn, "gpt-6-astra")).toBe(true);
+    expect(isConnectionCompatibleWithModel(conn, "gpt-6-sol")).toBe(true);
+    expect(isConnectionCompatibleWithModel(conn, "gpt-6-luna")).toBe(true);
     expect(isConnectionCompatibleWithModel(conn, "gpt-5.6-sol")).toBe(true);
     expect(isConnectionCompatibleWithModel(conn, "gpt-5.6-terra")).toBe(true);
     expect(isConnectionCompatibleWithModel(conn, "gpt-5.6-luna")).toBe(true);

package/src/providers/anthropic/client.ts

@@ -1004,7 +1004,7 @@ export class AnthropicProvider implements Provider {
       const isHaiku = effectiveModel.includes("haiku");
       // Effort support is per-model: Haiku and Sonnet 4.5 reject the param (see isEffortSupported).
       const supportsEffort = isEffortSupported(effectiveModel);
-      // opus-4-7 / opus-4-8 / opus-5 and sonnet-5 reject `temperature`,
+      // opus-4-7 / opus-4-8 / opus-5 / opus-5-5 and sonnet-5 reject `temperature`,
       // `top_p`, and `top_k` with a 400 "`temperature`/`top_p` is deprecated
       // for this model" — model-wide, not effort-conditional (verified
       // 2026-06-23). opus-4-6 / sonnet-4-6 / haiku-4-5 still accept them.
@@ -1016,7 +1016,7 @@ export class AnthropicProvider implements Provider {
       // suffix is what matches.
       const deprecatesSamplingParams =
         /claude-opus-4-[78]\b/.test(effectiveModel) ||
-        /claude-opus-5\b/.test(effectiveModel) ||
+        /claude-opus-5(-5)?\b/.test(effectiveModel) ||
         /claude-sonnet-5\b/.test(effectiveModel) ||
         /claude-fable-/.test(effectiveModel);
       const mergedOutputConfig = {

package/src/providers/model-catalog.ts

@@ -245,6 +245,24 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
           cacheReadPer1mTokens: 1,
         },
       },
+      {
+        id: "claude-opus-5-5",
+        displayName: "Claude Opus 5.5",
+        contextWindowTokens: 1000000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens: 200000,
+        supportsThinking: true,
+        adaptiveThinkingOnly: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        pricing: {
+          inputPer1mTokens: 4,
+          outputPer1mTokens: 20,
+          cacheWritePer1mTokens: 5,
+          cacheReadPer1mTokens: 0.2,
+        },
+      },
       {
         id: "claude-opus-5",
         displayName: "Claude Opus 5",
@@ -456,6 +474,66 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
           ],
         },
       },
+      {
+        id: "gpt-6-sol",
+        displayName: "GPT-6 Sol",
+        contextWindowTokens: 1050000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens:
+          OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+        supportsThinking: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        supportsPromptCacheBreakpoints: true,
+        maxEffort: "max",
+        supportedEfforts: ["low", "medium", "high", "xhigh", "max"],
+        pricing: {
+          inputPer1mTokens: 2,
+          outputPer1mTokens: 10,
+          cacheWritePer1mTokens: 2.5,
+          cacheReadPer1mTokens: 0.2,
+          tiers: [
+            {
+              inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+              inputPer1mTokens: 4,
+              outputPer1mTokens: 15,
+              cacheWritePer1mTokens: 5,
+              cacheReadPer1mTokens: 0.4,
+            },
+          ],
+        },
+      },
+      {
+        id: "gpt-6-luna",
+        displayName: "GPT-6 Luna",
+        contextWindowTokens: 1050000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens:
+          OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+        supportsThinking: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        supportsPromptCacheBreakpoints: true,
+        maxEffort: "max",
+        supportedEfforts: ["low", "medium", "high", "xhigh", "max"],
+        pricing: {
+          inputPer1mTokens: 0.1,
+          outputPer1mTokens: 0.5,
+          cacheWritePer1mTokens: 0.125,
+          cacheReadPer1mTokens: 0.01,
+          tiers: [
+            {
+              inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+              inputPer1mTokens: 0.2,
+              outputPer1mTokens: 0.75,
+              cacheWritePer1mTokens: 0.25,
+              cacheReadPer1mTokens: 0.02,
+            },
+          ],
+        },
+      },
       // GPT-5.6 family (Sol / Terra / Luna). cacheRead is the 90% cached-read
       // discount; cacheWrite is the 1.25x-input rate GPT-5.6+ bills for
       // prompt tokens written to the cache (reported as `cache_write_tokens`
@@ -1209,6 +1287,24 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
           cacheReadPer1mTokens: 1,
         },
       },
+      {
+        id: "anthropic/claude-opus-5.5",
+        displayName: "Claude Opus 5.5",
+        contextWindowTokens: 1000000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens: 200000,
+        supportsThinking: true,
+        adaptiveThinkingOnly: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        pricing: {
+          inputPer1mTokens: 4,
+          outputPer1mTokens: 20,
+          cacheWritePer1mTokens: 5,
+          cacheReadPer1mTokens: 0.2,
+        },
+      },
       {
         id: "anthropic/claude-opus-5",
         displayName: "Claude Opus 5",
@@ -1431,6 +1527,126 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
           ],
         },
       },
+      {
+        id: "openai/gpt-6-sol",
+        displayName: "GPT-6 Sol",
+        contextWindowTokens: 1050000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens:
+          OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+        supportsThinking: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        supportsPromptCacheBreakpoints: true,
+        maxEffort: "max",
+        supportedEfforts: ["low", "medium", "high", "xhigh", "max"],
+        pricing: {
+          inputPer1mTokens: 2,
+          outputPer1mTokens: 10,
+          cacheWritePer1mTokens: 2.5,
+          cacheReadPer1mTokens: 0.2,
+          tiers: [
+            {
+              inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+              inputPer1mTokens: 4,
+              outputPer1mTokens: 15,
+              cacheWritePer1mTokens: 5,
+              cacheReadPer1mTokens: 0.4,
+            },
+          ],
+        },
+      },
+      {
+        id: "openai/gpt-6-sol-pro",
+        displayName: "GPT-6 Sol Pro",
+        contextWindowTokens: 1050000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens:
+          OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+        supportsThinking: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        supportsPromptCacheBreakpoints: true,
+        maxEffort: "max",
+        supportedEfforts: ["low", "medium", "high", "xhigh", "max"],
+        pricing: {
+          inputPer1mTokens: 2,
+          outputPer1mTokens: 10,
+          cacheWritePer1mTokens: 2.5,
+          cacheReadPer1mTokens: 0.2,
+          tiers: [
+            {
+              inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+              inputPer1mTokens: 4,
+              outputPer1mTokens: 15,
+              cacheWritePer1mTokens: 5,
+              cacheReadPer1mTokens: 0.4,
+            },
+          ],
+        },
+      },
+      {
+        id: "openai/gpt-6-luna",
+        displayName: "GPT-6 Luna",
+        contextWindowTokens: 1050000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens:
+          OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+        supportsThinking: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        supportsPromptCacheBreakpoints: true,
+        maxEffort: "max",
+        supportedEfforts: ["low", "medium", "high", "xhigh", "max"],
+        pricing: {
+          inputPer1mTokens: 0.1,
+          outputPer1mTokens: 0.5,
+          cacheWritePer1mTokens: 0.125,
+          cacheReadPer1mTokens: 0.01,
+          tiers: [
+            {
+              inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+              inputPer1mTokens: 0.2,
+              outputPer1mTokens: 0.75,
+              cacheWritePer1mTokens: 0.25,
+              cacheReadPer1mTokens: 0.02,
+            },
+          ],
+        },
+      },
+      {
+        id: "openai/gpt-6-luna-pro",
+        displayName: "GPT-6 Luna Pro",
+        contextWindowTokens: 1050000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens:
+          OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+        supportsThinking: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        supportsPromptCacheBreakpoints: true,
+        maxEffort: "max",
+        supportedEfforts: ["low", "medium", "high", "xhigh", "max"],
+        pricing: {
+          inputPer1mTokens: 0.1,
+          outputPer1mTokens: 0.5,
+          cacheWritePer1mTokens: 0.125,
+          cacheReadPer1mTokens: 0.01,
+          tiers: [
+            {
+              inputTokenThreshold: OPENAI_LONG_CONTEXT_PRICING_THRESHOLD_TOKENS,
+              inputPer1mTokens: 0.2,
+              outputPer1mTokens: 0.75,
+              cacheWritePer1mTokens: 0.25,
+              cacheReadPer1mTokens: 0.02,
+            },
+          ],
+        },
+      },
       // GPT-5.6 family (Sol / Terra / Luna). The `*-pro` slugs are the same
       // underlying models served with `reasoning.mode: pro` at identical
       // rates. cacheWrite is the 1.25x-input rate GPT-5.6+ bills for prompt
@@ -2142,6 +2358,24 @@ const RAW_PROVIDER_CATALOG: ProviderCatalogEntry[] = [
           cacheReadPer1mTokens: 1,
         },
       },
+      {
+        id: "anthropic/claude-opus-5.5",
+        displayName: "Claude Opus 5.5",
+        contextWindowTokens: 1000000,
+        maxOutputTokens: 128000,
+        longContextPricingThresholdTokens: 200000,
+        supportsThinking: true,
+        adaptiveThinkingOnly: true,
+        supportsCaching: true,
+        supportsVision: true,
+        supportsToolUse: true,
+        pricing: {
+          inputPer1mTokens: 4,
+          outputPer1mTokens: 20,
+          cacheWritePer1mTokens: 5,
+          cacheReadPer1mTokens: 0.2,
+        },
+      },
       {
         id: "anthropic/claude-opus-5",
         displayName: "Claude Opus 5",

package/src/providers/openai/codex-models.ts

@@ -10,6 +10,8 @@
  */
 export const CODEX_SUBSCRIPTION_MODEL_IDS: ReadonlySet<string> = new Set([
   "gpt-6-astra",
+  "gpt-6-sol",
+  "gpt-6-luna",
   "gpt-5.6-sol",
   "gpt-5.6-terra",
   "gpt-5.6-luna",

package/src/runtime/routes/__tests__/plugins-routes.test.ts

@@ -165,8 +165,14 @@ const getCatalogSpy = mock(
   },
 );
 
+const getInstalledMetadataCatalogSpy = mock(
+  (ref: string, deps: SearchPluginsDeps): Promise<PluginCatalog> =>
+    getCatalogSpy(ref, deps),
+);
+
 mock.module("../../../cli/lib/plugin-catalog-cache.js", () => ({
   getPluginCatalog: getCatalogSpy,
+  getPluginCatalogForInstalledMetadata: getInstalledMetadataCatalogSpy,
 }));
 
 // Mock uninstallPlugin. The handler's error mapping is the wiring under
@@ -471,6 +477,7 @@ beforeEach(() => {
 describe("GET /v1/plugins", () => {
   beforeEach(() => {
     getCatalogSpy.mockClear();
+    getInstalledMetadataCatalogSpy.mockClear();
     // Default to an empty catalog so every installed plugin's category resolves
     // to `null` (bucketed under "system"). Tests that exercise real categories
     // override this.
@@ -849,6 +856,7 @@ describe("GET /v1/plugins", () => {
     );
 
     const map = await loadCategoryMapBounded();
+    expect(getInstalledMetadataCatalogSpy).toHaveBeenCalledTimes(1);
     expect(map.get("alpha")).toBe("productivity");
   });
 

package/src/runtime/routes/plugins-routes.ts

@@ -48,7 +48,10 @@ import {
   type InstalledPluginInfo,
   listInstalledPlugins,
 } from "../../cli/lib/list-installed-plugins.js";
-import { getPluginCatalog } from "../../cli/lib/plugin-catalog-cache.js";
+import {
+  getPluginCatalog,
+  getPluginCatalogForInstalledMetadata,
+} from "../../cli/lib/plugin-catalog-cache.js";
 import { resolvePluginSourceFromCatalog } from "../../cli/lib/plugin-catalog-resolve.js";
 import {
   DEFAULT_PIN_HISTORY_LIMIT,
@@ -969,7 +972,7 @@ export async function loadCategoryMapBounded(
   let timer: ReturnType<typeof setTimeout> | undefined;
   try {
     const catalog = await Promise.race([
-      getPluginCatalog(DEFAULT_PLUGIN_REF, {
+      getPluginCatalogForInstalledMetadata(DEFAULT_PLUGIN_REF, {
         fetch: globalThis.fetch.bind(globalThis),
       }),
       new Promise<null>((resolve) => {