@vellumai/assistant 0.12.4-staging.1 → 0.12.4-staging.3

package/src/config/schemas/llm.ts

@@ -14,12 +14,14 @@ import {
   parseVellumModel,
 } from "../../providers/vellum-model-routing.js";
 import {
-  BACKUP_PROFILE_KEYS,
   DEFAULT_PROFILE_KEYS,
   DEFAULT_PROFILE_PROVIDERS,
   FALLBACK_PROFILE_BY_KEY,
-  isBackupProfileKey,
   isDefaultProfileKey,
+  isManagedOnlyProfileKey,
+  JEV_MANAGED_PROFILE_CALL_SITES,
+  JEV_MANAGED_PROFILE_KEY,
+  MANAGED_ONLY_PROFILE_KEYS,
 } from "../default-profile-names.js";
 import { InputModalitiesSchema } from "../input-modalities.js";
 
@@ -766,10 +768,12 @@ function unresolvableProfileReason(
   name: string,
   backupsResolve: boolean,
 ): string {
-  return !backupsResolve &&
-    (BACKUP_PROFILE_KEYS as readonly string[]).includes(name)
-    ? "is a managed backup profile, which resolves only while llm.defaultProvider is the managed provider"
-    : "is not defined in llm.profiles";
+  if (backupsResolve || !isManagedOnlyProfileKey(name)) {
+    return "is not defined in llm.profiles";
+  }
+  return name === JEV_MANAGED_PROFILE_KEY
+    ? "is the managed Jev profile, which resolves only while llm.defaultProvider is the managed provider"
+    : "is a managed backup profile, which resolves only while llm.defaultProvider is the managed provider";
 }
 
 /**
@@ -800,7 +804,7 @@ function referenceableProfileKeys(
 ): string[] {
   return Object.entries(profiles ?? {})
     .filter(([name, value]) => {
-      if (backupsResolve || !isBackupProfileKey(name)) {
+      if (backupsResolve || !isManagedOnlyProfileKey(name)) {
         return true;
       }
       const entry =
@@ -854,7 +858,7 @@ export function collectFallbackProfileIssues(
   const profileNames = new Set([
     ...referenceableProfileKeys(profiles, backupsResolve),
     ...DEFAULT_PROFILE_KEYS,
-    ...(backupsResolve ? BACKUP_PROFILE_KEYS : []),
+    ...(backupsResolve ? MANAGED_ONLY_PROFILE_KEYS : []),
   ]);
   const mixProfileNames = new Set(
     entries
@@ -1004,6 +1008,8 @@ export const LLMSchema = z
     // they join the set only under a managed `llm.defaultProvider`: on a BYOK
     // or ChatGPT default provider they have no body to resolve to, and
     // keeping the reference would strand a selection the picker cannot show.
+    // The managed Jev profile (`JEV_MANAGED_PROFILE_KEY`) is scoped the same
+    // way, as a pin target for the judge call sites.
     // The flag-gated `os-beta` is excluded: it resolves only while a
     // workspace entry exists, so a reference to it is valid only when that
     // entry is present in `config.profiles`. A backup name materialized as a
@@ -1018,9 +1024,23 @@ export const LLMSchema = z
         backupsResolve,
       ),
       ...DEFAULT_PROFILE_KEYS,
-      ...(backupsResolve ? BACKUP_PROFILE_KEYS : []),
+      ...(backupsResolve ? MANAGED_ONLY_PROFILE_KEYS : []),
     ]);
     for (const [siteId, siteConfig] of Object.entries(config.callSites ?? {})) {
+      // The Jev profile answers with structured verdicts, so only the sites
+      // built to read one may pin it; a text-producing site would receive
+      // output it cannot use. (Mix arms are covered by the profile-level
+      // mix validation below, which bars the profile from any mix.)
+      if (
+        siteConfig?.profile === JEV_MANAGED_PROFILE_KEY &&
+        !(JEV_MANAGED_PROFILE_CALL_SITES as readonly string[]).includes(siteId)
+      ) {
+        ctx.addIssue({
+          code: "custom",
+          path: ["callSites", siteId, "profile"],
+          message: `Profile "${JEV_MANAGED_PROFILE_KEY}" returns structured answers rather than chat text, so it can only be pinned to ${JEV_MANAGED_PROFILE_CALL_SITES.join(", ")}`,
+        });
+      }
       if (siteConfig?.profile == null) {
         continue;
       }
@@ -1052,6 +1072,20 @@ export const LLMSchema = z
         message: `Profile "${config.advisorProfile}" referenced by llm.advisorProfile ${unresolvableProfileReason(config.advisorProfile, backupsResolve)}`,
       });
     }
+    // The managed Jev profile is a valid reference target for call-site pins
+    // only. Its model returns structured verdicts rather than chat text, so
+    // the conversation positions reject it here as well as at the write
+    // routes, which keeps a raw config write from handing the main agent or
+    // the advisor a model that cannot answer in prose.
+    for (const field of ["activeProfile", "advisorProfile"] as const) {
+      if (config[field] === JEV_MANAGED_PROFILE_KEY) {
+        ctx.addIssue({
+          code: "custom",
+          path: [field],
+          message: `Profile "${JEV_MANAGED_PROFILE_KEY}" referenced by llm.${field} returns structured answers rather than chat text, so it cannot be the conversation or advisor profile`,
+        });
+      }
+    }
 
     // --- Mix profile validation --------------------------------------------
     // Config keys a mix profile must NOT also set (a mix only references other
@@ -1105,6 +1139,17 @@ export const LLMSchema = z
           });
           continue;
         }
+        // A mix is a conversation profile, so an arm that returns structured
+        // verdicts rather than text would hand a chat turn to the decision
+        // model whenever the seeded pick lands on it.
+        if (arm.profile === JEV_MANAGED_PROFILE_KEY) {
+          ctx.addIssue({
+            code: "custom",
+            path: ["profiles", name, "mix", index, "profile"],
+            message: `Mix profile "${name}" references "${JEV_MANAGED_PROFILE_KEY}", which returns structured answers rather than chat text and cannot be a mix constituent.`,
+          });
+          continue;
+        }
         // (b) No nesting — a mix arm must be a standard (non-mix) profile.
         if (mixProfileNames.has(arm.profile)) {
           ctx.addIssue({

package/src/daemon/conversation-slash.ts

@@ -1,5 +1,5 @@
 import type { InterfaceId } from "../channels/types.js";
-import { getUserSelectableProfilesForProvider } from "../config/default-profile-catalog.js";
+import { getConversationProfilesForProvider } from "../config/default-profile-catalog.js";
 import { resolveEffectiveContextWindow } from "../config/llm-context-resolution.js";
 import { resolveCallSiteConfig } from "../config/llm-resolver.js";
 import {
@@ -148,7 +148,7 @@ async function resolveModelCommand(
   parse: ModelCommandParse,
 ): Promise<SlashResolution> {
   const config = getConfig();
-  const profiles = getUserSelectableProfilesForProvider(
+  const profiles = getConversationProfilesForProvider(
     config.llm.profiles,
     config.llm.defaultProvider ?? null,
   );

package/src/daemon/dictation-text-processing.ts

@@ -1,8 +1,8 @@
 /**
  * Pure text-processing functions for dictation profiles.
  *
- * - expandSnippets: pre-LLM, dictation mode only
- * - applyDictionary: post-LLM, dictation + command modes (including fallback paths)
+ * - expandSnippets: dictation mode only
+ * - applyDictionary: dictation and selected-text edits (including fallback paths)
  */
 
 import type {
@@ -33,7 +33,7 @@ function wrapWordBoundary(
 }
 
 /**
- * Expand snippet triggers in text before sending to the LLM.
+ * Expand user-configured snippet triggers in dictated text.
  *
  * - Only enabled snippets are considered
  * - Sorted by trigger length descending (longest match wins)
@@ -80,7 +80,7 @@ export function expandSnippets(
 }
 
 /**
- * Apply dictionary normalization to text after LLM processing (or on raw fallback).
+ * Apply user-configured dictionary replacements to dictated or edited text.
  *
  * - Sorted by spoken length descending
  * - Respects wholeWord (default true) and caseSensitive (default false) per entry

package/src/plugin-api/model-profiles.ts

@@ -1,4 +1,4 @@
-import { getUserSelectableProfilesForProvider } from "../config/default-profile-catalog.js";
+import { getConversationProfilesForProvider } from "../config/default-profile-catalog.js";
 import { getConfig } from "../config/loader.js";
 import { isDispatchableProfile } from "../config/profile-dispatchability.js";
 import { orderProfileKeys } from "../config/profile-order.js";
@@ -7,7 +7,9 @@ import type { ModelProfileInfo } from "./types.js";
 /**
  * List the workspace inference profiles a plugin can route to, in the order a
  * user-facing model picker presents them (`llm.profileOrder` first, then the
- * rest alphabetically). Managed backup routes are internal and omitted.
+ * rest alphabetically). Managed backup routes and the managed Jev profile
+ * are omitted with the rest of the conversation view: a plugin routes
+ * conversation turns, and a decision model cannot answer them.
  * Metadata-only entries without a provider, model, or mix are not routing
  * targets, so plugins never see them. Disabled profiles are included and flagged via
  * {@link ModelProfileInfo.isDisabled}; weighted "mix" profiles are included and
@@ -20,7 +22,7 @@ import type { ModelProfileInfo } from "./types.js";
 export function getModelProfiles(): ModelProfileInfo[] {
   const { llm } = getConfig();
   const { activeProfile } = llm;
-  const profiles = getUserSelectableProfilesForProvider(
+  const profiles = getConversationProfilesForProvider(
     llm.profiles,
     llm.defaultProvider ?? null,
   );

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

@@ -1,7 +1,9 @@
 import { PLATFORM_PROVIDER_META } from "./platform-proxy/constants.js";
 
 export type LongContextMode =
-  "native-model" | "provider-request-option" | "unsupported";
+  | "native-model"
+  | "provider-request-option"
+  | "unsupported";
 
 export interface CatalogModelPricingTier {
   /**
@@ -243,6 +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",
@@ -454,6 +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`
@@ -1207,6 +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",
@@ -1429,6 +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
@@ -2140,6 +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",
@@ -2602,6 +2838,16 @@ export const PROVIDER_CATALOG: ProviderCatalogEntry[] =
     supportsPlatformAuth: PLATFORM_PROVIDER_META[entry.id]?.managed === true,
   }));
 
+/**
+ * Providers that are routing identities rather than catalog owners. Mirrors
+ * `ROUTING_IDENTITY_PROVIDERS` in `inference/auth.ts`, which cannot be
+ * imported here without a cycle.
+ */
+const ROUTING_IDENTITY_PROVIDER_IDS: ReadonlySet<string> = new Set([
+  "vellum",
+  "chatgpt",
+]);
+
 /**
  * Whether a catalog model produces free-form chat text. Unlisted providers
  * and model ids default to true so custom endpoints and unknown snapshots
@@ -2614,9 +2860,16 @@ export function catalogModelSupportsText(
   if (typeof provider !== "string" || typeof modelId !== "string") {
     return true;
   }
-  const model = PROVIDER_CATALOG.find((p) => p.id === provider)?.models.find(
-    (m) => m.id === modelId,
-  );
+  // A routing identity ("vellum", "chatgpt") owns no catalog models of the
+  // upstream it dispatches to, so judge the model by its catalog owner: a
+  // managed Jev profile is `vellum` + `jev-latest`, and `jev-latest` opts out
+  // of text under `typesafe`.
+  const catalogProvider = ROUTING_IDENTITY_PROVIDER_IDS.has(provider)
+    ? (getCatalogProviderForModel(modelId) ?? provider)
+    : provider;
+  const model = PROVIDER_CATALOG.find(
+    (p) => p.id === catalogProvider,
+  )?.models.find((m) => m.id === modelId);
   return model?.supportsText !== false;
 }
 

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",