@vellumai/assistant 0.8.11 → 0.8.12-staging.2

package/src/workspace/adaptive-thinking-repair.ts

@@ -0,0 +1,113 @@
+import { existsSync, readFileSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+
+// Enable adaptive thinking on the managed "balanced" and "quality-optimized"
+// profiles.
+//
+// The assistant-side seed defaults (MANAGED_PROFILE_TEMPLATES in
+// seed-inference-profiles.ts) already ship thinking: { enabled: true,
+// streamThinking: true } for both profiles, which normalizes to
+// { type: "adaptive" } on the wire. Off-platform (BYOK) instances pick this
+// up on every boot because the seeder overwrites managed profiles. On-platform
+// instances preserve existing profiles (the platform overlay is authoritative),
+// so instances that were hatched before thinking was enabled in the templates
+// are stuck with thinking disabled or absent.
+//
+// Workspace migration 097 patches the on-disk config once, but it runs before
+// mergeDefaultWorkspaceConfig() which can overwrite the fix with overlay
+// profiles that have thinking disabled or absent. lifecycle.ts calls this
+// repair again after the overlay merge + profile seeding so the fix sticks.
+// The migration keeps its own frozen copy of this logic (migration files are
+// self-contained snapshots and must not be imported from).
+//
+// The repair patches both profiles, adding thinking: { enabled: true,
+// streamThinking: true } where it's missing or explicitly disabled. It skips
+// profiles that:
+//   - Don't exist (no profile to patch)
+//   - Already have thinking enabled (idempotent)
+//   - Are source: "user" (user-created profiles are untouched)
+//   - Have a non-managed, non-absent source (unknown origin)
+//   - Resolve to a non-Anthropic provider (adaptive thinking is
+//     Anthropic-specific). A profile with no explicit provider inherits
+//     llm.default.provider, so the check falls back to that — with a
+//     completely absent llm.default.provider treated as Anthropic, matching
+//     migration 052's own `?? "anthropic"` default. This keeps the legacy
+//     non-Anthropic empty `{}` shells seeded by migration 052 off the repair.
+
+const ADAPTIVE_THINKING = { enabled: true, streamThinking: true } as const;
+const TARGET_PROFILES = ["balanced", "quality-optimized"] as const;
+
+/**
+ * Patch managed Anthropic profiles that are missing adaptive thinking.
+ *
+ * Idempotent: profiles that already have thinking enabled are skipped.
+ */
+export function repairAdaptiveThinkingOnManagedProfiles(
+  workspaceDir: string,
+): void {
+  const configPath = join(workspaceDir, "config.json");
+  if (!existsSync(configPath)) return;
+
+  let config: Record<string, unknown>;
+  try {
+    const raw = JSON.parse(readFileSync(configPath, "utf-8"));
+    if (!raw || typeof raw !== "object" || Array.isArray(raw)) return;
+    config = raw as Record<string, unknown>;
+  } catch {
+    return;
+  }
+
+  const llm = readObject(config.llm);
+  if (llm === null) return;
+
+  const profiles = readObject(llm.profiles);
+  if (profiles === null) return;
+
+  // Profiles without an explicit provider inherit llm.default.provider at
+  // resolution time; an absent llm.default.provider resolves to Anthropic.
+  const defaultBlock = readObject(llm.default);
+  const defaultProvider =
+    typeof defaultBlock?.provider === "string"
+      ? defaultBlock.provider
+      : "anthropic";
+
+  let changed = false;
+
+  for (const name of TARGET_PROFILES) {
+    const profile = readObject(profiles[name]);
+    if (profile === null) continue;
+
+    // Only patch managed Anthropic profiles.
+    // Legacy profiles created before the `source` metadata field was introduced
+    // have source=undefined. Treat these as managed when the profile name is one
+    // of the canonical managed names (which TARGET_PROFILES already guarantees)
+    // and the effective provider — explicit, or inherited from llm.default — is
+    // Anthropic. Explicit `source: "user"` profiles are always skipped.
+    if (profile.source === "user") continue;
+    if (profile.source !== undefined && profile.source !== "managed") continue;
+    const effectiveProvider =
+      typeof profile.provider === "string" ? profile.provider : defaultProvider;
+    if (effectiveProvider !== "anthropic") continue;
+
+    // Skip if thinking is already enabled.
+    const thinking = readObject(profile.thinking);
+    if (thinking !== null && thinking.enabled === true) continue;
+
+    profile.thinking = { ...ADAPTIVE_THINKING };
+    profiles[name] = profile;
+    changed = true;
+  }
+
+  if (changed) {
+    llm.profiles = profiles;
+    config.llm = llm;
+    writeFileSync(configPath, JSON.stringify(config, null, 2) + "\n");
+  }
+}
+
+function readObject(value: unknown): Record<string, unknown> | null {
+  if (value === null || typeof value !== "object" || Array.isArray(value)) {
+    return null;
+  }
+  return value as Record<string, unknown>;
+}

package/src/workspace/migrations/097-enable-adaptive-thinking-managed-profiles.ts

@@ -22,87 +22,90 @@ import type { WorkspaceMigration } from "./types.js";
 //   - Already have thinking enabled (idempotent)
 //   - Are source: "user" (user-created profiles are untouched)
 //   - Have a non-managed, non-absent source (unknown origin)
-//   - Use a non-Anthropic provider (adaptive thinking is Anthropic-specific)
+//   - Resolve to a non-Anthropic provider (adaptive thinking is
+//     Anthropic-specific). A profile with no explicit provider inherits
+//     llm.default.provider, so the check falls back to that — with a
+//     completely absent llm.default.provider treated as Anthropic, matching
+//     migration 052's own `?? "anthropic"` default. This keeps the legacy
+//     non-Anthropic empty `{}` shells seeded by migration 052 off the repair.
+//
+// Note: the live startup path re-runs the same repair after the platform
+// overlay merge — see assistant/src/workspace/adaptive-thinking-repair.ts.
+// This file keeps its own frozen copy of the logic because migration modules
+// are self-contained snapshots and must never be imported by other code.
 
 const ADAPTIVE_THINKING = { enabled: true, streamThinking: true } as const;
 const TARGET_PROFILES = ["balanced", "quality-optimized"] as const;
 
-/**
- * Patch managed Anthropic profiles that are missing adaptive thinking.
- *
- * Exported so lifecycle.ts can re-run the repair after mergeDefaultWorkspaceConfig()
- * and seedInferenceProfiles(). On-platform instances with a config overlay have their
- * profiles overwritten by the overlay merge (which runs after workspace migrations),
- * so the migration alone is insufficient — the post-overlay call ensures the repair
- * sticks even when the overlay supplies profiles without thinking enabled.
- *
- * The function is idempotent: profiles that already have thinking enabled are skipped.
- */
-export function repairAdaptiveThinkingOnManagedProfiles(
-  workspaceDir: string,
-): void {
-  const configPath = join(workspaceDir, "config.json");
-  if (!existsSync(configPath)) return;
+export const enableAdaptiveThinkingManagedProfilesMigration: WorkspaceMigration =
+  {
+    id: "097-enable-adaptive-thinking-managed-profiles",
+    description:
+      "Enable adaptive thinking on managed balanced and quality-optimized profiles",
+    run(workspaceDir: string): void {
+      const configPath = join(workspaceDir, "config.json");
+      if (!existsSync(configPath)) return;
 
-  let config: Record<string, unknown>;
-  try {
-    const raw = JSON.parse(readFileSync(configPath, "utf-8"));
-    if (!raw || typeof raw !== "object" || Array.isArray(raw)) return;
-    config = raw as Record<string, unknown>;
-  } catch {
-    return;
-  }
+      let config: Record<string, unknown>;
+      try {
+        const raw = JSON.parse(readFileSync(configPath, "utf-8"));
+        if (!raw || typeof raw !== "object" || Array.isArray(raw)) return;
+        config = raw as Record<string, unknown>;
+      } catch {
+        return;
+      }
 
-  const llm = readObject(config.llm);
-  if (llm === null) return;
+      const llm = readObject(config.llm);
+      if (llm === null) return;
 
-  const profiles = readObject(llm.profiles);
-  if (profiles === null) return;
+      const profiles = readObject(llm.profiles);
+      if (profiles === null) return;
 
-  let changed = false;
+      // Profiles without an explicit provider inherit llm.default.provider at
+      // resolution time; an absent llm.default.provider resolves to Anthropic.
+      const defaultBlock = readObject(llm.default);
+      const defaultProvider =
+        typeof defaultBlock?.provider === "string"
+          ? defaultBlock.provider
+          : "anthropic";
 
-  for (const name of TARGET_PROFILES) {
-    const profile = readObject(profiles[name]);
-    if (profile === null) continue;
+      let changed = false;
 
-    // Only patch managed Anthropic profiles.
-    // Legacy profiles created before the `source` metadata field was introduced
-    // have source=undefined. Treat these as managed when the profile name is one
-    // of the canonical managed names (which TARGET_PROFILES already guarantees)
-    // and the provider is Anthropic (or absent). Explicit `source: "user"`
-    // profiles are always skipped.
-    if (profile.source === "user") continue;
-    if (profile.source !== undefined && profile.source !== "managed") continue;
-    if (
-      typeof profile.provider === "string" &&
-      profile.provider !== "anthropic"
-    ) {
-      continue;
-    }
+      for (const name of TARGET_PROFILES) {
+        const profile = readObject(profiles[name]);
+        if (profile === null) continue;
 
-    // Skip if thinking is already enabled.
-    const thinking = readObject(profile.thinking);
-    if (thinking !== null && thinking.enabled === true) continue;
+        // Only patch managed Anthropic profiles.
+        // Legacy profiles created before the `source` metadata field was
+        // introduced have source=undefined. Treat these as managed when the
+        // profile name is one of the canonical managed names (which
+        // TARGET_PROFILES already guarantees) and the effective provider —
+        // explicit, or inherited from llm.default — is Anthropic. Explicit
+        // `source: "user"` profiles are always skipped.
+        if (profile.source === "user") continue;
+        if (profile.source !== undefined && profile.source !== "managed") {
+          continue;
+        }
+        const effectiveProvider =
+          typeof profile.provider === "string"
+            ? profile.provider
+            : defaultProvider;
+        if (effectiveProvider !== "anthropic") continue;
 
-    profile.thinking = { ...ADAPTIVE_THINKING };
-    profiles[name] = profile;
-    changed = true;
-  }
+        // Skip if thinking is already enabled.
+        const thinking = readObject(profile.thinking);
+        if (thinking !== null && thinking.enabled === true) continue;
 
-  if (changed) {
-    llm.profiles = profiles;
-    config.llm = llm;
-    writeFileSync(configPath, JSON.stringify(config, null, 2) + "\n");
-  }
-}
+        profile.thinking = { ...ADAPTIVE_THINKING };
+        profiles[name] = profile;
+        changed = true;
+      }
 
-export const enableAdaptiveThinkingManagedProfilesMigration: WorkspaceMigration =
-  {
-    id: "097-enable-adaptive-thinking-managed-profiles",
-    description:
-      "Enable adaptive thinking on managed balanced and quality-optimized profiles",
-    run(workspaceDir: string): void {
-      repairAdaptiveThinkingOnManagedProfiles(workspaceDir);
+      if (changed) {
+        llm.profiles = profiles;
+        config.llm = llm;
+        writeFileSync(configPath, JSON.stringify(config, null, 2) + "\n");
+      }
     },
     down(_workspaceDir: string): void {
       // Forward-only.

package/src/workspace/migrations/098-remove-stale-updates-bulletin-file.ts

@@ -0,0 +1,31 @@
+import { rmSync } from "node:fs";
+import { join } from "node:path";
+
+import type { WorkspaceMigration } from "./types.js";
+
+const MIGRATION_ID = "098-remove-stale-updates-bulletin-file";
+
+/**
+ * Delete `<workspace>/UPDATES.md`.
+ *
+ * Release notes used to be appended to the file by workspace migrations and
+ * processed by a background conversation dispatched at daemon startup. That
+ * feature has been removed — nothing consumes the file's contents anymore,
+ * and it was system-written bulletin material, so it is deleted outright.
+ *
+ * Idempotent: `rmSync` with `force` is a no-op when the file is absent.
+ */
+export const removeStaleUpdatesBulletinFileMigration: WorkspaceMigration = {
+  id: MIGRATION_ID,
+  description:
+    "Delete leftover UPDATES.md release-note bulletins (processing job removed)",
+
+  run(workspaceDir: string): void {
+    rmSync(join(workspaceDir, "UPDATES.md"), { force: true });
+  },
+
+  down(_workspaceDir: string): void {
+    // Forward-only: the removed content was pending release-note bulletins
+    // for a feature that no longer exists.
+  },
+};

package/src/workspace/migrations/099-disable-cache-one-shot-callsites.ts

@@ -0,0 +1,81 @@
+import { existsSync, readFileSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+
+import type { WorkspaceMigration } from "./types.js";
+
+/**
+ * Add `disableCache: true` to existing on-disk entries for the one-shot LLM
+ * call sites whose shipped defaults now opt out of prompt caching.
+ *
+ * The resolver prefers an explicit `llm.callSites.<id>` entry over
+ * CALL_SITE_DEFAULTS, and earlier migrations (054 recall, 072 replySuggestion,
+ * and platform overlays for the others) seeded such entries. Without this
+ * migration, upgraded workspaces would keep paying prompt-cache writes on
+ * call sites whose prompts never produce a cache read — only fresh installs
+ * would pick up the new shipped default.
+ *
+ * Idempotent and customization-preserving: an entry that already carries a
+ * `disableCache` key (either value — the user may have opted back in) is
+ * left untouched, and only the one key is added to entries we patch. Entries
+ * that don't exist on disk are skipped — those workspaces already resolve
+ * the shipped default.
+ */
+const ONE_SHOT_CALL_SITES = [
+  "recall",
+  "replySuggestion",
+  "homeGreeting",
+  "homeSuggestedPrompts",
+  "conversationTitle",
+  "memoryConsolidation",
+] as const;
+
+export const disableCacheOneShotCallsitesMigration: WorkspaceMigration = {
+  id: "099-disable-cache-one-shot-callsites",
+  description:
+    "Add disableCache: true to existing one-shot call-site entries so upgraded workspaces match the new shipped defaults",
+  run(workspaceDir: string): void {
+    if (process.env.VELLUM_DEFAULT_WORKSPACE_CONFIG_PATH) return;
+
+    const configPath = join(workspaceDir, "config.json");
+    if (!existsSync(configPath)) return;
+
+    let config: Record<string, unknown> = {};
+    try {
+      const raw = JSON.parse(readFileSync(configPath, "utf-8"));
+      if (!raw || typeof raw !== "object" || Array.isArray(raw)) return;
+      config = raw as Record<string, unknown>;
+    } catch {
+      return;
+    }
+
+    const llm = readObject(config.llm);
+    if (llm === null) return;
+    const callSites = readObject(llm.callSites);
+    if (callSites === null) return;
+
+    let changed = false;
+    for (const callSite of ONE_SHOT_CALL_SITES) {
+      const entry = readObject(callSites[callSite]);
+      if (entry === null) continue;
+      if ("disableCache" in entry) continue;
+      entry.disableCache = true;
+      callSites[callSite] = entry;
+      changed = true;
+    }
+    if (!changed) return;
+
+    llm.callSites = callSites;
+    config.llm = llm;
+    writeFileSync(configPath, JSON.stringify(config, null, 2) + "\n");
+  },
+  down(_workspaceDir: string): void {
+    // Forward-only.
+  },
+};
+
+function readObject(value: unknown): Record<string, unknown> | null {
+  if (value === null || typeof value !== "object" || Array.isArray(value)) {
+    return null;
+  }
+  return value as Record<string, unknown>;
+}

package/src/workspace/migrations/100-upgrade-quality-profile-to-fable-5.ts

@@ -0,0 +1,86 @@
+import { existsSync, readFileSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+
+import type { WorkspaceMigration } from "./types.js";
+
+// Upgrade quality-optimized profiles to Claude Fable 5.
+//
+// The quality-optimized model intent moved to claude-fable-5 (#34498), but
+// the intent map is only consulted when a profile is seeded: off-platform
+// installs reseed managed profiles on every boot, while on-platform
+// workspaces keep the profile they were hatched with — the seeder skips
+// existing profiles there, and the hatch overlay is consumed exactly once.
+// Existing platform assistants are therefore stuck on whichever model the
+// quality intent pointed at when they hatched (claude-opus-4-7 or
+// claude-opus-4-8) until a migration rewrites it.
+//
+// Only the managed quality-optimized profile is touched — the user-owned
+// custom-quality-optimized copy is theirs to manage. Within it, only a model
+// still on a previous quality-intent default is upgraded — matching by model
+// value also covers the OpenRouter-prefixed ids. A profile whose model the
+// user changed to anything else is left untouched.
+
+const TARGET_PROFILES = ["quality-optimized"];
+
+const MODEL_UPGRADES: Record<string, string> = {
+  "claude-opus-4-7": "claude-fable-5",
+  "claude-opus-4-8": "claude-fable-5",
+  "anthropic/claude-opus-4.7": "anthropic/claude-fable-5",
+  "anthropic/claude-opus-4.8": "anthropic/claude-fable-5",
+};
+
+export const upgradeQualityProfileToFable5Migration: WorkspaceMigration = {
+  id: "100-upgrade-quality-profile-to-fable-5",
+  description:
+    "Upgrade the managed quality-optimized profile from Opus 4.7/4.8 to Claude Fable 5",
+  run(workspaceDir: string): void {
+    const configPath = join(workspaceDir, "config.json");
+    if (!existsSync(configPath)) return;
+
+    let config: Record<string, unknown>;
+    try {
+      const raw = JSON.parse(readFileSync(configPath, "utf-8"));
+      if (!raw || typeof raw !== "object" || Array.isArray(raw)) return;
+      config = raw as Record<string, unknown>;
+    } catch {
+      return;
+    }
+
+    const llm = readObject(config.llm);
+    if (llm === null) return;
+
+    const profiles = readObject(llm.profiles);
+    if (profiles === null) return;
+
+    let changed = false;
+
+    for (const name of TARGET_PROFILES) {
+      const profile = readObject(profiles[name]);
+      if (profile === null) continue;
+      if (typeof profile.model !== "string") continue;
+
+      const upgraded = MODEL_UPGRADES[profile.model];
+      if (upgraded === undefined) continue;
+
+      profile.model = upgraded;
+      profiles[name] = profile;
+      changed = true;
+    }
+
+    if (changed) {
+      llm.profiles = profiles;
+      config.llm = llm;
+      writeFileSync(configPath, JSON.stringify(config, null, 2) + "\n");
+    }
+  },
+  down(_workspaceDir: string): void {
+    // Forward-only.
+  },
+};
+
+function readObject(value: unknown): Record<string, unknown> | null {
+  if (value === null || typeof value !== "object" || Array.isArray(value)) {
+    return null;
+  }
+  return value as Record<string, unknown>;
+}

package/src/workspace/migrations/101-upgrade-balanced-economy-to-minimax-m3.ts

@@ -0,0 +1,70 @@
+import { existsSync, readFileSync, writeFileSync } from "node:fs";
+import { join } from "node:path";
+
+import type { WorkspaceMigration } from "./types.js";
+
+// Switch the Balanced Economy profile from Kimi K2.6 to MiniMax M3.
+//
+// The Fireworks balanced model intent moved to MiniMax M3 (#34726), but the
+// intent map is only consulted when a profile is seeded: off-platform
+// installs reseed managed profiles on every boot, while on-platform
+// workspaces keep the profile they were hatched with — the seeder skips
+// existing profiles there. Existing platform assistants would otherwise stay
+// on Kimi K2.6 (including the now-dropped `suppress-cjk` logit-bias preset)
+// until a migration rewrites the profile.
+//
+// Only the managed balanced-economy profile still pointing at the old
+// default model is touched — a profile whose model the user changed is left
+// alone. The rewrite mirrors the new seed template: MiniMax M3, 32K max
+// output tokens, no logit-bias preset.
+
+const OLD_MODEL = "accounts/fireworks/models/kimi-k2p6";
+const NEW_MODEL = "accounts/fireworks/models/minimax-m3";
+
+export const upgradeBalancedEconomyToMinimaxM3Migration: WorkspaceMigration = {
+  id: "101-upgrade-balanced-economy-to-minimax-m3",
+  description:
+    "Switch the managed Balanced Economy profile from Kimi K2.6 to MiniMax M3",
+  run(workspaceDir: string): void {
+    const configPath = join(workspaceDir, "config.json");
+    if (!existsSync(configPath)) return;
+
+    let config: Record<string, unknown>;
+    try {
+      const raw = JSON.parse(readFileSync(configPath, "utf-8"));
+      if (!raw || typeof raw !== "object" || Array.isArray(raw)) return;
+      config = raw as Record<string, unknown>;
+    } catch {
+      return;
+    }
+
+    const llm = readObject(config.llm);
+    if (llm === null) return;
+
+    const profiles = readObject(llm.profiles);
+    if (profiles === null) return;
+
+    const profile = readObject(profiles["balanced-economy"]);
+    if (profile === null || profile.model !== OLD_MODEL) return;
+
+    profile.model = NEW_MODEL;
+    profile.description =
+      "Strong open model (MiniMax M3) at a lower price point";
+    profile.maxTokens = 32000;
+    delete profile.logitBias;
+    profiles["balanced-economy"] = profile;
+    llm.profiles = profiles;
+    config.llm = llm;
+    writeFileSync(configPath, JSON.stringify(config, null, 2) + "\n");
+  },
+  down(_workspaceDir: string): void {
+    // Forward-only.
+  },
+};
+
+function readObject(value: unknown): Record<string, unknown> | null {
+  if (value === null || typeof value !== "object" || Array.isArray(value)) {
+    return null;
+  }
+  return value as Record<string, unknown>;
+}

package/src/workspace/migrations/registry.ts

@@ -95,6 +95,10 @@ import { seedAvatarManifestMigration } from "./094-seed-avatar-manifest.js";
 import { bumpHeartbeatInterval30mTo60mMigration } from "./095-bump-heartbeat-interval-30m-to-60m.js";
 import { reduceQualityProfileEffortMigration } from "./096-reduce-quality-profile-effort.js";
 import { enableAdaptiveThinkingManagedProfilesMigration } from "./097-enable-adaptive-thinking-managed-profiles.js";
+import { removeStaleUpdatesBulletinFileMigration } from "./098-remove-stale-updates-bulletin-file.js";
+import { disableCacheOneShotCallsitesMigration } from "./099-disable-cache-one-shot-callsites.js";
+import { upgradeQualityProfileToFable5Migration } from "./100-upgrade-quality-profile-to-fable-5.js";
+import { upgradeBalancedEconomyToMinimaxM3Migration } from "./101-upgrade-balanced-economy-to-minimax-m3.js";
 import { migrateToWorkspaceVolumeMigration } from "./migrate-to-workspace-volume.js";
 import type { WorkspaceMigration } from "./types.js";
 
@@ -201,4 +205,8 @@ export const WORKSPACE_MIGRATIONS: WorkspaceMigration[] = [
   bumpHeartbeatInterval30mTo60mMigration,
   reduceQualityProfileEffortMigration,
   enableAdaptiveThinkingManagedProfilesMigration,
+  removeStaleUpdatesBulletinFileMigration,
+  disableCacheOneShotCallsitesMigration,
+  upgradeQualityProfileToFable5Migration,
+  upgradeBalancedEconomyToMinimaxM3Migration,
 ];