MocA-Love · MocA-Love · Apr 18, 2026 · Apr 18, 2026
diff --git a/apps/desktop/package.json b/apps/desktop/package.json
@@ -73,6 +73,7 @@
 		"@hono/node-server": "^1.14.1",
 		"@hookform/resolvers": "^5.2.2",
 		"@lezer/highlight": "^1.2.3",
+		"@mastra/core": "1.25.0",
 		"@parcel/watcher": "^2.5.6",
 		"@pierre/diffs": "1.1.3",
 		"@radix-ui/react-dialog": "^1.1.15",
@@ -200,7 +201,7 @@
 		"lowdb": "^7.0.1",
 		"lowlight": "^3.3.0",
 		"lucide-react": "^0.563.0",
-		"mastracode": "0.9.2",
+		"mastracode": "0.14.0",
 		"nanoid": "^5.1.6",
 		"node-addon-api": "^7.1.0",
 		"node-pty": "1.1.0",

diff --git a/apps/desktop/plans/done/20260415-v2-host-service-ai-branch-naming-test-plan.md b/apps/desktop/plans/done/20260415-v2-host-service-ai-branch-naming-test-plan.md
@@ -0,0 +1,86 @@
+# Manual Testing Plan — PR #3517
+
+## Prerequisites
+- Desktop dev running (`bun dev` from apps/desktop, or full `bun dev` from root)
+- At least one project configured with a git repo
+
+## 1. v1 AI Branch Naming (API key path)
+
+**Setup**: `ANTHROPIC_API_KEY` or `OPENAI_API_KEY` set in env (or stored via Settings > Models).
+
+| Step | Expected |
+|---|---|
+| Open v1 new-workspace modal (Cmd+N) | Modal opens |
+| Type a prompt: "fix dropdown alignment bug" | Text entered |
+| Submit (Enter or click Create) | Modal closes, pending workspace shows "Generating branch…" briefly |
+| Wait for workspace to initialize | Branch name is AI-generated kebab-case (e.g. `fix-dropdown-alignment`), not random words |
+| Check worktree | Branch exists locally |
+
+## 2. v1 AI Branch Naming (no credentials)
+
+**Setup**: unset `ANTHROPIC_API_KEY` and `OPENAI_API_KEY` from env. No stored API keys in Settings > Models.
+
+| Step | Expected |
+|---|---|
+| Create workspace with prompt | Branch name falls back to random friendly name (e.g. `pickle-streetcar`) or prompt-derived slug |
+| No error toast | Degradation is silent |
+
+## 3. v1 Workspace Auto-Rename
+
+**Setup**: API key available.
+
+| Step | Expected |
+|---|---|
+| Create workspace with prompt "refactor auth middleware" | Workspace title updates to AI-generated name (e.g. "Refactor Auth Middleware") after a few seconds |
+| If no API key available | Title falls back to prompt text or friendly name |
+
+## 4. Anthropic OAuth Auto-Refresh (from #3510)
+
+**Setup**: Anthropic OAuth configured (Claude Max). Requires waiting for token expiry or manual simulation.
+
+| Step | Expected |
+|---|---|
+| Sign in to Anthropic via OAuth in Settings > Models | "Active" badge appears |
+| Force-expire: edit `~/Library/Application Support/mastracode/auth.json`, set `anthropic.expires` to a past timestamp | — |
+| Send a chat message | Chat succeeds silently (token auto-refreshed via `authStorage.getApiKey`). No "Reconnect" prompt. |
+| If refresh token is also invalid | Falls to expired state, "Expired" badge + "Reconnect" button appears |
+| Check terminal for `[chat-service] Anthropic OAuth refresh failed` | Logged if refresh fails |
+
+## 5. Settings > Models Page
+
+| Step | Expected |
+|---|---|
+| Navigate to Settings > Models | Page loads with Anthropic + OpenAI sections, each with provider icon in header |
+| Each provider shows a single card with OAuth row + API Key row | OAuth row: label + badge + action. API Key row: input + contextual buttons |
+| **Disconnected state** | "Not connected" badge, primary "Connect" button, no Save/Clear buttons |
+| **API key flow**: type key → Save appears → click Save | "API key updated" toast, "Active" badge, "Logout" button appears |
+| **API key flow**: click Clear | Key removed, badge reverts to "Not connected" |
+| **OAuth flow**: click Connect → complete in browser | "Active" badge, "Logout" button |
+| **OAuth flow**: click Logout | Badge reverts, Connect button returns |
+| **API key + OAuth**: set API key, then connect OAuth, then disconnect OAuth | API key should survive the OAuth cycle (backup/restore workaround) |
+| **OpenAI dialog** auto-opens browser on Connect | No manual "Open browser" step needed |
+| **Copy URL** button shows "Copied!" feedback for 2s | — |
+
+## 6. Production Build
+
+| Step | Expected |
+|---|---|
+| `bun run compile:app` (from apps/desktop) | Succeeds. `get-small-model` chunk ~1.2 MB, no 20 MB chunk. |
+| `bun run copy:native-modules` | Succeeds |
+| `bun run validate:native-runtime` | All checks pass |
+| `npx electron dist/main/index.js` | Main process boots (renderer 404 expected in non-packaged mode). No onnxruntime error. |
+
+## 7. Host-Service Procedure (dormant — future v2)
+
+Not yet wired to UI. Verify via tRPC playground or direct call if available:
+
+| Step | Expected |
+|---|---|
+| Call `workspaceCreation.generateBranchName({ projectId, prompt: "fix auth bug" })` | Returns `{ branchName: "fix-auth-bug" }` or similar (requires API key in host-service env) |
+| Call with empty prompt | Returns `{ branchName: null }` |
+| Call with no API key in env | Returns `{ branchName: null }` (graceful fallback) |
+
+## Known Regressions (documented, accepted)
+
+- **OAuth-only users** (Claude Max / OpenAI Codex without stored API key) get random branch names and prompt-derived workspace titles for small-model tasks. Main chat retains full OAuth.
+- **Upstream dependency**: API key storage slot collision with OAuth is worked around via backup/restore. Proper fix tracked at mastra-ai/mastra#15483.
diff --git a/apps/desktop/plans/done/20260415-v2-host-service-ai-branch-naming.md b/apps/desktop/plans/done/20260415-v2-host-service-ai-branch-naming.md
@@ -0,0 +1,168 @@
+# V2 Workspace Modal — Host-Service AI Branch Naming
+
+Port v1's AI branch-name generation into v2's workspace modal, routed through host-service. Approach: **use upstream `mastracode`'s `resolveModel`** via a lightweight `createMastraCode({ disableMcp: true, disableHooks: true })` singleton. Delete our small-model abstraction; keep OAuth parity (Claude Max + Codex) because mastracode handles it internally.
+
+## Completed
+
+- ✅ Bumped `mastracode` 0.9.2 → **0.14.0** (+ transitive `@mastra/core` 1.16 → 1.25). Typecheck + tests green. Removed `minimumReleaseAge` from `bunfig.toml`.
+
+## Target architecture
+
+```
+v2 useSubmitWorkspace
+  └─> client.workspaceCreation.generateBranchName.mutate({ projectId, prompt })
+        └─> generateBranchNameFromPrompt(...)                       [host-service]
+              └─> getSmallModel()                                    [shared helper]
+                    └─> resolveModel(modelId)                        from mastracode
+                          (full auth: API-key + keychain + OAuth middleware)
+```
+
+Desktop v1's existing `ai-branch-name.ts` migrates to the same `getSmallModel` helper — single implementation, two consumers.
+
+## Shared helper
+
+`packages/chat/src/server/shared/small-model/get-small-model.ts`:
+
+```ts
+import { createAuthStorage, createMastraCode } from "mastracode";
+import type { MastraLanguageModel } from "@mastra/core/llm";
+
+const ANTHROPIC_SMALL = "anthropic/claude-haiku-4-5-20251001";
+const OPENAI_SMALL = "openai/gpt-4o-mini";
+
+type Resolver = Awaited<ReturnType<typeof createMastraCode>>["resolveModel"];
+let initPromise: Promise<Resolver> | null = null;
+
+function getResolver(): Promise<Resolver> {
+  if (!initPromise) {
+    initPromise = createMastraCode({ disableMcp: true, disableHooks: true })
+      .then((r) => r.resolveModel);
+  }
+  return initPromise;
+}
+
+function pickSmallModelId(): string | null {
+  const auth = createAuthStorage();
+  auth.reload();
+  if (auth.has("anthropic")) return ANTHROPIC_SMALL;
+  if (auth.has("openai")) return OPENAI_SMALL;
+  return null;
+}
+
+export async function getSmallModel(): Promise<MastraLanguageModel | null> {
+  const modelId = pickSmallModelId();
+  if (!modelId) return null;
+  const resolveModel = await getResolver();
+  return resolveModel(modelId) as MastraLanguageModel;
+}
+```
+
+Module-level promise caches the mastracode init (one-time cost per process). Credential check is per-call (cheap, in-memory).
+
+## Code-removal budget
+
+| File | LOC | Fate |
+|---|---|---|
+| `apps/desktop/src/lib/ai/call-small-model.ts` | 184 | delete |
+| `apps/desktop/src/lib/ai/call-small-model.test.ts` | 399 | delete |
+| `apps/desktop/src/lib/ai/provider-diagnostics.ts` | 89 | delete if no other consumer |
+| `packages/chat/src/server/desktop/small-model/small-model.ts` | 146 | delete |
+| `packages/chat/src/server/desktop/small-model/small-model.test.ts` | 391 | delete |
+| `packages/chat/src/server/desktop/title-generation/title-generation.ts` | 99 | trim (~50, drop streaming variant) |
+| `packages/chat/src/server/desktop/auth/anthropic/anthropic.ts` | 232 | trim (~50, keep OAuth login helpers chat-service uses) |
+| `packages/chat/src/server/desktop/auth/openai/openai.ts` | 99 | trim (~30) |
+| `apps/desktop/src/lib/trpc/routers/workspaces/utils/ai-branch-name.ts` | 117 | rewrite → ~60 |
+| New `shared/small-model/get-small-model.ts` | — | +50 |
+
+Net: **~1200 lines removed**.
+
+---
+
+## Step 1 — Shared helper + migrate v1 branch naming
+
+### Actionable tasks
+1. Create `packages/chat/src/server/shared/small-model/{get-small-model.ts, index.ts}` with the helper above.
+2. Update `packages/chat/src/server/desktop/index.ts` barrel if needed; new helper lives in `shared/` and is imported directly from `@superset/chat/server/shared/small-model` — no re-export from desktop.
+3. Rewrite `apps/desktop/src/lib/trpc/routers/workspaces/utils/ai-branch-name.ts`:
+   - Replace `callSmallModel` + provider branching with `getSmallModel()` + `generateText({ model, system, prompt })`.
+   - Keep `BRANCH_NAME_INSTRUCTIONS`, `resolveConflict`, `sanitizeBranchNameWithMaxLength`.
+4. Grep for `callSmallModel`, `SmallModelProvider`, `getDefaultSmallModelProviders`, `generateTitleFromMessageWithStreamingModel`:
+   - Rewrite each consumer to `getSmallModel` + `generateText` (or Mastra Agent if the caller wants tracing).
+5. Delete:
+   - `apps/desktop/src/lib/ai/call-small-model.ts` + test.
+   - `packages/chat/src/server/desktop/small-model/small-model.ts` + test + `index.ts`.
+   - `generateTitleFromMessageWithStreamingModel` from `title-generation.ts`.
+6. `apps/desktop/src/lib/ai/provider-diagnostics.ts` — grep for consumers; delete if only `call-small-model.ts` uses it. Otherwise leave.
+7. Audit `auth/anthropic` and `auth/openai`: keep exports chat-service uses for OAuth login UI; delete any credential-resolution helpers used only for small-model.
+8. Run `bun run typecheck` + focused tests (chat-service, ai-branch-name). Fix breaks.
+9. Smoke: launch desktop, create v1 workspace with a prompt, verify AI branch naming still works (both API key and OAuth paths).
+
+### Risks (step 1)
+- **mastracode init side effects**: `createMastraCode` with disabled MCP/hooks still initializes storage, auto-detects project, etc. Confirm startup stays under ~200ms and doesn't create unwanted files. If it tries to touch a DB/libsql, pass an explicit `storage` config.
+- **Second init conflict**: chat service already calls `createMastraCode` for its runtime. Running a second one for small-model might duplicate auth-storage singletons or compete for files. Mitigation: verify `createMastraCode` is side-effect-safe when called twice; if not, share the existing chat runtime's resolver.
+- **Credential regression**: `authStorage.has("anthropic")` must cover all the "credential present" cases our current `getAnthropicCredentialsFromAnySource` covers (env vars, stored API keys, OAuth). Audit before replacing.
+
+---
+
+## Step 2 — Host-service procedure
+
+### Actionable tasks
+1. Port `sanitizeBranchNameWithMaxLength` (`apps/desktop/src/shared/utils/branch.ts`) and `resolveBranchPrefix` (`apps/desktop/src/lib/trpc/routers/workspaces/utils/branch-prefix.ts`) into `packages/host-service/src/trpc/router/workspace-creation/utils/`.
+2. Create `packages/host-service/src/trpc/router/workspace-creation/utils/ai-branch-name.ts` — same helper as desktop's rewritten v1, imports `getSmallModel` from `@superset/chat/server/shared/small-model`.
+3. Add to `workspace-creation.ts`:
+   ```ts
+   generateBranchName: publicProcedure
+     .input(z.object({ projectId: z.string(), prompt: z.string() }))
+     .mutation(async ({ input }) => {
+       const trimmed = input.prompt.trim();
+       if (!trimmed) return { branchName: null };
+       const project = /* existing project lookup */;
+       const existingBranches = /* existing branch listing */;
+       const prefix = await resolveBranchPrefix(project, existingBranches);
+       const branchName = await generateBranchNameFromPrompt(trimmed, existingBranches, prefix);
+       return { branchName };
+     }),
+   ```
+4. Delete `packages/host-service/src/providers/model-providers/LocalModelProvider/utils/resolveAnthropicCredential.ts` + `resolveOpenAICredential.ts` if unused after step (LocalModelProvider no longer needs them since auth flows through mastracode).
+5. Run typecheck + host-service tests.
+
+---
+
+## Step 3 — Wire v2
+
+### Actionable tasks
+1. Update `apps/desktop/src/renderer/routes/_authenticated/components/DashboardNewWorkspaceModal/components/DashboardNewWorkspaceForm/PromptGroup/hooks/useSubmitWorkspace/useSubmitWorkspace.ts`:
+   - Compute `willGenerateAIName = !draft.branchNameEdited && !!trimmedPrompt && !draft.linkedPR`.
+   - Fallback via `resolveNames(draft)` (unchanged).
+   - Insert pending row with status `"generating-branch"` if `willGenerateAIName`.
+   - Close + navigate (unchanged).
+   - If `willGenerateAIName`, race `client.workspaceCreation.generateBranchName.mutate(...)` vs 30s timeout:
+     - success → update pending row `branchName` + status `"creating"`.
+     - auth error → toast + abort + remove pending row.
+     - other/timeout → toast `"Using random branch name..."`, keep fallback name.
+   - Call `client.workspaceCreation.create(...)` with resolved `branchName`.
+2. Add `"generating-branch"` to `pendingWorkspaces` status union (`packages/local-db/src/schema/schema.ts`). Drizzle migration.
+3. Update pending page UI (`apps/desktop/src/renderer/routes/_authenticated/_dashboard/pending/$pendingId/page.tsx`) to render "Naming your branch…" for that status.
+
+---
+
+## Effort
+
+| Step | Effort |
+|---|---|
+| 0. mastracode upgrade | ✅ done |
+| 1. Shared helper + v1 migration + deletions | 2–3 hrs |
+| 2. Host-service procedure | 1–1.5 hrs |
+| 3. v2 wiring + pending UI | 1–2 hrs |
+| **Remaining** | **~4–6.5 hrs** |
+
+## Risks
+
+- **mastracode init side effects** at singleton init (see step 1).
+- **Remote host-service API-key availability**: remote hosts need `ANTHROPIC_API_KEY` / `OPENAI_API_KEY` set; otherwise v2 on remote hosts falls back to random-name. Document.
+- **OAuth parity in host-service**: host-service can't do an interactive OAuth flow. `createAuthStorage().loadStoredApiKeysIntoEnv(...)` loads stored API keys but NOT OAuth tokens into env. For host-service, OAuth-only users get random names.
+- **Diagnostics UI**: removing `provider-diagnostics.ts` removes mid-call `reportProviderIssue` signals. Audit settings UI for providers; they may source signals from chat-service regardless.
+
+## Out of scope
+- Live/debounced ghost suggestion in v2 branch-name input.
+- Retiring v1's desktop-tRPC `generateBranchName` procedure (it becomes a proxy over the shared helper; deleting it is a follow-up).
diff --git a/apps/desktop/plans/done/20260417-fix-api-key-storage-slot.md b/apps/desktop/plans/done/20260417-fix-api-key-storage-slot.md
@@ -0,0 +1,57 @@
+# Fix: API keys overwritten by OAuth connect/disconnect cycle
+
+## Problem
+
+Settings > Models "API key" field writes to the same auth.json slot as OAuth. When a user:
+1. Saves an API key → `authStorage.set("anthropic", { type: "api_key", key: "sk-..." })`
+2. Connects OAuth → `authStorage.login("anthropic", ...)` overwrites with `{ type: "oauth", ... }`
+3. Disconnects OAuth → `authStorage.remove("anthropic")` deletes everything
+
+The API key is lost. The model picker shows "disabled" even though the user saved a key.
+
+Chat still works because `createMastraCode`'s model resolution reads from env vars / external config independently of this status check.
+
+## Root cause
+
+`setApiKeyForProvider` uses `authStorage.set(providerId, credential)` which writes to the main provider slot. OAuth also writes to the same slot. They collide.
+
+mastracode's `AuthStorage` has **two separate storage mechanisms**:
+- `set(providerId, credential)` / `get(providerId)` → main slot (`"anthropic"` in auth.json)
+- `setStoredApiKey(providerId, key)` / `getStoredApiKey(providerId)` → dedicated API key slot (`"apikey:anthropic"` in auth.json)
+
+We're using the wrong one for API keys.
+
+## Fix
+
+### `auth-storage-utils.ts`
+
+**`setApiKeyForProvider`**: switch from `authStorage.set()` to `authStorage.setStoredApiKey()`.
+
+**`clearApiKeyForProvider`**: clear the `apikey:` slot. Use `authStorage.set("apikey:<providerId>", ...)` with a removal, or check `hasStoredApiKey` and handle accordingly. Since mastracode doesn't expose `removeStoredApiKey`, use `authStorage.remove("apikey:<providerId>")`.
+
+**`resolveAuthMethodForProvider`**: after checking the main slot, also check `authStorage.hasStoredApiKey(providerId)` as a fallback → return `"api_key"`.
+
+### `chat-service.ts`
+
+No changes needed — `getAnthropicAuthStatus` and `getOpenAIAuthStatus` already delegate to `resolveAuthMethodForProvider` which will now find stored API keys.
+
+The `setStoredAnthropicApiKeyFromEnvVariables` helper in `disconnectAnthropicOAuth` should also use `setStoredApiKey` for consistency, but it's less critical since it reads from the env config file.
+
+## Behavior after fix
+
+| Action | `"anthropic"` (main) | `"apikey:anthropic"` (dedicated) |
+|---|---|---|
+| Save API key (Settings) | unchanged | written |
+| Connect OAuth | overwritten with OAuth | survives |
+| Disconnect OAuth | removed | survives |
+| Auth status check | reads both | ← |
+
+## Side effect: small-model tasks
+
+`getSmallModel` reads `apikey:anthropic` from auth.json directly. Currently, API keys saved via Settings go to the main `"anthropic"` slot, so `getSmallModel` doesn't find them. After this fix, saved API keys land in `apikey:anthropic` where `getSmallModel` already looks → branch naming works for Settings-saved keys without any additional change.
+
+## Scope
+
+- `packages/chat/src/server/desktop/chat-service/auth-storage-utils.ts` (~15 LOC changed)
+- `packages/chat/src/server/desktop/chat-service/chat-service.ts` — `setStoredAnthropicApiKeyFromEnvVariables` updated for consistency (~2 LOC)
+- Tests in `chat-service.test.ts` if any mock `setApiKeyForProvider` behavior
diff --git a/apps/desktop/runtime-dependencies.ts b/apps/desktop/runtime-dependencies.ts
@@ -105,6 +105,11 @@ const packagedSupportModules = [
 export const mainExternalizedDependencies = [
 	...externalizedRuntimeModules.map((module) => module.specifier),
 	"pg-native",
+	// mastracode transitively loads @mastra/fastembed → onnxruntime-node, whose
+	// native binding is loaded via a dynamic `require` that @rollup/plugin-commonjs
+	// can't resolve at bundle time. Externalizing lets Node handle the require at
+	// runtime from node_modules. Also keeps the bundle size sane (~20 MB chunk).
+	"mastracode",
 ];
 
 export const packagedNodeModuleCopies = [