Merge branch 'codex/investigate-desktop-first-token-latency'

desktop/src/stores/settingsStore.ts

@@ -39,7 +39,7 @@ type SettingsStore = {
 export const useSettingsStore = create<SettingsStore>((set, get) => ({
   permissionMode: 'default',
   currentModel: null,
-  effortLevel: 'high',
+  effortLevel: 'medium',
   availableModels: [],
   activeProviderName: null,
   locale: getStoredLocale(),

src/main.tsx

@@ -2723,16 +2723,30 @@ async function run(): Promise<CommanderCommand> {
           }));
         }, configs).catch(err => logForDebugging(`[MCP] ${label} connect error: ${err}`));
       };
-      // Await all MCP configs — print mode is often single-turn, so
-      // "late-connecting servers visible next turn" doesn't help. SDK init
-      // message and turn-1 tool list both need configured MCP tools present.
+      // Await all MCP configs for regular print-mode sessions — they're
+      // often single-turn, so "late-connecting servers visible next turn"
+      // doesn't help. SDK init message and turn-1 tool list should include
+      // configured MCP tools when running plain `claude -p`.
+      //
+      // Desktop/bridge sessions (`--sdk-url`) are different: the user is
+      // waiting on the first visible assistant token, not a fully populated
+      // MCP inventory. Blocking startup on slow or failing MCP servers (for
+      // example a local `chatlog` HTTP endpoint timing out) makes the whole
+      // chat feel frozen before the first word appears. In that mode we still
+      // kick off the connections immediately, but let them settle in the
+      // background so headlessStore picks them up after the session starts.
+      //
       // Zero-server case is free via the early return in connectMcpBatch.
       // Connectors parallelize inside getMcpToolsCommandsAndResources
       // (processBatched with Promise.all). claude.ai is awaited too — its
       // fetch was kicked off early (line ~2558) so only residual time blocks
       // here. --bare skips claude.ai entirely for perf-sensitive scripts.
       profileCheckpoint('before_connectMcp');
-      await connectMcpBatch(regularMcpConfigs, 'regular');
+      if (sdkUrl) {
+        void connectMcpBatch(regularMcpConfigs, 'regular');
+      } else {
+        await connectMcpBatch(regularMcpConfigs, 'regular');
+      }
       profileCheckpoint('after_connectMcp');
       // Dedup: suppress plugin MCP servers that duplicate a claude.ai
       // connector (connector wins), then connect claude.ai servers.

src/server/__tests__/conversation-service.test.ts

@@ -170,4 +170,18 @@ describe('ConversationService', () => {
       expect(args[0]).toContain(path.join('bin', 'claude-haha'))
     }
   })
+
+  test('buildSessionCliArgs enables partial assistant messages for desktop streaming', () => {
+    const service = new ConversationService() as any
+    const args = service.buildSessionCliArgs(
+      '123e4567-e89b-12d3-a456-426614174000',
+      'ws://127.0.0.1:3456/sdk/test-session?token=test-token',
+      false,
+      { permissionMode: 'bypassPermissions' },
+    ) as string[]
+
+    expect(args).toContain('--include-partial-messages')
+    expect(args).toContain('--sdk-url')
+    expect(args).toContain('--replay-user-messages')
+  })
 })

src/server/api/models.ts

@@ -38,7 +38,7 @@ const DEFAULT_MODELS = [
 const EFFORT_LEVELS = ['low', 'medium', 'high', 'max'] as const
 
 const DEFAULT_MODEL = 'claude-opus-4-7'
-const DEFAULT_EFFORT = 'max'
+const DEFAULT_EFFORT = 'medium'
 
 const settingsService = new SettingsService()
 const providerService = new ProviderService()

src/server/services/conversationService.ts

@@ -64,6 +64,33 @@ export class ConversationStartupError extends Error {
 export class ConversationService {
   private sessions = new Map<string, SessionProcess>()
 
+  private buildSessionCliArgs(
+    sessionId: string,
+    sdkUrl: string,
+    shouldResume: boolean,
+    options?: SessionStartOptions,
+  ): string[] {
+    const dangerousMode = process.env.CLAUDE_DANGEROUS_MODE === '1'
+    return this.resolveCliArgs([
+      '--print',
+      '--verbose',
+      '--sdk-url',
+      sdkUrl,
+      '--enable-auth-status',
+      '--input-format',
+      'stream-json',
+      '--output-format',
+      'stream-json',
+      // Desktop chat depends on partial assistant deltas; without this the
+      // server only sees the completed assistant message at turn end.
+      '--include-partial-messages',
+      ...(shouldResume ? ['--resume', sessionId] : ['--session-id', sessionId]),
+      '--replay-user-messages',
+      ...this.getRuntimeArgs(options),
+      ...this.getPermissionArgs(options?.permissionMode, dangerousMode),
+    ])
+  }
+
   async startSession(
     sessionId: string,
     workDir: string,
@@ -88,22 +115,12 @@ export class ConversationService {
       )
     }
 
-    const dangerousMode = process.env.CLAUDE_DANGEROUS_MODE === '1'
-    const args = this.resolveCliArgs([
-      '--print',
-      '--verbose',
-      '--sdk-url',
+    const args = this.buildSessionCliArgs(
+      sessionId,
       sdkUrl,
-      '--enable-auth-status',
-      '--input-format',
-      'stream-json',
-      '--output-format',
-      'stream-json',
-      ...(shouldResume ? ['--resume', sessionId] : ['--session-id', sessionId]),
-      '--replay-user-messages',
-      ...this.getRuntimeArgs(options),
-      ...this.getPermissionArgs(options?.permissionMode, dangerousMode),
-    ])
+      shouldResume,
+      options,
+    )
 
     console.log(
       `[ConversationService] Starting CLI for ${sessionId}, cwd: ${workDir} (process.cwd()=${process.cwd()}, CALLER_DIR will be pinned to workDir)`,