From fd49334c3369af8805b167ed99a3d958133d92f0 Mon Sep 17 00:00:00 2001
From: Arjun <6592213+arkml@users.noreply.github.com>
Date: Sun, 5 Jul 2026 21:30:04 +0530
Subject: [PATCH] move local-model concurrency to a turn-level defer lock

Review feedback: drop the model-layer scheduler (per-call-site
interactive/classifier/background priorities, local-provider hostname
heuristic) in favor of app-level orchestration:

- deferBackgroundTasks flag in models.json, surfaced as a settings
  toggle; auto-enabled once (UI logic) when the user connects Ollama
- ChatActivity counter marked by both chat runtimes (sessions layer and
  legacy AgentRuntime.trigger)
- startWhenPossible/runWhenPossible wrappers around the headless agent
  runner; all background invocations (knowledge pipeline, live notes,
  background tasks, scheduled + prebuilt agents) go through them and
  wait for chat-idle when the flag is set

createLanguageModel keeps only the Ollama context-window middleware;
the LM Studio capability probe now keys off the provider flavor instead
of hostname sniffing.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 .../src/components/settings-dialog.tsx        |  37 ++-
 .../core/src/agent-schedule/runner.ts         |   4 +-
 .../packages/core/src/agents/headless-app.ts  |  28 ++
 apps/x/packages/core/src/agents/runtime.ts    |   8 +-
 .../core/src/application/lib/builtin-tools.ts |   3 +-
 .../src/application/lib/chat-activity.test.ts |  50 ++++
 .../core/src/application/lib/chat-activity.ts |  39 +++
 .../src/background-tasks/event-consumer.ts    |   2 +-
 .../core/src/background-tasks/runner.ts       |  10 +-
 .../core/src/knowledge/agent_notes.ts         |   4 +-
 .../core/src/knowledge/build_graph.ts         |   6 +-
 .../core/src/knowledge/classify_thread.ts     |   2 +-
 .../knowledge/email_importance_feedback.ts    |   2 +-
 .../core/src/knowledge/inline_tasks.ts        |   8 +-
 .../core/src/knowledge/label_emails.ts        |   4 +-
 .../src/knowledge/live-note/event-consumer.ts |   2 +-
 .../core/src/knowledge/live-note/runner.ts    |  10 +-
 .../core/src/knowledge/meeting_prep_brief.ts  |   2 +-
 .../core/src/knowledge/summarize_meeting.ts   |   2 +-
 .../packages/core/src/knowledge/tag_notes.ts  |   4 +-
 apps/x/packages/core/src/models/defaults.ts   |  10 +
 apps/x/packages/core/src/models/local.test.ts | 123 +-------
 apps/x/packages/core/src/models/local.ts      | 278 +++---------------
 apps/x/packages/core/src/models/models.ts     |  17 +-
 apps/x/packages/core/src/models/repo.ts       |   5 +-
 apps/x/packages/core/src/pre_built/runner.ts  |   4 +-
 .../security/auto-permission-classifier.ts    |   2 +-
 apps/x/packages/core/src/sessions/sessions.ts |   9 +
 .../src/turns/bridges/real-model-registry.ts  | 224 +++++++-------
 .../packages/core/src/turns/model-registry.ts |   4 -
 apps/x/packages/core/src/turns/runtime.ts     |   5 -
 apps/x/packages/shared/src/ipc.ts             |   1 +
 apps/x/packages/shared/src/models.ts          |   5 +
 33 files changed, 373 insertions(+), 541 deletions(-)
 create mode 100644 apps/x/packages/core/src/application/lib/chat-activity.test.ts
 create mode 100644 apps/x/packages/core/src/application/lib/chat-activity.ts

diff --git a/apps/x/apps/renderer/src/components/settings-dialog.tsx b/apps/x/apps/renderer/src/components/settings-dialog.tsx
index 657b66b9..248c6cc6 100644
--- a/apps/x/apps/renderer/src/components/settings-dialog.tsx
+++ b/apps/x/apps/renderer/src/components/settings-dialog.tsx
@@ -364,6 +364,11 @@ function ModelSettings({ dialogOpen, rowboatConnected = false }: { dialogOpen: b
   const [testState, setTestState] = useState<{ status: "idle" | "testing" | "success" | "error"; error?: string }>({ status: "idle" })
   const [configLoading, setConfigLoading] = useState(true)
   const [showMoreProviders, setShowMoreProviders] = useState(false)
+  // "Defer background tasks while a chat is running" — a top-level
+  // models.json flag. deferExplicit tracks whether the user (or the Ollama
+  // auto-enable) has ever set it, so we only auto-enable once.
+  const [deferBackgroundTasks, setDeferBackgroundTasks] = useState(false)
+  const [deferExplicit, setDeferExplicit] = useState(false)
 
   const activeConfig = providerConfigs[provider]
   const showApiKey = provider === "openai" || provider === "anthropic" || provider === "google" || provider === "openrouter" || provider === "aigateway" || provider === "openai-compatible"
@@ -406,6 +411,8 @@ function ModelSettings({ dialogOpen, rowboatConnected = false }: { dialogOpen: b
           path: "config/models.json",
         })
         const parsed = JSON.parse(result.data)
+        setDeferBackgroundTasks(parsed?.deferBackgroundTasks === true)
+        setDeferExplicit(typeof parsed?.deferBackgroundTasks === "boolean")
         if (parsed?.provider?.flavor && parsed?.model) {
           const flavor = parsed.provider.flavor as LlmProviderFlavor
           setProvider(flavor)
@@ -462,6 +469,17 @@ function ModelSettings({ dialogOpen, rowboatConnected = false }: { dialogOpen: b
     loadCurrentConfig()
   }, [dialogOpen])
 
+  const handleDeferToggle = useCallback(async (value: boolean) => {
+    setDeferBackgroundTasks(value)
+    setDeferExplicit(true)
+    try {
+      await window.ipc.invoke("models:updateConfig", { deferBackgroundTasks: value })
+      window.dispatchEvent(new Event("models-config-changed"))
+    } catch {
+      toast.error("Failed to save setting")
+    }
+  }, [])
+
   // Load models catalog
   useEffect(() => {
     if (!dialogOpen) return
@@ -532,6 +550,12 @@ function ModelSettings({ dialogOpen, rowboatConnected = false }: { dialogOpen: b
         setDefaultProvider(provider)
         setTestState({ status: "success" })
         window.dispatchEvent(new Event('models-config-changed'))
+        // Local models compete with background agents for the same hardware:
+        // when the user connects Ollama and has never touched the defer
+        // flag, enable it for them (they can switch it off below).
+        if (provider === "ollama" && !deferExplicit && !deferBackgroundTasks) {
+          void handleDeferToggle(true)
+        }
         // Capability probe caveats (local models): saved, but the user should
         // know when the model can't do tools or has a too-small context.
         const warnings: string[] = result.warnings ?? []
@@ -551,7 +575,7 @@ function ModelSettings({ dialogOpen, rowboatConnected = false }: { dialogOpen: b
       setTestState({ status: "error", error: "Connection test failed" })
       toast.error("Connection test failed")
     }
-  }, [canTest, provider, activeConfig, rowboatConnected])
+  }, [canTest, provider, activeConfig, rowboatConnected, deferExplicit, deferBackgroundTasks, handleDeferToggle])
 
   const handleSetDefault = useCallback(async (prov: LlmProviderFlavor) => {
     const config = providerConfigs[prov]
@@ -938,6 +962,17 @@ function ModelSettings({ dialogOpen, rowboatConnected = false }: { dialogOpen: b
         </div>
       )}
 
+      {/* Defer background tasks while chatting */}
+      <div className="flex items-center justify-between gap-4 rounded-md border px-3 py-2.5">
+        <div className="min-w-0">
+          <div className="text-sm font-medium">Defer background tasks while chatting</div>
+          <div className="text-xs text-muted-foreground mt-0.5">
+            Background agents (knowledge sync, live notes, tasks) wait until your chat finishes. Recommended for local models.
+          </div>
+        </div>
+        <Switch checked={deferBackgroundTasks} onCheckedChange={handleDeferToggle} />
+      </div>
+
       {/* Test & Save button */}
       <Button
         onClick={handleTestAndSave}
diff --git a/apps/x/packages/core/src/agent-schedule/runner.ts b/apps/x/packages/core/src/agent-schedule/runner.ts
index bd818388..7db29a99 100644
--- a/apps/x/packages/core/src/agent-schedule/runner.ts
+++ b/apps/x/packages/core/src/agent-schedule/runner.ts
@@ -4,7 +4,7 @@ import { IAgentScheduleRepo } from "./repo.js";
 import { IAgentScheduleStateRepo } from "./state-repo.js";
 import { AgentScheduleConfig, AgentScheduleEntry } from "@x/shared/dist/agent-schedule.js";
 import { AgentScheduleState, AgentScheduleStateEntry } from "@x/shared/dist/agent-schedule-state.js";
-import { startHeadlessAgent } from "../agents/headless-app.js";
+import { startWhenPossible } from "../agents/headless-app.js";
 import { withUseCase } from "../analytics/use_case.js";
 import z from "zod";
 
@@ -165,7 +165,7 @@ async function runAgent(
         const startingMessage = entry.startingMessage ?? DEFAULT_STARTING_MESSAGE;
         const handle = await withUseCase(
             { useCase: 'copilot_chat', subUseCase: 'scheduled' },
-            () => startHeadlessAgent({
+            () => startWhenPossible({
                 agentId: agentName,
                 message: startingMessage,
                 signal: AbortSignal.timeout(TIMEOUT_MS),
diff --git a/apps/x/packages/core/src/agents/headless-app.ts b/apps/x/packages/core/src/agents/headless-app.ts
index d9b78f3a..7dd0e127 100644
--- a/apps/x/packages/core/src/agents/headless-app.ts
+++ b/apps/x/packages/core/src/agents/headless-app.ts
@@ -1,4 +1,6 @@
 import container from "../di/container.js";
+import { chatActivity } from "../application/lib/chat-activity.js";
+import { shouldDeferBackgroundTasks } from "../models/defaults.js";
 import {
     type HeadlessAgentHandle,
     type HeadlessAgentOptions,
@@ -22,4 +24,30 @@ export function runHeadlessAgent(
     return runner().run(options);
 }
 
+// When the user enabled "Defer background tasks while a chat is running"
+// (recommended for local models — a background run competes with the chat
+// for the same hardware), wait for the chat to go idle before starting.
+// Re-check after each wake: another chat turn may have started meanwhile.
+async function waitForChatIdleIfConfigured(): Promise<void> {
+    while ((await shouldDeferBackgroundTasks()) && chatActivity.activeCount > 0) {
+        await chatActivity.waitUntilIdle();
+    }
+}
+
+/** startHeadlessAgent for background work: honors the defer-while-chatting setting. */
+export async function startWhenPossible(
+    options: HeadlessAgentOptions,
+): Promise<HeadlessAgentHandle> {
+    await waitForChatIdleIfConfigured();
+    return startHeadlessAgent(options);
+}
+
+/** runHeadlessAgent for background work: honors the defer-while-chatting setting. */
+export async function runWhenPossible(
+    options: HeadlessAgentOptions,
+): Promise<HeadlessAgentResult & { turnId: string }> {
+    await waitForChatIdleIfConfigured();
+    return runHeadlessAgent(options);
+}
+
 export { toolInputPaths } from "./headless.js";
diff --git a/apps/x/packages/core/src/agents/runtime.ts b/apps/x/packages/core/src/agents/runtime.ts
index 357e0ff3..3793e858 100644
--- a/apps/x/packages/core/src/agents/runtime.ts
+++ b/apps/x/packages/core/src/agents/runtime.ts
@@ -20,6 +20,7 @@ import container from "../di/container.js";
 import { notifyIfEnabled } from "../application/notification/notifier.js";
 import { IModelConfigRepo } from "../models/repo.js";
 import { createLanguageModel } from "../models/models.js";
+import { chatActivity } from "../application/lib/chat-activity.js";
 import { resolveProviderConfig } from "../models/defaults.js";
 import { IAgentsRepo } from "./repo.js";
 import { IMonotonicallyIncreasingIdGenerator } from "../application/lib/id-gen.js";
@@ -496,6 +497,9 @@ export class AgentRuntime implements IAgentRuntime {
             return;
         }
         const signal = this.abortRegistry.createForRun(runId);
+        // Legacy runs are user-facing chats: mark activity so background
+        // agents can defer (see agents/headless-app.ts runWhenPossible).
+        chatActivity.enter();
         try {
             await this.bus.publish({
                 runId,
@@ -610,6 +614,7 @@ export class AgentRuntime implements IAgentRuntime {
             await this.runsRepo.appendEvents(runId, [errorEvent]);
             await this.bus.publish(errorEvent);
         } finally {
+            chatActivity.exit();
             this.abortRegistry.cleanup(runId);
             await this.runsLock.release(runId);
             await this.bus.publish({
@@ -1365,8 +1370,7 @@ export async function* streamAgent({
     }
     const modelId = state.runModel;
     const providerConfig = await resolveProviderConfig(state.runProvider);
-    // Legacy runs are user-facing chats: interactive priority on local models.
-    const model = createLanguageModel(providerConfig, modelId, { priority: "interactive" });
+    const model = createLanguageModel(providerConfig, modelId);
     logger.log(`using model: ${modelId} (provider: ${state.runProvider})`);
 
     // Install use-case context for tool-internal LLM calls (e.g. parseFile)
diff --git a/apps/x/packages/core/src/application/lib/builtin-tools.ts b/apps/x/packages/core/src/application/lib/builtin-tools.ts
index 8d4faad4..47b6b0c3 100644
--- a/apps/x/packages/core/src/application/lib/builtin-tools.ts
+++ b/apps/x/packages/core/src/application/lib/builtin-tools.ts
@@ -584,8 +584,7 @@ export const BuiltinTools: z.infer<typeof BuiltinToolsSchema> = {
 
                 const { model: modelId, provider: providerName } = await getDefaultModelAndProvider();
                 const providerConfig = await resolveProviderConfig(providerName);
-                // Runs as a tool inside a chat turn more often than not.
-                const model = createLanguageModel(providerConfig, modelId, { priority: 'interactive' });
+                const model = createLanguageModel(providerConfig, modelId);
 
                 const userPrompt = prompt || 'Convert this file to well-structured markdown.';
 
diff --git a/apps/x/packages/core/src/application/lib/chat-activity.test.ts b/apps/x/packages/core/src/application/lib/chat-activity.test.ts
new file mode 100644
index 00000000..4d9ea7f3
--- /dev/null
+++ b/apps/x/packages/core/src/application/lib/chat-activity.test.ts
@@ -0,0 +1,50 @@
+import { describe, expect, it } from "vitest";
+import { ChatActivity } from "./chat-activity.js";
+
+const tick = () => new Promise<void>((r) => setTimeout(r, 0));
+
+describe("ChatActivity", () => {
+    it("resolves waiters immediately when idle", async () => {
+        const activity = new ChatActivity();
+        await activity.waitUntilIdle();
+        expect(activity.activeCount).toBe(0);
+    });
+
+    it("blocks waiters until every active chat exits", async () => {
+        const activity = new ChatActivity();
+        activity.enter();
+        activity.enter();
+
+        let released = false;
+        const waiting = activity.waitUntilIdle().then(() => {
+            released = true;
+        });
+
+        activity.exit();
+        await tick();
+        expect(released).toBe(false);
+
+        activity.exit();
+        await waiting;
+        expect(released).toBe(true);
+    });
+
+    it("wakes all pending waiters at once", async () => {
+        const activity = new ChatActivity();
+        activity.enter();
+        const results: number[] = [];
+        const a = activity.waitUntilIdle().then(() => results.push(1));
+        const b = activity.waitUntilIdle().then(() => results.push(2));
+        activity.exit();
+        await Promise.all([a, b]);
+        expect(results.sort()).toEqual([1, 2]);
+    });
+
+    it("tolerates unbalanced exits", () => {
+        const activity = new ChatActivity();
+        activity.exit();
+        expect(activity.activeCount).toBe(0);
+        activity.enter();
+        expect(activity.activeCount).toBe(1);
+    });
+});
diff --git a/apps/x/packages/core/src/application/lib/chat-activity.ts b/apps/x/packages/core/src/application/lib/chat-activity.ts
new file mode 100644
index 00000000..9718b5e8
--- /dev/null
+++ b/apps/x/packages/core/src/application/lib/chat-activity.ts
@@ -0,0 +1,39 @@
+// Tracks whether any user-facing chat turn is currently being processed.
+// Both chat runtimes mark their turns here (sessions/sessions.ts for the
+// turns runtime, agents/runtime.ts trigger() for legacy runs). Background
+// agent invocations consult it via startWhenPossible/runWhenPossible in
+// agents/headless-app.ts when the user enabled "Defer background tasks
+// while a chat is running" — useful on local models, where a background run
+// competes with the chat for the same hardware.
+export class ChatActivity {
+    private active = 0;
+    private waiters: Array<() => void> = [];
+
+    enter(): void {
+        this.active++;
+    }
+
+    exit(): void {
+        this.active = Math.max(0, this.active - 1);
+        if (this.active === 0) {
+            const waiters = this.waiters.splice(0);
+            for (const waiter of waiters) {
+                waiter();
+            }
+        }
+    }
+
+    get activeCount(): number {
+        return this.active;
+    }
+
+    /** Resolves immediately when no chat turn is running. */
+    waitUntilIdle(): Promise<void> {
+        if (this.active === 0) {
+            return Promise.resolve();
+        }
+        return new Promise((resolve) => this.waiters.push(resolve));
+    }
+}
+
+export const chatActivity = new ChatActivity();
diff --git a/apps/x/packages/core/src/background-tasks/event-consumer.ts b/apps/x/packages/core/src/background-tasks/event-consumer.ts
index 53f010f2..e9212470 100644
--- a/apps/x/packages/core/src/background-tasks/event-consumer.ts
+++ b/apps/x/packages/core/src/background-tasks/event-consumer.ts
@@ -12,7 +12,7 @@ async function resolveRoutingModel() {
     const { model: modelId, provider } = await getBackgroundTaskAgentModel();
     const config = await resolveProviderConfig(provider);
     return {
-        model: createLanguageModel(config, modelId, { priority: 'classifier' }),
+        model: createLanguageModel(config, modelId),
         modelId,
         providerName: provider,
     };
diff --git a/apps/x/packages/core/src/background-tasks/runner.ts b/apps/x/packages/core/src/background-tasks/runner.ts
index b1a0bc8f..810b9aac 100644
--- a/apps/x/packages/core/src/background-tasks/runner.ts
+++ b/apps/x/packages/core/src/background-tasks/runner.ts
@@ -2,7 +2,7 @@ import type { BackgroundTask, BackgroundTaskTriggerType } from '@x/shared/dist/b
 import { PrefixLogger } from '@x/shared/dist/prefix-logger.js';
 import { fetchTask, patchTask, prependRunId } from './fileops.js';
 import { getBackgroundTaskAgentModel } from '../models/defaults.js';
-import { startHeadlessAgent } from '../agents/headless-app.js';
+import { startHeadlessAgent, startWhenPossible } from '../agents/headless-app.js';
 import { buildTriggerBlock } from '../agents/build-trigger-block.js';
 import { backgroundTaskBus } from './bus.js';
 import { withUseCase } from '../analytics/use_case.js';
@@ -143,13 +143,19 @@ export async function runBackgroundTask(
         const selection = await getBackgroundTaskAgentModel();
         const model = task.model || selection.model;
         const provider = task.provider ?? (task.model ? undefined : selection.provider);
+        // Manual runs are user-requested (the Run button, or the copilot's
+        // run-background-task-agent tool mid-chat) and must NOT wait for
+        // chat-idle: the requesting chat turn holds the chat-activity lock,
+        // so deferring here would deadlock the turn. Only autonomous
+        // triggers (cron/window/event) defer.
+        const start = trigger === 'manual' ? startHeadlessAgent : startWhenPossible;
         // Establish the use-case context for the whole turn so every tool the
         // agent calls (notably notify-user) reads `background_task_agent` via
         // getCurrentUseCase(); the AsyncLocalStorage context set here flows
         // through the turn's async execution chain.
         const handle = await withUseCase(
             { useCase: 'background_task_agent', subUseCase: trigger },
-            () => startHeadlessAgent({
+            () => start({
                 agentId: 'background-task-agent',
                 message: buildMessage(slug, task, trigger, context, codeProject),
                 model,
diff --git a/apps/x/packages/core/src/knowledge/agent_notes.ts b/apps/x/packages/core/src/knowledge/agent_notes.ts
index 892e3f36..3f25002d 100644
--- a/apps/x/packages/core/src/knowledge/agent_notes.ts
+++ b/apps/x/packages/core/src/knowledge/agent_notes.ts
@@ -2,7 +2,7 @@ import fs from 'fs';
 import path from 'path';
 import { google } from 'googleapis';
 import { WorkDir } from '../config/config.js';
-import { runHeadlessAgent } from '../agents/headless-app.js';
+import { runWhenPossible } from '../agents/headless-app.js';
 import { getKgModel } from '../models/defaults.js';
 import { getErrorDetails } from '../agents/utils.js';
 import { serviceLogger } from '../services/service_logger.js';
@@ -281,7 +281,7 @@ async function processAgentNotes(): Promise<void> {
         const timestamp = new Date().toISOString();
         const message = `Current timestamp: ${timestamp}\n\nProcess the following source material and update the Agent Notes folder accordingly.\n\n${messageParts.join('\n\n')}`;
 
-        await runHeadlessAgent({
+        await runWhenPossible({
             agentId: AGENT_ID,
             message,
             ...(await getKgModel()),
diff --git a/apps/x/packages/core/src/knowledge/build_graph.ts b/apps/x/packages/core/src/knowledge/build_graph.ts
index 741082fb..b50852ad 100644
--- a/apps/x/packages/core/src/knowledge/build_graph.ts
+++ b/apps/x/packages/core/src/knowledge/build_graph.ts
@@ -2,7 +2,7 @@ import fs from 'fs';
 import path from 'path';
 import { WorkDir } from '../config/config.js';
 import { getKgModel } from '../models/defaults.js';
-import { runHeadlessAgent, toolInputPaths } from '../agents/headless-app.js';
+import { runWhenPossible, toolInputPaths } from '../agents/headless-app.js';
 import { getErrorDetails } from '../agents/utils.js';
 import { serviceLogger, type ServiceRunContext } from '../services/service_logger.js';
 import {
@@ -385,7 +385,7 @@ async function createNotesFromBatch(
         message += `(3) No placeholder text ("Unknown"/"-") and no links between entities that didn't co-occur in one source file.\n`;
     }
 
-    const { turnId, state } = await runHeadlessAgent({
+    const { turnId, state } = await runWhenPossible({
         agentId: NOTE_CREATION_AGENT,
         message,
         ...(await getKgModel()),
@@ -943,7 +943,7 @@ export async function curateNotes(): Promise<void> {
             message += `Curate the following knowledge note per your instructions. Rewrite it in place with a single file-writeText to the SAME path.\n\n`;
             message += `**Note path:** ${relPath}\n\n`;
             message += `**Current content:**\n\n${content}\n`;
-            await runHeadlessAgent({
+            await runWhenPossible({
                 agentId: CURATION_AGENT,
                 message,
                 ...(await getKgModel()),
diff --git a/apps/x/packages/core/src/knowledge/classify_thread.ts b/apps/x/packages/core/src/knowledge/classify_thread.ts
index c83da43d..95728255 100644
--- a/apps/x/packages/core/src/knowledge/classify_thread.ts
+++ b/apps/x/packages/core/src/knowledge/classify_thread.ts
@@ -247,7 +247,7 @@ export async function classifyThread(
 
         const { model: modelId, provider } = await getKgModel();
         const config = await resolveProviderConfig(provider);
-        const model = createLanguageModel(config, modelId, { priority: 'background' });
+        const model = createLanguageModel(config, modelId);
 
         let systemPrompt = options.skipDraft
             ? `${SYSTEM_PROMPT}\n\n# Skip the draft\n\nThe user already has their own draft in progress for this thread — DO NOT generate a draftResponse. Always omit the draftResponse field.`
diff --git a/apps/x/packages/core/src/knowledge/email_importance_feedback.ts b/apps/x/packages/core/src/knowledge/email_importance_feedback.ts
index febdccb7..203f3403 100644
--- a/apps/x/packages/core/src/knowledge/email_importance_feedback.ts
+++ b/apps/x/packages/core/src/knowledge/email_importance_feedback.ts
@@ -147,7 +147,7 @@ export async function maybeDistillImportanceRules(): Promise<void> {
     try {
         const { model: modelId, provider } = await getKgModel();
         const config = await resolveProviderConfig(provider);
-        const model = createLanguageModel(config, modelId, { priority: 'background' });
+        const model = createLanguageModel(config, modelId);
 
         const correctionLines = fb.corrections.map(c =>
             `- From: ${c.from} | Subject: "${c.subject}" | classifier said ${c.agentVerdict}, user corrected to ${c.userVerdict}`
diff --git a/apps/x/packages/core/src/knowledge/inline_tasks.ts b/apps/x/packages/core/src/knowledge/inline_tasks.ts
index e856b3f6..71bdf38f 100644
--- a/apps/x/packages/core/src/knowledge/inline_tasks.ts
+++ b/apps/x/packages/core/src/knowledge/inline_tasks.ts
@@ -3,7 +3,7 @@ import path from 'path';
 import { CronExpressionParser } from 'cron-parser';
 import { generateText } from 'ai';
 import { WorkDir } from '../config/config.js';
-import { runHeadlessAgent } from '../agents/headless-app.js';
+import { runWhenPossible } from '../agents/headless-app.js';
 import { getKgModel } from '../models/defaults.js';
 import container from '../di/container.js';
 import type { IModelConfigRepo } from '../models/repo.js';
@@ -480,7 +480,7 @@ async function processInlineTasks(): Promise<void> {
                     '```',
                 ].join('\n');
 
-                const { summary: result } = await runHeadlessAgent({
+                const { summary: result } = await runWhenPossible({
                     agentId: INLINE_TASK_AGENT,
                     message,
                     ...(await getKgModel()),
@@ -559,7 +559,7 @@ export async function processRowboatInstruction(
         '```',
     ].join('\n');
 
-    const { summary: rawResponse } = await runHeadlessAgent({
+    const { summary: rawResponse } = await runWhenPossible({
         agentId: INLINE_TASK_AGENT,
         message,
         ...(await getKgModel()),
@@ -613,7 +613,7 @@ export async function processRowboatInstruction(
 export async function classifySchedule(instruction: string): Promise<InlineTaskSchedule | null> {
     const repo = container.resolve<IModelConfigRepo>('modelConfigRepo');
     const config = await repo.getConfig();
-    const model = createLanguageModel(config.provider, config.model, { priority: 'classifier' });
+    const model = createLanguageModel(config.provider, config.model);
 
     const now = new Date();
     const defaultEnd = new Date(now.getTime() + 7 * 24 * 60 * 60 * 1000);
diff --git a/apps/x/packages/core/src/knowledge/label_emails.ts b/apps/x/packages/core/src/knowledge/label_emails.ts
index cd71e826..bac565a1 100644
--- a/apps/x/packages/core/src/knowledge/label_emails.ts
+++ b/apps/x/packages/core/src/knowledge/label_emails.ts
@@ -1,7 +1,7 @@
 import fs from 'fs';
 import path from 'path';
 import { WorkDir } from '../config/config.js';
-import { runHeadlessAgent, toolInputPaths } from '../agents/headless-app.js';
+import { runWhenPossible, toolInputPaths } from '../agents/headless-app.js';
 import { getKgModel } from '../models/defaults.js';
 import { getErrorDetails } from '../agents/utils.js';
 import { serviceLogger } from '../services/service_logger.js';
@@ -84,7 +84,7 @@ async function labelEmailBatch(
         message += `\n\n---\n\n`;
     }
 
-    const { turnId, state } = await runHeadlessAgent({
+    const { turnId, state } = await runWhenPossible({
         agentId: LABELING_AGENT,
         message,
         ...(await getKgModel()),
diff --git a/apps/x/packages/core/src/knowledge/live-note/event-consumer.ts b/apps/x/packages/core/src/knowledge/live-note/event-consumer.ts
index 2b15bb23..e2aa7718 100644
--- a/apps/x/packages/core/src/knowledge/live-note/event-consumer.ts
+++ b/apps/x/packages/core/src/knowledge/live-note/event-consumer.ts
@@ -10,7 +10,7 @@ async function resolveRoutingModel() {
     const { model: modelId, provider } = await getLiveNoteAgentModel();
     const config = await resolveProviderConfig(provider);
     return {
-        model: createLanguageModel(config, modelId, { priority: 'classifier' }),
+        model: createLanguageModel(config, modelId),
         modelId,
         providerName: provider,
     };
diff --git a/apps/x/packages/core/src/knowledge/live-note/runner.ts b/apps/x/packages/core/src/knowledge/live-note/runner.ts
index 6a9442e1..e192a08e 100644
--- a/apps/x/packages/core/src/knowledge/live-note/runner.ts
+++ b/apps/x/packages/core/src/knowledge/live-note/runner.ts
@@ -1,7 +1,7 @@
 import type { LiveNote, LiveNoteTriggerType } from '@x/shared/dist/live-note.js';
 import { fetchLiveNote, patchLiveNote, readNoteBody } from './fileops.js';
 import { getLiveNoteAgentModel } from '../../models/defaults.js';
-import { startHeadlessAgent } from '../../agents/headless-app.js';
+import { startHeadlessAgent, startWhenPossible } from '../../agents/headless-app.js';
 import { withUseCase } from '../../analytics/use_case.js';
 import { buildTriggerBlock } from '../../agents/build-trigger-block.js';
 import { liveNoteBus } from './bus.js';
@@ -114,12 +114,18 @@ export async function runLiveNoteAgent(
         const selection = await getLiveNoteAgentModel();
         const model = live.model ?? selection.model;
         const provider = live.provider ?? (live.model ? undefined : selection.provider);
+        // Manual runs are user-requested (the Run button, or the copilot's
+        // run-live-note-agent tool mid-chat) and must NOT wait for chat-idle:
+        // the requesting chat turn holds the chat-activity lock, so deferring
+        // here would deadlock the turn. Only autonomous triggers
+        // (cron/window/event) defer.
+        const start = trigger === 'manual' ? startHeadlessAgent : startWhenPossible;
         // The use-case context propagates to every tool the agent calls; the
         // granular trigger doubles as the sub-use-case (manual / cron /
         // window / event) so dashboards can break down what woke the agent.
         const handle = await withUseCase(
             { useCase: 'live_note_agent', subUseCase: trigger },
-            () => startHeadlessAgent({
+            () => start({
                 agentId: 'live-note-agent',
                 message: buildMessage(filePath, live, trigger, context),
                 model,
diff --git a/apps/x/packages/core/src/knowledge/meeting_prep_brief.ts b/apps/x/packages/core/src/knowledge/meeting_prep_brief.ts
index d841c980..275d1d14 100644
--- a/apps/x/packages/core/src/knowledge/meeting_prep_brief.ts
+++ b/apps/x/packages/core/src/knowledge/meeting_prep_brief.ts
@@ -177,7 +177,7 @@ async function generateBrief(event: CalendarEvent, ctx: Awaited<ReturnType<typeo
 
     const { model: modelId, provider: providerName } = await getMeetingNotesModel();
     const providerConfig = await resolveProviderConfig(providerName);
-    const model = createLanguageModel(providerConfig, modelId, { priority: 'background' });
+    const model = createLanguageModel(providerConfig, modelId);
 
     const result = await withUseCase({ useCase: 'meeting_prep' }, () => generateText({
         model,
diff --git a/apps/x/packages/core/src/knowledge/summarize_meeting.ts b/apps/x/packages/core/src/knowledge/summarize_meeting.ts
index 44bccdab..29f3cf92 100644
--- a/apps/x/packages/core/src/knowledge/summarize_meeting.ts
+++ b/apps/x/packages/core/src/knowledge/summarize_meeting.ts
@@ -139,7 +139,7 @@ function loadCalendarEventContext(calendarEventJson: string): string {
 export async function summarizeMeeting(transcript: string, meetingStartTime?: string, calendarEventJson?: string): Promise<string> {
     const { model: modelId, provider: providerName } = await getMeetingNotesModel();
     const providerConfig = await resolveProviderConfig(providerName);
-    const model = createLanguageModel(providerConfig, modelId, { priority: 'background' });
+    const model = createLanguageModel(providerConfig, modelId);
 
     // If a specific calendar event was linked, use it directly.
     // Otherwise fall back to scanning events within ±3 hours.
diff --git a/apps/x/packages/core/src/knowledge/tag_notes.ts b/apps/x/packages/core/src/knowledge/tag_notes.ts
index faed39f5..be87b849 100644
--- a/apps/x/packages/core/src/knowledge/tag_notes.ts
+++ b/apps/x/packages/core/src/knowledge/tag_notes.ts
@@ -1,7 +1,7 @@
 import fs from 'fs';
 import path from 'path';
 import { WorkDir } from '../config/config.js';
-import { runHeadlessAgent, toolInputPaths } from '../agents/headless-app.js';
+import { runWhenPossible, toolInputPaths } from '../agents/headless-app.js';
 import { getKgModel } from '../models/defaults.js';
 import { getErrorDetails } from '../agents/utils.js';
 import { serviceLogger } from '../services/service_logger.js';
@@ -97,7 +97,7 @@ async function tagNoteBatch(
         message += `\n\n---\n\n`;
     }
 
-    const { turnId, state } = await runHeadlessAgent({
+    const { turnId, state } = await runWhenPossible({
         agentId: NOTE_TAGGING_AGENT,
         message,
         ...(await getKgModel()),
diff --git a/apps/x/packages/core/src/models/defaults.ts b/apps/x/packages/core/src/models/defaults.ts
index 82f178d5..da616c39 100644
--- a/apps/x/packages/core/src/models/defaults.ts
+++ b/apps/x/packages/core/src/models/defaults.ts
@@ -54,6 +54,16 @@ export async function getDefaultModelAndProvider(): Promise<{ model: string; pro
     return { model: cfg.model, provider: cfg.provider.flavor };
 }
 
+/**
+ * "Defer background tasks while a chat is running" (settings checkbox,
+ * models.json `deferBackgroundTasks`). Read at each background invocation so
+ * toggling takes effect immediately.
+ */
+export async function shouldDeferBackgroundTasks(): Promise<boolean> {
+    const cfg = await readConfig();
+    return cfg?.deferBackgroundTasks === true;
+}
+
 /**
  * Resolve a provider name (as stored on a run, an agent, or returned by
  * getDefaultModelAndProvider) into the full LlmProvider config that
diff --git a/apps/x/packages/core/src/models/local.test.ts b/apps/x/packages/core/src/models/local.test.ts
index d441164b..7973edf8 100644
--- a/apps/x/packages/core/src/models/local.test.ts
+++ b/apps/x/packages/core/src/models/local.test.ts
@@ -1,37 +1,5 @@
 import { afterEach, describe, expect, it, vi } from "vitest";
-import {
-    LocalLlmScheduler,
-    isLocalProvider,
-    makeOllamaThinkFetch,
-    resolveThinkValue,
-} from "./local.js";
-
-function deferred() {
-    let resolve!: () => void;
-    const promise = new Promise<void>((r) => {
-        resolve = r;
-    });
-    return { promise, resolve };
-}
-
-const tick = () => new Promise<void>((r) => setTimeout(r, 0));
-
-describe("isLocalProvider", () => {
-    it("treats ollama as local", () => {
-        expect(isLocalProvider({ flavor: "ollama" })).toBe(true);
-    });
-
-    it("treats loopback openai-compatible endpoints as local", () => {
-        expect(isLocalProvider({ flavor: "openai-compatible", baseURL: "http://localhost:1234/v1" })).toBe(true);
-        expect(isLocalProvider({ flavor: "openai-compatible", baseURL: "http://127.0.0.1:8080/v1" })).toBe(true);
-    });
-
-    it("treats remote openai-compatible endpoints and cloud flavors as non-local", () => {
-        expect(isLocalProvider({ flavor: "openai-compatible", baseURL: "https://api.together.xyz/v1" })).toBe(false);
-        expect(isLocalProvider({ flavor: "openai" })).toBe(false);
-        expect(isLocalProvider({ flavor: "rowboat" })).toBe(false);
-    });
-});
+import { makeOllamaThinkFetch, resolveThinkValue } from "./local.js";
 
 describe("resolveThinkValue", () => {
     it("passes effort levels straight through for gpt-oss variants", () => {
@@ -111,92 +79,3 @@ describe("makeOllamaThinkFetch", () => {
         expect(calls.some((c) => c.url.endsWith("/api/tags"))).toBe(true);
     });
 });
-
-describe("LocalLlmScheduler", () => {
-    it("serves waiters by priority, FIFO within a priority", async () => {
-        const scheduler = new LocalLlmScheduler(1);
-        const order: string[] = [];
-        const first = deferred();
-
-        const running = scheduler.run("background", undefined, async () => {
-            await first.promise;
-            order.push("initial");
-        });
-        await tick();
-
-        const bg1 = scheduler.run("background", undefined, async () => {
-            order.push("bg1");
-        });
-        const bg2 = scheduler.run("background", undefined, async () => {
-            order.push("bg2");
-        });
-        const chat = scheduler.run("interactive", undefined, async () => {
-            order.push("chat");
-        });
-        const classifier = scheduler.run("classifier", undefined, async () => {
-            order.push("classifier");
-        });
-        await tick();
-
-        first.resolve();
-        await Promise.all([running, bg1, bg2, chat, classifier]);
-        expect(order).toEqual(["initial", "chat", "classifier", "bg1", "bg2"]);
-    });
-
-    it("releases the slot when a task throws", async () => {
-        const scheduler = new LocalLlmScheduler(1);
-        await expect(
-            scheduler.run("interactive", undefined, async () => {
-                throw new Error("boom");
-            }),
-        ).rejects.toThrow("boom");
-        // Slot must be free again.
-        await scheduler.run("interactive", undefined, async () => undefined);
-    });
-
-    it("rejects queued waiters whose signal aborts, without leaking the slot", async () => {
-        const scheduler = new LocalLlmScheduler(1);
-        const gate = deferred();
-        const running = scheduler.run("background", undefined, () => gate.promise);
-        await tick();
-
-        const controller = new AbortController();
-        const waiting = scheduler.acquire("interactive", controller.signal);
-        controller.abort();
-        await expect(waiting).rejects.toThrow(/abort/i);
-
-        gate.resolve();
-        await running;
-        // Queue is clean: a fresh acquire succeeds immediately.
-        const release = await scheduler.acquire("background");
-        release();
-    });
-
-    it("rejects immediately when acquiring with an already-aborted signal", async () => {
-        const scheduler = new LocalLlmScheduler(1);
-        const controller = new AbortController();
-        controller.abort();
-        await expect(scheduler.acquire("interactive", controller.signal)).rejects.toThrow(/abort/i);
-    });
-
-    it("allows up to maxConcurrent tasks at once", async () => {
-        const scheduler = new LocalLlmScheduler(2);
-        const gateA = deferred();
-        const gateB = deferred();
-        let active = 0;
-        let peak = 0;
-        const track = async (gate: Promise<void>) => {
-            active++;
-            peak = Math.max(peak, active);
-            await gate;
-            active--;
-        };
-        const a = scheduler.run("background", undefined, () => track(gateA.promise));
-        const b = scheduler.run("background", undefined, () => track(gateB.promise));
-        await tick();
-        expect(peak).toBe(2);
-        gateA.resolve();
-        gateB.resolve();
-        await Promise.all([a, b]);
-    });
-});
diff --git a/apps/x/packages/core/src/models/local.ts b/apps/x/packages/core/src/models/local.ts
index 43a574f9..f1ebc3bb 100644
--- a/apps/x/packages/core/src/models/local.ts
+++ b/apps/x/packages/core/src/models/local.ts
@@ -1,9 +1,6 @@
 import { wrapLanguageModel, type LanguageModel } from "ai";
 import type { z } from "zod";
 import type { LlmProvider } from "@x/shared/dist/models.js";
-import { PrefixLogger } from "@x/shared";
-
-const log = new PrefixLogger("LocalLlm");
 
 // Ollama's server-side default context window (~4k tokens) is far below what
 // Rowboat's agents need (the copilot's system prompt + tool schemas alone are
@@ -13,218 +10,6 @@ const log = new PrefixLogger("LocalLlm");
 // `contextLength` in models.json.
 export const DEFAULT_OLLAMA_CONTEXT_LENGTH = 32768;
 
-const LOOPBACK_HOSTS = new Set(["localhost", "127.0.0.1", "0.0.0.0", "::1", "[::1]"]);
-
-/**
- * Whether requests to this provider are served by a model running on the
- * user's own machine (and therefore need scheduling: local runtimes process
- * requests mostly serially, so background work must not starve chat).
- */
-export function isLocalProvider(config: z.infer<typeof LlmProvider>): boolean {
-    if (config.flavor === "ollama") {
-        return true;
-    }
-    if (config.flavor === "openai-compatible") {
-        // LM Studio, llama.cpp server, vLLM on the same machine, etc.
-        try {
-            const host = new URL(config.baseURL ?? "").hostname;
-            return LOOPBACK_HOSTS.has(host) || host.endsWith(".local");
-        } catch {
-            return false;
-        }
-    }
-    return false;
-}
-
-export type LlmPriority = "interactive" | "classifier" | "background";
-
-const PRIORITY_ORDER: Record<LlmPriority, number> = {
-    interactive: 0,
-    classifier: 1,
-    background: 2,
-};
-
-interface Waiter {
-    priority: LlmPriority;
-    seq: number;
-    resolve: (release: () => void) => void;
-    reject: (error: Error) => void;
-    signal?: AbortSignal;
-    onAbort?: () => void;
-}
-
-function abortError(): Error {
-    const error = new Error("local LLM slot acquisition aborted");
-    error.name = "AbortError";
-    return error;
-}
-
-/**
- * Serializes access to local LLM runtimes. One slot by default: local servers
- * effectively process one request at a time, and interleaving requests
- * destroys their KV-cache reuse. Waiters are served strictly by priority
- * (interactive chat > lightweight classifiers > background knowledge
- * pipeline), FIFO within a priority — so a queued email-labeling job can
- * never delay the user's chat by more than the one request already running.
- */
-export class LocalLlmScheduler {
-    private active = 0;
-    private seq = 0;
-    private readonly waiting: Waiter[] = [];
-
-    constructor(private readonly maxConcurrent = 1) {}
-
-    async acquire(priority: LlmPriority, signal?: AbortSignal): Promise<() => void> {
-        if (signal?.aborted) {
-            throw abortError();
-        }
-        if (this.active < this.maxConcurrent) {
-            this.active++;
-            return this.makeRelease();
-        }
-        return new Promise<() => void>((resolve, reject) => {
-            const waiter: Waiter = { priority, seq: this.seq++, resolve, reject, signal };
-            if (signal) {
-                waiter.onAbort = () => {
-                    const index = this.waiting.indexOf(waiter);
-                    if (index >= 0) {
-                        this.waiting.splice(index, 1);
-                        reject(abortError());
-                    }
-                };
-                signal.addEventListener("abort", waiter.onAbort, { once: true });
-            }
-            this.waiting.push(waiter);
-            if (this.waiting.length === 1 || this.waiting.length % 10 === 0) {
-                log.log(`queueing ${priority} request (${this.waiting.length} waiting, ${this.active} active)`);
-            }
-        });
-    }
-
-    /** Run `fn` while holding a slot. */
-    async run<T>(priority: LlmPriority, signal: AbortSignal | undefined, fn: () => PromiseLike<T>): Promise<T> {
-        const release = await this.acquire(priority, signal);
-        try {
-            return await fn();
-        } finally {
-            release();
-        }
-    }
-
-    get queueDepth(): number {
-        return this.waiting.length;
-    }
-
-    private makeRelease(): () => void {
-        let released = false;
-        return () => {
-            if (released) {
-                return;
-            }
-            released = true;
-            this.active--;
-            this.dequeue();
-        };
-    }
-
-    private dequeue(): void {
-        while (this.active < this.maxConcurrent && this.waiting.length > 0) {
-            let best = 0;
-            for (let i = 1; i < this.waiting.length; i++) {
-                const a = this.waiting[i];
-                const b = this.waiting[best];
-                if (
-                    PRIORITY_ORDER[a.priority] < PRIORITY_ORDER[b.priority] ||
-                    (PRIORITY_ORDER[a.priority] === PRIORITY_ORDER[b.priority] && a.seq < b.seq)
-                ) {
-                    best = i;
-                }
-            }
-            const [waiter] = this.waiting.splice(best, 1);
-            if (waiter.signal && waiter.onAbort) {
-                waiter.signal.removeEventListener("abort", waiter.onAbort);
-            }
-            this.active++;
-            waiter.resolve(this.makeRelease());
-        }
-    }
-}
-
-function envConcurrency(): number {
-    const raw = process.env.ROWBOAT_LOCAL_LLM_CONCURRENCY;
-    const parsed = raw ? Number.parseInt(raw, 10) : NaN;
-    return Number.isFinite(parsed) && parsed > 0 ? parsed : 1;
-}
-
-// One queue for all local providers: users run one local server, and even
-// with several, the machine's compute is the shared resource.
-export const localLlmScheduler = new LocalLlmScheduler(envConcurrency());
-
-/**
- * Wrap a language model so every call requests an explicit context window
- * from Ollama (merged under the caller's providerOptions — an explicit
- * caller value wins) and, when `priority` is set and the provider is local,
- * goes through the shared scheduler.
- */
-export function applyLocalModelSettings(
-    model: LanguageModel,
-    providerConfig: z.infer<typeof LlmProvider>,
-    priority: LlmPriority | null,
-): LanguageModel {
-    if (typeof model === "string") {
-        // Bare model-id strings resolve through the global registry; local
-        // providers never take this path.
-        return model;
-    }
-    const local = isLocalProvider(providerConfig);
-    const wantsNumCtx = providerConfig.flavor === "ollama";
-    if (!wantsNumCtx && !(local && priority)) {
-        return model;
-    }
-    const numCtx = providerConfig.contextLength ?? DEFAULT_OLLAMA_CONTEXT_LENGTH;
-    const schedule = local && priority ? priority : null;
-    return wrapLanguageModel({
-        model,
-        middleware: {
-            ...(wantsNumCtx
-                ? {
-                      transformParams: async ({ params }) => {
-                          const providerOptions = (params.providerOptions ?? {}) as Record<string, Record<string, unknown>>;
-                          const ollama = (providerOptions.ollama ?? {}) as Record<string, unknown>;
-                          const options = (ollama.options ?? {}) as Record<string, unknown>;
-                          return {
-                              ...params,
-                              providerOptions: {
-                                  ...providerOptions,
-                                  ollama: {
-                                      ...ollama,
-                                      options: { num_ctx: numCtx, ...options },
-                                  },
-                              },
-                          };
-                      },
-                  }
-                : {}),
-            ...(schedule
-                ? {
-                      wrapGenerate: async ({ doGenerate, params }) =>
-                          localLlmScheduler.run(schedule, params.abortSignal, () => doGenerate()),
-                      wrapStream: async ({ doStream, params }) => {
-                          const release = await localLlmScheduler.acquire(schedule, params.abortSignal);
-                          try {
-                              const { stream, ...rest } = await doStream();
-                              return { ...rest, stream: releaseOnSettled(stream, release) };
-                          } catch (error) {
-                              release();
-                              throw error;
-                          }
-                      },
-                  }
-                : {}),
-        },
-    });
-}
-
 export type ReasoningEffort = "low" | "medium" | "high";
 
 // Local models default to snappy: gpt-oss at medium effort spends ~3x the
@@ -233,6 +18,43 @@ export type ReasoningEffort = "low" | "medium" | "high";
 // false, which thinking models like gpt-oss simply ignore).
 export const DEFAULT_OLLAMA_REASONING_EFFORT: ReasoningEffort = "low";
 
+/**
+ * Wrap a language model so every call requests an explicit context window
+ * from Ollama (merged under the caller's providerOptions — an explicit
+ * caller value wins). Non-Ollama providers pass through untouched.
+ */
+export function applyLocalModelSettings(
+    model: LanguageModel,
+    providerConfig: z.infer<typeof LlmProvider>,
+): LanguageModel {
+    if (typeof model === "string" || providerConfig.flavor !== "ollama") {
+        // Bare model-id strings resolve through the global registry; local
+        // providers never take this path.
+        return model;
+    }
+    const numCtx = providerConfig.contextLength ?? DEFAULT_OLLAMA_CONTEXT_LENGTH;
+    return wrapLanguageModel({
+        model,
+        middleware: {
+            transformParams: async ({ params }) => {
+                const providerOptions = (params.providerOptions ?? {}) as Record<string, Record<string, unknown>>;
+                const ollama = (providerOptions.ollama ?? {}) as Record<string, unknown>;
+                const options = (ollama.options ?? {}) as Record<string, unknown>;
+                return {
+                    ...params,
+                    providerOptions: {
+                        ...providerOptions,
+                        ollama: {
+                            ...ollama,
+                            options: { num_ctx: numCtx, ...options },
+                        },
+                    },
+                };
+            },
+        },
+    });
+}
+
 /**
  * Map a configured effort to the Ollama `think` value for a given model.
  * - gpt-oss accepts effort levels directly ("low" | "medium" | "high").
@@ -316,29 +138,3 @@ export function makeOllamaThinkFetch(
         return fetch(input, init);
     };
 }
-
-// Hold the slot until the provider stream drains, errors, or is cancelled —
-// a streaming response occupies the local server for its full duration.
-function releaseOnSettled<T>(stream: ReadableStream<T>, release: () => void): ReadableStream<T> {
-    const reader = stream.getReader();
-    return new ReadableStream<T>({
-        async pull(controller) {
-            try {
-                const { done, value } = await reader.read();
-                if (done) {
-                    release();
-                    controller.close();
-                } else {
-                    controller.enqueue(value);
-                }
-            } catch (error) {
-                release();
-                controller.error(error);
-            }
-        },
-        cancel(reason) {
-            release();
-            return reader.cancel(reason);
-        },
-    });
-}
diff --git a/apps/x/packages/core/src/models/models.ts b/apps/x/packages/core/src/models/models.ts
index 624d08a9..c14452aa 100644
--- a/apps/x/packages/core/src/models/models.ts
+++ b/apps/x/packages/core/src/models/models.ts
@@ -14,11 +14,9 @@ import { getChatModelIds } from "./models-dev.js";
 import { withUseCase } from "../analytics/use_case.js";
 import {
     applyLocalModelSettings,
-    isLocalProvider,
     makeOllamaThinkFetch,
     DEFAULT_OLLAMA_CONTEXT_LENGTH,
     DEFAULT_OLLAMA_REASONING_EFFORT,
-    type LlmPriority,
 } from "./local.js";
 
 export const Provider = LlmProvider;
@@ -90,18 +88,14 @@ export function createProvider(config: z.infer<typeof Provider>): ProviderV2 {
 
 /**
  * The one place model instances are created. Applies local-runtime settings
- * (explicit Ollama context window) and, when `priority` is given, routes the
- * call through the shared local scheduler so background work cannot starve
- * interactive chat. Pass `priority: null` when the caller does its own
- * scheduling (the turn runtime's model registry).
+ * (explicit Ollama context window) on top of the raw provider model.
  */
 export function createLanguageModel(
     providerConfig: z.infer<typeof Provider>,
     modelId: string,
-    opts: { priority?: LlmPriority | null } = {},
 ): LanguageModel {
     const model = createProvider(providerConfig).languageModel(modelId);
-    return applyLocalModelSettings(model, providerConfig, opts.priority ?? null);
+    return applyLocalModelSettings(model, providerConfig);
 }
 
 export interface ModelCapabilities {
@@ -151,9 +145,10 @@ export async function probeModelCapabilities(
             }
             return result;
         }
-        if (providerConfig.flavor === "openai-compatible" && isLocalProvider(providerConfig)) {
+        if (providerConfig.flavor === "openai-compatible") {
             // LM Studio's enhanced REST API lives at /api/v0 on the same
-            // origin as the OpenAI-compatible /v1 endpoint.
+            // origin as the OpenAI-compatible /v1 endpoint. Non-LM Studio
+            // endpoints just 404 here, which reports as "unknown".
             const origin = new URL(providerConfig.baseURL ?? "").origin;
             const res = await fetch(`${origin}/api/v0/models`, {
                 headers: providerConfig.headers ?? {},
@@ -342,7 +337,7 @@ export async function generateOneShot(opts: GenerateTextOptions): Promise<Genera
         const modelId = opts.model || def.model;
         const providerName = opts.provider || def.provider;
         const providerConfig = await resolveProviderConfig(providerName);
-        const languageModel = createLanguageModel(providerConfig, modelId, { priority: "interactive" });
+        const languageModel = createLanguageModel(providerConfig, modelId);
         const result = await withUseCase(
             { useCase: "copilot_chat", subUseCase: "email_compose" },
             () => generateText({
diff --git a/apps/x/packages/core/src/models/repo.ts b/apps/x/packages/core/src/models/repo.ts
index 223c9434..3a3a5b77 100644
--- a/apps/x/packages/core/src/models/repo.ts
+++ b/apps/x/packages/core/src/models/repo.ts
@@ -10,7 +10,8 @@ export type ModelConfigPatch = {
         | "knowledgeGraphModel"
         | "meetingNotesModel"
         | "liveNoteAgentModel"
-        | "autoPermissionDecisionModel"]?: z.infer<typeof ModelConfig>[K] | null;
+        | "autoPermissionDecisionModel"
+        | "deferBackgroundTasks"]?: z.infer<typeof ModelConfig>[K] | null;
 };
 
 export interface IModelConfigRepo {
@@ -85,7 +86,7 @@ export class FSModelConfigRepo implements IModelConfigRepo {
             const raw = await fs.readFile(this.configPath, "utf8");
             const existing = JSON.parse(raw);
             existingSelections = Object.fromEntries(
-                ["defaultSelection", "knowledgeGraphModel", "meetingNotesModel", "liveNoteAgentModel", "autoPermissionDecisionModel"]
+                ["defaultSelection", "knowledgeGraphModel", "meetingNotesModel", "liveNoteAgentModel", "autoPermissionDecisionModel", "deferBackgroundTasks"]
                     .filter((key) => existing[key] !== undefined && (config as Record<string, unknown>)[key] === undefined)
                     .map((key) => [key, existing[key]]),
             );
diff --git a/apps/x/packages/core/src/pre_built/runner.ts b/apps/x/packages/core/src/pre_built/runner.ts
index 7c6d7be5..60243c4b 100644
--- a/apps/x/packages/core/src/pre_built/runner.ts
+++ b/apps/x/packages/core/src/pre_built/runner.ts
@@ -1,7 +1,7 @@
 import fs from 'fs';
 import path from 'path';
 import { WorkDir } from '../config/config.js';
-import { runHeadlessAgent } from '../agents/headless-app.js';
+import { runWhenPossible } from '../agents/headless-app.js';
 import { getKgModel } from '../models/defaults.js';
 import {
     loadConfig,
@@ -52,7 +52,7 @@ Process new items and use the user context above to identify yourself when draft
         // The agent file is expected to be in the agents directory with
         // the same name. Waits for the turn to settle (errors tolerated,
         // matching the old no-throwOnError wait).
-        await runHeadlessAgent({
+        await runWhenPossible({
             agentId: agentName,
             message,
             ...(await getKgModel()),
diff --git a/apps/x/packages/core/src/security/auto-permission-classifier.ts b/apps/x/packages/core/src/security/auto-permission-classifier.ts
index 1ce2d1c9..42bdbf0f 100644
--- a/apps/x/packages/core/src/security/auto-permission-classifier.ts
+++ b/apps/x/packages/core/src/security/auto-permission-classifier.ts
@@ -83,7 +83,7 @@ export async function classifyToolPermissions(input: {
 
     const { model: modelId, provider: providerName } = await getAutoPermissionDecisionModel();
     const providerConfig = await resolveProviderConfig(providerName);
-    const model = createLanguageModel(providerConfig, modelId, { priority: "classifier" });
+    const model = createLanguageModel(providerConfig, modelId);
 
     const result = await withUseCase(
         {
diff --git a/apps/x/packages/core/src/sessions/sessions.ts b/apps/x/packages/core/src/sessions/sessions.ts
index ad9d8c98..e4c262db 100644
--- a/apps/x/packages/core/src/sessions/sessions.ts
+++ b/apps/x/packages/core/src/sessions/sessions.ts
@@ -17,6 +17,7 @@ import {
     reduceTurn,
 } from "@x/shared/dist/turns.js";
 import type { IMonotonicallyIncreasingIdGenerator } from "../application/lib/id-gen.js";
+import { chatActivity } from "../application/lib/chat-activity.js";
 import type {
     ITurnRuntime,
     Turn,
@@ -339,9 +340,17 @@ export class SessionsImpl implements ISessions {
         input?: TurnExternalInput,
     ): TurnExecution {
         const controller = new AbortController();
+        // A session turn is a user-facing chat: mark it active so background
+        // agents can defer (see agents/headless-app.ts runWhenPossible).
+        if (sessionId !== null) {
+            chatActivity.enter();
+        }
         const execution = this.turnRuntime.advanceTurn(turnId, input, {
             signal: controller.signal,
         });
+        if (sessionId !== null) {
+            void execution.outcome.catch(() => undefined).finally(() => chatActivity.exit());
+        }
         this.active.set(turnId, { sessionId, controller, execution });
 
         void (async () => {
diff --git a/apps/x/packages/core/src/turns/bridges/real-model-registry.ts b/apps/x/packages/core/src/turns/bridges/real-model-registry.ts
index dbbda950..36935d8a 100644
--- a/apps/x/packages/core/src/turns/bridges/real-model-registry.ts
+++ b/apps/x/packages/core/src/turns/bridges/real-model-registry.ts
@@ -14,11 +14,7 @@ import type { JsonValue, ModelDescriptor, TurnUsage } from "@x/shared/dist/turns
 import { convertFromMessages } from "../../agents/runtime.js";
 import { resolveProviderConfig } from "../../models/defaults.js";
 import { createProvider } from "../../models/models.js";
-import {
-    applyLocalModelSettings,
-    isLocalProvider,
-    localLlmScheduler,
-} from "../../models/local.js";
+import { applyLocalModelSettings } from "../../models/local.js";
 import type {
     IModelRegistry,
     LlmStreamEvent,
@@ -70,15 +66,11 @@ export class RealModelRegistry implements IModelRegistry {
     ): Promise<ResolvedModel> {
         const providerConfig = await this.resolveProvider(descriptor.provider);
         const provider = this.createProviderImpl(providerConfig);
-        // Local settings (Ollama context window) are applied here, but
-        // scheduling happens per-step in run() where the turn's priority is
-        // known — so priority stays null.
+        // Local settings (Ollama context window) are applied here.
         const model = applyLocalModelSettings(
             provider.languageModel(descriptor.model),
             providerConfig,
-            null,
         );
-        const local = isLocalProvider(providerConfig);
         return {
             descriptor,
             // The structural -> wire conversion the app uses today: weaves
@@ -87,14 +79,13 @@ export class RealModelRegistry implements IModelRegistry {
             // per-message, so composed requests are byte-stable.
             encodeMessages: (messages) =>
                 convertFromMessages(messages) as unknown as JsonValue[],
-            stream: (request) => this.run(model, request, local),
+            stream: (request) => this.run(model, request),
         };
     }
 
     private async *run(
         model: LanguageModel,
         request: ModelStreamRequest,
-        local: boolean,
     ): AsyncGenerator<LlmStreamEvent, void, void> {
         const tools: ToolSet = {};
         for (const descriptor of request.tools) {
@@ -123,15 +114,6 @@ export class RealModelRegistry implements IModelRegistry {
                 : {}),
         };
 
-        // One scheduler slot per model step: local runtimes serve requests
-        // serially, and headless turns must not starve interactive chat.
-        const release = local
-            ? await localLlmScheduler.acquire(
-                  request.priority ?? "interactive",
-                  request.signal,
-              )
-            : null;
-
         const parts: Array<z.infer<typeof AssistantContentPart>> = [];
         let textBuffer = "";
         let reasoningBuffer = "";
@@ -139,112 +121,108 @@ export class RealModelRegistry implements IModelRegistry {
         let usage: z.infer<typeof TurnUsage> = {};
         let providerMetadata: JsonValue | undefined;
 
-        try {
-            const result = this.invoke({
-                model,
-                system: request.systemPrompt,
-                messages: request.messages as ModelMessage[],
-                tools,
-                abortSignal: request.signal,
-                ...generationParams,
-            });
+        const result = this.invoke({
+            model,
+            system: request.systemPrompt,
+            messages: request.messages as ModelMessage[],
+            tools,
+            abortSignal: request.signal,
+            ...generationParams,
+        });
 
-            for await (const raw of result.fullStream) {
-                request.signal.throwIfAborted();
-                const event = raw as {
-                    type: string;
-                    text?: string;
-                    toolCallId?: string;
-                    toolName?: string;
-                    input?: unknown;
-                    finishReason?: string;
-                    usage?: Record<string, number | undefined>;
-                    providerMetadata?: unknown;
-                    error?: unknown;
-                };
-                switch (event.type) {
-                    case "text-start":
-                        textBuffer = "";
-                        yield { type: "step_event", event: { type: "text_start" } };
-                        break;
-                    case "text-delta": {
-                        const delta = event.text ?? "";
-                        textBuffer += delta;
-                        const last = parts[parts.length - 1];
-                        if (last?.type === "text") {
-                            last.text += delta;
-                        } else {
-                            parts.push({ type: "text", text: delta });
-                        }
-                        yield { type: "text_delta", delta };
-                        break;
+        for await (const raw of result.fullStream) {
+            request.signal.throwIfAborted();
+            const event = raw as {
+                type: string;
+                text?: string;
+                toolCallId?: string;
+                toolName?: string;
+                input?: unknown;
+                finishReason?: string;
+                usage?: Record<string, number | undefined>;
+                providerMetadata?: unknown;
+                error?: unknown;
+            };
+            switch (event.type) {
+                case "text-start":
+                    textBuffer = "";
+                    yield { type: "step_event", event: { type: "text_start" } };
+                    break;
+                case "text-delta": {
+                    const delta = event.text ?? "";
+                    textBuffer += delta;
+                    const last = parts[parts.length - 1];
+                    if (last?.type === "text") {
+                        last.text += delta;
+                    } else {
+                        parts.push({ type: "text", text: delta });
                     }
-                    case "text-end":
-                        yield {
-                            type: "step_event",
-                            event: { type: "text_end", text: textBuffer },
-                        };
-                        break;
-                    case "reasoning-start":
-                        reasoningBuffer = "";
-                        yield { type: "step_event", event: { type: "reasoning_start" } };
-                        break;
-                    case "reasoning-delta": {
-                        const delta = event.text ?? "";
-                        reasoningBuffer += delta;
-                        const last = parts[parts.length - 1];
-                        if (last?.type === "reasoning") {
-                            last.text += delta;
-                        } else {
-                            parts.push({ type: "reasoning", text: delta });
-                        }
-                        yield { type: "reasoning_delta", delta };
-                        break;
-                    }
-                    case "reasoning-end":
-                        yield {
-                            type: "step_event",
-                            event: { type: "reasoning_end", text: reasoningBuffer },
-                        };
-                        break;
-                    case "tool-call": {
-                        const toolCall = {
-                            type: "tool-call" as const,
-                            toolCallId: String(event.toolCallId),
-                            toolName: String(event.toolName),
-                            arguments: event.input,
-                        };
-                        parts.push(toolCall);
-                        yield { type: "step_event", event: { type: "tool_call", toolCall } };
-                        break;
-                    }
-                    case "finish-step": {
-                        finishReason = event.finishReason ?? "unknown";
-                        usage = mapUsage(event.usage);
-                        providerMetadata = toJsonValue(event.providerMetadata);
-                        yield {
-                            type: "step_event",
-                            event: {
-                                type: "finish_step",
-                                finishReason,
-                                usage,
-                                ...(providerMetadata === undefined
-                                    ? {}
-                                    : { providerMetadata }),
-                            },
-                        };
-                        break;
-                    }
-                    case "error":
-                        throw event.error instanceof Error
-                            ? event.error
-                            : new Error(formatStreamError(event.error));
-                    default:
-                        break;
+                    yield { type: "text_delta", delta };
+                    break;
                 }
+                case "text-end":
+                    yield {
+                        type: "step_event",
+                        event: { type: "text_end", text: textBuffer },
+                    };
+                    break;
+                case "reasoning-start":
+                    reasoningBuffer = "";
+                    yield { type: "step_event", event: { type: "reasoning_start" } };
+                    break;
+                case "reasoning-delta": {
+                    const delta = event.text ?? "";
+                    reasoningBuffer += delta;
+                    const last = parts[parts.length - 1];
+                    if (last?.type === "reasoning") {
+                        last.text += delta;
+                    } else {
+                        parts.push({ type: "reasoning", text: delta });
+                    }
+                    yield { type: "reasoning_delta", delta };
+                    break;
+                }
+                case "reasoning-end":
+                    yield {
+                        type: "step_event",
+                        event: { type: "reasoning_end", text: reasoningBuffer },
+                    };
+                    break;
+                case "tool-call": {
+                    const toolCall = {
+                        type: "tool-call" as const,
+                        toolCallId: String(event.toolCallId),
+                        toolName: String(event.toolName),
+                        arguments: event.input,
+                    };
+                    parts.push(toolCall);
+                    yield { type: "step_event", event: { type: "tool_call", toolCall } };
+                    break;
+                }
+                case "finish-step": {
+                    finishReason = event.finishReason ?? "unknown";
+                    usage = mapUsage(event.usage);
+                    providerMetadata = toJsonValue(event.providerMetadata);
+                    yield {
+                        type: "step_event",
+                        event: {
+                            type: "finish_step",
+                            finishReason,
+                            usage,
+                            ...(providerMetadata === undefined
+                                ? {}
+                                : { providerMetadata }),
+                        },
+                    };
+                    break;
+                }
+                case "error":
+                    throw event.error instanceof Error
+                        ? event.error
+                        : new Error(formatStreamError(event.error));
+                default:
+                    break;
             }
-        } finally {
-            release?.();
         }
 
         yield {
diff --git a/apps/x/packages/core/src/turns/model-registry.ts b/apps/x/packages/core/src/turns/model-registry.ts
index 65631d87..d4c83a76 100644
--- a/apps/x/packages/core/src/turns/model-registry.ts
+++ b/apps/x/packages/core/src/turns/model-registry.ts
@@ -34,10 +34,6 @@ export interface ModelStreamRequest {
     tools: Array<z.infer<typeof ToolDescriptor>>;
     parameters: Record<string, JsonValue>;
     signal: AbortSignal;
-    // Scheduling class for local-model queueing: interactive turns (a human
-    // is watching) preempt queued headless/background work. Defaults to
-    // "interactive" so an unset priority never slows a user down.
-    priority?: "interactive" | "background";
 }
 
 export interface ResolvedModel {
diff --git a/apps/x/packages/core/src/turns/runtime.ts b/apps/x/packages/core/src/turns/runtime.ts
index 90b72d7b..da2362a4 100644
--- a/apps/x/packages/core/src/turns/runtime.ts
+++ b/apps/x/packages/core/src/turns/runtime.ts
@@ -951,11 +951,6 @@ class TurnAdvance {
                 tools: composed.tools,
                 parameters: composed.parameters,
                 signal: this.signal,
-                // Headless turns (no human waiting) yield the local-model
-                // queue to interactive chat.
-                priority: this.definition.config.humanAvailable
-                    ? "interactive"
-                    : "background",
             })) {
                 switch (event.type) {
                     case "text_delta":
diff --git a/apps/x/packages/shared/src/ipc.ts b/apps/x/packages/shared/src/ipc.ts
index 2886e3a3..3ad66e0b 100644
--- a/apps/x/packages/shared/src/ipc.ts
+++ b/apps/x/packages/shared/src/ipc.ts
@@ -623,6 +623,7 @@ const ipcSchemas = {
       meetingNotesModel: ModelOverride.nullable().optional(),
       liveNoteAgentModel: ModelOverride.nullable().optional(),
       autoPermissionDecisionModel: ModelOverride.nullable().optional(),
+      deferBackgroundTasks: z.boolean().nullable().optional(),
     }),
     res: z.object({
       success: z.literal(true),
diff --git a/apps/x/packages/shared/src/models.ts b/apps/x/packages/shared/src/models.ts
index 2ec107d3..7d51a089 100644
--- a/apps/x/packages/shared/src/models.ts
+++ b/apps/x/packages/shared/src/models.ts
@@ -37,6 +37,11 @@ export const LlmModelConfig = z.object({
   // the signed-in curated default and the legacy top-level provider/model
   // pair — this is what lets signed-in users default to a BYOK model.
   defaultSelection: ModelRef.optional(),
+  // When true, background agent runs (knowledge pipeline, live notes,
+  // background tasks) wait until no chat turn is running before starting.
+  // Surfaced as a settings checkbox; recommended for local models, where a
+  // background run competes with the chat for the same hardware.
+  deferBackgroundTasks: z.boolean().optional(),
   providers: z.record(z.string(), z.object({
     apiKey: z.string().optional(),
     baseURL: z.string().optional(),