@agent-native/core 0.5.0-dev.b51eaae → 0.6.1

package/dist/server/agent-chat-plugin.js

@@ -1,11 +1,25 @@
-import { createProductionAgentHandler, getActiveRunForThreadAsync, abortRun, subscribeToRun, } from "../agent/production-agent.js";
+import { getH3App } from "./framework-request-handler.js";
+import { createProductionAgentHandler, runAgentLoop, actionsToEngineTools, getActiveRunForThreadAsync, abortRun, subscribeToRun, } from "../agent/production-agent.js";
+import { resolveEngine, createAnthropicEngine } from "../agent/engine/index.js";
+import { discoverAgents } from "./agent-discovery.js";
+import { loadSchemaPromptBlock } from "./schema-prompt.js";
 import { buildAssistantMessage, extractThreadMeta, } from "../agent/thread-data-builder.js";
-import { defineEventHandler, readBody, setResponseStatus, setResponseHeader, getMethod, getQuery, } from "h3";
+import { defineEventHandler, setResponseStatus, setResponseHeader, getMethod, getQuery, } from "h3";
 import { getSession } from "./auth.js";
 import { createThread, getThread, listThreads, searchThreads, updateThreadData, deleteThread, } from "../chat-threads/store.js";
 import { resourceListAccessible, resourceList, resourceGet, resourceGetByPath, ensurePersonalDefaults, SHARED_OWNER, } from "../resources/store.js";
-import fs from "node:fs";
 import nodePath from "node:path";
+import { readBody } from "./h3-helpers.js";
+// Lazy fs — loaded via dynamic import() on first use.
+// This avoids require() which bundlers convert to createRequire(import.meta.url)
+// that crashes on CF Workers where import.meta.url is undefined.
+let _fs;
+async function lazyFs() {
+    if (!_fs) {
+        _fs = await import("node:fs");
+    }
+    return _fs;
+}
 /**
  * Wraps a core CLI script (that writes to console.log) as a ActionEntry
  * by capturing stdout.
@@ -157,16 +171,88 @@ async function createResourceScriptEntries() {
         return {};
     }
 }
+/**
+ * Creates chat management ActionEntries (search-chats, open-chat).
+ */
+async function createChatScriptEntries() {
+    try {
+        const [searchMod, openMod] = await Promise.all([
+            import("../scripts/chat/search-chats.js"),
+            import("../scripts/chat/open-chat.js"),
+        ]);
+        return {
+            "search-chats": wrapCliScript({
+                description: "Search or list past agent chat threads. Use this to find previous conversations by keyword.",
+                parameters: {
+                    type: "object",
+                    properties: {
+                        query: {
+                            type: "string",
+                            description: "Search term to find chats by title, preview, or content",
+                        },
+                        limit: {
+                            type: "string",
+                            description: "Max number of results (default: 20)",
+                        },
+                        format: {
+                            type: "string",
+                            description: "Output format",
+                            enum: ["json", "text"],
+                        },
+                    },
+                },
+            }, searchMod.default),
+            "open-chat": wrapCliScript({
+                description: "Open a chat thread in the UI as a new tab and focus it. Use search-chats first to find the thread ID.",
+                parameters: {
+                    type: "object",
+                    properties: {
+                        id: {
+                            type: "string",
+                            description: "The chat thread ID to open",
+                        },
+                    },
+                    required: ["id"],
+                },
+            }, openMod.default),
+        };
+    }
+    catch {
+        return {};
+    }
+}
+/**
+ * Creates agent engine management tools (list-agent-engines, set-agent-engine,
+ * test-agent-engine). Let the agent inspect and configure the active LLM engine.
+ */
+async function createAgentEngineScriptEntries() {
+    try {
+        const [listMod, setMod, testMod] = await Promise.all([
+            import("../scripts/agent-engines/list-agent-engines.js"),
+            import("../scripts/agent-engines/set-agent-engine.js"),
+            import("../scripts/agent-engines/test-agent-engine.js"),
+        ]);
+        return {
+            "list-agent-engines": { tool: listMod.tool, run: listMod.run },
+            "set-agent-engine": { tool: setMod.tool, run: setMod.run },
+            "test-agent-engine": { tool: testMod.tool, run: testMod.run },
+        };
+    }
+    catch {
+        return {};
+    }
+}
 /**
  * Creates the call-agent ActionEntry for cross-agent A2A communication.
+ * Binds selfAppId so the agent cannot call itself via call-agent.
  */
-async function createCallAgentScriptEntry() {
+async function createCallAgentScriptEntry(selfAppId) {
     try {
         const mod = await import("../scripts/call-agent.js");
         return {
             "call-agent": {
                 tool: mod.tool,
-                run: mod.run,
+                run: (args, context) => mod.run(args, context, selfAppId),
             },
         };
     }
@@ -174,6 +260,185 @@ async function createCallAgentScriptEntry() {
         return {};
     }
 }
+/**
+ * Creates agent team orchestration tools (spawn-task, task-status, read-task-result).
+ * These let the main agent spawn sub-agents and coordinate work.
+ */
+function createTeamTools(deps) {
+    return {
+        "spawn-task": {
+            tool: {
+                description: "Spawn a sub-agent to handle a task in the background. The sub-agent runs independently with its own conversation thread. Use this to delegate work so the main chat stays free for new requests. A live preview card will appear in the chat showing the sub-agent's progress.",
+                parameters: {
+                    type: "object",
+                    properties: {
+                        task: {
+                            type: "string",
+                            description: "Clear description of what the sub-agent should accomplish",
+                        },
+                        instructions: {
+                            type: "string",
+                            description: "Optional additional instructions or context for the sub-agent",
+                        },
+                        name: {
+                            type: "string",
+                            description: "Short name for the sub-agent tab (e.g. 'Research', 'Draft email'). If omitted, derived from the task.",
+                        },
+                    },
+                    required: ["task"],
+                },
+            },
+            run: async (args) => {
+                // Capture the send function NOW (at spawn time) so that
+                // concurrent runs don't clobber each other's send reference.
+                const capturedSend = deps.getSend();
+                const { spawnTask } = await import("./agent-teams.js");
+                // Filter out team orchestration tools so sub-agents can't spawn sub-agents
+                const teamToolNames = new Set([
+                    "spawn-task",
+                    "task-status",
+                    "read-task-result",
+                    "send-to-task",
+                    "list-tasks",
+                ]);
+                const subAgentActions = Object.fromEntries(Object.entries(deps.getActions()).filter(([name]) => !teamToolNames.has(name)));
+                const task = await spawnTask({
+                    description: args.task,
+                    instructions: args.instructions,
+                    ownerEmail: deps.getOwner(),
+                    systemPrompt: deps.getSystemPrompt(),
+                    actions: subAgentActions,
+                    engine: deps.getEngine(),
+                    model: deps.getModel(),
+                    parentThreadId: deps.getParentThreadId(),
+                    parentSend: (event) => {
+                        if (capturedSend)
+                            capturedSend(event);
+                    },
+                });
+                return JSON.stringify({
+                    taskId: task.taskId,
+                    threadId: task.threadId,
+                    status: task.status,
+                    description: task.description,
+                    name: args.name || "",
+                });
+            },
+        },
+        "task-status": {
+            tool: {
+                description: "Check the status of a sub-agent task. Returns current status, preview of output, and current step.",
+                parameters: {
+                    type: "object",
+                    properties: {
+                        taskId: {
+                            type: "string",
+                            description: "The task ID returned by spawn-task",
+                        },
+                    },
+                    required: ["taskId"],
+                },
+            },
+            run: async (args) => {
+                const { getTask } = await import("./agent-teams.js");
+                const task = await getTask(args.taskId);
+                if (!task)
+                    return JSON.stringify({ error: "Task not found" });
+                return JSON.stringify({
+                    taskId: task.taskId,
+                    threadId: task.threadId,
+                    status: task.status,
+                    description: task.description,
+                    preview: task.preview,
+                    currentStep: task.currentStep,
+                    summary: task.summary,
+                });
+            },
+        },
+        "read-task-result": {
+            tool: {
+                description: "Read the result of a completed sub-agent task. Returns the full output summary.",
+                parameters: {
+                    type: "object",
+                    properties: {
+                        taskId: {
+                            type: "string",
+                            description: "The task ID returned by spawn-task",
+                        },
+                    },
+                    required: ["taskId"],
+                },
+            },
+            run: async (args) => {
+                const { getTask } = await import("./agent-teams.js");
+                const task = await getTask(args.taskId);
+                if (!task)
+                    return JSON.stringify({ error: "Task not found" });
+                if (task.status === "running") {
+                    return JSON.stringify({
+                        status: "running",
+                        preview: task.preview,
+                        message: "Task is still running. Check back later.",
+                    });
+                }
+                return JSON.stringify({
+                    taskId: task.taskId,
+                    status: task.status,
+                    summary: task.summary,
+                    preview: task.preview,
+                });
+            },
+        },
+        "send-to-task": {
+            tool: {
+                description: "Send a message or update to a running sub-agent. Use this to redirect, add context, or give feedback to a sub-agent while it's working.",
+                parameters: {
+                    type: "object",
+                    properties: {
+                        taskId: {
+                            type: "string",
+                            description: "The task ID returned by spawn-task",
+                        },
+                        message: {
+                            type: "string",
+                            description: "Message to send to the sub-agent",
+                        },
+                    },
+                    required: ["taskId", "message"],
+                },
+            },
+            run: async (args) => {
+                const { sendToTask } = await import("./agent-teams.js");
+                const result = await sendToTask(args.taskId, args.message);
+                return JSON.stringify(result);
+            },
+        },
+        "list-tasks": {
+            tool: {
+                description: "List all sub-agent tasks and their current status. Use this to see what's running, completed, or failed.",
+                parameters: {
+                    type: "object",
+                    properties: {},
+                },
+            },
+            run: async () => {
+                const { listTasks } = await import("./agent-teams.js");
+                const tasks = await listTasks();
+                if (tasks.length === 0) {
+                    return "No sub-agent tasks.";
+                }
+                return JSON.stringify(tasks.map((t) => ({
+                    taskId: t.taskId,
+                    threadId: t.threadId,
+                    description: t.description,
+                    status: t.status,
+                    currentStep: t.currentStep,
+                    hasResult: t.summary.length > 0,
+                })), null, 2);
+            },
+        },
+    };
+}
 /**
  * Framework-level instructions injected into every agent's system prompt.
  * This is the single source of truth for the core philosophy, rules, and patterns.
@@ -187,7 +452,7 @@ const FRAMEWORK_CORE = `
 ### Core Rules
 
 1. **Data lives in SQL** — All app state is in a SQL database (could be SQLite, Postgres, Turso, or Cloudflare D1 — never assume which). Use the available database tools.
-2. **Context awareness** — Before taking action, understand what the user is looking at. Use the \`view-screen\` tool if available — it returns the current UI state (which page, which item is focused, what's selected).
+2. **Context awareness** — The user's current screen state is automatically included in each message as a \`<current-screen>\` block. Use it to understand what the user is looking at. You can still call \`view-screen\` for a more detailed snapshot if needed, but you should NOT need to call it before every action.
 3. **Navigate the UI** — Use the \`navigate\` tool to switch views, open items, or focus elements for the user.
 4. **Application state** — Ephemeral UI state (drafts, selections, navigation) lives in \`application_state\`. Use \`readAppState\`/\`writeAppState\` to read and write it. When you write state, the UI updates automatically.
 5. **Resources for memory** — Use the Resources system for persistent notes and context. Update LEARNINGS.md when you learn user preferences or corrections. Update the shared AGENTS.md for instructions that should apply to all users.
@@ -204,6 +469,80 @@ When the user gives instructions that should apply to all users/sessions, update
 ### Navigation Rule
 
 When the user says "show me", "go to", "open", "switch to", or similar navigation language, ALWAYS use the \`navigate\` action to update the UI. The user expects to SEE the result in the main app, not just read it in chat. Navigate first, then fetch/display data.
+
+### Chat History
+
+You can search and restore previous chat conversations:
+- \`search-chats\` — Search or list past chat threads by keyword
+- \`open-chat\` — Open a chat thread in the UI as a new tab and focus it
+
+When the user asks to find a previous conversation, use \`search-chats\` first to find matching threads, then \`open-chat\` to restore the one they want.
+
+### Agent Teams — Orchestration
+
+You are an orchestrator. For complex or multi-step tasks, delegate to sub-agents:
+- \`spawn-task\` — Spawn a sub-agent for a task. It runs in its own thread while you stay available. A live preview card appears in the chat.
+- \`task-status\` — Check the progress of a running sub-agent.
+- \`read-task-result\` — Read the result when a sub-agent finishes.
+
+**When to delegate vs do directly:**
+- **Delegate** when the task involves multiple tool calls, research, content generation, or anything that takes more than a few seconds. Examples: "create a deck about X", "analyze the data and write a report", "look up Y and draft an email about it".
+- **Do directly** for quick single-step tasks like navigation, reading state, or answering simple questions.
+- **Spawn multiple sub-agents** when the user asks for multiple independent things — they'll run in parallel.
+
+**How to orchestrate:**
+1. When the user asks for something complex, spawn a sub-agent with a clear task description.
+2. Tell the user what you've started ("I'm having a sub-agent research that for you").
+3. You can keep chatting — sub-agents run independently.
+4. Use \`read-task-result\` to check results when needed, or the user can see live progress in the card.
+5. If the user's request has multiple steps, you can spawn one sub-agent per step, or chain them.
+
+Sub-agents have access to all template tools but **cannot spawn sub-agents themselves** — only you (the orchestrator) can do that. Give the sub-agent a specific, actionable task description — it will figure out which tools to use.
+
+### Recurring Jobs
+
+You can create recurring jobs that run on a cron schedule. Jobs are resource files under \`jobs/\`. Each job has a cron schedule and instructions that the agent executes automatically.
+
+- \`create-job\` — Create a new recurring job with a cron schedule and instructions
+- \`list-jobs\` — List all recurring jobs and their status (schedule, last run, next run, errors)
+- \`update-job\` — Update a job's schedule, instructions, or toggle enabled/disabled
+- Delete a job with \`resource-delete --path jobs/<name>.md\`
+
+When the user asks for something recurring ("every morning", "daily at 9am", "weekly on Mondays"), create a job. Convert natural language to 5-field cron format:
+- "every morning" / "daily at 9am" → \`0 9 * * *\`
+- "every weekday at 9am" → \`0 9 * * 1-5\`
+- "every hour" → \`0 * * * *\`
+- "every 30 minutes" → \`*/30 * * * *\`
+- "every monday at 9am" → \`0 9 * * 1\`
+- "twice a day" / "morning and evening" → \`0 9,17 * * *\`
+
+Job instructions should be self-contained — include which actions to call, what conditions to check, and what to do with results. The agent executing the job has access to all the same tools you do.
+
+### call-agent — External Apps Only
+
+The \`call-agent\` tool sends a message to a DIFFERENT, separately-deployed app's agent (A2A protocol). It is **not** for calling actions within the current app.
+
+**NEVER use \`call-agent\` to:**
+- Call your own app by name (if you are the "macros" agent, never do \`call-agent(agent="macros")\`)
+- Perform tasks you can accomplish with your own registered tools
+- Wrap your own actions in an A2A round-trip
+
+**ONLY use \`call-agent\` when:**
+- The user explicitly asks you to communicate with a different app (e.g., "ask the mail agent to...")
+- You need data that only another deployed app can provide
+- You are coordinating across genuinely separate apps
+
+If \`call-agent\` returns an error saying the agent is yourself — stop and use your own tools instead.
+
+### Auto-Memory
+
+Proactively update \`LEARNINGS.md\` when you learn something important during conversations:
+- User corrects your approach → capture the correct way
+- User shares preferences (tone, style, workflow) → capture them
+- You discover a non-obvious pattern or gotcha → capture it
+- User provides personal context (contacts, team info, domain knowledge) → capture it
+
+**Don't ask permission — just save it.** Keep entries concise (one line each, grouped by category). Don't save things that are obvious from reading the code or that are temporary.
 `;
 const PROD_FRAMEWORK_PROMPT = `## Agent-Native Framework — Production Mode
 
@@ -212,6 +551,21 @@ You are an AI agent in an agent-native application, running in **production mode
 The agent and the UI are equal partners — everything the UI can do, you can do via your tools, and vice versa. They share the same SQL database and stay in sync automatically.
 
 **In production mode, you operate through registered actions exposed as tools.** These are your capabilities — use them to read data, take actions, and help the user. You cannot edit source code or access the filesystem directly. Your tools are the app's API.
+
+### Code Changes Not Available
+
+When the user asks you to change the UI, modify code, add features, fix bugs in the app itself, change styles, or do anything that requires editing source files — you MUST respond with this exact message (replacing APP_NAME with the current app name from the URL hostname):
+
+> This app is running in hosted mode and I can't make code changes here. To customize this app:
+>
+> **Option 1:** Connect to [Builder.io](https://www.builder.io/m/agent-native) for a full development environment
+>
+> **Option 2:** Clone and run locally:
+> \`\`\`
+> npx agent-native create APP_NAME
+> \`\`\`
+
+Do NOT attempt to work around this limitation. Do NOT say "I can't do that" without providing the options above. Always include both the Builder.io link and the CLI command.
 ${FRAMEWORK_CORE}`;
 const DEV_FRAMEWORK_PROMPT = `## Agent-Native Framework — Development Mode
 
@@ -219,7 +573,17 @@ You are an AI agent in an agent-native application, running in **development mod
 
 The agent and the UI are equal partners — everything the UI can do, you can do via tools/scripts, and vice versa. They share the same SQL database and stay in sync automatically.
 
-**In development mode, you have full access to the project filesystem, shell, and database** — in addition to all the app's production tools. You can edit source code, run commands, install packages, and modify the app directly.
+**In development mode, you have UNRESTRICTED access.** You can:
+- Run ANY shell command via the \`shell\` tool (bash, node, curl, pnpm, etc.)
+- Execute arbitrary code: \`shell({ command: 'node -e "console.log(1+1)"' })\`
+- Read/write any file on the filesystem
+- Query and modify the database
+- Call external APIs (via shell with curl, or via scripts)
+- Edit source code, install packages, modify the app
+
+**There are NO restrictions in dev mode.** If a dedicated tool/action doesn't exist for what you need, use \`shell\` to run any command. For example: \`shell({ command: 'curl -s https://api.example.com/data' })\`
+
+**Template-specific actions are invoked via shell, NOT as direct tools.** In dev mode, the only tools registered as native tool calls are framework-level utilities (shell, file ops, resources, chat, teams, jobs). Anything from the template's \`actions/\` directory must be run through shell: \`shell({ command: 'pnpm action <name> --arg value' })\`. The "Available Actions" section below shows the exact CLI syntax for each one — copy that command verbatim and pass it to \`shell\`. Do not try to call template actions by name as if they were tools; they will not appear in your tool list.
 
 When editing code, follow the agent-native architecture:
 - Every feature needs all four areas: UI + scripts + skills/instructions + application-state sync
@@ -229,59 +593,177 @@ When editing code, follow the agent-native architecture:
 ${FRAMEWORK_CORE}`;
 const DEFAULT_SYSTEM_PROMPT = PROD_FRAMEWORK_PROMPT;
 /**
- * Pre-load AGENTS.md and LEARNINGS.md (personal + shared) and append to the system prompt.
- * This ensures the agent always has the user's context without needing to call tools first.
+ * Pre-load the agent's context: AGENTS.md (template instructions), the skills
+ * index, and LEARNINGS.md (user notes). These all get appended to the system
+ * prompt so the agent has everything it needs from the first turn — no tool
+ * calls required to figure out "what is this app".
+ *
+ * Three sources are layered:
+ *
+ *   1. `<template>` — AGENTS.md + skills index from the `virtual:agents-bundle`
+ *      module (inlined at build time by the Vite plugin, falls back to a
+ *      filesystem read from `process.cwd()` in dev). Canonical source for
+ *      "what is this app, what can it do, what skills are available".
+ *   2. `<shared>` — LEARNINGS.md from the SQL shared scope. Team-level notes.
+ *   3. `<personal>` — LEARNINGS.md from the SQL personal scope. The current
+ *      user's own notes.
+ *
+ * Each source is read independently — no copying between them. Editing
+ * AGENTS.md and restarting the server is all it takes; Vite HMR invalidates
+ * the bundle in dev so changes land instantly.
  */
 async function loadResourcesForPrompt(owner) {
     await ensurePersonalDefaults(owner);
-    const resourceNames = ["AGENTS.md", "LEARNINGS.md"];
     const sections = [];
-    for (const name of resourceNames) {
-        // Read shared
+    // 1. Template AGENTS.md + skills index — from the virtual bundle.
+    try {
+        const { loadAgentsBundle, generateSkillsPromptBlock } = await import("./agents-bundle.js");
+        const bundle = await loadAgentsBundle();
+        if (bundle.agentsMd.trim()) {
+            sections.push(`<resource name="AGENTS.md" scope="template">\n${bundle.agentsMd.trim()}\n</resource>`);
+        }
+        const skillsBlock = generateSkillsPromptBlock(bundle);
+        if (skillsBlock)
+            sections.push(skillsBlock);
+    }
+    catch { }
+    // LEARNINGS.md from SQL (template-level instructions are in AGENTS.md above).
+    // 2. Shared SQL scope
+    try {
+        const shared = await resourceGetByPath(SHARED_OWNER, "LEARNINGS.md");
+        if (shared?.content?.trim()) {
+            sections.push(`<resource name="LEARNINGS.md" scope="shared">\n${shared.content.trim()}\n</resource>`);
+        }
+    }
+    catch { }
+    // 3. Personal SQL scope (skip if owner is the shared sentinel)
+    if (owner !== SHARED_OWNER) {
         try {
-            const shared = await resourceGetByPath(SHARED_OWNER, name);
-            if (shared?.content?.trim()) {
-                sections.push(`<resource name="${name}" scope="shared">\n${shared.content.trim()}\n</resource>`);
+            const personal = await resourceGetByPath(owner, "LEARNINGS.md");
+            if (personal?.content?.trim()) {
+                sections.push(`<resource name="LEARNINGS.md" scope="personal">\n${personal.content.trim()}\n</resource>`);
             }
         }
         catch { }
-        // Read personal (skip if owner is the shared sentinel)
-        if (owner !== SHARED_OWNER) {
-            try {
-                const personal = await resourceGetByPath(owner, name);
-                if (personal?.content?.trim()) {
-                    sections.push(`<resource name="${name}" scope="personal">\n${personal.content.trim()}\n</resource>`);
-                }
-            }
-            catch { }
-        }
     }
     if (sections.length === 0)
         return "";
     return ("\n\nThe following resources contain template-specific instructions and user context. Use the information in them to help the user.\n\n" +
         sections.join("\n\n"));
 }
+/**
+ * Build the per-request SQL-schema context block. Reads AGENT_ORG_ID live
+ * from the environment so scheduler/A2A/HTTP call sites all see whatever
+ * org was just resolved for this request.
+ */
+async function buildSchemaBlock(owner, hasRawDbTools) {
+    try {
+        return await loadSchemaPromptBlock({
+            owner,
+            orgId: process.env.AGENT_ORG_ID ?? null,
+            hasRawDbTools,
+        });
+    }
+    catch {
+        return "";
+    }
+}
 /** @deprecated Kept for backward compat — dev prompt is now part of DEV_FRAMEWORK_PROMPT */
 const DEFAULT_DEV_PROMPT = "";
 /**
  * Generates a system prompt section describing registered template actions.
  * This helps the agent prefer template-specific actions over raw db-query/db-exec.
+ *
+ * Two output modes:
+ *
+ *   - `"tool"` — used in production, where template actions are registered
+ *     as native Anthropic tools. Output reads `name(arg*: type; ...) — desc`.
+ *   - `"cli"` — used in dev, where template actions are NOT registered as
+ *     native tools and must be invoked via `shell(command="pnpm action ...")`.
+ *     Output reads `pnpm action name --arg <type> [--opt <type>] — desc`.
  */
-function generateActionsPrompt(registry) {
+function generateActionsPrompt(registry, mode = "tool") {
     if (!registry || Object.keys(registry).length === 0)
         return "";
     const lines = Object.entries(registry).map(([name, entry]) => {
         const desc = entry.tool.description;
         const params = entry.tool.parameters?.properties;
+        const requiredFields = new Set(entry.tool.parameters?.required ?? []);
+        if (mode === "cli") {
+            // CLI mode: emit `pnpm action <name> --required <type> [--optional <type>]`
+            if (!params || Object.keys(params).length === 0) {
+                return `- \`pnpm action ${name}\` — ${desc}`;
+            }
+            const entries = Object.entries(params);
+            // Required first (alphabetical), then optional (alphabetical)
+            entries.sort(([a], [b]) => {
+                const ar = requiredFields.has(a) ? 0 : 1;
+                const br = requiredFields.has(b) ? 0 : 1;
+                if (ar !== br)
+                    return ar - br;
+                return a.localeCompare(b);
+            });
+            const required = [];
+            const optional = [];
+            const requiredNames = [];
+            for (const [k, v] of entries) {
+                const type = v.type ?? "any";
+                const flag = `--${k} <${type}>`;
+                if (requiredFields.has(k)) {
+                    required.push(flag);
+                    requiredNames.push(`--${k}`);
+                }
+                else {
+                    optional.push(`[${flag}]`);
+                }
+            }
+            const cmd = ["pnpm action " + name, ...required, ...optional].join(" ");
+            const requiredNote = requiredNames.length > 0
+                ? ` Required: ${requiredNames.join(", ")}.`
+                : "";
+            return `- \`${cmd}\` — ${desc}.${requiredNote}`;
+        }
+        // tool mode (production / native tool calls)
         if (params) {
-            const paramList = Object.entries(params)
-                .map(([k, v]) => `--${k}${v.description ? ` (${v.description})` : ""}`)
-                .join(", ");
-            return `- \`${name}\` — ${desc} Args: ${paramList}`;
+            // Order required params first, then optional. Mark required with "*"
+            // and include type + description so the agent knows exactly how to call.
+            const entries = Object.entries(params);
+            entries.sort(([a], [b]) => {
+                const ar = requiredFields.has(a) ? 0 : 1;
+                const br = requiredFields.has(b) ? 0 : 1;
+                if (ar !== br)
+                    return ar - br;
+                return a.localeCompare(b);
+            });
+            const paramList = entries
+                .map(([k, v]) => {
+                const isRequired = requiredFields.has(k);
+                const type = v.type ?? "any";
+                const marker = isRequired ? "*" : "?";
+                const descPart = v.description ? ` — ${v.description}` : "";
+                return `${k}${marker}: ${type}${descPart}`;
+            })
+                .join("; ");
+            return `- \`${name}\`(${paramList}) — ${desc}`;
         }
-        return `- \`${name}\` — ${desc}`;
+        return `- \`${name}\`() — ${desc}`;
     });
-    return `\n\n## Available Actions\n\nThese are your registered template actions. ALWAYS prefer these over raw db-query/db-exec when a matching action exists:\n\n${lines.join("\n")}`;
+    if (mode === "cli") {
+        return `\n\n## Available Actions
+
+**These template actions are NOT exposed as direct tools in dev mode. To run any of them, use the \`shell\` tool with the exact command shown below.** Example: \`shell(command="pnpm action add-slide --deckId abc --content 'Hello'")\`.
+
+Do NOT try to call these by name as if they were tools — they will not exist in your tool list. Always go through \`shell\`.
+
+${lines.join("\n")}`;
+    }
+    return `\n\n## Available Actions
+
+**Use these actions directly to accomplish tasks. Do NOT use \`db-schema\`, \`search-files\`, or \`shell\` to explore the app — these actions already connect to the correct database and services.**
+
+Parameter notation: \`name*\` = required, \`name?\` = optional. Always pass the tool's parameters as a JSON object to the tool_use call — never via shell or string-concatenated CLI flags.
+
+${lines.join("\n")}`;
 }
 /**
  * Creates a Nitro plugin that mounts the agent chat endpoint.
@@ -292,7 +774,7 @@ function generateActionsPrompt(registry) {
  * Usage in templates:
  * ```ts
  * // server/plugins/agent-chat.ts
- * import { createAgentChatPlugin } from "@agent-native/core/server";
+ * import { readBody, createAgentChatPlugin } from "@agent-native/core/server";
  * import { scriptRegistry } from "../../scripts/registry.js";
  *
  * export default createAgentChatPlugin({
@@ -301,7 +783,7 @@ function generateActionsPrompt(registry) {
  * });
  * ```
  */
-function collectFiles(dir, prefix, depth, results) {
+async function collectFiles(dir, prefix, depth, results) {
     if (depth > 4 || results.length >= 500)
         return;
     const skip = new Set([
@@ -316,6 +798,7 @@ function collectFiles(dir, prefix, depth, results) {
     ]);
     let entries;
     try {
+        const fs = await lazyFs();
         entries = fs.readdirSync(dir, { withFileTypes: true });
     }
     catch {
@@ -334,7 +817,7 @@ function collectFiles(dir, prefix, depth, results) {
             type: isDir ? "folder" : "file",
         });
         if (isDir)
-            collectFiles(nodePath.join(dir, entry.name), relPath, depth + 1, results);
+            await collectFiles(nodePath.join(dir, entry.name), relPath, depth + 1, results);
     }
 }
 function parseSkillFrontmatter(content) {
@@ -358,6 +841,11 @@ function isLocalhost(event) {
 }
 export function createAgentChatPlugin(options) {
     return async (nitroApp) => {
+        // Wait for default framework plugins (auth, core-routes, integrations, ...)
+        // to finish mounting their middleware before we register our own. Without
+        // this, requests can race ahead of the bootstrap and hit the SSR catch-all.
+        const { awaitBootstrap } = await import("./framework-request-handler.js");
+        await awaitBootstrap(nitroApp);
         const env = process.env.NODE_ENV;
         // AGENT_MODE=production forces production agent constraints even in dev
         const canToggle = (env === "development" || env === "test") &&
@@ -368,18 +856,163 @@ export function createAgentChatPlugin(options) {
         const templateScripts = typeof rawActions === "function"
             ? await rawActions()
             : (rawActions ?? {});
-        // Resource and cross-agent scripts are available in both prod and dev modes
+        // Resource, chat, and cross-agent scripts are available in both prod and dev modes
         const resourceScripts = await createResourceScriptEntries();
-        const callAgentScript = await createCallAgentScriptEntry();
+        const engineScripts = await createAgentEngineScriptEntries();
+        const chatScripts = {
+            ...(await createChatScriptEntries()),
+            ...engineScripts,
+        };
+        const callAgentScript = await createCallAgentScriptEntry(options?.appId);
         // Auto-mount A2A protocol endpoints so every app is discoverable
         // and callable by other agents via the standard protocol.
-        // The custom handler calls the local agent-chat endpoint to process
-        // A2A messages using the same production agent pipeline.
-        const allScripts = {
-            ...templateScripts,
-            ...resourceScripts,
-            ...callAgentScript,
-        };
+        // In dev mode, include dev scripts (filesystem-discovered) so the A2A agent
+        // has access to the same tools as the interactive agent.
+        let devScriptsForA2A = {};
+        let discoveredActions = {};
+        if (canToggle) {
+            try {
+                const { createDevScriptRegistry } = await import("../scripts/dev/index.js");
+                devScriptsForA2A = await createDevScriptRegistry();
+            }
+            catch { }
+            // Auto-discover template action files and register as shell-based tools.
+            // This ensures templates without a custom agent-chat plugin (e.g., analytics)
+            // still have their domain actions available as tools.
+            try {
+                const fs = await import("fs");
+                const pathMod = await import("path");
+                const cwd = process.cwd();
+                const skipFiles = new Set([
+                    "helpers",
+                    "run",
+                    "registry",
+                    "_utils",
+                    "db-connect",
+                    "db-status",
+                ]);
+                for (const dir of ["actions", "scripts"]) {
+                    const actionsDir = pathMod.join(cwd, dir);
+                    const _fs = await lazyFs();
+                    if (!_fs.existsSync(actionsDir))
+                        continue;
+                    const files = _fs
+                        .readdirSync(actionsDir)
+                        .filter((f) => f.endsWith(".ts") &&
+                        !f.startsWith("_") &&
+                        !skipFiles.has(f.replace(/\.ts$/, "")));
+                    for (const file of files) {
+                        const name = file.replace(/\.ts$/, "");
+                        if (templateScripts[name] || devScriptsForA2A[name])
+                            continue;
+                        // Try to load the action module directly so we get the real
+                        // run function (not a shell wrapper). This makes HTTP endpoints
+                        // work correctly. Only fall back to shell wrapper if the import
+                        // fails (e.g., CLI-style scripts that throw at top level).
+                        const filePath = pathMod.join(actionsDir, file);
+                        try {
+                            const mod = await import(/* @vite-ignore */ filePath);
+                            const def = mod.default && typeof mod.default === "object"
+                                ? mod.default
+                                : mod;
+                            if (def?.tool && typeof def.run === "function") {
+                                discoveredActions[name] = {
+                                    tool: def.tool,
+                                    run: def.run,
+                                    ...(def.http !== undefined ? { http: def.http } : {}),
+                                };
+                                continue;
+                            }
+                        }
+                        catch {
+                            // Fall through to shell wrapper for CLI-style scripts
+                            // (and .ts files Node can't parse natively).
+                        }
+                        // Static-parse the source for `http: false` or
+                        // `http: { method: "GET" }` so the shell-wrapper fallback still
+                        // mounts HTTP routes with the correct method. We can't load the
+                        // .ts module to read the real defineAction object in this Node
+                        // context, so this regex sniff is the best we can do until the
+                        // discovery is moved into a Vite-aware codepath.
+                        let httpConfig;
+                        try {
+                            const src = _fs.readFileSync(filePath, "utf-8");
+                            if (/\bhttp\s*:\s*false\b/.test(src)) {
+                                httpConfig = false;
+                            }
+                            else {
+                                const httpStart = src.search(/\bhttp\s*:\s*\{/);
+                                if (httpStart >= 0) {
+                                    const window = src.slice(httpStart, httpStart + 200);
+                                    const m = window.match(/method\s*:\s*['"`](GET|POST|PUT|DELETE)['"`]/);
+                                    const p = window.match(/path\s*:\s*['"`]([^'"`]+)['"`]/);
+                                    if (m || p) {
+                                        httpConfig = {
+                                            ...(m
+                                                ? {
+                                                    method: m[1],
+                                                }
+                                                : {}),
+                                            ...(p ? { path: p[1] } : {}),
+                                        };
+                                    }
+                                }
+                            }
+                        }
+                        catch {
+                            // File read failed — leave httpConfig undefined (default POST)
+                        }
+                        // Fallback: shell-based wrapper for CLI-style scripts
+                        discoveredActions[name] = {
+                            tool: {
+                                description: `Run the ${name} action. Use: pnpm action ${name} --arg=value`,
+                                parameters: {
+                                    type: "object",
+                                    properties: {
+                                        args: {
+                                            type: "string",
+                                            description: "CLI arguments as a string (e.g., --metrics=sessions --days=7)",
+                                        },
+                                    },
+                                },
+                            },
+                            run: async (input) => {
+                                const shellEntry = devScriptsForA2A["shell"];
+                                if (!shellEntry)
+                                    return "Error: shell not available";
+                                return shellEntry.run({
+                                    command: `pnpm action ${name} ${input.args || ""}`.trim(),
+                                });
+                            },
+                            ...(httpConfig !== undefined ? { http: httpConfig } : {}),
+                        };
+                    }
+                }
+                if (Object.keys(discoveredActions).length > 0 && process.env.DEBUG)
+                    console.log(`[agent-chat] Auto-discovered ${Object.keys(discoveredActions).length} action(s): ${Object.keys(discoveredActions).join(", ")}`);
+            }
+            catch { }
+        }
+        // In dev mode, template actions (templateScripts and discoveredActions) are
+        // NOT registered as native tools — the agent invokes them via shell instead.
+        // This avoids degenerate empty-object tool calls that Anthropic models
+        // sometimes emit for actions with complex schemas. Production keeps the
+        // native registration since it has no shell access.
+        const allScripts = canToggle
+            ? {
+                ...resourceScripts,
+                ...chatScripts,
+                ...callAgentScript,
+                ...devScriptsForA2A,
+            }
+            : {
+                ...discoveredActions,
+                ...templateScripts,
+                ...resourceScripts,
+                ...chatScripts,
+                ...callAgentScript,
+                ...devScriptsForA2A,
+            };
         const { mountA2A } = await import("../a2a/server.js");
         mountA2A(nitroApp, {
             name: options?.appId
@@ -394,12 +1027,9 @@ export function createAgentChatPlugin(options) {
             streaming: true,
             handler: async function* (message, context) {
                 // Resolve the caller's identity for user-scoped data access.
-                // Dev: use most recent session from local DB (single user, no verification).
-                // Prod: verify Google OAuth token from caller, check email domain.
                 const isDev = process.env.NODE_ENV !== "production";
                 let userEmail;
                 if (isDev) {
-                    // Dev mode: trust the caller's claimed email, or auto-detect from local sessions
                     userEmail = context.metadata?.userEmail || undefined;
                     if (!userEmail) {
                         try {
@@ -416,7 +1046,6 @@ export function createAgentChatPlugin(options) {
                     }
                 }
                 else {
-                    // Prod mode: verify identity via Google OAuth token
                     const googleToken = context.metadata?.googleToken;
                     if (googleToken) {
                         try {
@@ -447,109 +1076,156 @@ export function createAgentChatPlugin(options) {
                     };
                     return;
                 }
-                const Anthropic = (await import("@anthropic-ai/sdk")).default;
-                const apiKey = options?.apiKey ?? process.env.ANTHROPIC_API_KEY;
-                if (!apiKey) {
-                    yield {
-                        role: "agent",
-                        parts: [
-                            {
-                                type: "text",
-                                text: "Anthropic API key is not configured. Set ANTHROPIC_API_KEY in .env.",
-                            },
-                        ],
-                    };
-                    return;
-                }
-                const client = new Anthropic({ apiKey });
-                const model = options?.model ?? "claude-sonnet-4-6";
-                // Load shared AGENTS.md resource so the agent knows how to use its tools
+                // Use the SAME agent setup as the interactive chat — identical tools,
+                // prompt, and capabilities. The A2A agent IS the app's agent.
+                const a2aEngine = await resolveEngine({
+                    engineOption: options?.engine,
+                    apiKey: options?.apiKey,
+                });
+                // Use the same handler (dev or prod) that the interactive chat uses
+                const handler = canToggle && devHandler ? devHandler : prodHandler;
+                // Build the same system prompt the interactive agent uses
                 const owner = userEmail || "local@localhost";
                 const resources = await loadResourcesForPrompt(owner);
-                const a2aSystemPrompt = `You are the ${options?.appId ?? "app"} agent responding to a request from another agent. Be concise and helpful. Use your tools to look up data or take actions as needed. Do not ask for clarification — use your tools to find the answer.` +
-                    resources;
-                const tools = Object.entries(templateScripts).map(([name, entry]) => ({
-                    name,
-                    description: entry.tool.description,
-                    input_schema: entry.tool.parameters ?? {
-                        type: "object",
-                        properties: {},
-                    },
-                }));
-                const msgs = [{ role: "user", content: text }];
+                const schemaBlock = await buildSchemaBlock(owner, canToggle);
+                const systemPrompt = canToggle
+                    ? devPrompt + resources + schemaBlock
+                    : basePrompt + resources + schemaBlock;
+                const model = options?.model ??
+                    (canToggle ? "claude-sonnet-4-6" : "claude-haiku-4-5-20251001");
+                // Build tools — same as interactive handler but WITHOUT call-agent
+                // to prevent infinite recursive A2A loops (agent calling itself).
+                // In dev mode, template actions are invoked via shell (not native tools),
+                // so they're omitted from the tool registry — see allScripts comment.
+                const a2aActions = canToggle
+                    ? {
+                        ...resourceScripts,
+                        ...chatScripts,
+                        ...devScriptsForA2A,
+                    }
+                    : {
+                        ...templateScripts,
+                        ...resourceScripts,
+                        ...chatScripts,
+                    };
+                const a2aTools = actionsToEngineTools(a2aActions);
+                const a2aMessages = [
+                    { role: "user", content: [{ type: "text", text }] },
+                ];
+                // Run the SAME agent loop, collect text events, yield as A2A messages
                 let accumulatedText = "";
-                for (let i = 0; i < 10; i++) {
-                    // Stream the Anthropic response so we can yield text chunks
-                    const apiStream = client.messages.stream({
-                        model,
-                        max_tokens: 4096,
-                        system: a2aSystemPrompt,
-                        tools: tools.length > 0 ? tools : undefined,
-                        messages: msgs,
-                    });
-                    // Yield text deltas as they arrive
-                    for await (const event of apiStream) {
-                        if (event.type === "content_block_delta" &&
-                            event.delta?.type === "text_delta") {
-                            accumulatedText += event.delta.text;
-                            yield {
-                                role: "agent",
-                                parts: [{ type: "text", text: accumulatedText }],
-                            };
+                const controller = new AbortController();
+                console.log(`[A2A] Starting agent loop: ${a2aTools.length} tools, prompt ${systemPrompt.length} chars`);
+                await runAgentLoop({
+                    engine: a2aEngine,
+                    model,
+                    systemPrompt,
+                    tools: a2aTools,
+                    messages: a2aMessages,
+                    actions: a2aActions,
+                    send: (event) => {
+                        if (event.type === "text") {
+                            accumulatedText += event.text;
                         }
-                    }
-                    // Get the complete message for tool_use detection
-                    const finalMessage = await apiStream.finalMessage();
-                    const toolUseBlocks = finalMessage.content.filter((b) => b.type === "tool_use");
-                    if (toolUseBlocks.length === 0 ||
-                        finalMessage.stop_reason !== "tool_use") {
-                        break;
-                    }
-                    // Execute tools and continue the loop
-                    msgs.push({ role: "assistant", content: finalMessage.content });
-                    const toolResults = [];
-                    for (const tb of toolUseBlocks) {
-                        const script = allScripts[tb.name];
-                        let result = `Unknown tool: ${tb.name}`;
-                        if (script) {
-                            try {
-                                result = await script.run(tb.input);
-                            }
-                            catch (err) {
-                                result = `Error: ${err?.message}`;
-                            }
+                        else if (event.type === "tool_start") {
+                            console.log(`[A2A] Tool call: ${event.tool}`);
                         }
-                        toolResults.push({
-                            type: "tool_result",
-                            tool_use_id: tb.id,
-                            content: result,
-                        });
-                    }
-                    msgs.push({ role: "user", content: toolResults });
-                }
-                // Final yield if nothing was yielded yet
-                if (!accumulatedText) {
-                    yield {
-                        role: "agent",
-                        parts: [{ type: "text", text: "(no response)" }],
-                    };
-                }
+                        else if (event.type === "error") {
+                            console.error(`[A2A] Error: ${event.error}`);
+                        }
+                        else if (event.type === "done") {
+                            console.log(`[A2A] Done. Response: ${accumulatedText.length} chars`);
+                        }
+                    },
+                    signal: controller.signal,
+                });
+                console.log(`[A2A] Loop complete. Text: ${accumulatedText.slice(0, 100)}...`);
+                // Yield the final accumulated text
+                yield {
+                    role: "agent",
+                    parts: [
+                        {
+                            type: "text",
+                            text: accumulatedText || "(no response)",
+                        },
+                    ],
+                };
             },
         });
         // Generate an "Available Actions" section from template-specific actions
-        // so the agent knows to use them instead of raw SQL
-        const actionsPrompt = generateActionsPrompt(templateScripts);
+        // so the agent knows to use them instead of raw SQL.
+        //
+        // Production: actions are native tools — emit `name(arg*: type) — desc`
+        // Dev: actions are invoked via shell — emit `pnpm action name --arg <type>`
+        //      and include discoveredActions too, since those are also missing
+        //      from the dev tool registry.
+        const prodActionsPrompt = generateActionsPrompt(templateScripts, "tool");
+        const devActionsPrompt = generateActionsPrompt({ ...discoveredActions, ...templateScripts }, "cli");
         // Build system prompts — dynamic functions that pre-load resources per-request.
         // Production gets PROD_FRAMEWORK_PROMPT, dev gets DEV_FRAMEWORK_PROMPT.
         // Custom systemPrompt from options overrides the framework default entirely.
-        const prodPrompt = (options?.systemPrompt ?? PROD_FRAMEWORK_PROMPT) + actionsPrompt;
+        const prodPrompt = (options?.systemPrompt ?? PROD_FRAMEWORK_PROMPT) + prodActionsPrompt;
         const devPrompt = (options?.devSystemPrompt
             ? options.devSystemPrompt +
                 (options?.systemPrompt ?? PROD_FRAMEWORK_PROMPT)
-            : DEV_FRAMEWORK_PROMPT) + actionsPrompt;
+            : DEV_FRAMEWORK_PROMPT) + devActionsPrompt;
         // Keep legacy names for the composition below
         const basePrompt = prodPrompt;
         const devPrefix = options?.devSystemPrompt ?? DEFAULT_DEV_PROMPT;
+        // Mount MCP remote server — same action registry as A2A + agent chat
+        const { mountMCP } = await import("../mcp/server.js");
+        mountMCP(nitroApp, {
+            name: options?.appId
+                ? options.appId.charAt(0).toUpperCase() + options.appId.slice(1)
+                : "Agent",
+            description: `Agent-native ${options?.appId ?? "app"} agent`,
+            actions: allScripts,
+            askAgent: async (message) => {
+                const mcpEngine = await resolveEngine({
+                    engineOption: options?.engine,
+                    apiKey: options?.apiKey,
+                });
+                const model = options?.model ??
+                    (canToggle ? "claude-sonnet-4-6" : "claude-haiku-4-5-20251001");
+                // Same actions as A2A — without call-agent to prevent loops.
+                // In dev mode, template actions go through shell, not native tools.
+                const mcpActions = canToggle
+                    ? {
+                        ...resourceScripts,
+                        ...chatScripts,
+                        ...devScriptsForA2A,
+                    }
+                    : {
+                        ...templateScripts,
+                        ...resourceScripts,
+                        ...chatScripts,
+                    };
+                const mcpTools = actionsToEngineTools(mcpActions);
+                const resources = await loadResourcesForPrompt("local@localhost");
+                const schemaBlock = await buildSchemaBlock("local@localhost", canToggle);
+                const systemPrompt = canToggle
+                    ? devPrompt + resources + schemaBlock
+                    : basePrompt + resources + schemaBlock;
+                let accumulatedText = "";
+                const controller = new AbortController();
+                await runAgentLoop({
+                    engine: mcpEngine,
+                    model,
+                    systemPrompt,
+                    tools: mcpTools,
+                    messages: [
+                        { role: "user", content: [{ type: "text", text: message }] },
+                    ],
+                    actions: mcpActions,
+                    send: (event) => {
+                        if (event.type === "text")
+                            accumulatedText += event.text;
+                    },
+                    signal: controller.signal,
+                });
+                return accumulatedText || "(no response)";
+            },
+        });
         // Resolve owner from the H3 event's session — matches how resources are created
         const getOwnerFromEvent = async (event) => {
             try {
@@ -560,6 +1236,20 @@ export function createAgentChatPlugin(options) {
                 return "local@localhost";
             }
         };
+        // Auto-mount template actions as HTTP endpoints under /_agent-native/actions/
+        // Include engine management scripts so the UI can call list/set/test-agent-engine.
+        const httpActions = {
+            ...discoveredActions,
+            ...templateScripts,
+            ...engineScripts,
+        };
+        if (Object.keys(httpActions).length > 0) {
+            const { mountActionRoutes } = await import("./action-routes.js");
+            mountActionRoutes(nitroApp, httpActions, {
+                getOwnerFromEvent,
+                resolveOrgId: options?.resolveOrgId,
+            });
+        }
         // Callback to persist agent response when run finishes (even if client disconnected).
         // Reconstructs the assistant message from buffered events and appends to thread_data.
         const onRunComplete = async (run, threadId) => {
@@ -588,12 +1278,24 @@ export function createAgentChatPlugin(options) {
                 if (!Array.isArray(repo.messages))
                     repo.messages = [];
                 const lastMsg = repo.messages[repo.messages.length - 1];
-                if (lastMsg?.role === "assistant") {
+                // Check both wrapped ({ message: { role } }) and unwrapped ({ role }) formats
+                const lastRole = lastMsg?.message?.role ?? lastMsg?.role;
+                if (lastRole === "assistant") {
                     // Frontend already saved the assistant response — just bump timestamp
                     await updateThreadData(threadId, thread.threadData, thread.title, thread.preview, thread.messageCount);
                     return;
                 }
-                repo.messages.push(assistantMsg);
+                // Determine if repo uses wrapped format ({ message, parentId }) or flat format
+                const isWrapped = lastMsg && "message" in lastMsg;
+                if (isWrapped) {
+                    const parentId = repo.messages.length > 0
+                        ? (repo.messages[repo.messages.length - 1].message?.id ?? null)
+                        : null;
+                    repo.messages.push({ message: assistantMsg, parentId });
+                }
+                else {
+                    repo.messages.push(assistantMsg);
+                }
                 const meta = extractThreadMeta(repo);
                 await updateThreadData(threadId, JSON.stringify(repo), meta.title || thread.title, meta.preview || thread.preview, repo.messages.length);
             }
@@ -601,38 +1303,128 @@ export function createAgentChatPlugin(options) {
                 // Best-effort — don't break cleanup
             }
         };
-        // Always build the production handler (includes resource tools + call-agent)
+        // ─── Agent Teams: per-run send reference ─────────────────────────
+        // Team tools need to emit events to the parent chat's SSE stream.
+        // Each run gets its own send function, keyed by threadId so concurrent
+        // requests for different threads don't clobber each other.
+        const _runSendByThread = new Map();
+        let _currentRunOwner = "local@localhost";
+        let _currentRunThreadId = "";
+        let _currentRunSystemPrompt = basePrompt;
+        // Default to Haiku in production mode to manage costs for hosted apps
+        const resolvedModel = options?.model ??
+            (canToggle ? "claude-sonnet-4-6" : "claude-haiku-4-5-20251001");
+        const teamTools = createTeamTools({
+            getOwner: () => _currentRunOwner,
+            getSystemPrompt: () => _currentRunSystemPrompt,
+            getActions: () => canToggle
+                ? {
+                    // Sub-agents spawned in dev mode also invoke template actions
+                    // via shell, so omit them from the native tool registry.
+                    ...resourceScripts,
+                    ...chatScripts,
+                    ...devScriptsForA2A,
+                }
+                : {
+                    ...templateScripts,
+                    ...resourceScripts,
+                    ...chatScripts,
+                },
+            getEngine: () => createAnthropicEngine({
+                apiKey: options?.apiKey ?? process.env.ANTHROPIC_API_KEY,
+            }),
+            getModel: () => resolvedModel,
+            getParentThreadId: () => _currentRunThreadId,
+            getSend: () => {
+                // Return the send for the current run's thread
+                const send = _runSendByThread.get(_currentRunThreadId);
+                return send ?? null;
+            },
+        });
+        // Hook into the run lifecycle to set/clear the send reference.
+        // Job management tools (create-job, list-jobs, update-job)
+        let jobTools = {};
+        try {
+            const { createJobTools } = await import("../jobs/tools.js");
+            jobTools = createJobTools();
+        }
+        catch { }
+        const prodActions = {
+            ...templateScripts,
+            ...resourceScripts,
+            ...chatScripts,
+            ...callAgentScript,
+            ...teamTools,
+            ...jobTools,
+        };
+        // Always build the production handler (includes resource tools + call-agent + team tools)
+        // In production mode (!canToggle), enable usage tracking and limits
+        const isHostedProd = !canToggle;
         const prodHandler = createProductionAgentHandler({
-            actions: { ...templateScripts, ...resourceScripts, ...callAgentScript },
+            actions: prodActions,
             systemPrompt: async (event) => {
                 const owner = await getOwnerFromEvent(event);
+                _currentRunOwner = owner;
                 const resources = await loadResourcesForPrompt(owner);
-                return basePrompt + resources;
+                const schemaBlock = await buildSchemaBlock(owner, false);
+                _currentRunSystemPrompt = basePrompt + resources + schemaBlock;
+                return _currentRunSystemPrompt;
             },
-            model: options?.model,
+            model: options?.model ??
+                (isHostedProd ? "claude-haiku-4-5-20251001" : undefined),
             apiKey: options?.apiKey,
-            onRunComplete,
+            onRunStart: (send, threadId) => {
+                _runSendByThread.set(threadId, send);
+                _currentRunThreadId = threadId;
+            },
+            onRunComplete: async (run, threadId) => {
+                if (threadId)
+                    _runSendByThread.delete(threadId);
+                await onRunComplete(run, threadId);
+            },
+            // Usage tracking for hosted production deployments
+            trackUsage: isHostedProd,
+            resolveOwnerEmail: isHostedProd ? getOwnerFromEvent : undefined,
         });
         // Build the dev handler (with filesystem/shell/db tools) if environment allows toggling
         let devHandler = null;
         if (canToggle) {
             const { createDevScriptRegistry } = await import("../scripts/dev/index.js");
+            // Dev mode: template actions (templateScripts and discoveredActions) are
+            // intentionally OMITTED from the native tool registry. The agent invokes
+            // them via `shell(command="pnpm action <name> ...")` instead. This mirrors
+            // how Claude Code works locally and dramatically reduces the rate of
+            // degenerate empty-object tool calls. The CLI syntax for each action is
+            // listed in the dev system prompt's "Available Actions" section.
             const devActions = {
-                ...templateScripts,
                 ...resourceScripts,
+                ...chatScripts,
                 ...callAgentScript,
+                ...teamTools,
+                ...jobTools,
                 ...(await createDevScriptRegistry()),
             };
             devHandler = createProductionAgentHandler({
                 actions: devActions,
                 systemPrompt: async (event) => {
                     const owner = await getOwnerFromEvent(event);
+                    _currentRunOwner = owner;
                     const resources = await loadResourcesForPrompt(owner);
-                    return devPrompt + resources;
+                    const schemaBlock = await buildSchemaBlock(owner, true);
+                    _currentRunSystemPrompt = devPrompt + resources + schemaBlock;
+                    return _currentRunSystemPrompt;
                 },
                 model: options?.model,
                 apiKey: options?.apiKey,
-                onRunComplete,
+                onRunStart: (send, threadId) => {
+                    _runSendByThread.set(threadId, send);
+                    _currentRunThreadId = threadId;
+                },
+                onRunComplete: async (run, threadId) => {
+                    if (threadId)
+                        _runSendByThread.delete(threadId);
+                    await onRunComplete(run, threadId);
+                },
             });
         }
         // Resolve mention providers
@@ -643,7 +1435,7 @@ export function createAgentChatPlugin(options) {
         // Mutable mode flag — starts in dev if environment allows
         let currentDevMode = canToggle;
         // Mount mode endpoint — GET returns current mode, POST toggles it (localhost only)
-        nitroApp.h3App.use(`${routePath}/mode`, defineEventHandler(async (event) => {
+        getH3App(nitroApp).use(`${routePath}/mode`, defineEventHandler(async (event) => {
             if (getMethod(event) === "POST") {
                 if (!canToggle) {
                     setResponseStatus(event, 403);
@@ -666,7 +1458,7 @@ export function createAgentChatPlugin(options) {
         }));
         // Mount save-key BEFORE the prefix handler so it isn't shadowed
         // Only functional in Node.js environments (writes to .env file)
-        nitroApp.h3App.use(`${routePath}/save-key`, defineEventHandler(async (event) => {
+        getH3App(nitroApp).use(`${routePath}/save-key`, defineEventHandler(async (event) => {
             if (getMethod(event) !== "POST") {
                 setResponseStatus(event, 405);
                 return { error: "Method not allowed" };
@@ -682,7 +1474,7 @@ export function createAgentChatPlugin(options) {
                 const path = await import("path");
                 const { upsertEnvFile } = await import("./create-server.js");
                 const envPath = path.join(process.cwd(), ".env");
-                upsertEnvFile(envPath, [
+                await upsertEnvFile(envPath, [
                     { key: "ANTHROPIC_API_KEY", value: trimmedKey },
                 ]);
             }
@@ -694,7 +1486,7 @@ export function createAgentChatPlugin(options) {
             return { ok: true };
         }));
         // Mount file search endpoint
-        nitroApp.h3App.use(`${routePath}/files`, defineEventHandler(async (event) => {
+        getH3App(nitroApp).use(`${routePath}/files`, defineEventHandler(async (event) => {
             if (getMethod(event) !== "GET") {
                 setResponseStatus(event, 405);
                 return { error: "Method not allowed" };
@@ -707,7 +1499,7 @@ export function createAgentChatPlugin(options) {
             if (currentDevMode) {
                 const codebaseFiles = [];
                 try {
-                    collectFiles(process.cwd(), "", 0, codebaseFiles);
+                    await collectFiles(process.cwd(), "", 0, codebaseFiles);
                 }
                 catch {
                     // Filesystem access failed — skip
@@ -751,7 +1543,7 @@ export function createAgentChatPlugin(options) {
             return { files: filtered.slice(0, 30) };
         }));
         // Mount skills listing endpoint
-        nitroApp.h3App.use(`${routePath}/skills`, defineEventHandler(async (event) => {
+        getH3App(nitroApp).use(`${routePath}/skills`, defineEventHandler(async (event) => {
             if (getMethod(event) !== "GET") {
                 setResponseStatus(event, 405);
                 return { error: "Method not allowed" };
@@ -761,8 +1553,9 @@ export function createAgentChatPlugin(options) {
             // In dev mode, scan .agents/skills/ directory
             if (currentDevMode) {
                 try {
+                    const _fs = await lazyFs();
                     const skillsDir = nodePath.join(process.cwd(), ".agents", "skills");
-                    const entries = fs.readdirSync(skillsDir, {
+                    const entries = _fs.readdirSync(skillsDir, {
                         withFileTypes: true,
                     });
                     for (const entry of entries) {
@@ -772,7 +1565,7 @@ export function createAgentChatPlugin(options) {
                         if (entry.isDirectory()) {
                             // Subdirectory layout: .agents/skills/<name>/SKILL.md
                             const candidate = nodePath.join(skillsDir, entry.name, "SKILL.md");
-                            if (!fs.existsSync(candidate))
+                            if (!_fs.existsSync(candidate))
                                 continue;
                             skillFilePath = candidate;
                             skillRelPath = `.agents/skills/${entry.name}/SKILL.md`;
@@ -786,7 +1579,7 @@ export function createAgentChatPlugin(options) {
                             continue;
                         }
                         try {
-                            const content = fs.readFileSync(skillFilePath, "utf-8");
+                            const content = _fs.readFileSync(skillFilePath, "utf-8");
                             const fm = parseSkillFrontmatter(content);
                             const skillName = fm.name || entry.name.replace(/\.md$/, "");
                             if (!seenNames.has(skillName)) {
@@ -851,107 +1644,144 @@ export function createAgentChatPlugin(options) {
             return result;
         }));
         // Mount unified mentions endpoint (files + resources + custom providers)
-        nitroApp.h3App.use(`${routePath}/mentions`, defineEventHandler(async (event) => {
+        getH3App(nitroApp).use(`${routePath}/mentions`, defineEventHandler(async (event) => {
             if (getMethod(event) !== "GET") {
                 setResponseStatus(event, 405);
                 return { error: "Method not allowed" };
             }
             const query = getQuery(event);
             const q = typeof query.q === "string" ? query.q.toLowerCase() : "";
-            const items = [];
-            // 1. Built-in: files from codebase (dev mode only)
-            if (currentDevMode) {
-                const codebaseFiles = [];
-                try {
-                    collectFiles(process.cwd(), "", 0, codebaseFiles);
-                }
-                catch { }
-                for (const f of codebaseFiles) {
-                    items.push({
-                        id: `codebase:${f.path}`,
-                        label: f.name,
-                        description: f.path !== f.name ? f.path : undefined,
-                        icon: f.type,
-                        source: "codebase",
-                        refType: "file",
-                        refPath: f.path,
-                        section: "Files",
-                    });
-                }
-            }
-            // 2. Built-in: resources from SQL
-            try {
-                const resources = currentDevMode
-                    ? await resourceListAccessible("local@localhost")
-                    : await resourceList(SHARED_OWNER);
-                for (const r of resources) {
-                    const isShared = r.owner === SHARED_OWNER;
-                    items.push({
-                        id: `resource:${r.path}`,
-                        label: r.path.split("/").pop() || r.path,
-                        description: r.path,
-                        icon: "file",
-                        source: isShared ? "resource:shared" : "resource:private",
-                        refType: "file",
-                        refPath: r.path,
-                        section: "Files",
-                    });
-                }
-            }
-            catch { }
-            // 3. Custom mention providers
-            const providerResults = await Promise.all(Object.entries(mentionProviders).map(async ([key, provider]) => {
-                try {
-                    const providerItems = await provider.search(q);
-                    return providerItems.map((item) => ({
-                        id: item.id,
-                        label: item.label,
-                        description: item.description,
-                        icon: item.icon || provider.icon || "file",
-                        source: key,
-                        refType: item.refType,
-                        refPath: item.refPath,
-                        refId: item.refId,
-                        section: provider.label,
-                    }));
-                }
-                catch {
-                    return [];
-                }
-            }));
-            for (const batch of providerResults) {
-                items.push(...batch);
-            }
-            // 4. Discovered peer agents
-            try {
-                const { discoverAgents } = await import("./agent-discovery.js");
-                const agents = discoverAgents(options?.appId);
-                for (const agent of agents) {
-                    items.push({
-                        id: `agent:${agent.id}`,
-                        label: agent.name,
-                        description: agent.description,
-                        icon: "agent",
-                        source: "agent",
-                        refType: "agent",
-                        refPath: agent.url,
-                        refId: agent.id,
-                        section: "Agents",
-                    });
-                }
-            }
-            catch {
-                // Agent discovery not available — skip
-            }
-            // Filter by query and limit
-            const filtered = q
-                ? items.filter((item) => item.label.toLowerCase().includes(q) ||
-                    (item.description?.toLowerCase().includes(q) ?? false))
-                : items;
-            return { items: filtered.slice(0, 30) };
+            const matchesQuery = (item) => !q ||
+                item.label.toLowerCase().includes(q) ||
+                (item.description?.toLowerCase().includes(q) ?? false);
+            const enc = new TextEncoder();
+            // Stream NDJSON — each source flushes its batch as soon as it's ready.
+            setResponseHeader(event, "Content-Type", "application/x-ndjson");
+            setResponseHeader(event, "Cache-Control", "no-cache");
+            const stream = new ReadableStream({
+                async start(controller) {
+                    const MAX_RESULTS = 50;
+                    let totalSent = 0;
+                    let cancelled = false;
+                    const flush = (batch) => {
+                        if (cancelled)
+                            return;
+                        const filtered = batch.filter(matchesQuery);
+                        if (filtered.length === 0)
+                            return;
+                        const remaining = MAX_RESULTS - totalSent;
+                        const toSend = filtered.slice(0, remaining);
+                        if (toSend.length > 0) {
+                            totalSent += toSend.length;
+                            try {
+                                controller.enqueue(enc.encode(JSON.stringify({ items: toSend }) + "\n"));
+                            }
+                            catch {
+                                // Stream was closed by client
+                                cancelled = true;
+                            }
+                        }
+                    };
+                    // All sources run in parallel; each flushes independently.
+                    const sources = [];
+                    // 1. Resources from SQL (fast — flush first)
+                    sources.push((async () => {
+                        try {
+                            const resources = currentDevMode
+                                ? await resourceListAccessible("local@localhost")
+                                : await resourceList(SHARED_OWNER);
+                            flush(resources.map((r) => {
+                                const isShared = r.owner === SHARED_OWNER;
+                                return {
+                                    id: `resource:${r.path}`,
+                                    label: r.path.split("/").pop() || r.path,
+                                    description: r.path,
+                                    icon: "file",
+                                    source: isShared
+                                        ? "resource:shared"
+                                        : "resource:private",
+                                    refType: "file",
+                                    refPath: r.path,
+                                    section: "Files",
+                                };
+                            }));
+                        }
+                        catch { }
+                    })());
+                    // 2. Codebase files (dev mode only — can be slow on large repos)
+                    if (currentDevMode) {
+                        sources.push((async () => {
+                            const codebaseFiles = [];
+                            try {
+                                await collectFiles(process.cwd(), "", 0, codebaseFiles);
+                            }
+                            catch { }
+                            flush(codebaseFiles.map((f) => ({
+                                id: `codebase:${f.path}`,
+                                label: f.name,
+                                description: f.path !== f.name ? f.path : undefined,
+                                icon: f.type,
+                                source: "codebase",
+                                refType: "file",
+                                refPath: f.path,
+                                section: "Files",
+                            })));
+                        })());
+                    }
+                    // 3. Custom mention providers (each flushes independently)
+                    for (const [key, provider] of Object.entries(mentionProviders)) {
+                        sources.push((async () => {
+                            try {
+                                const providerItems = await provider.search(q, event);
+                                flush(providerItems.map((item) => ({
+                                    id: item.id,
+                                    label: item.label,
+                                    description: item.description,
+                                    icon: item.icon || provider.icon || "file",
+                                    source: key,
+                                    refType: item.refType,
+                                    refPath: item.refPath,
+                                    refId: item.refId,
+                                    section: provider.label,
+                                })));
+                            }
+                            catch (e) {
+                                console.error(`[agent-native] Mention provider "${key}" failed:`, e);
+                            }
+                        })());
+                    }
+                    // 4. Peer agent discovery (network call — often slowest)
+                    sources.push((async () => {
+                        try {
+                            const agents = await discoverAgents(options?.appId);
+                            flush(agents.map((agent) => ({
+                                id: `agent:${agent.id}`,
+                                label: agent.name,
+                                description: agent.description,
+                                icon: "agent",
+                                source: "agent",
+                                refType: "agent",
+                                refPath: agent.url,
+                                refId: agent.id,
+                                section: "Agents",
+                            })));
+                        }
+                        catch (e) {
+                            console.error("[agent-native] Agent discovery failed:", e);
+                        }
+                    })());
+                    await Promise.all(sources);
+                    if (!cancelled)
+                        controller.close();
+                },
+                cancel() {
+                    // Client disconnected — stop enqueuing
+                },
+            });
+            return stream;
         }));
         // ─── Generate thread title ──────────────────────────────────────────
-        nitroApp.h3App.use(`${routePath}/generate-title`, defineEventHandler(async (event) => {
+        getH3App(nitroApp).use(`${routePath}/generate-title`, defineEventHandler(async (event) => {
             if (getMethod(event) !== "POST") {
                 setResponseStatus(event, 405);
                 return { error: "Method not allowed" };
@@ -963,10 +1793,12 @@ export function createAgentChatPlugin(options) {
                 setResponseStatus(event, 400);
                 return { error: "message is required" };
             }
+            // Strip mention markup: @[Name|type] → @Name
+            const cleanMessage = message.replace(/@\[([^\]|]+)\|[^\]]*\]/g, "@$1");
             const apiKey = process.env.ANTHROPIC_API_KEY;
             if (!apiKey) {
                 // Fallback: truncate the message
-                return { title: message.trim().slice(0, 60) };
+                return { title: cleanMessage.trim().slice(0, 60) };
             }
             try {
                 const res = await fetch("https://api.anthropic.com/v1/messages", {
@@ -982,38 +1814,42 @@ export function createAgentChatPlugin(options) {
                         messages: [
                             {
                                 role: "user",
-                                content: `Generate a very short title (3-6 words, no quotes) for a chat that starts with this message:\n\n${message.slice(0, 500)}`,
+                                content: `Generate a very short title (3-6 words, no quotes) for a chat that starts with this message:\n\n${cleanMessage.slice(0, 500)}`,
                             },
                         ],
                     }),
                 });
                 if (!res.ok) {
-                    return { title: message.trim().slice(0, 60) };
+                    return { title: cleanMessage.trim().slice(0, 60) };
                 }
                 const data = (await res.json());
                 const text = data.content?.[0]?.text?.trim();
-                return { title: text || message.trim().slice(0, 60) };
+                return { title: text || cleanMessage.trim().slice(0, 60) };
             }
             catch {
-                return { title: message.trim().slice(0, 60) };
+                return { title: cleanMessage.trim().slice(0, 60) };
             }
         }));
         // ─── Run management endpoints (for hot-reload resilience) ─────────────
         // GET /runs/active?threadId=X — check if there's an active run for a thread
-        nitroApp.h3App.use(`${routePath}/runs`, defineEventHandler(async (event) => {
+        getH3App(nitroApp).use(`${routePath}/runs`, defineEventHandler(async (event) => {
             // Auth check — ensure the user is authenticated
             await getOwnerFromEvent(event);
             const method = getMethod(event);
             const url = event.node?.req?.url || event.path || "";
             // Route: POST /runs/:id/abort
-            const abortMatch = url.match(/\/runs\/([^/?]+)\/abort/);
+            // Match both full URL (/runs/{id}/abort) and h3 prefix-stripped (/{id}/abort)
+            const abortMatch = url.match(/\/runs\/([^/?]+)\/abort/) ||
+                url.match(/^\/([^/?]+)\/abort/);
             if (abortMatch && method === "POST") {
                 const runId = decodeURIComponent(abortMatch[1]);
                 abortRun(runId); // Aborts in-memory + marks aborted in SQL
                 return { ok: true };
             }
             // Route: GET /runs/:id/events?after=N
-            const eventsMatch = url.match(/\/runs\/([^/?]+)\/events/);
+            // Match both full URL (/runs/{id}/events) and h3 prefix-stripped (/{id}/events)
+            const eventsMatch = url.match(/\/runs\/([^/?]+)\/events/) ||
+                url.match(/^\/([^/?]+)\/events/);
             if (eventsMatch && method === "GET") {
                 const runId = decodeURIComponent(eventsMatch[1]);
                 const query = getQuery(event);
@@ -1054,7 +1890,7 @@ export function createAgentChatPlugin(options) {
         // ─── Thread management endpoints ──────────────────────────────────────
         // Single handler for /threads and /threads/:id — h3's use() does prefix
         // matching so we can't reliably split them into separate handlers.
-        nitroApp.h3App.use(`${routePath}/threads`, defineEventHandler(async (event) => {
+        getH3App(nitroApp).use(`${routePath}/threads`, defineEventHandler(async (event) => {
             const owner = await getOwnerFromEvent(event);
             const method = getMethod(event);
             // Determine if this is a specific-thread request.
@@ -1116,6 +1952,7 @@ export function createAgentChatPlugin(options) {
             if (method === "POST") {
                 const body = await readBody(event);
                 const thread = await createThread(owner, {
+                    id: body?.id,
                     title: body?.title ?? "",
                 });
                 return thread;
@@ -1127,7 +1964,7 @@ export function createAgentChatPlugin(options) {
         // This is mounted last because h3's use() is prefix-based, meaning /_agent-native/agent-chat
         // also matches /_agent-native/agent-chat/threads/... — we skip sub-path requests here so the
         // earlier-mounted handlers (mode, save-key, files, skills, mentions, threads) handle them.
-        nitroApp.h3App.use(routePath, defineEventHandler(async (event) => {
+        getH3App(nitroApp).use(routePath, defineEventHandler(async (event) => {
             // Skip sub-path requests — they're handled by earlier-mounted handlers
             const url = event.node?.req?.url || event.path || "";
             const afterBase = url.slice(url.indexOf(routePath) + routePath.length);
@@ -1141,9 +1978,65 @@ export function createAgentChatPlugin(options) {
             // created by users who authenticated via OAuth (e.g., Gmail).
             const owner = await getOwnerFromEvent(event);
             process.env.AGENT_USER_EMAIL = owner;
+            // Set AGENT_ORG_ID so db-query/db-exec scope by org_id when applicable.
+            // Priority: explicit resolveOrgId callback > session.orgId from Better Auth
+            let resolvedOrgId = null;
+            if (options?.resolveOrgId) {
+                resolvedOrgId = await options.resolveOrgId(event);
+            }
+            else {
+                try {
+                    const session = await getSession(event);
+                    resolvedOrgId = session?.orgId ?? null;
+                }
+                catch {
+                    // Session not available
+                }
+            }
+            if (resolvedOrgId) {
+                process.env.AGENT_ORG_ID = resolvedOrgId;
+            }
+            else {
+                delete process.env.AGENT_ORG_ID;
+            }
             const handler = currentDevMode && devHandler ? devHandler : prodHandler;
             return handler(event);
         }));
+        // ─── Recurring Jobs Scheduler ──────────────────────────────────────
+        // Poll every 60 seconds for due recurring jobs and execute them.
+        // Uses setInterval so it works in all deployment environments without
+        // requiring Nitro experimental tasks configuration.
+        try {
+            const { processRecurringJobs } = await import("../jobs/scheduler.js");
+            const schedulerDeps = {
+                getActions: () => ({
+                    ...templateScripts,
+                    ...resourceScripts,
+                    ...chatScripts,
+                    ...jobTools,
+                }),
+                getSystemPrompt: async (owner) => {
+                    const resources = await loadResourcesForPrompt(owner);
+                    const schemaBlock = await buildSchemaBlock(owner, false);
+                    return basePrompt + resources + schemaBlock;
+                },
+                apiKey: options?.apiKey ?? process.env.ANTHROPIC_API_KEY,
+                model: resolvedModel,
+            };
+            // Start after a 10-second delay to let the server fully initialize
+            setTimeout(() => {
+                setInterval(() => {
+                    processRecurringJobs(schedulerDeps).catch((err) => {
+                        console.error("[recurring-jobs] Scheduler error:", err?.message);
+                    });
+                }, 60_000);
+                if (process.env.DEBUG)
+                    console.log("[recurring-jobs] Scheduler started (60s interval)");
+            }, 10_000);
+        }
+        catch (err) {
+            // Jobs module not available — skip silently
+        }
     };
 }
 /**