Beta TUI, Grok-only setup, and multi-agent kernel

docs/cli.md

@@ -44,6 +44,10 @@ Sessions are stored as `<cwd>/.polycode/sessions/<id>.json` and saved after each
 | `/context` | Token breakdown vs the model window. |
 | `/review` (`/cranky`) | Spawn a depth-1 reviewer child on the local git diff. |
 | `/explore <q>` | Spawn a read-only explore child. |
+| `/team <task>` | Parallel explore (code + verify), then implement in a git worktree, then review. `/worktree apply` to copy back. |
+| `/dashboard` (`/agents`, `Ctrl+\`) | Child agents. Enter/`p` peek live output · `a` attach into the transcript · `x` kill. |
+| `/loop [5m] <prompt>` | Recurring prompt (min 15s). `/loop` lists, `/loop stop [id]` cancels. Skips a tick if a turn is already running. |
+| `/personas` | Persona overlays from `.polycode/personas/*.md`. Pass `persona` on `task`. |
 | `/skills` | List loaded skills (bundled forme, matsumura-style, deep-research). |
 | `/mcp` | MCP servers from `.polycode/mcp.json` (fail-soft connect). Notes deferred schemas. |
 | `/plugins` | Loaded plugin bundles (`.polycode/plugins/<name>`). |
@@ -67,8 +71,8 @@ Sessions are stored as `<cwd>/.polycode/sessions/<id>.json` and saved after each
 | `/rewind` (`/undo`) | Restore files from the last mutating-tool checkpoint and drop that turn. |
 | `/settings` (or `/login`) | Open the API-key Settings screen (paste / import-env / open-page / validate / agentic). |
 | `/clear` | Clear the transcript. |
-| `/help` | List commands. |
-| `/exit` (or `/quit`) | Quit. |
+| `/help` | Command overlay (also `/`). Tab completes; ↑↓ history. |
+| `/exit` (or `/quit`) | Quit. Ctrl+C interrupts a turn; twice quits. |
 
 ## Permission prompt
 
@@ -79,6 +83,8 @@ class, and the call. Answer with:
 - `a` — allow this tool for the rest of the session (sticky; `plan` mode still overrides)
 - `n` / `Esc` — deny
 
+While a turn is running: **Esc** interrupts (kills foreground children). **Ctrl+B** detaches running children and frees the composer (`/dashboard` to peek/attach).
+
 A denied call is reported back to the model so it can adapt.
 
 ## Server endpoints (hosted mode)

docs/html/cli.html

@@ -189,6 +189,22 @@ <h2>Slash commands (in the TUI)</h2>
 <td>Spawn a read-only explore child.</td>
 </tr>
 <tr>
+<td><code>/team &lt;task&gt;</code></td>
+<td>Parallel explore (code + verify), then implement in a git worktree, then review. <code>/worktree apply</code> to copy back.</td>
+</tr>
+<tr>
+<td><code>/dashboard</code> (<code>/agents</code>, <code>Ctrl+\</code>)</td>
+<td>Child agents. Enter/<code>p</code> peek live output · <code>a</code> attach into the transcript · <code>x</code> kill.</td>
+</tr>
+<tr>
+<td><code>/loop [5m] &lt;prompt&gt;</code></td>
+<td>Recurring prompt (min 15s). <code>/loop</code> lists, <code>/loop stop [id]</code> cancels. Skips a tick if a turn is already running.</td>
+</tr>
+<tr>
+<td><code>/personas</code></td>
+<td>Persona overlays from <code>.polycode/personas/*.md</code>. Pass <code>persona</code> on <code>task</code>.</td>
+</tr>
+<tr>
 <td><code>/skills</code></td>
 <td>List loaded skills (bundled forme, matsumura-style, deep-research).</td>
 </tr>
@@ -282,11 +298,11 @@ <h2>Slash commands (in the TUI)</h2>
 </tr>
 <tr>
 <td><code>/help</code></td>
-<td>List commands.</td>
+<td>Command overlay (also <code>/</code>). Tab completes; ↑↓ history.</td>
 </tr>
 <tr>
 <td><code>/exit</code> (or <code>/quit</code>)</td>
-<td>Quit.</td>
+<td>Quit. Ctrl+C interrupts a turn; twice quits.</td>
 </tr>
 </tbody></table>
 <h2>Permission prompt</h2>
@@ -297,6 +313,7 @@ <h2>Permission prompt</h2>
 <li><code>a</code> — allow this tool for the rest of the session (sticky; <code>plan</code> mode still overrides)</li>
 <li><code>n</code> / <code>Esc</code> — deny</li>
 </ul>
+<p>While a turn is running: <strong>Esc</strong> interrupts (kills foreground children). <strong>Ctrl+B</strong> detaches running children and frees the composer (<code>/dashboard</code> to peek/attach).</p>
 <p>A denied call is reported back to the model so it can adapt.</p>
 <h2>Server endpoints (hosted mode)</h2>
 <table>

examples/personas/terse.md

@@ -0,0 +1,3 @@
+# Terse
+
+Reply in at most 5 bullets. No preamble. Cite file paths.

packages/cli/src/index.ts

@@ -10,6 +10,7 @@ import {
   mergeHookSets,
   registerExtraChildren,
   listExtraChildren,
+  loadPersonas,
   CHILD_TYPES,
   type CompactConfig,
   type GenerateRequest,
@@ -40,7 +41,7 @@ import {
 import { loadWorkflows } from "@polycode/workflows";
 import { loadPlugins } from "@polycode/plugins";
 import { createLspTool } from "@polycode/lsp";
-import { hydrateEnv, getKey, ENV_VAR, type ProviderId } from "@polycode/secrets";
+import { hydrateEnv, loadDotEnvFiles, envFileCandidates, getKey, ENV_VAR, type ProviderId } from "@polycode/secrets";
 import { tools as allTools } from "@polycode/tools";
 import { connectMcpServers, createMcpSearchTool, loadMcpConfig, type McpServerStatus } from "@polycode/mcp";
 import { expandSkill, loadSkills, skillPromptBlock } from "@polycode/skills";
@@ -85,7 +86,7 @@ const DEFAULT_CONFIG: AppConfig = {
     long: { provider: "google", model: "gemini-2.5-pro" },
   },
   system:
-    "You are polycode, a terminal coding agent. Be concise. Use tools to inspect and edit the project.",
+    "You are polycode, a terminal coding agent. Be concise. Use tools to inspect and edit the project. When work splits, call several task tools in one turn so they run in parallel: explore/researcher for read-only surveys, isolation=worktree for implementers. background=true returns an id (task_wait collects). resume_from continues a finished child. persona= applies .polycode/personas. Depth 1. Summarize child returns; do not paste raw dumps.",
 };
 
 function loadConfig(): AppConfig {
@@ -124,7 +125,7 @@ async function main(): Promise<void> {
     allowPositionals: true,
   });
 
-  // Pull keys from the OS keychain / file store into the env the SDK reads.
+  loadDotEnvFiles(envFileCandidates(process.cwd()));
   hydrateEnv();
 
   const cfg = loadConfig();
@@ -363,6 +364,8 @@ async function main(): Promise<void> {
     route,
     autoRoute: router.strategy === "model",
     validate: (p) => validateKey(cfg, p),
+    specForProvider: (p) => ({ provider: p, model: modelFor(cfg, p as ProviderId) }),
+    personas: loadPersonas(cwd),
     onAgentic: (p) =>
       `agentic provisioning for ${p} is not wired yet — coming soon (MCP/tool flow). Use paste / import-env / open-page for now.`,
     initialMessages,

packages/core/src/agent.test.ts

@@ -102,6 +102,80 @@ describe("Agent", () => {
     expect(toolMessage?.content.map((p) => "name" in p ? p.name : "")).toEqual(["slow", "fast"]);
   });
 
+  it("runs consecutive explore task children concurrently", async () => {
+    let active = 0;
+    let maxActive = 0;
+    const task: ToolSpec = {
+      name: "task",
+      description: "task",
+      parameters: {},
+      permission: "mutating",
+      parallelSafe: true,
+      async run() {
+        active++;
+        maxActive = Math.max(maxActive, active);
+        await new Promise((r) => setTimeout(r, 25));
+        active--;
+        return { output: "ok" };
+      },
+    };
+    const call = (id: string, type: string) =>
+      ({
+        type: "tool_call" as const,
+        call: {
+          type: "tool_call" as const,
+          id,
+          name: "task",
+          input: { description: id, prompt: "p", subagent_type: type },
+        },
+      });
+    const provider = new ScriptedProvider([
+      [call("1", "explore"), call("2", "explore"), { type: "stop", reason: "tool_use" }],
+      [{ type: "stop", reason: "end_turn" }],
+    ]);
+    const agent = new Agent(provider, [task], new PermissionEngine("ask", async () => "deny"), { sandbox });
+    agent.pushUser("fan out");
+    await collect(agent);
+    expect(maxActive).toBe(2);
+  });
+
+  it("auto-isolates two writable task children onto worktrees", async () => {
+    const seen: unknown[] = [];
+    const task: ToolSpec = {
+      name: "task",
+      description: "task",
+      parameters: {},
+      permission: "mutating",
+      parallelSafe: true,
+      async run(input) {
+        seen.push(input);
+        return { output: "ok" };
+      },
+    };
+    const call = (id: string) =>
+      ({
+        type: "tool_call" as const,
+        call: {
+          type: "tool_call" as const,
+          id,
+          name: "task",
+          input: { description: id, prompt: "edit", subagent_type: "general" },
+        },
+      });
+    const provider = new ScriptedProvider([
+      [call("1"), call("2"), { type: "stop", reason: "tool_use" }],
+      [{ type: "stop", reason: "end_turn" }],
+    ]);
+    const agent = new Agent(provider, [task], new PermissionEngine("yolo", async () => "once"), {
+      sandbox,
+      openWorktree: async () => ({ sandbox, path: "/tmp/wt" }),
+    });
+    agent.pushUser("two writers");
+    await collect(agent);
+    expect(seen).toHaveLength(2);
+    expect(seen.every((s) => (s as { isolation?: string }).isolation === "worktree")).toBe(true);
+  });
+
   it("stops a runaway tool loop at maxSteps", async () => {
     const call = { type: "tool_call", call: { type: "tool_call", id: "1", name: "read", input: {} } } as const;
     const provider = new ScriptedProvider([[call, { type: "stop", reason: "tool_use" }]]);