Harden multi-agent kernel and add local beta logger

docs/beta-test.md

@@ -0,0 +1,78 @@
+# Beta test — buddy playbook
+
+polycode is a **terminal** coding agent (not a website). Your buddy runs it on their
+machine, or sits at yours. Results of automated checks + a fill-in checklist are logged
+**locally** under `.polycode/beta-logs/` (gitignored — never committed, no API keys).
+
+## You (host)
+
+From the repo:
+
+```powershell
+corepack pnpm -C "C:\Users\gasan\OneDrive\Dev\Projects\polycode" install
+corepack pnpm -C "C:\Users\gasan\OneDrive\Dev\Projects\polycode" beta -- --tester you
+```
+
+That runs typecheck + the unit suite and writes two local files (gitignored):
+
+- `.polycode/beta-logs/<timestamp>_you.md`
+- `.polycode/beta-logs/LATEST.md` (last **automated** run — send this one; checklist-only does not overwrite it)
+
+Live model round-trip (needs a Grok key already saved):
+
+```powershell
+corepack pnpm -C "C:\Users\gasan\OneDrive\Dev\Projects\polycode" beta -- --tester you --smoke xai:grok-4.3
+```
+
+Checklist-only (no test run — just a blank log for a session):
+
+```powershell
+corepack pnpm -C "C:\Users\gasan\OneDrive\Dev\Projects\polycode" beta -- --tester buddy --checklist-only
+```
+
+Send your buddy **that markdown file** after they fill the checkboxes. Not `.env`, not the keychain.
+
+## Your buddy (first 10 minutes)
+
+1. Node 20+ installed.
+2. In the repo:
+
+```powershell
+corepack pnpm install
+corepack pnpm dev
+```
+
+3. Settings: arrow to **xAI (Grok)** (not Meta Muse). Enter → paste `xai-…` key → Enter.
+   Chat should open. If stuck, they have a Grok key in `.env` as `XAI_API_KEY=` or `GROK_API_KEY=` and press **i**.
+4. Work through the checklist in the log file (or below). Mark pass/fail. No keys in notes.
+5. Quit with `/exit`.
+
+### What to try
+
+| Try | Expect |
+|---|---|
+| `/help` | Overlay of commands, Esc closes |
+| `/mo` + Tab | Completes toward `/mode` or `/model` |
+| “what is this repo?” | Tools run; answer mentions polycode |
+| `/mode plan` then “edit README” | Writes refused |
+| `/mode ask` then a tiny edit | `y` / `a` / `n` prompt |
+| `/team add a comment in README` | Explorers, then a worktree implement (git repo) |
+| `/dashboard` | Child list; Enter peek; `a` attach |
+| Long turn + **Ctrl+B** | Composer comes back; kids still in dashboard |
+| `/loop 15s say ping` | Recurring line; `/loop stop` ends it |
+| `/hepl` | Error “did you mean /help?” — not sent to Grok |
+
+## Safety for a buddy on your box
+
+- Use **ask** mode (`/mode ask`), not yolo.
+- They should not need `--serve` (hosted HTTP).
+- Their key: they paste their own Grok key, or you temporarily set one and `/settings` when they leave.
+- Logs in `.polycode/beta-logs/` are local only.
+
+## After the session
+
+```powershell
+corepack pnpm beta -- --tester buddy-name --checklist-only
+```
+
+Fill the checklist in the new file, keep it in `.polycode/beta-logs/`, and tell you what failed.

package.json

@@ -10,6 +10,7 @@
     "dev": "tsx packages/cli/src/index.ts",
     "serve": "tsx packages/cli/src/index.ts --serve",
     "smoke": "tsx scripts/smoke.ts",
+    "beta": "tsx scripts/beta-log.ts",
     "route-check": "tsx scripts/route-check.ts",
     "test": "vitest run",
     "test:watch": "vitest",

packages/core/src/agent.test.ts

@@ -385,6 +385,48 @@ describe("Agent", () => {
     expect(agent.history()).toHaveLength(0);
   });
 
+  it("stamps worktree isolation on tool_call before the UI sees the call", async () => {
+    const task: ToolSpec = {
+      name: "task",
+      description: "task",
+      parameters: {},
+      permission: "mutating",
+      parallelSafe: false,
+      async run(input: { description: string; prompt: string }, ctx) {
+        if (!ctx.spawnChild) return { output: "no spawn", isError: true };
+        return ctx.spawnChild({ ...input, subagent_type: "general" });
+      },
+    };
+    const provider = new ScriptedProvider([
+      [
+        {
+          type: "tool_call",
+          call: { type: "tool_call", id: "1", name: "task", input: { description: "a", prompt: "a" } },
+        },
+        {
+          type: "tool_call",
+          call: { type: "tool_call", id: "2", name: "task", input: { description: "b", prompt: "b" } },
+        },
+        { type: "stop", reason: "tool_use" },
+      ],
+      [{ type: "text_delta", text: "ca" }, { type: "stop", reason: "end_turn" }],
+      [{ type: "text_delta", text: "cb" }, { type: "stop", reason: "end_turn" }],
+      [{ type: "text_delta", text: "parent" }, { type: "stop", reason: "end_turn" }],
+    ]);
+    const agent = new Agent(provider, [task], new PermissionEngine("yolo", async () => "once"), {
+      sandbox,
+      openWorktree: async () => ({ sandbox, path: "/tmp/wt-stamp" }),
+    });
+    agent.pushUser("go");
+    const events = await collect(agent);
+    const calls = events.filter((e) => e.type === "tool_call");
+    expect(calls).toHaveLength(2);
+    for (const ev of calls) {
+      if (ev.type !== "tool_call") continue;
+      expect((ev.call.input as { isolation?: string }).isolation).toBe("worktree");
+    }
+  });
+
   it("records worktree paths from spawnChild", async () => {
     const provider = new ScriptedProvider([
       [{ type: "text_delta", text: "ok" }, { type: "stop", reason: "end_turn" }],