diff --git a/.dockerignore b/.dockerignore
index 59bc06c..1c09996 100644
--- a/.dockerignore
+++ b/.dockerignore
@@ -1,6 +1,5 @@
 .git/
 .venv/
-.entire/
 .claude/
 .secret.*
 *.wormhole
@@ -14,4 +13,4 @@ dist/
 build/
 *.db
 .env
-docs/plans/
+docs/
diff --git a/README.md b/README.md
index 5225020..e0b2083 100644
--- a/README.md
+++ b/README.md
@@ -83,7 +83,7 @@ Sessions auto-transition based on inactivity: `active → idle → suspended →
 |---|---|---|
 | `WARREN_ADMIN_TOKEN` | (required) | Token for admin operations (pairing, config) |
 | `WARREN_INTERNAL_SECRET` | (recommended) | Token required of agents connecting to `/api/agent/ws` |
-| `WARREN_BACKEND` | `websocket` | Backend adapter: `websocket` or `orca` |
+| `WARREN_BACKEND` | `websocket` | Backend adapter: `websocket` (default) or `orca` (experimental, single local terminal) |
 | `WARREN_CONFIG` | `config.yaml` | Path to YAML config file |
 | `WARREN_DB` | `warren.db` | Path to SQLite database |
 | `WARREN_HOST` | `0.0.0.0` | Server bind address |
@@ -139,7 +139,7 @@ uv run pytest tests/ -v         # Unit tests
 |---|---|
 | `server/src/warren/app.py` | FastAPI app, endpoints, SSE streaming |
 | `server/src/warren/adapters/websocket.py` | WebSocket agent proxy adapter (default) |
-| `server/src/warren/adapters/orca.py` | Local Orca terminal adapter |
+| `server/src/warren/adapters/orca.py` | Local Orca terminal adapter (experimental, single-session) |
 | `server/src/warren/voice.py` | Voice WebSocket handler (OpenClaw protocol) |
 | `server/src/warren/bus.py` | SessionBus pub/sub (multi-consumer events) |
 | `server/src/warren/lifecycle.py` | Background session lifecycle sweep |
diff --git a/creation/index.html b/creation/index.html
index f033835..268872d 100644
--- a/creation/index.html
+++ b/creation/index.html
@@ -49,8 +49,6 @@
       <span class="title">NEW SESSION</span>
     </div>
     <div class="form-area">
-      <label for="repo-select">REPO</label>
-      <select id="repo-select" class="input"></select>
       <label for="initial-msg">PROMPT</label>
       <input id="initial-msg" type="text" placeholder="enter command..." class="input">
     </div>
@@ -75,10 +73,6 @@
         <div class="monitor-section-title">AGENTS</div>
         <div id="monitor-agents"></div>
       </div>
-      <div class="monitor-section">
-        <div class="monitor-section-title">TASKS</div>
-        <div id="monitor-tasks"></div>
-      </div>
     </div>
     <div class="footer">
       <button id="btn-monitor-refresh" class="btn-primary">[ REFRESH ]</button>
diff --git a/creation/js/api.js b/creation/js/api.js
index e11529e..3e14ef4 100644
--- a/creation/js/api.js
+++ b/creation/js/api.js
@@ -36,7 +36,7 @@ const Warren = (() => {
 
   /**
    * Initialize from URL token (?token=DEVICE_TOKEN).
-   * Returns { ok, reason } for debugging on screen.
+   * Returns { ok, reason }.
    */
   function initFromUrl() {
     const token = new URLSearchParams(window.location.search).get("token");
@@ -46,10 +46,6 @@ const Warren = (() => {
     return { ok: true };
   }
 
-  async function fetchRepos() {
-    return _fetch("/repos");
-  }
-
   async function listSessions() {
     return _fetch("/sessions");
   }
@@ -131,22 +127,9 @@ const Warren = (() => {
     return _fetch("/monitor/agents");
   }
 
-  async function fetchMonitorTasks() {
-    return _fetch("/monitor/tasks");
-  }
-
-  async function pauseTask(taskId) {
-    return _fetch(`/monitor/tasks/${taskId}/pause`, { method: "POST" });
-  }
-
-  async function resumeTask(taskId) {
-    return _fetch(`/monitor/tasks/${taskId}/resume`, { method: "POST" });
-  }
-
   return {
     init,
     initFromUrl,
-    fetchRepos,
     listSessions,
     createSession,
     sendMessage,
@@ -158,9 +141,6 @@ const Warren = (() => {
     fetchSessionMessages,
     fetchMonitorHealth,
     fetchMonitorAgents,
-    fetchMonitorTasks,
-    pauseTask,
-    resumeTask,
     get serverUrl() { return _serverUrl; },
     get connected() { return !!_token; },
   };
diff --git a/creation/js/app.js b/creation/js/app.js
index 4e2463b..8bfc262 100644
--- a/creation/js/app.js
+++ b/creation/js/app.js
@@ -27,12 +27,10 @@
   const msgInput = document.getElementById("msg-input");
   const pttIndicator = document.getElementById("ptt-indicator");
   const connDot = document.getElementById("connection-dot");
-  const repoSelect = document.getElementById("repo-select");
   const initialMsg = document.getElementById("initial-msg");
   const monitorContent = document.getElementById("monitor-content");
   const monitorHealth = document.getElementById("monitor-health");
   const monitorAgents = document.getElementById("monitor-agents");
-  const monitorTasks = document.getElementById("monitor-tasks");
   const monitorStatus = document.getElementById("monitor-status");
 
   // -- Helpers --
@@ -70,7 +68,7 @@
     selectedIndex = 0;
 
     if (name === "sessions") refreshSessions();
-    if (name === "new") loadRepos();
+    if (name === "new") initialMsg.focus();
     if (name === "monitor") refreshMonitor();
   }
 
@@ -226,7 +224,7 @@
         }
       });
     } catch (err) {
-      console.error("Failed to load message history:", err);
+      // History is best-effort; ignore load failures (no console on the R1).
     }
     renderStepProgress();
 
@@ -360,14 +358,12 @@
   async function refreshMonitor() {
     monitorStatus.className = "dot dot-yellow";
     try {
-      const [health, agents, tasks] = await Promise.all([
+      const [health, agents] = await Promise.all([
         Warren.fetchMonitorHealth(),
         Warren.fetchMonitorAgents(),
-        Warren.fetchMonitorTasks(),
       ]);
       renderMonitorHealth(health);
       renderMonitorAgents(agents);
-      renderMonitorTasks(tasks);
       monitorStatus.className = "dot dot-green";
     } catch (err) {
       monitorStatus.className = "dot dot-red";
@@ -436,45 +432,6 @@
     });
   }
 
-  function renderMonitorTasks(data) {
-    monitorTasks.textContent = "";
-    const taskList = data.tasks || [];
-    if (taskList.length === 0) {
-      const empty = document.createElement("div");
-      empty.className = "monitor-empty";
-      empty.textContent = "No scheduled tasks";
-      monitorTasks.appendChild(empty);
-      return;
-    }
-    taskList.forEach((task) => {
-      const item = document.createElement("div");
-      item.className = "monitor-item";
-
-      const dot = document.createElement("span");
-      dot.className = "dot " + (task.status === "active" ? "dot-green" : "dot-yellow");
-
-      const label = document.createElement("span");
-      label.className = "label";
-      label.textContent = task.id;
-
-      const value = document.createElement("span");
-      value.className = "value";
-      if (task.status === "paused") {
-        value.textContent = "paused";
-      } else if (task.next_run) {
-        const next = new Date(task.next_run);
-        value.textContent = next.toLocaleTimeString([], { hour: "2-digit", minute: "2-digit" });
-      } else {
-        value.textContent = task.schedule_value;
-      }
-
-      item.appendChild(dot);
-      item.appendChild(label);
-      item.appendChild(value);
-      monitorTasks.appendChild(item);
-    });
-  }
-
   // -- Command Palette --
 
   const paletteOverlay = document.getElementById("view-palette");
@@ -708,15 +665,7 @@
 
   // -- New Session View --
 
-  function loadRepos() {
-    const repoLabel = repoSelect.previousElementSibling;
-    repoSelect.style.display = "none";
-    if (repoLabel) repoLabel.style.display = "none";
-    initialMsg.focus();
-  }
-
   async function startSession() {
-    const repo = repoSelect.value || "";
     const msg = initialMsg.value.trim();
     if (!msg) return;
 
@@ -730,7 +679,7 @@
     }
 
     try {
-      const data = await Warren.createSession(repo, finalMsg);
+      const data = await Warren.createSession("", finalMsg);
       initialMsg.value = "";
       openSession({ id: data.session_id, name: data.name, repo: data.repo });
     } catch (err) {
@@ -915,15 +864,8 @@
     sessionList.textContent = "";
     const empty = document.createElement("div");
     empty.className = "empty-state";
-    empty.textContent = "ERR: " + result.reason;
+    empty.textContent = "NOT PAIRED — re-scan the pairing QR code";
     sessionList.appendChild(empty);
-
-    const debug = document.createElement("div");
-    debug.className = "empty-state";
-    debug.style.fontSize = "10px";
-    debug.style.wordBreak = "break-all";
-    debug.textContent = window.location.href;
-    sessionList.appendChild(debug);
   }
 
   initialize();
diff --git a/docker-compose.yml b/docker-compose.yml
index fddfb4c..b6d62dd 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -6,26 +6,17 @@ services:
     ports:
       - "4030:4030"
     volumes:
-      - /home/deploy/repos:/home/deploy/repos
-      - /home/deploy/.ssh:/home/appuser/.ssh:ro
-      - /home/deploy/.gitconfig:/home/appuser/.gitconfig:ro
-      - /home/deploy/.claude:/home/appuser/.claude
       - warren-data:/data
+      # Optional: mount host SSH/git config if the orca backend needs repo access.
+      # - ${HOME}/.ssh:/home/appuser/.ssh:ro
     environment:
-      - ANTHROPIC_API_KEY=${ANTHROPIC_API_KEY}
       - WARREN_CONFIG=/data/config.yaml
       - WARREN_DB=/data/warren.db
       - WARREN_ADMIN_TOKEN=${WARREN_ADMIN_TOKEN}
+      - WARREN_INTERNAL_SECRET=${WARREN_INTERNAL_SECRET}
       - WARREN_VOICE_ENABLED=${WARREN_VOICE_ENABLED:-false}
       - WARREN_VOICE_PORT=${WARREN_VOICE_PORT:-443}
-      - WARREN_INTERNAL_SECRET=${WARREN_INTERNAL_SECRET}
     restart: unless-stopped
-    networks:
-      - dokploy-network
 
 volumes:
   warren-data:
-
-networks:
-  dokploy-network:
-    external: true
diff --git a/docs/plans/2026-02-19-ux-context-improvements-design.md b/docs/plans/2026-02-19-ux-context-improvements-design.md
deleted file mode 100644
index ac278aa..0000000
--- a/docs/plans/2026-02-19-ux-context-improvements-design.md
+++ /dev/null
@@ -1,87 +0,0 @@
-# Warren UX + Context Improvements Design
-
-## Goal
-
-Reduce friction in daily Warren usage on the R1 by improving session creation flow, output readability, command palette reach, and Claude session context.
-
-## Architecture
-
-Four independent improvements that touch server config, Claude CLI invocation, and creation-side UI. No new dependencies. Changes are additive and backwards-compatible.
-
-## 1. Per-Repo Configuration (model, system prompt)
-
-Add optional fields to each repo entry in `config.yaml`:
-
-```yaml
-repos:
-  warren:
-    path: /home/user/warren
-    git_url: git@github.com:user/warren.git
-    model: opus
-    append_system_prompt: |
-      This is Warren, a FastAPI server for Rabbit R1.
-      The creation/ dir is a 240x282 WebView with CRT styling.
-```
-
-**Server changes:**
-- `config.py`: Add `model: str = ""` and `append_system_prompt: str = ""` to `RepoConfig`
-- `claude_runner.py`: `_build_command` accepts optional `model` and `append_system_prompt` params. Passes `--model` and `--append-system-prompt` to the CLI when set.
-- `app.py`: Pass repo config fields through to `ClaudeRunner.run()`
-
-Uses `--append-system-prompt` (adds to Claude Code's built-in prompt) rather than `--system-prompt` (which would replace it entirely).
-
-## 2. Simplified Session Creation
-
-**Current flow:** tap NEW -> pick repo -> type name -> type message -> tap EXEC (4 steps, 2 typing)
-
-**New flow:** tap NEW -> pick repo -> type message -> tap EXEC (3 steps, 1 typing). Single-repo configs skip the picker (2 steps).
-
-**Changes:**
-- Remove "NAME (OPT)" field from new session form in `index.html`
-- Server auto-generates name from first ~30 chars of message (already supports `name: None`)
-- If only one repo, auto-select and hide dropdown — go straight to prompt input
-- Replace `<select>` dropdown with a tap list (matches session list UX, easier on R1 touchscreen)
-
-## 3. Output Visual Hierarchy + Density
-
-**Problem:** All message types look nearly identical — same padding, same font size, no visual separation. Walls of green text on 240x282.
-
-**Visual separation:**
-- User messages: dim background bar (`#0a1a0a`) + `> ` prefix to distinguish input from output
-- Tool events: left border accent (`2px solid #0a6e0a`) to visually group as a "work block"
-- Text blocks: small top margin (4px) between messages for breathing room
-- Result lines: brighter background bar (`#0a2a0a`), bump to 12px — these are the headlines
-- Error lines: red left border accent matching the tool pattern
-
-**Density reduction:**
-- Collapse consecutive tool events into single line: `[3 tools] Read, Edit, Bash` — tap to expand
-- Cap text blocks at 4 lines with `...` — tap to expand
-- Auto-scroll so `[OK]` result lands near top of viewport
-
-**Goal:** Glance at screen and immediately see "what I asked" (user msg), "work happened" (tool block), "here's the answer" (result bar).
-
-All CSS/JS changes in creation — no server changes needed.
-
-## 4. Command Palette as Primary Input
-
-**Problem:** Palette only accessible inside chat sessions. Typing on R1 is painful for anything beyond short messages.
-
-**Changes:**
-- Add `[CMD]` button to sessions list view footer (alongside `[+ NEW]`)
-- Actions from session list auto-create a new session (pick default/only repo, auto-name from action label)
-- Quick-action flow: open Warren -> tap CMD -> tap "RUN TESTS" -> session created and running. Zero typing, two taps.
-
-## Files Affected
-
-**Server:**
-- `server/src/warren/config.py` — RepoConfig fields
-- `server/src/warren/claude_runner.py` — _build_command params
-- `server/src/warren/app.py` — pass config through to runner
-
-**Creation:**
-- `creation/index.html` — form changes, CMD button in sessions footer
-- `creation/css/styles.css` — visual hierarchy styles
-- `creation/js/app.js` — tool collapsing, text truncation, auto-scroll, palette from sessions, simplified creation flow
-
-**Config:**
-- `server/config.yaml.example` — document new repo fields
diff --git a/docs/plans/2026-02-19-ux-context-improvements-plan.md b/docs/plans/2026-02-19-ux-context-improvements-plan.md
deleted file mode 100644
index da01f50..0000000
--- a/docs/plans/2026-02-19-ux-context-improvements-plan.md
+++ /dev/null
@@ -1,886 +0,0 @@
-# Warren UX + Context Improvements Implementation Plan
-
-> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
-
-**Goal:** Reduce friction in daily Warren usage on the R1 by adding per-repo model/prompt config, simplifying session creation, improving output readability, and making the command palette the primary input method.
-
-**Architecture:** Four independent feature areas. Tasks 1-3 are server-side (config + runner). Tasks 4-6 are creation-side (HTML/CSS/JS). Each task is self-contained and testable independently.
-
-**Tech Stack:** Python/FastAPI (server), vanilla HTML/CSS/JS (creation), pytest + httpx (tests)
-
----
-
-### Task 1: Add model and append_system_prompt to RepoConfig
-
-**Files:**
-- Modify: `server/src/warren/config.py:12-14`
-- Modify: `server/tests/test_config.py`
-
-**Step 1: Write the failing test**
-
-Add to `server/tests/test_config.py`:
-
-```python
-def test_repo_config_with_model_and_prompt(tmp_path):
-    """Repo config accepts optional model and append_system_prompt fields."""
-    from warren.config import load_config
-
-    config_file = tmp_path / "config.yaml"
-    config_file.write_text(yaml.dump({
-        "repos": {
-            "my-project": {
-                "path": "/home/user/project",
-                "model": "opus",
-                "append_system_prompt": "This is a Python web app.",
-            },
-            "simple-repo": {
-                "path": "/tmp/simple",
-            },
-        },
-    }))
-
-    config = load_config(str(config_file))
-    assert config.repos["my-project"].model == "opus"
-    assert config.repos["my-project"].append_system_prompt == "This is a Python web app."
-    # Defaults to empty when not specified
-    assert config.repos["simple-repo"].model == ""
-    assert config.repos["simple-repo"].append_system_prompt == ""
-```
-
-**Step 2: Run test to verify it fails**
-
-Run: `cd server && uv run pytest tests/test_config.py::test_repo_config_with_model_and_prompt -v`
-Expected: FAIL — `RepoConfig` has no `model` or `append_system_prompt` attribute
-
-**Step 3: Write minimal implementation**
-
-In `server/src/warren/config.py`, update the `RepoConfig` class:
-
-```python
-class RepoConfig(BaseModel):
-    path: str
-    git_url: str = ""
-    model: str = ""
-    append_system_prompt: str = ""
-```
-
-**Step 4: Run test to verify it passes**
-
-Run: `cd server && uv run pytest tests/test_config.py -v`
-Expected: ALL PASS
-
-**Step 5: Commit**
-
-```bash
-git add server/src/warren/config.py server/tests/test_config.py
-git commit -m "feat: add model and append_system_prompt to RepoConfig"
-```
-
----
-
-### Task 2: Pass model and system prompt through ClaudeRunner
-
-**Files:**
-- Modify: `server/src/warren/claude_runner.py:63-79`
-- Modify: `server/tests/test_claude_runner.py`
-
-**Step 1: Write the failing tests**
-
-Add to `server/tests/test_claude_runner.py` inside `TestBuildCommand`:
-
-```python
-    async def test_model_flag(self):
-        from warren.claude_runner import ClaudeRunner
-
-        runner = ClaudeRunner(claude_bin="claude")
-        cmd = runner._build_command(
-            message="hello",
-            repo_path="/home/user/project",
-            model="opus",
-        )
-        assert "--model" in cmd
-        idx = cmd.index("--model")
-        assert cmd[idx + 1] == "opus"
-
-    async def test_append_system_prompt_flag(self):
-        from warren.claude_runner import ClaudeRunner
-
-        runner = ClaudeRunner(claude_bin="claude")
-        cmd = runner._build_command(
-            message="hello",
-            repo_path="/home/user/project",
-            append_system_prompt="You are a Python expert.",
-        )
-        assert "--append-system-prompt" in cmd
-        idx = cmd.index("--append-system-prompt")
-        assert cmd[idx + 1] == "You are a Python expert."
-
-    async def test_no_flags_when_empty(self):
-        from warren.claude_runner import ClaudeRunner
-
-        runner = ClaudeRunner(claude_bin="claude")
-        cmd = runner._build_command(
-            message="hello",
-            repo_path="/home/user/project",
-            model="",
-            append_system_prompt="",
-        )
-        assert "--model" not in cmd
-        assert "--append-system-prompt" not in cmd
-```
-
-**Step 2: Run tests to verify they fail**
-
-Run: `cd server && uv run pytest tests/test_claude_runner.py::TestBuildCommand -v`
-Expected: FAIL — `_build_command` doesn't accept `model` or `append_system_prompt`
-
-**Step 3: Write minimal implementation**
-
-Update `_build_command` in `server/src/warren/claude_runner.py`:
-
-```python
-    def _build_command(
-        self,
-        message: str,
-        repo_path: str,
-        resume_session_id: str | None = None,
-        model: str = "",
-        append_system_prompt: str = "",
-    ) -> list[str]:
-        """Build the claude CLI command."""
-        cmd = [
-            self._claude_bin,
-            "-p", message,
-            "--output-format", "stream-json",
-            "--verbose",
-            "--dangerously-skip-permissions",
-        ]
-        if resume_session_id is not None:
-            cmd.extend(["--resume", resume_session_id])
-        if model:
-            cmd.extend(["--model", model])
-        if append_system_prompt:
-            cmd.extend(["--append-system-prompt", append_system_prompt])
-        return cmd
-```
-
-Also update the `run` method signature and pass-through:
-
-```python
-    async def run(
-        self,
-        message: str,
-        repo_path: str,
-        resume_session_id: str | None = None,
-        run_id: str | None = None,
-        model: str = "",
-        append_system_prompt: str = "",
-    ) -> AsyncGenerator[dict, None]:
-```
-
-And update the `_build_command` call inside `run`:
-
-```python
-            cmd = self._build_command(
-                message, repo_path, resume_session_id,
-                model=model, append_system_prompt=append_system_prompt,
-            )
-```
-
-**Step 4: Run tests to verify they pass**
-
-Run: `cd server && uv run pytest tests/test_claude_runner.py -v`
-Expected: ALL PASS
-
-**Step 5: Commit**
-
-```bash
-git add server/src/warren/claude_runner.py server/tests/test_claude_runner.py
-git commit -m "feat: pass --model and --append-system-prompt to Claude CLI"
-```
-
----
-
-### Task 3: Wire repo config through app endpoints + update example
-
-**Files:**
-- Modify: `server/src/warren/app.py:286-291,346-352`
-- Modify: `server/tests/test_app.py`
-- Modify: `server/config.yaml.example`
-
-**Step 1: Write the failing test**
-
-Add to `server/tests/test_app.py`. The test needs to verify that when a session is created for a repo with `model` and `append_system_prompt` configured, those values are passed to the runner. Use a mock to capture the args:
-
-```python
-class TestRepoConfig:
-    async def test_create_session_passes_model_to_runner(self, authed_client):
-        """Session creation passes repo model config to ClaudeRunner."""
-        captured = {}
-
-        async def mock_run(*args, **kwargs):
-            captured.update(kwargs)
-            yield {"type": "status", "state": "working"}
-            yield {"type": "result", "session_id": "mock-123", "summary": "Done"}
-            yield {"type": "status", "state": "waiting"}
-
-        with patch("warren.app.ClaudeRunner.run", side_effect=mock_run):
-            resp = await authed_client.post("/sessions", json={
-                "repo": "test-repo",
-                "message": "hello",
-            })
-            assert resp.status_code == 200
-
-        assert captured.get("model") == "opus"
-        assert "test system prompt" in captured.get("append_system_prompt", "")
-```
-
-This test needs the `authed_client` fixture to include a repo with model/prompt config. Update the fixture's config yaml to include:
-
-```python
-    config_file.write_text(yaml.dump({
-        "repos": {"test-repo": {
-            "path": str(tmp_path),
-            "model": "opus",
-            "append_system_prompt": "test system prompt",
-        }},
-    }))
-```
-
-**Important:** The existing `authed_client` fixture in `test_app.py` already creates a `test-repo` — update that fixture to add the new fields so all existing tests continue to work (model/prompt are optional, so they won't break existing assertions).
-
-**Step 2: Run test to verify it fails**
-
-Run: `cd server && uv run pytest tests/test_app.py::TestRepoConfig -v`
-Expected: FAIL — `runner.run()` not receiving `model` kwarg
-
-**Step 3: Write minimal implementation**
-
-In `server/src/warren/app.py`, update the `create_session` endpoint. After line 273 (`repo_path = config.repos[body.repo].path`), extract the repo config:
-
-```python
-        repo_config = config.repos[body.repo]
-        repo_path = repo_config.path
-```
-
-Then update the `runner.run()` call inside `_run()`:
-
-```python
-                async for event in runner.run(
-                    message=body.message,
-                    repo_path=repo_path,
-                    model=repo_config.model,
-                    append_system_prompt=repo_config.append_system_prompt,
-                ):
-```
-
-Do the same for `send_message` endpoint — after getting `repo_path`, also get the repo config and pass model/prompt:
-
-```python
-        repo_config = config.repos[session["repo"]]
-        repo_path = repo_config.path
-        ...
-                async for event in runner.run(
-                    message=body.message,
-                    repo_path=repo_path,
-                    resume_session_id=claude_sid,
-                    model=repo_config.model,
-                    append_system_prompt=repo_config.append_system_prompt,
-                ):
-```
-
-**Step 4: Update config.yaml.example**
-
-Add the new fields to `server/config.yaml.example`:
-
-```yaml
-repos:
-  my-project:
-    path: /home/user/projects/my-project
-    git_url: git@github.com:user/my-project.git
-    # Optional: override the Claude model (default: sonnet).
-    # Values: opus, sonnet, haiku, or a full model ID.
-    model: opus
-    # Optional: extra context appended to Claude's system prompt.
-    # Use this to describe the project, constraints, or conventions.
-    append_system_prompt: |
-      This is a FastAPI web application.
-      Always run tests after making changes.
-```
-
-**Step 5: Run full test suite**
-
-Run: `cd server && uv run pytest tests/ -v`
-Expected: ALL PASS
-
-Run: `cd server && uv run ruff check src/ tests/`
-Expected: clean
-
-**Step 6: Commit**
-
-```bash
-git add server/src/warren/app.py server/tests/test_app.py server/config.yaml.example
-git commit -m "feat: wire per-repo model and system prompt through to sessions"
-```
-
----
-
-### Task 4: Simplify session creation in the R1 UI
-
-**Files:**
-- Modify: `creation/index.html:39-54`
-- Modify: `creation/js/app.js:387-420,500-512`
-
-**Step 1: Remove the name field from the new session form**
-
-In `creation/index.html`, remove the NAME field from the new session form. The form should become:
-
-```html
-  <!-- New Session View -->
-  <div id="view-new" class="view">
-    <div class="header">
-      <button id="btn-back-new" class="btn-back">&lt;-</button>
-      <span class="title">NEW SESSION</span>
-    </div>
-    <div class="form-area">
-      <label for="repo-select">REPO</label>
-      <select id="repo-select" class="input"></select>
-      <label for="initial-msg">PROMPT</label>
-      <input id="initial-msg" type="text" placeholder="enter command..." class="input">
-      <button id="btn-start" class="btn-primary">[ EXEC ]</button>
-    </div>
-  </div>
-```
-
-**Step 2: Update app.js — remove sessionName references**
-
-In `creation/js/app.js`:
-
-1. Remove the `sessionName` DOM ref (line 29: `const sessionName = document.getElementById("session-name");`)
-
-2. Update `startSession()` — remove `sessionName.value.trim()` and pass `undefined` for name:
-
-```javascript
-  async function startSession() {
-    const repo = repoSelect.value;
-    const msg = initialMsg.value.trim();
-    if (!repo || !msg) return;
-
-    try {
-      const data = await Warren.createSession(repo, msg);
-      initialMsg.value = "";
-      openSession({ id: data.session_id, name: data.name, repo: data.repo });
-    } catch (err) {
-      appendMsg("msg-error", "Failed to create session");
-    }
-  }
-```
-
-3. Update `loadRepos()` — auto-select and skip to prompt if only one repo:
-
-```javascript
-  async function loadRepos() {
-    try {
-      const repos = await Warren.fetchRepos();
-      const names = Object.keys(repos);
-      repoSelect.textContent = "";
-      names.forEach((name) => {
-        const opt = document.createElement("option");
-        opt.value = name;
-        opt.textContent = name;
-        repoSelect.appendChild(opt);
-      });
-      // Auto-hide repo picker if only one repo
-      const repoLabel = repoSelect.previousElementSibling;
-      if (names.length <= 1) {
-        repoSelect.style.display = "none";
-        repoLabel.style.display = "none";
-      } else {
-        repoSelect.style.display = "";
-        repoLabel.style.display = "";
-      }
-      // Focus the prompt input
-      initialMsg.focus();
-    } catch (err) {
-      repoSelect.textContent = "";
-      const opt = document.createElement("option");
-      opt.textContent = "Error loading repos";
-      repoSelect.appendChild(opt);
-    }
-  }
-```
-
-**Step 3: Verify manually**
-
-The creation has no automated test suite — test by opening `creation/index.html` in a browser:
-- The "NEW SESSION" view should show only REPO (if >1) and PROMPT fields
-- With a single repo, only the PROMPT input and EXEC button are visible
-- Session name auto-generates server-side from the message
-
-**Step 4: Commit**
-
-```bash
-git add creation/index.html creation/js/app.js
-git commit -m "feat: simplify session creation — remove name field, auto-hide single repo"
-```
-
----
-
-### Task 5: Output visual hierarchy and density
-
-**Files:**
-- Modify: `creation/css/styles.css:243-297`
-- Modify: `creation/js/app.js:199-234`
-
-**Step 1: Update CSS for visual hierarchy**
-
-In `creation/css/styles.css`, replace the message styles (`.msg` through `.msg-result::before`) with:
-
-```css
-/* Chat messages — visual hierarchy */
-.msg {
-  padding: 3px 8px;
-  font-size: 11px;
-  line-height: 1.3;
-}
-
-.msg-user {
-  background: #0a1a0a;
-  padding: 4px 8px;
-  margin-top: 6px;
-  font-size: 11px;
-  color: #33ff33;
-  text-shadow: 0 0 4px rgba(51, 255, 51, 0.3);
-}
-
-.msg-user::before {
-  content: "> ";
-  font-weight: 700;
-}
-
-.msg-text {
-  white-space: pre-wrap;
-  word-break: break-word;
-  text-shadow: 0 0 4px rgba(51, 255, 51, 0.3);
-  margin-top: 2px;
-}
-
-.msg-text.truncated {
-  max-height: 4.4em; /* ~4 lines at 1.1em line-height */
-  overflow: hidden;
-  position: relative;
-}
-
-.msg-text.truncated::after {
-  content: "...";
-  position: absolute;
-  bottom: 0;
-  right: 8px;
-  background: #0a0a0a;
-  padding-left: 4px;
-  color: #0a6e0a;
-}
-
-.msg-text.expanded {
-  max-height: none;
-}
-
-.msg-text.expanded::after {
-  display: none;
-}
-
-.msg-tool {
-  color: #0a6e0a;
-  font-size: 10px;
-  padding: 2px 8px 2px 10px;
-  border-left: 2px solid #0a6e0a;
-  margin-left: 6px;
-}
-
-.msg-tool::before {
-  content: "[TOOL] ";
-  color: #33ff33;
-}
-
-.msg-tool-group {
-  color: #0a6e0a;
-  font-size: 10px;
-  padding: 2px 8px 2px 10px;
-  border-left: 2px solid #0a6e0a;
-  margin-left: 6px;
-  cursor: pointer;
-}
-
-.msg-tool-group::before {
-  content: "[TOOLS] ";
-  color: #33ff33;
-}
-
-.msg-error {
-  color: #ff3333;
-  font-size: 10px;
-  padding: 2px 8px 2px 10px;
-  border-left: 2px solid #ff3333;
-  margin-left: 6px;
-  text-shadow: 0 0 4px rgba(255, 51, 51, 0.3);
-}
-
-.msg-error::before {
-  content: "[ERR] ";
-}
-
-.msg-result {
-  background: #0a2a0a;
-  color: #33ff33;
-  font-size: 12px;
-  padding: 4px 8px;
-  margin-top: 4px;
-  font-weight: 700;
-  text-shadow: 0 0 6px rgba(51, 255, 51, 0.5);
-}
-
-.msg-result::before {
-  content: "[OK] ";
-}
-```
-
-**Step 2: Update app.js — user messages, tool collapsing, text truncation, auto-scroll**
-
-In `creation/js/app.js`, update `handleEvent` and `appendMsg` plus add new helpers:
-
-1. Add a tool batching buffer above the `handleEvent` function:
-
-```javascript
-  let toolBuffer = [];
-  let toolFlushTimer = null;
-```
-
-2. Add helper functions for tool collapsing and text truncation:
-
-```javascript
-  function flushToolBuffer() {
-    if (toolBuffer.length === 0) return;
-    if (toolBuffer.length === 1) {
-      appendRawMsg("msg-tool", toolBuffer[0]);
-    } else {
-      const names = toolBuffer.map((t) => t.split(" ")[0]);
-      const unique = [...new Set(names)];
-      const summary = `${toolBuffer.length} tools: ${unique.join(", ")}`;
-      const group = document.createElement("div");
-      group.className = "msg msg-tool-group";
-      group.textContent = summary;
-      // Tap to expand
-      const details = toolBuffer.slice();
-      group.addEventListener("click", () => {
-        if (group.nextElementSibling?.classList.contains("msg-tool")) return; // already expanded
-        details.forEach((t) => {
-          const el = document.createElement("div");
-          el.className = "msg msg-tool";
-          el.textContent = t;
-          group.after(el);
-        });
-      });
-      chatMessages.appendChild(group);
-      chatMessages.scrollTop = chatMessages.scrollHeight;
-    }
-    toolBuffer = [];
-  }
-```
-
-3. Replace `handleEvent` to use the buffer and add user messages:
-
-```javascript
-  function handleEvent(event) {
-    switch (event.type) {
-      case "text":
-        flushToolBuffer();
-        appendTruncatedMsg("msg-text", event.content);
-        break;
-      case "tool": {
-        const label = event.file
-          ? event.action + " " + event.file
-          : event.cmd
-          ? "$ " + event.cmd
-          : event.action;
-        toolBuffer.push(label);
-        clearTimeout(toolFlushTimer);
-        toolFlushTimer = setTimeout(flushToolBuffer, 300);
-        break;
-      }
-      case "result":
-        flushToolBuffer();
-        if (event.summary) appendRawMsg("msg-result", event.summary);
-        updateChatStatus("waiting");
-        Notify.fire();
-        // Auto-scroll so result is near top of viewport
-        setTimeout(() => {
-          const results = chatMessages.querySelectorAll(".msg-result");
-          const last = results[results.length - 1];
-          if (last) last.scrollIntoView({ block: "start" });
-        }, 50);
-        break;
-      case "status":
-        flushToolBuffer();
-        updateChatStatus(event.state);
-        break;
-      case "error":
-        flushToolBuffer();
-        appendRawMsg("msg-error", event.message || "Unknown error");
-        updateChatStatus("error");
-        break;
-    }
-  }
-```
-
-4. Rename old `appendMsg` to `appendRawMsg` and add `appendTruncatedMsg` and `appendUserMsg`:
-
-```javascript
-  function appendRawMsg(cls, text) {
-    const div = document.createElement("div");
-    div.className = "msg " + cls;
-    div.textContent = text;
-    chatMessages.appendChild(div);
-    chatMessages.scrollTop = chatMessages.scrollHeight;
-  }
-
-  function appendTruncatedMsg(cls, text) {
-    const div = document.createElement("div");
-    div.className = "msg " + cls + " truncated";
-    div.textContent = text;
-    // Tap to expand
-    div.addEventListener("click", () => {
-      div.classList.toggle("truncated");
-      div.classList.toggle("expanded");
-    });
-    chatMessages.appendChild(div);
-    chatMessages.scrollTop = chatMessages.scrollHeight;
-  }
-
-  function appendUserMsg(text) {
-    const div = document.createElement("div");
-    div.className = "msg msg-user";
-    div.textContent = text;
-    chatMessages.appendChild(div);
-    chatMessages.scrollTop = chatMessages.scrollHeight;
-  }
-```
-
-5. Update `sendMsg()` to use `appendUserMsg` instead of `appendMsg("msg-text", "> " + text)`:
-
-```javascript
-    appendUserMsg(text);
-```
-
-6. Update ALL other calls from `appendMsg(` to `appendRawMsg(` — search the file for remaining `appendMsg(` calls and rename them. These are in:
-   - `openPalette` catch block: `appendMsg("msg-error", ...)` → `appendRawMsg("msg-error", ...)`
-   - `sendMsg` catch block: `appendMsg("msg-error", ...)` → `appendRawMsg("msg-error", ...)`
-   - `startSession` catch block: `appendMsg("msg-error", ...)` → `appendRawMsg("msg-error", ...)`
-   - STT spike handlers (longPressStart): all `appendMsg(` → `appendRawMsg(`
-
-**Step 3: Verify manually**
-
-Open the creation in a browser. Check that:
-- User messages have `>` prefix and dim background
-- Tool events get left border accent
-- Multiple consecutive tools collapse into a single "[N tools]" line
-- Text blocks truncate at ~4 lines with "..." — tap to expand
-- Result `[OK]` lines are brighter and larger
-- After a result arrives, it scrolls near the top of the view
-
-**Step 4: Commit**
-
-```bash
-git add creation/css/styles.css creation/js/app.js
-git commit -m "feat: output visual hierarchy — user msgs, tool collapsing, text truncation"
-```
-
----
-
-### Task 6: Command palette from session list
-
-**Files:**
-- Modify: `creation/index.html:17-19`
-- Modify: `creation/js/app.js:255-327,369-385,500-514`
-- Modify: `creation/js/api.js`
-
-**Step 1: Add CMD button to sessions footer**
-
-In `creation/index.html`, update the sessions footer to include both buttons:
-
-```html
-    <div class="footer">
-      <div class="session-footer-row">
-        <button id="btn-new-session" class="btn-primary">[ + NEW ]</button>
-        <button id="btn-cmd-sessions" class="btn-cmd-lg">[ CMD ]</button>
-      </div>
-    </div>
-```
-
-**Step 2: Add CSS for the footer row**
-
-In `creation/css/styles.css`, add after the `.btn-primary:active` rule:
-
-```css
-.session-footer-row {
-  display: flex;
-  gap: 4px;
-}
-
-.session-footer-row .btn-primary {
-  flex: 1;
-}
-
-.btn-cmd-lg {
-  background: transparent;
-  border: 1px solid #0a6e0a;
-  color: #0a6e0a;
-  font-family: "Courier New", "Lucida Console", monospace;
-  font-size: 11px;
-  font-weight: 700;
-  padding: 6px 8px;
-  cursor: pointer;
-  text-transform: uppercase;
-  letter-spacing: 1px;
-}
-
-.btn-cmd-lg:active {
-  border-color: #33ff33;
-  color: #33ff33;
-}
-```
-
-**Step 3: Update app.js — palette from sessions creates new session**
-
-In `creation/js/app.js`:
-
-1. Add a flag to track where the palette was opened from:
-
-```javascript
-  let paletteFromSessions = false;
-```
-
-2. Update `openPalette` to accept an optional source parameter — no changes needed to the function body itself, just track the source:
-
-```javascript
-  async function openPalette(fromSessions) {
-    paletteFromSessions = !!fromSessions;
-    // ... rest of existing openPalette code unchanged ...
-  }
-```
-
-3. Update `selectPaletteItem` to handle session creation when opened from sessions list:
-
-```javascript
-  function selectPaletteItem(index) {
-    const item = paletteItems[index];
-    if (!item) return;
-    closePalette();
-
-    if (paletteFromSessions) {
-      // Create a new session with this prompt
-      createSessionFromPalette(item);
-      return;
-    }
-
-    if (item.type === "template") {
-      // Pre-fill input with template, select placeholder for overtype
-      const match = item.prompt.match(/\{[^}]+\}/);
-      if (match) {
-        msgInput.value = item.prompt;
-        msgInput.focus();
-        const start = item.prompt.indexOf(match[0]);
-        msgInput.setSelectionRange(start, start + match[0].length);
-      } else {
-        msgInput.value = item.prompt;
-        sendMsg();
-      }
-    } else {
-      // Actions and recents — send immediately
-      msgInput.value = item.prompt;
-      sendMsg();
-    }
-  }
-```
-
-4. Add the `createSessionFromPalette` function:
-
-```javascript
-  async function createSessionFromPalette(item) {
-    // Templates from session list: go to new session view with pre-filled prompt
-    if (item.type === "template") {
-      showView("new");
-      initialMsg.value = item.prompt;
-      const match = item.prompt.match(/\{[^}]+\}/);
-      if (match) {
-        initialMsg.focus();
-        const start = item.prompt.indexOf(match[0]);
-        initialMsg.setSelectionRange(start, start + match[0].length);
-      }
-      return;
-    }
-
-    // Actions and recents: auto-create session with default/first repo
-    try {
-      const repos = await Warren.fetchRepos();
-      const repoName = Object.keys(repos)[0];
-      if (!repoName) {
-        appendRawMsg("msg-error", "No repos configured");
-        return;
-      }
-      const data = await Warren.createSession(repoName, item.prompt);
-      openSession({ id: data.session_id, name: data.name, repo: data.repo });
-    } catch (err) {
-      appendRawMsg("msg-error", "Failed to create session");
-    }
-  }
-```
-
-5. Wire up the button handlers. Update the existing `btn-cmd` handler and add the new one:
-
-```javascript
-  document.getElementById("btn-cmd").addEventListener("click", () => openPalette(false));
-  document.getElementById("btn-cmd-sessions").addEventListener("click", () => openPalette(true));
-```
-
-**Step 4: Verify manually**
-
-Open creation in a browser:
-- Session list view shows `[+ NEW]` and `[CMD]` side by side
-- Tapping CMD from sessions opens the palette
-- Selecting an action auto-creates a session and opens chat
-- Selecting a template navigates to new session view with prompt pre-filled
-- CMD from chat still works as before (sends to current session)
-
-**Step 5: Commit**
-
-```bash
-git add creation/index.html creation/css/styles.css creation/js/app.js
-git commit -m "feat: command palette accessible from session list for zero-typing actions"
-```
-
----
-
-### Task 7: Final verification
-
-**Step 1: Run full test suite and lint**
-
-```bash
-cd server && uv run ruff check src/ tests/
-cd server && uv run pytest tests/ -v
-```
-
-Expected: all pass, lint clean.
-
-**Step 2: Verify no regressions in existing functionality**
-
-Check that:
-- Existing session creation with name still works (server accepts `name: null`)
-- Sessions list, chat, SSE streaming all work
-- Command palette in chat still works
-- Admin endpoints still work
-
-**Step 3: Commit any fixes if needed, then push**
-
-```bash
-git push origin main
-```
diff --git a/docs/plans/2026-02-20-settings-dashboard-design.md b/docs/plans/2026-02-20-settings-dashboard-design.md
deleted file mode 100644
index e3a5eaa..0000000
--- a/docs/plans/2026-02-20-settings-dashboard-design.md
+++ /dev/null
@@ -1,82 +0,0 @@
-# Warren Settings Dashboard Design
-
-## Goal
-
-Add a browser-accessible settings dashboard so config changes (commands, per-repo model/prompt) don't require SSH + editing config.yaml manually.
-
-## Architecture
-
-Admin-gated SPA-style dashboard served at `/settings`. Vanilla HTML/CSS/JS with fetch-based saves to API endpoints. Config persisted to YAML and updated in memory. Matches existing admin cookie auth flow.
-
-## 1. API Endpoints
-
-### `GET /config` (admin-gated)
-
-Returns all editable config sections:
-
-```json
-{
-  "commands": {
-    "actions": [{"label": "RUN TESTS", "prompt": "run the test suite"}],
-    "templates": [{"label": "FIX BUG", "prompt": "fix the bug in {description}"}]
-  },
-  "repos": {
-    "warren": {"model": "opus", "append_system_prompt": "This is Warren..."},
-    "other-project": {"model": "", "append_system_prompt": ""}
-  }
-}
-```
-
-### `PUT /config/repos` (admin-gated)
-
-Updates `model` and `append_system_prompt` per repo. Does NOT add/remove repos or change paths/git_urls.
-
-```json
-{
-  "warren": {"model": "opus", "append_system_prompt": "..."},
-  "other-project": {"model": "sonnet", "append_system_prompt": ""}
-}
-```
-
-### `PUT /config/commands` (existing)
-
-Already implemented. Updates actions and templates.
-
-## 2. Settings Dashboard UI
-
-Standalone HTML page at `/settings`. Not inside the R1 Creation (240x282 is too small for editing).
-
-**Auth flow:** Page loads, tries `GET /config`. If 403, redirects to `/admin` login page. After login, cookie is set, redirects back to `/settings`.
-
-**Layout:** Dark CRT theme (matching Warren aesthetic), full-screen friendly. Two sections:
-
-**Commands section:**
-- Action list: label + prompt fields per row, delete button, "Add action" button
-- Template list: label + prompt fields per row, delete button, "Add template" button
-- Save button PUTs to `/config/commands`
-
-**Repos section:**
-- Per repo: name (read-only), model dropdown (empty/opus/sonnet/haiku), system prompt textarea
-- Save button PUTs to `/config/repos`
-
-**Tech:** Single HTML file with inline CSS/JS. Vanilla JS, fetch-based. No framework.
-
-## 3. Config Persistence
-
-Add `save_repos()` to `config.py` mirroring existing `save_commands()`:
-- Reads YAML, updates only `model` and `append_system_prompt` per repo
-- Preserves `path`, `git_url`, and all other config fields
-- Updates `app.state.config` in memory after save (no restart needed)
-
-## Files Affected
-
-**Server:**
-- `server/src/warren/config.py` — add `save_repos()` helper
-- `server/src/warren/app.py` — add `GET /config`, `PUT /config/repos`, serve settings page, admin redirect
-
-**Dashboard:**
-- `server/settings.html` (new) — standalone settings dashboard
-
-**Tests:**
-- `server/tests/test_config.py` — test `save_repos()`
-- `server/tests/test_app.py` — test new endpoints (auth, GET, PUT round-trip)
diff --git a/docs/plans/2026-02-20-settings-dashboard-plan.md b/docs/plans/2026-02-20-settings-dashboard-plan.md
deleted file mode 100644
index 3831589..0000000
--- a/docs/plans/2026-02-20-settings-dashboard-plan.md
+++ /dev/null
@@ -1,344 +0,0 @@
-# Settings Dashboard Implementation Plan
-
-> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
-
-**Goal:** Add a browser-accessible admin settings dashboard for editing commands and per-repo settings without SSH.
-
-**Architecture:** Admin-gated SPA-style dashboard at `/settings`. Vanilla HTML/CSS/JS with fetch-based saves. New `save_repos()` helper mirrors existing `save_commands()`. Two new API endpoints: `GET /config` and `PUT /config/repos`. Admin cookie auth (same flow as `/admin` login).
-
-**Tech Stack:** FastAPI, pydantic, PyYAML, vanilla HTML/CSS/JS
-
----
-
-### Task 1: `save_repos()` helper
-
-**Files:**
-- Modify: `server/src/warren/config.py:59-66` (add `save_repos` next to `save_commands`)
-- Test: `server/tests/test_config.py`
-
-**Step 1: Write the failing test**
-
-Add to `server/tests/test_config.py`:
-
-```python
-def test_save_repos_updates_model_and_prompt(tmp_path):
-    """save_repos() updates repo model/prompt in YAML and preserves other fields."""
-    from warren.config import save_repos
-
-    config_file = tmp_path / "config.yaml"
-    config_file.write_text(yaml.dump({
-        "repos": {
-            "my-project": {
-                "path": "/home/user/project",
-                "git_url": "git@github.com:user/project.git",
-                "model": "",
-                "append_system_prompt": "",
-            },
-            "other": {
-                "path": "/tmp/other",
-            },
-        },
-        "max_concurrent_sessions": 4,
-        "commands": {"actions": [{"label": "TEST", "prompt": "run tests"}], "templates": []},
-    }))
-
-    save_repos(str(config_file), {
-        "my-project": {"model": "opus", "append_system_prompt": "This is a web app."},
-        "other": {"model": "haiku", "append_system_prompt": ""},
-    })
-
-    with open(config_file) as f:
-        data = yaml.safe_load(f)
-
-    # Repo settings updated
-    assert data["repos"]["my-project"]["model"] == "opus"
-    assert data["repos"]["my-project"]["append_system_prompt"] == "This is a web app."
-    assert data["repos"]["other"]["model"] == "haiku"
-    # path, git_url preserved
-    assert data["repos"]["my-project"]["path"] == "/home/user/project"
-    assert data["repos"]["my-project"]["git_url"] == "git@github.com:user/project.git"
-    assert data["repos"]["other"]["path"] == "/tmp/other"
-    # Other config fields preserved
-    assert data["max_concurrent_sessions"] == 4
-    assert data["commands"]["actions"][0]["label"] == "TEST"
-```
-
-**Step 2: Run test to verify it fails**
-
-Run: `cd server && uv run pytest tests/test_config.py::test_save_repos_updates_model_and_prompt -v`
-Expected: FAIL with `ImportError: cannot import name 'save_repos'`
-
-**Step 3: Write minimal implementation**
-
-Add to `server/src/warren/config.py` after `save_commands()`:
-
-```python
-def save_repos(path: str, repos: dict[str, dict[str, str]]) -> None:
-    """Update model and append_system_prompt for repos in the YAML config file."""
-    p = Path(path)
-    with open(p) as f:
-        data = yaml.safe_load(f) or {}
-    existing = data.get("repos", {})
-    for name, updates in repos.items():
-        if name in existing:
-            existing[name]["model"] = updates.get("model", "")
-            existing[name]["append_system_prompt"] = updates.get("append_system_prompt", "")
-    data["repos"] = existing
-    with open(p, "w") as f:
-        yaml.dump(data, f, default_flow_style=False)
-```
-
-**Step 4: Run test to verify it passes**
-
-Run: `cd server && uv run pytest tests/test_config.py -v`
-Expected: All PASS
-
-**Step 5: Commit**
-
-```bash
-git add server/src/warren/config.py server/tests/test_config.py
-git commit -m "feat: add save_repos() helper for persisting repo settings"
-```
-
----
-
-### Task 2: `GET /config` and `PUT /config/repos` endpoints
-
-**Files:**
-- Modify: `server/src/warren/app.py:26` (add `save_repos` import)
-- Modify: `server/src/warren/app.py:230-242` (add endpoints after existing `PUT /config/commands`)
-- Test: `server/tests/test_app.py`
-
-**Step 1: Write the failing tests**
-
-Add to `server/tests/test_app.py`:
-
-```python
-class TestConfigEndpoints:
-    async def test_get_config_requires_admin(self, client):
-        resp = await client.get("/config")
-        assert resp.status_code == 403
-
-    async def test_get_config_returns_commands_and_repos(self, client):
-        resp = await client.get("/config", headers=ADMIN_HEADERS)
-        assert resp.status_code == 200
-        data = resp.json()
-        # Commands section
-        assert "commands" in data
-        assert len(data["commands"]["actions"]) == 1
-        assert data["commands"]["actions"][0]["label"] == "RUN TESTS"
-        # Repos section (model + prompt only, no path/git_url)
-        assert "repos" in data
-        assert "test-repo" in data["repos"]
-        assert data["repos"]["test-repo"]["model"] == "opus"
-        assert "path" not in data["repos"]["test-repo"]
-
-    async def test_put_repos_requires_admin(self, client):
-        resp = await client.put("/config/repos", json={
-            "test-repo": {"model": "haiku", "append_system_prompt": ""},
-        })
-        assert resp.status_code == 403
-
-    async def test_put_repos_updates_and_returns(self, client):
-        resp = await client.put(
-            "/config/repos",
-            headers=ADMIN_HEADERS,
-            json={
-                "test-repo": {"model": "haiku", "append_system_prompt": "Updated prompt."},
-            },
-        )
-        assert resp.status_code == 200
-        data = resp.json()
-        assert data["test-repo"]["model"] == "haiku"
-        assert data["test-repo"]["append_system_prompt"] == "Updated prompt."
-
-    async def test_put_repos_round_trip(self, client):
-        """PUT then GET reflects updated repo settings."""
-        await client.put(
-            "/config/repos",
-            headers=ADMIN_HEADERS,
-            json={
-                "test-repo": {"model": "sonnet", "append_system_prompt": "New prompt."},
-            },
-        )
-        resp = await client.get("/config", headers=ADMIN_HEADERS)
-        assert resp.status_code == 200
-        data = resp.json()
-        assert data["repos"]["test-repo"]["model"] == "sonnet"
-        assert data["repos"]["test-repo"]["append_system_prompt"] == "New prompt."
-```
-
-**Step 2: Run tests to verify they fail**
-
-Run: `cd server && uv run pytest tests/test_app.py::TestConfigEndpoints -v`
-Expected: FAIL (404 — endpoints don't exist yet)
-
-**Step 3: Write minimal implementation**
-
-In `server/src/warren/app.py`, update the import line:
-
-```python
-from warren.config import CommandsConfig, Settings, load_config, save_commands, save_repos
-```
-
-Add these endpoints after the existing `PUT /config/commands` block (after line 242):
-
-```python
-    @app.get("/config", dependencies=[Depends(require_admin)])
-    async def get_config(request: Request):
-        config = request.app.state.config
-        return {
-            "commands": {
-                "actions": [
-                    {"label": a.label, "prompt": a.prompt}
-                    for a in config.commands.actions
-                ],
-                "templates": [
-                    {"label": t.label, "prompt": t.prompt}
-                    for t in config.commands.templates
-                ],
-            },
-            "repos": {
-                name: {
-                    "model": repo.model,
-                    "append_system_prompt": repo.append_system_prompt,
-                }
-                for name, repo in config.repos.items()
-            },
-        }
-
-    @app.put("/config/repos", dependencies=[Depends(require_admin)])
-    async def update_repos(request: Request):
-        body = await request.json()
-        config_path = request.app.state.settings.config
-        save_repos(config_path, body)
-        # Update in-memory config
-        for name, updates in body.items():
-            if name in request.app.state.config.repos:
-                request.app.state.config.repos[name].model = updates.get("model", "")
-                request.app.state.config.repos[name].append_system_prompt = updates.get("append_system_prompt", "")
-        return {
-            name: {
-                "model": repo.model,
-                "append_system_prompt": repo.append_system_prompt,
-            }
-            for name, repo in request.app.state.config.repos.items()
-            if name in body
-        }
-```
-
-**Step 4: Run tests to verify they pass**
-
-Run: `cd server && uv run pytest tests/test_app.py -v`
-Expected: All PASS
-
-**Step 5: Lint**
-
-Run: `cd server && uv run ruff check src/ tests/`
-Expected: Clean
-
-**Step 6: Commit**
-
-```bash
-git add server/src/warren/app.py server/tests/test_app.py
-git commit -m "feat: add GET /config and PUT /config/repos admin endpoints"
-```
-
----
-
-### Task 3: Settings dashboard HTML page
-
-**Files:**
-- Create: `server/settings.html`
-- Modify: `server/src/warren/app.py` (serve the settings page)
-
-**Step 1: Create the settings page**
-
-Create `server/settings.html`. This is a standalone SPA-style page with inline CSS/JS:
-
-- Dark CRT theme matching Warren's aesthetic (green on black, monospace)
-- On load: fetch `GET /config`, populate forms
-- Commands section: editable list of actions (label + prompt), editable list of templates (label + prompt), add/delete buttons, save button → `PUT /config/commands`
-- Repos section: per-repo model dropdown (empty/opus/sonnet/haiku) + system prompt textarea, save button → `PUT /config/repos`
-- Auth: all fetches include credentials (cookie). If any fetch returns 403, redirect to `/admin`
-- Success/error feedback inline (no alerts)
-
-The full HTML content is provided at implementation time since it's a complete file. Key structure:
-
-```html
-<!DOCTYPE html>
-<html>
-<head>
-  <title>Warren Settings</title>
-  <style>/* CRT dark theme, form styles */</style>
-</head>
-<body>
-  <h1>WARREN:// SETTINGS</h1>
-  <section id="commands-section"><!-- actions + templates lists --></section>
-  <section id="repos-section"><!-- per-repo model + prompt --></section>
-  <script>/* fetch, render, save logic */</script>
-</body>
-</html>
-```
-
-**Step 2: Serve the settings page from app.py**
-
-Add a route in `server/src/warren/app.py` that serves the HTML file with admin auth. Add after the `admin_login` endpoint:
-
-```python
-    @app.get("/settings", dependencies=[Depends(require_admin)])
-    async def settings_page():
-        settings_candidates = [
-            Path("/app/server/settings.html"),
-            Path(__file__).parent.parent.parent / "settings.html",
-        ]
-        for candidate in settings_candidates:
-            if candidate.is_file():
-                return HTMLResponse(content=candidate.read_text())
-        raise HTTPException(status_code=404, detail="Settings page not found")
-```
-
-**Step 3: Update admin login redirect**
-
-Currently `/admin/login` redirects to `/pair/qr`. We should keep that as default, but the settings page should handle its own auth check by trying `GET /config` and redirecting to `/admin` on 403. No change to the login endpoint needed — the settings JS handles the redirect.
-
-**Step 4: Run all tests**
-
-Run: `cd server && uv run pytest tests/ -v`
-Expected: All PASS
-
-**Step 5: Lint**
-
-Run: `cd server && uv run ruff check src/ tests/`
-Expected: Clean
-
-**Step 6: Commit**
-
-```bash
-git add server/settings.html server/src/warren/app.py
-git commit -m "feat: add settings dashboard page"
-```
-
----
-
-### Task 4: Final verification
-
-**Step 1: Run full test suite**
-
-Run: `cd server && uv run pytest tests/ -v`
-Expected: All PASS
-
-**Step 2: Lint**
-
-Run: `cd server && uv run ruff check src/ tests/`
-Expected: Clean
-
-**Step 3: Manual smoke test description**
-
-1. Start server: `cd server && uv run python -m warren`
-2. Open browser to `http://localhost:4030/settings` — should redirect to `/admin`
-3. Enter admin token, should redirect to `/pair/qr` (existing behavior)
-4. Navigate to `http://localhost:4030/settings` — should load dashboard with current config
-5. Edit a command label, save — should show success
-6. Edit a repo model, save — should show success
-7. Reload page — changes should persist
diff --git a/docs/plans/2026-02-20-warren-gateway-design.md b/docs/plans/2026-02-20-warren-gateway-design.md
deleted file mode 100644
index ef811e4..0000000
--- a/docs/plans/2026-02-20-warren-gateway-design.md
+++ /dev/null
@@ -1,317 +0,0 @@
-# Warren Universal AI Gateway Design
-
-## Summary
-
-Warren becomes a universal AI gateway. Instead of building container
-orchestration, scheduling, and swarm infrastructure natively, Warren
-uses a backend adapter interface. The R1 talks to Warren. Warren talks
-to whatever backend handles AI execution.
-
-NanoClaw becomes the first managed backend via file-based IPC. Warren's
-current bare Claude CLI mode becomes the "direct" backend adapter.
-AI-native patterns (skills, self-setup, self-healing) are built in
-Warren's `.claude/skills/` to be extracted to a reusable library later.
-
-**Supersedes:** `2026-02-20-nanoclaw-integration-design.md` (native build
-approach). That design assumed Warren would port NanoClaw's internals to
-Python. This design keeps NanoClaw intact and makes Warren a thin adapter.
-
----
-
-## Architecture
-
-```
-R1 Device
-  |
-  | HTTPS / SSE
-  v
-+----------------------------------------------+
-|  Warren (FastAPI)                            |
-|                                              |
-|  R1 Interface:        Backend Adapters:      |
-|  - QR pairing         - DirectAdapter        |
-|  - Session routing       (bare claude CLI)   |
-|  - SSE streaming       - NanoClawAdapter     |
-|  - Command palette       (file-based IPC)    |
-|  - Creation serving    - [Future adapters]   |
-|                                              |
-|  AI-Native:                                  |
-|  - .claude/skills/*                          |
-|  - Self-setup (/setup)                       |
-|  - Self-healing (/debug)                     |
-|  - Self-extending (/customize)               |
-+----------------------------------------------+
-       |                        |
-       | subprocess             | file IPC
-       v                        v
-  claude CLI             NanoClaw (Node.js)
-  (local dev)               |
-                             | docker run
-                             v
-                        +----------+  +----------+
-                        | Agent A  |  | Agent B  |
-                        +----------+  +----------+
-```
-
-**Warren owns:**
-- R1 UI, device management, auth, pairing
-- Session routing and protocol translation
-- Backend adapter interface and adapter implementations
-- AI-native skills for self-setup, customization, debugging
-- Config (which backend, backend-specific settings)
-
-**Warren does NOT own:**
-- Container orchestration (NanoClaw)
-- Cron scheduling (NanoClaw)
-- Swarm orchestration (NanoClaw)
-- Notifications (NanoClaw)
-- AI execution (backend)
-
----
-
-## Part 1: Backend Adapter Interface
-
-Message-based, not RPC. Warren sends messages into the adapter and
-receives events back. The adapter is a bidirectional event pipe.
-
-### Protocol
-
-```python
-class BackendAdapter(Protocol):
-    async def send(self, message: BackendMessage) -> None:
-        """Send a message to the backend."""
-        ...
-
-    def events(self) -> AsyncIterator[BackendEvent]:
-        """Stream of events from the backend."""
-        ...
-
-    async def start(self) -> None:
-        """Initialize the adapter."""
-        ...
-
-    async def stop(self) -> None:
-        """Graceful shutdown."""
-        ...
-```
-
-### Message and Event Types
-
-```python
-@dataclass
-class BackendMessage:
-    kind: str
-    session_id: str | None = None
-    payload: dict = field(default_factory=dict)
-
-@dataclass
-class BackendEvent:
-    kind: str
-    session_id: str
-    data: dict = field(default_factory=dict)
-```
-
-Message kinds: `new_session`, `user_message`, `cancel`
-
-Event kinds: `text`, `tool_use`, `result`, `status`, `error`
-
-Adapters translate between Warren's vocabulary and whatever the backend
-speaks. Warren's app.py never knows which backend it's talking to.
-
----
-
-## Part 2: DirectAdapter
-
-Refactors the current `ClaudeRunner` into the adapter interface. This is
-what Warren does today, wrapped in the new abstraction.
-
-- `new_session` message -> spawn `claude` subprocess with `--print`,
-  `--output-format stream-json`
-- Subprocess stdout -> parsed into `BackendEvent`s via existing
-  `parse_stream_line()`
-- `user_message` message -> spawn subprocess with `--resume`
-- `cancel` message -> terminate subprocess
-
-No behavioral change from the user's perspective. Existing R1 sessions
-work exactly as before. This adapter is the default for local
-development and simple deployments.
-
----
-
-## Part 3: NanoClawAdapter
-
-Communicates with NanoClaw via its file-based IPC protocol. Both
-services run on the same VPS.
-
-### IPC Flow
-
-1. Warren writes a task request file to the shared IPC directory
-2. NanoClaw's IPC watcher picks it up and spawns a container
-3. Container runs Claude CLI, writes output with sentinel markers
-4. NanoClaw writes result files to IPC directory
-5. Warren's adapter polls/watches for result files
-6. Adapter translates results to `BackendEvent`s for SSE streaming
-
-### Session Mapping
-
-The adapter maintains a mapping between Warren session IDs and
-NanoClaw task/container IDs. Stored in Warren's SQLite for
-persistence across restarts.
-
-### NanoClaw IPC Protocol
-
-Exact file format and polling mechanics TBD during implementation --
-requires detailed reading of NanoClaw's `src/ipc.ts`. The high-level
-pattern is file-based message passing with structured JSON payloads.
-
----
-
-## Part 4: AI-Native Skills
-
-Built directly in Warren's `.claude/skills/`. Follows NanoClaw's
-philosophy: "Fix it. Don't tell the user to go fix it themselves."
-
-### Skills
-
-```
-.claude/skills/
-  setup/
-    SKILL.md             # VPS deployment wizard
-    scripts/
-      01-check-env.sh    # Detect OS, Python, Docker, Node
-      02-install-deps.sh # uv sync, NanoClaw setup
-      03-configure.sh    # config.yaml, .env, Dokploy volumes
-      04-verify.sh       # End-to-end health check
-  customize/
-    SKILL.md             # Open-ended: "add X to Warren"
-  debug/
-    SKILL.md             # Diagnose adapter, IPC, containers
-```
-
-### Structured Status Blocks
-
-Health check scripts emit JSON:
-```json
-{"component": "warren", "status": "ok", "version": "0.2.0"}
-{"component": "nanoclaw", "status": "ok", "version": "1.0.0"}
-{"component": "ipc", "status": "error", "detail": "shared volume not mounted"}
-{"component": "docker", "status": "ok", "version": "24.0.7"}
-```
-
-The `/debug` skill reads these and fixes problems automatically.
-
-### Self-Documenting CLAUDE.md
-
-Warren's CLAUDE.md expands with extension points:
-
-```
-## Extension Points
-
-To add a backend adapter: create server/src/warren/adapters/{name}.py
-implementing BackendAdapter protocol. Register in config.py.
-
-To add an API endpoint: add route in app.py inside create_app().
-
-To add a skill: create .claude/skills/{name}/SKILL.md.
-```
-
-### Future: Extractable Patterns
-
-These patterns (structured status blocks, skill anatomy, self-healing
-loops, fix-it philosophy) are designed to be extracted into a standalone
-patterns library later. Warren is the proving ground.
-
----
-
-## Part 5: Deployment
-
-Two Dokploy services on the same VPS:
-
-### Warren Service
-- FastAPI container (current setup, mostly unchanged)
-- Shared volume: `/opt/warren-data/ipc/` for NanoClaw IPC
-- Environment: `WARREN_BACKEND=nanoclaw` (or `direct` for standalone)
-
-### NanoClaw Service
-- Node.js container
-- Docker socket mounted: `/var/run/docker.sock`
-- Shared volume: `/opt/warren-data/ipc/` for Warren IPC
-- NanoClaw's own config for repos, container profiles, scheduled tasks
-
-### Setup Flow (guided by /setup skill)
-
-1. Deploy Warren via Dokploy (existing workflow)
-2. Deploy NanoClaw via Dokploy (new service)
-3. Configure shared volume in both services
-4. Mount Docker socket in NanoClaw service
-5. Set `WARREN_BACKEND=nanoclaw` in Warren
-6. Configure NanoClaw repos, tasks, and container profiles
-7. Verify end-to-end: R1 -> Warren -> NanoClaw -> container -> response
-
----
-
-## Part 6: Config Changes
-
-### Settings (env vars)
-
-```python
-class Settings(BaseSettings):
-    # ... existing ...
-    backend: str = "direct"       # "direct" or "nanoclaw"
-    data_dir: str = "data"        # shared data directory
-```
-
-### Config (YAML)
-
-```yaml
-backend_config:
-  direct:
-    claude_bin: claude
-    max_concurrent: 4
-  nanoclaw:
-    ipc_dir: /opt/warren-data/ipc
-```
-
----
-
-## Part 7: API Surface
-
-R1-facing API is mostly unchanged. The backend adapter is invisible
-to clients.
-
-Existing endpoints work identically:
-- `POST /sessions` -> creates session via whichever backend
-- `POST /sessions/{id}/msg` -> sends message via adapter
-- `GET /sessions/{id}/stream` -> SSE from adapter events
-- All auth, pairing, config endpoints unchanged
-
-Optional future additions (proxy to NanoClaw):
-- `GET /api/tasks` -> scheduled task status
-- `GET /api/monitor/health` -> system health
-- These would be thin proxies reading NanoClaw's state
-
----
-
-## New/Modified Files
-
-```
-server/src/warren/
-  adapters/
-    __init__.py          # BackendAdapter protocol, message/event types
-    direct.py            # DirectAdapter (refactored ClaudeRunner)
-    nanoclaw.py          # NanoClawAdapter (file-based IPC)
-  config.py              # + backend, data_dir settings
-
-.claude/skills/
-  setup/SKILL.md + scripts/
-  customize/SKILL.md
-  debug/SKILL.md
-
-CLAUDE.md                # + extension points
-```
-
-## Dependencies
-
-- No new Python dependencies for the adapter interface
-- NanoClaw installed separately (Node.js)
-- Docker on VPS (for NanoClaw's container runtime)
diff --git a/docs/plans/2026-02-20-warren-gateway-plan.md b/docs/plans/2026-02-20-warren-gateway-plan.md
deleted file mode 100644
index f1a1c10..0000000
--- a/docs/plans/2026-02-20-warren-gateway-plan.md
+++ /dev/null
@@ -1,1388 +0,0 @@
-# Warren Universal AI Gateway Implementation Plan
-
-> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
-
-**Goal:** Refactor Warren from a tightly-coupled Claude CLI wrapper into a universal AI gateway with pluggable backend adapters, then add AI-native skills for self-setup, customization, and debugging.
-
-**Architecture:** Message-based BackendAdapter protocol. DirectAdapter wraps existing ClaudeRunner. NanoClawAdapter (stub) prepares for file-based IPC integration. Global event pump in app.py replaces per-session background tasks. AI-native skills in `.claude/skills/`.
-
-**Tech Stack:** Python 3.11, FastAPI, aiosqlite, pydantic-settings, asyncio
-
-**Design Doc:** `docs/plans/2026-02-20-warren-gateway-design.md`
-
----
-
-## Phase 1: Backend Adapter Interface
-
-### Task 1: Define adapter types (protocol, messages, events)
-
-**Files:**
-- Create: `server/src/warren/adapters/__init__.py`
-- Create: `server/tests/test_adapters.py`
-
-**Step 1: Write failing tests**
-
-Create `server/tests/test_adapters.py`:
-
-```python
-"""Tests for backend adapter types."""
-
-
-def test_backend_message_defaults():
-    from warren.adapters import BackendMessage
-
-    msg = BackendMessage(kind="new_session")
-    assert msg.kind == "new_session"
-    assert msg.session_id is None
-    assert msg.payload == {}
-
-
-def test_backend_event_creation():
-    from warren.adapters import BackendEvent
-
-    event = BackendEvent(kind="text", session_id="abc", data={"content": "hello"})
-    assert event.kind == "text"
-    assert event.session_id == "abc"
-    assert event.data["content"] == "hello"
-
-
-def test_backend_message_with_payload():
-    from warren.adapters import BackendMessage
-
-    msg = BackendMessage(
-        kind="new_session",
-        session_id="abc",
-        payload={"message": "hello", "repo": "test"},
-    )
-    assert msg.session_id == "abc"
-    assert msg.payload["message"] == "hello"
-```
-
-**Step 2: Run tests to verify they fail**
-
-Run: `cd server && uv run pytest tests/test_adapters.py -v`
-Expected: FAIL — `ModuleNotFoundError: No module named 'warren.adapters'`
-
-**Step 3: Write minimal implementation**
-
-Create `server/src/warren/adapters/__init__.py`:
-
-```python
-"""Backend adapter interface for Warren.
-
-Warren uses a message-based adapter protocol to communicate with AI backends.
-Send BackendMessages in, receive BackendEvents out. The adapter translates
-between Warren's vocabulary and whatever the backend speaks.
-"""
-
-from __future__ import annotations
-
-from collections.abc import AsyncIterator
-from dataclasses import dataclass, field
-from typing import Protocol, runtime_checkable
-
-
-@dataclass
-class BackendMessage:
-    """Message sent from Warren to a backend adapter."""
-
-    kind: str  # "new_session", "user_message", "cancel"
-    session_id: str | None = None
-    payload: dict = field(default_factory=dict)
-
-
-@dataclass
-class BackendEvent:
-    """Event emitted by a backend adapter back to Warren."""
-
-    kind: str  # "text", "tool", "result", "status", "error"
-    session_id: str
-    data: dict = field(default_factory=dict)
-
-
-@runtime_checkable
-class BackendAdapter(Protocol):
-    """Protocol that all backend adapters must satisfy."""
-
-    async def send(self, message: BackendMessage) -> None: ...
-    def events(self) -> AsyncIterator[BackendEvent]: ...
-    async def start(self) -> None: ...
-    async def stop(self) -> None: ...
-```
-
-**Step 4: Run tests to verify they pass**
-
-Run: `cd server && uv run pytest tests/test_adapters.py -v`
-Expected: 3 passed
-
-**Step 5: Commit**
-
-```bash
-git add server/src/warren/adapters/__init__.py server/tests/test_adapters.py
-git commit -m "feat: add backend adapter protocol and message types"
-```
-
----
-
-### Task 2: Implement DirectAdapter
-
-**Files:**
-- Create: `server/src/warren/adapters/direct.py`
-- Modify: `server/tests/test_adapters.py`
-
-**Step 1: Write failing tests**
-
-Append to `server/tests/test_adapters.py`:
-
-```python
-import asyncio
-from unittest.mock import patch
-
-import pytest
-
-
-class TestDirectAdapter:
-    async def test_new_session_produces_events(self):
-        from warren.adapters import BackendMessage
-        from warren.adapters.direct import DirectAdapter
-
-        adapter = DirectAdapter(claude_bin="claude")
-
-        events_to_yield = [
-            {"type": "status", "state": "working"},
-            {"type": "text", "content": "Hello"},
-            {"type": "result", "summary": "Done", "session_id": "claude-123"},
-            {"type": "status", "state": "waiting"},
-        ]
-
-        async def mock_run(**kwargs):
-            for e in events_to_yield:
-                yield e
-
-        with patch.object(adapter._runner, "run", side_effect=mock_run):
-            await adapter.send(BackendMessage(
-                kind="new_session",
-                session_id="test-session",
-                payload={"message": "hello", "repo_path": "/tmp/test"},
-            ))
-
-            collected = []
-            async for event in adapter.events():
-                collected.append(event)
-                if event.kind == "status" and event.data.get("state") == "waiting":
-                    break
-
-        assert len(collected) == 4
-        assert all(e.session_id == "test-session" for e in collected)
-        assert collected[0].kind == "status"
-        assert collected[1].kind == "text"
-        assert collected[2].kind == "result"
-        assert collected[3].kind == "status"
-
-    async def test_user_message_passes_resume_id(self):
-        from warren.adapters import BackendMessage
-        from warren.adapters.direct import DirectAdapter
-
-        adapter = DirectAdapter(claude_bin="claude")
-        captured = {}
-
-        async def mock_run(**kwargs):
-            captured.update(kwargs)
-            yield {"type": "status", "state": "working"}
-            yield {"type": "status", "state": "waiting"}
-
-        with patch.object(adapter._runner, "run", side_effect=mock_run):
-            await adapter.send(BackendMessage(
-                kind="user_message",
-                session_id="test-session",
-                payload={
-                    "message": "continue",
-                    "repo_path": "/tmp/test",
-                    "resume_session_id": "claude-abc",
-                },
-            ))
-
-            # Drain events
-            async for event in adapter.events():
-                if event.kind == "status" and event.data.get("state") == "waiting":
-                    break
-
-        assert captured["resume_session_id"] == "claude-abc"
-
-    async def test_cancel_removes_task(self):
-        from warren.adapters import BackendMessage
-        from warren.adapters.direct import DirectAdapter
-
-        adapter = DirectAdapter(claude_bin="claude")
-
-        # Manually create a tracked task
-        async def long_running():
-            await asyncio.sleep(100)
-
-        adapter._tasks["test-session"] = asyncio.create_task(long_running())
-
-        await adapter.send(BackendMessage(kind="cancel", session_id="test-session"))
-        assert "test-session" not in adapter._tasks
-
-    async def test_error_produces_error_event(self):
-        from warren.adapters import BackendMessage
-        from warren.adapters.direct import DirectAdapter
-
-        adapter = DirectAdapter(claude_bin="claude")
-
-        async def mock_run(**kwargs):
-            raise FileNotFoundError("/tmp/missing")
-            yield  # make it a generator  # noqa: RUF027
-
-        with patch.object(adapter._runner, "run", side_effect=mock_run):
-            await adapter.send(BackendMessage(
-                kind="new_session",
-                session_id="test-session",
-                payload={"message": "hello", "repo_path": "/tmp/missing"},
-            ))
-
-            collected = []
-            async for event in adapter.events():
-                collected.append(event)
-                if event.kind == "status" and event.data.get("state") == "waiting":
-                    break
-
-        assert any(e.kind == "error" for e in collected)
-        assert collected[-1].kind == "status"
-        assert collected[-1].data["state"] == "waiting"
-
-    async def test_stop_cancels_all_tasks(self):
-        from warren.adapters.direct import DirectAdapter
-
-        adapter = DirectAdapter(claude_bin="claude")
-
-        async def long_running():
-            await asyncio.sleep(100)
-
-        adapter._tasks["a"] = asyncio.create_task(long_running())
-        adapter._tasks["b"] = asyncio.create_task(long_running())
-
-        await adapter.stop()
-        assert len(adapter._tasks) == 0
-```
-
-**Step 2: Run tests to verify they fail**
-
-Run: `cd server && uv run pytest tests/test_adapters.py::TestDirectAdapter -v`
-Expected: FAIL — `ModuleNotFoundError: No module named 'warren.adapters.direct'`
-
-**Step 3: Write implementation**
-
-Create `server/src/warren/adapters/direct.py`:
-
-```python
-"""Direct backend adapter -- runs Claude CLI via subprocess.
-
-Wraps the existing ClaudeRunner, adapting its async-generator interface
-to the message/event BackendAdapter protocol.
-"""
-
-from __future__ import annotations
-
-import asyncio
-import logging
-from collections.abc import AsyncIterator
-
-from warren.adapters import BackendEvent, BackendMessage
-from warren.claude_runner import ClaudeRunner
-
-logger = logging.getLogger(__name__)
-
-
-class DirectAdapter:
-    """Backend adapter that runs Claude CLI directly via subprocess."""
-
-    def __init__(self, claude_bin: str = "claude", max_concurrent: int = 4):
-        self._runner = ClaudeRunner(claude_bin=claude_bin, max_concurrent=max_concurrent)
-        self._event_queue: asyncio.Queue[BackendEvent] = asyncio.Queue()
-        self._tasks: dict[str, asyncio.Task] = {}
-
-    async def start(self) -> None:
-        """No-op for direct adapter (no external connections)."""
-
-    async def stop(self) -> None:
-        """Cancel all running sessions."""
-        for task in self._tasks.values():
-            task.cancel()
-        self._tasks.clear()
-
-    async def send(self, message: BackendMessage) -> None:
-        """Handle an incoming message from Warren."""
-        if message.kind in ("new_session", "user_message"):
-            self._spawn_run(message)
-        elif message.kind == "cancel":
-            task = self._tasks.pop(message.session_id, None)
-            if task:
-                task.cancel()
-
-    async def events(self) -> AsyncIterator[BackendEvent]:
-        """Yield events from all running sessions."""
-        while True:
-            event = await self._event_queue.get()
-            yield event
-
-    def _spawn_run(self, message: BackendMessage) -> None:
-        """Start a Claude CLI subprocess for a session."""
-        session_id = message.session_id
-        payload = message.payload
-
-        async def _run() -> None:
-            try:
-                async for event in self._runner.run(
-                    message=payload["message"],
-                    repo_path=payload["repo_path"],
-                    resume_session_id=payload.get("resume_session_id"),
-                    model=payload.get("model", ""),
-                    append_system_prompt=payload.get("append_system_prompt", ""),
-                ):
-                    await self._event_queue.put(
-                        BackendEvent(
-                            kind=event["type"],
-                            session_id=session_id,
-                            data=event,
-                        )
-                    )
-            except FileNotFoundError as exc:
-                repo_path = payload.get("repo_path", "unknown")
-                await self._event_queue.put(BackendEvent(
-                    kind="error",
-                    session_id=session_id,
-                    data={"type": "error", "message": f"Repo not found: {exc.filename or repo_path}"},
-                ))
-                await self._event_queue.put(BackendEvent(
-                    kind="status",
-                    session_id=session_id,
-                    data={"type": "status", "state": "waiting"},
-                ))
-            except Exception as exc:
-                logger.exception("Direct adapter: session %s failed", session_id)
-                await self._event_queue.put(BackendEvent(
-                    kind="error",
-                    session_id=session_id,
-                    data={"type": "error", "message": str(exc)},
-                ))
-                await self._event_queue.put(BackendEvent(
-                    kind="status",
-                    session_id=session_id,
-                    data={"type": "status", "state": "waiting"},
-                ))
-            finally:
-                self._tasks.pop(session_id, None)
-
-        self._tasks[session_id] = asyncio.create_task(_run())
-```
-
-**Step 4: Run tests to verify they pass**
-
-Run: `cd server && uv run pytest tests/test_adapters.py -v`
-Expected: 8 passed (3 type tests + 5 DirectAdapter tests)
-
-**Step 5: Run all existing tests to verify no regressions**
-
-Run: `cd server && uv run pytest tests/ -v`
-Expected: All existing tests still pass (no changes to existing code)
-
-**Step 6: Commit**
-
-```bash
-git add server/src/warren/adapters/direct.py server/tests/test_adapters.py
-git commit -m "feat: add DirectAdapter wrapping ClaudeRunner"
-```
-
----
-
-### Task 3: Add backend selection to config
-
-**Files:**
-- Modify: `server/src/warren/config.py`
-- Modify: `server/tests/test_config.py`
-
-**Step 1: Write failing tests**
-
-Append to `server/tests/test_config.py`:
-
-```python
-def test_settings_backend_defaults():
-    """Settings defaults to 'direct' backend."""
-    from warren.config import Settings
-
-    s = Settings(anthropic_api_key="sk-test")
-    assert s.backend == "direct"
-    assert s.data_dir == "data"
-
-
-def test_config_with_backend_config(tmp_path):
-    """backend_config section parsed from YAML."""
-    from warren.config import load_config
-
-    config_file = tmp_path / "config.yaml"
-    config_file.write_text(yaml.dump({
-        "repos": {"test": {"path": "/tmp"}},
-        "backend_config": {
-            "direct": {"claude_bin": "claude", "max_concurrent": 2},
-            "nanoclaw": {"ipc_dir": "/opt/warren-data/ipc"},
-        },
-    }))
-
-    config = load_config(str(config_file))
-    assert config.backend_config["direct"]["claude_bin"] == "claude"
-    assert config.backend_config["nanoclaw"]["ipc_dir"] == "/opt/warren-data/ipc"
-
-
-def test_config_without_backend_config(tmp_path):
-    """backend_config is optional -- defaults to empty dict."""
-    from warren.config import load_config
-
-    config_file = tmp_path / "config.yaml"
-    config_file.write_text(yaml.dump({
-        "repos": {"test": {"path": "/tmp"}},
-    }))
-
-    config = load_config(str(config_file))
-    assert config.backend_config == {}
-```
-
-**Step 2: Run tests to verify they fail**
-
-Run: `cd server && uv run pytest tests/test_config.py::test_settings_backend_defaults tests/test_config.py::test_config_with_backend_config tests/test_config.py::test_config_without_backend_config -v`
-Expected: FAIL — `AttributeError: 'Settings' object has no attribute 'backend'`
-
-**Step 3: Add fields to config models**
-
-In `server/src/warren/config.py`, add to `WarrenConfig`:
-
-```python
-backend_config: dict[str, dict] = {}
-```
-
-Add to `Settings`:
-
-```python
-backend: str = "direct"
-data_dir: str = "data"
-```
-
-**Step 4: Run tests to verify they pass**
-
-Run: `cd server && uv run pytest tests/test_config.py -v`
-Expected: All config tests pass
-
-**Step 5: Run full test suite**
-
-Run: `cd server && uv run pytest tests/ -v`
-Expected: All tests pass
-
-**Step 6: Commit**
-
-```bash
-git add server/src/warren/config.py server/tests/test_config.py
-git commit -m "feat: add backend selection to config"
-```
-
----
-
-### Task 4: Wire adapter into app.py
-
-This is the critical refactoring task. Replace `ClaudeRunner` usage in `app.py`
-with the adapter interface. Add a global event pump.
-
-**Files:**
-- Modify: `server/src/warren/app.py`
-- Modify: `server/tests/test_app.py`
-
-**Step 1: Add adapter factory function to app.py**
-
-Add after the existing imports in `server/src/warren/app.py`:
-
-```python
-from warren.adapters import BackendAdapter, BackendEvent, BackendMessage
-from warren.adapters.direct import DirectAdapter
-```
-
-Add a factory function (before `create_app`):
-
-```python
-def create_adapter(settings: Settings, config) -> DirectAdapter:
-    """Create the backend adapter based on settings."""
-    if settings.backend == "direct":
-        backend_cfg = config.backend_config.get("direct", {})
-        return DirectAdapter(
-            claude_bin=backend_cfg.get("claude_bin", settings.claude_bin),
-            max_concurrent=backend_cfg.get("max_concurrent", config.max_concurrent_sessions),
-        )
-    raise ValueError(f"Unknown backend: {settings.backend}")
-```
-
-**Step 2: Add event pump function**
-
-Add in `server/src/warren/app.py` (before `create_app`):
-
-```python
-async def event_pump(
-    adapter: BackendAdapter,
-    event_queues: dict[str, asyncio.Queue],
-    db: Database,
-) -> None:
-    """Route backend events to per-session SSE queues."""
-    async for event in adapter.events():
-        queue = event_queues.get(event.session_id)
-        if queue:
-            await queue.put(event.data)
-            if (
-                event.kind == "result"
-                and event.data.get("session_id")
-            ):
-                await db.update_claude_session_id(
-                    event.session_id, event.data["session_id"]
-                )
-```
-
-**Step 3: Update lifespan**
-
-Replace the `ClaudeRunner` setup in the `lifespan` context manager:
-
-Before (lines 73-77):
-```python
-        app.state.runner = ClaudeRunner(
-            claude_bin=settings.claude_bin,
-            max_concurrent=config.max_concurrent_sessions,
-        )
-```
-
-After:
-```python
-        adapter = create_adapter(settings, config)
-        await adapter.start()
-        app.state.adapter = adapter
-```
-
-Also replace (line 79):
-```python
-        # session_id -> asyncio.Queue for SSE events
-        app.state.event_queues: dict[str, asyncio.Queue] = {}
-```
-
-With:
-```python
-        app.state.event_queues: dict[str, asyncio.Queue] = {}
-
-        pump_task = asyncio.create_task(event_pump(
-            adapter, app.state.event_queues, db,
-        ))
-```
-
-Update teardown (around line 91). Before:
-```python
-        lifecycle_task.cancel()
-```
-
-After:
-```python
-        pump_task.cancel()
-        lifecycle_task.cancel()
-        await adapter.stop()
-```
-
-**Step 4: Update create_session endpoint**
-
-Replace the `_run` background task in `create_session` (lines 330-374).
-
-Before:
-```python
-        runner: ClaudeRunner = request.app.state.runner
-        queues: dict[str, asyncio.Queue] = request.app.state.event_queues
-
-        session_id = uuid.uuid4().hex
-        name = body.name or f"{body.repo} session"
-        await db.create_session(session_id, name, body.repo)
-        await db.save_message(session_id, body.message)
-
-        # Create event queue and run Claude in background (non-blocking)
-        queue: asyncio.Queue = asyncio.Queue()
-        queues[session_id] = queue
-
-        async def _run():
-            try:
-                async for event in runner.run(
-                    message=body.message,
-                    repo_path=repo_path,
-                    model=repo_config.model,
-                    append_system_prompt=repo_config.append_system_prompt,
-                ):
-                    await queue.put(event)
-                    if (
-                        event.get("type") == "result"
-                        and event.get("session_id")
-                    ):
-                        await db.update_claude_session_id(
-                            session_id, event["session_id"]
-                        )
-            except FileNotFoundError as e:
-                await queue.put({
-                    "type": "error",
-                    "message": f"Repo not found: {e.filename or repo_path}",
-                })
-                await queue.put({"type": "status", "state": "waiting"})
-            except Exception as e:
-                logger.exception("Claude runner failed")
-                await queue.put({
-                    "type": "error",
-                    "message": str(e),
-                })
-                await queue.put({"type": "status", "state": "waiting"})
-
-        asyncio.create_task(_run())
-        return {"session_id": session_id, "name": name, "repo": body.repo}
-```
-
-After:
-```python
-        adapter = request.app.state.adapter
-        queues: dict[str, asyncio.Queue] = request.app.state.event_queues
-
-        session_id = uuid.uuid4().hex
-        name = body.name or f"{body.repo} session"
-        await db.create_session(session_id, name, body.repo)
-        await db.save_message(session_id, body.message)
-
-        queues[session_id] = asyncio.Queue()
-
-        await adapter.send(BackendMessage(
-            kind="new_session",
-            session_id=session_id,
-            payload={
-                "message": body.message,
-                "repo_path": repo_path,
-                "model": repo_config.model,
-                "append_system_prompt": repo_config.append_system_prompt,
-            },
-        ))
-
-        return {"session_id": session_id, "name": name, "repo": body.repo}
-```
-
-**Step 5: Update send_message endpoint**
-
-Replace the `_run` background task in `send_message` (lines 376-429).
-
-Before:
-```python
-        runner: ClaudeRunner = request.app.state.runner
-        queues: dict[str, asyncio.Queue] = request.app.state.event_queues
-        ...
-        claude_sid = session.get("claude_session_id")
-
-        async def _run():
-            try:
-                async for event in runner.run(
-                    message=body.message,
-                    repo_path=repo_path,
-                    resume_session_id=claude_sid,
-                    model=repo_config.model,
-                    append_system_prompt=repo_config.append_system_prompt,
-                ):
-                    await queue.put(event)
-                    ...
-            except Exception as e:
-                ...
-
-        asyncio.create_task(_run())
-        return {"status": "streaming", "session_id": session_id}
-```
-
-After:
-```python
-        adapter = request.app.state.adapter
-        queues: dict[str, asyncio.Queue] = request.app.state.event_queues
-
-        if session_id not in queues:
-            queues[session_id] = asyncio.Queue()
-
-        await db.touch(session_id)
-        await db.save_message(session_id, body.message)
-        await db.update_status(session_id, "active")
-
-        claude_sid = session.get("claude_session_id")
-
-        await adapter.send(BackendMessage(
-            kind="user_message",
-            session_id=session_id,
-            payload={
-                "message": body.message,
-                "repo_path": repo_path,
-                "resume_session_id": claude_sid,
-                "model": repo_config.model,
-                "append_system_prompt": repo_config.append_system_prompt,
-            },
-        ))
-
-        return {"status": "streaming", "session_id": session_id}
-```
-
-**Step 6: Update delete_session endpoint**
-
-Add cancel message. After `queues.pop(session_id, None)`:
-
-```python
-        adapter = request.app.state.adapter
-        await adapter.send(BackendMessage(kind="cancel", session_id=session_id))
-```
-
-**Step 7: Remove ClaudeRunner import**
-
-Remove from app.py imports:
-```python
-from warren.claude_runner import ClaudeRunner
-```
-
-**Step 8: Update test fixture**
-
-In `server/tests/test_app.py`, update the `client` fixture.
-
-Replace (line 50):
-```python
-    app.state.runner = ClaudeRunner(claude_bin="claude", max_concurrent=2)
-```
-
-With:
-```python
-    from warren.adapters.direct import DirectAdapter
-    adapter = DirectAdapter(claude_bin="claude", max_concurrent=2)
-    await adapter.start()
-    app.state.adapter = adapter
-```
-
-Remove the `ClaudeRunner` import from the fixture (line 14):
-```python
-    from warren.claude_runner import ClaudeRunner
-```
-
-**Step 9: Update TestRepoConfig test**
-
-Replace `test_create_session_passes_model_to_runner` (lines 383-401):
-
-Before:
-```python
-    async def test_create_session_passes_model_to_runner(self, authed_client):
-        captured = {}
-
-        async def mock_run(*args, **kwargs):
-            captured.update(kwargs)
-            yield {"type": "status", "state": "working"}
-            yield {"type": "result", "session_id": "mock-123", "summary": "Done"}
-            yield {"type": "status", "state": "waiting"}
-
-        with patch("warren.app.ClaudeRunner.run", side_effect=mock_run):
-            resp = await authed_client.post("/sessions", json={
-                "repo": "test-repo",
-                "message": "hello",
-            })
-            assert resp.status_code == 200
-
-        assert captured.get("model") == "opus"
-        assert "test system prompt" in captured.get("append_system_prompt", "")
-```
-
-After:
-```python
-    async def test_create_session_passes_config_to_adapter(self, authed_client):
-        captured_messages = []
-        original_send = authed_client.app.state.adapter.send
-
-        async def mock_send(message):
-            captured_messages.append(message)
-
-        with patch.object(authed_client.app.state.adapter, "send", side_effect=mock_send):
-            resp = await authed_client.post("/sessions", json={
-                "repo": "test-repo",
-                "message": "hello",
-            })
-            assert resp.status_code == 200
-
-        assert len(captured_messages) == 1
-        msg = captured_messages[0]
-        assert msg.payload["model"] == "opus"
-        assert "test system prompt" in msg.payload.get("append_system_prompt", "")
-```
-
-**Step 10: Run full test suite**
-
-Run: `cd server && uv run pytest tests/ -v`
-Expected: All tests pass
-
-**Step 11: Lint**
-
-Run: `cd server && uv run ruff check src/ tests/`
-Expected: No errors (fix any import ordering issues)
-
-**Step 12: Commit**
-
-```bash
-git add server/src/warren/app.py server/tests/test_app.py
-git commit -m "refactor: wire BackendAdapter into app.py, replace direct ClaudeRunner usage"
-```
-
----
-
-### Task 5: Add NanoClaw adapter stub
-
-**Files:**
-- Create: `server/src/warren/adapters/nanoclaw.py`
-- Modify: `server/tests/test_adapters.py`
-- Modify: `server/src/warren/app.py` (adapter factory)
-
-**Step 1: Write test**
-
-Append to `server/tests/test_adapters.py`:
-
-```python
-class TestNanoClawAdapter:
-    async def test_instantiation(self):
-        from warren.adapters.nanoclaw import NanoClawAdapter
-
-        adapter = NanoClawAdapter(ipc_dir="/tmp/ipc")
-        assert adapter._ipc_dir == "/tmp/ipc"
-
-    async def test_send_raises_not_implemented(self):
-        from warren.adapters import BackendMessage
-        from warren.adapters.nanoclaw import NanoClawAdapter
-
-        adapter = NanoClawAdapter(ipc_dir="/tmp/ipc")
-        with pytest.raises(NotImplementedError):
-            await adapter.send(BackendMessage(kind="new_session", session_id="x"))
-```
-
-**Step 2: Run tests to verify they fail**
-
-Run: `cd server && uv run pytest tests/test_adapters.py::TestNanoClawAdapter -v`
-Expected: FAIL — `ModuleNotFoundError`
-
-**Step 3: Write stub implementation**
-
-Create `server/src/warren/adapters/nanoclaw.py`:
-
-```python
-"""NanoClaw backend adapter -- file-based IPC with NanoClaw.
-
-Stub implementation. Full IPC integration requires studying NanoClaw's
-src/ipc.ts protocol and implementing file-based message passing.
-
-See: https://github.com/qwibitai/nanoclaw
-"""
-
-from __future__ import annotations
-
-from collections.abc import AsyncIterator
-
-from warren.adapters import BackendEvent, BackendMessage
-
-
-class NanoClawAdapter:
-    """Backend adapter that communicates with NanoClaw via file-based IPC."""
-
-    def __init__(self, ipc_dir: str):
-        self._ipc_dir = ipc_dir
-
-    async def start(self) -> None:
-        raise NotImplementedError("NanoClaw adapter not yet implemented")
-
-    async def stop(self) -> None:
-        raise NotImplementedError("NanoClaw adapter not yet implemented")
-
-    async def send(self, message: BackendMessage) -> None:
-        raise NotImplementedError("NanoClaw adapter not yet implemented")
-
-    async def events(self) -> AsyncIterator[BackendEvent]:
-        raise NotImplementedError("NanoClaw adapter not yet implemented")
-        yield  # pragma: no cover
-```
-
-**Step 4: Update adapter factory in app.py**
-
-In `server/src/warren/app.py`, update `create_adapter`:
-
-```python
-def create_adapter(settings: Settings, config):
-    """Create the backend adapter based on settings."""
-    if settings.backend == "direct":
-        backend_cfg = config.backend_config.get("direct", {})
-        return DirectAdapter(
-            claude_bin=backend_cfg.get("claude_bin", settings.claude_bin),
-            max_concurrent=backend_cfg.get("max_concurrent", config.max_concurrent_sessions),
-        )
-    if settings.backend == "nanoclaw":
-        from warren.adapters.nanoclaw import NanoClawAdapter
-        backend_cfg = config.backend_config.get("nanoclaw", {})
-        return NanoClawAdapter(ipc_dir=backend_cfg.get("ipc_dir", "data/ipc"))
-    raise ValueError(f"Unknown backend: {settings.backend}")
-```
-
-**Step 5: Run tests**
-
-Run: `cd server && uv run pytest tests/test_adapters.py -v`
-Expected: All adapter tests pass
-
-**Step 6: Commit**
-
-```bash
-git add server/src/warren/adapters/nanoclaw.py server/tests/test_adapters.py server/src/warren/app.py
-git commit -m "feat: add NanoClaw adapter stub"
-```
-
----
-
-## Phase 2: AI-Native Foundation (Skills)
-
-Skills have no code dependencies. They are documentation + scripts that guide
-Claude Code through complex operations on Warren itself.
-
-### Task 6: Create /setup skill
-
-**Files:**
-- Create: `.claude/skills/setup/SKILL.md`
-- Create: `.claude/skills/setup/scripts/01-check-env.sh`
-- Create: `.claude/skills/setup/scripts/02-install-deps.sh`
-- Create: `.claude/skills/setup/scripts/03-configure.sh`
-- Create: `.claude/skills/setup/scripts/04-verify.sh`
-
-**Step 1: Write SKILL.md**
-
-Create `.claude/skills/setup/SKILL.md`:
-
-```markdown
----
-name: setup
-description: Deploy Warren on a VPS with AI-guided setup. Checks environment, installs dependencies, configures services, and verifies end-to-end.
----
-
-# Warren Setup
-
-Deploy Warren to a Linux VPS. This skill guides you through the full process,
-fixing problems as they arise.
-
-**Principle:** Fix it. Don't tell the user to go fix it themselves — unless it
-genuinely requires their manual action (API keys, domain config).
-
-## Prerequisites
-
-Before starting, ask the user for:
-- Target VPS hostname or confirm we're running on the target machine
-- Whether they use Dokploy, systemd, or docker-compose for deployment
-- Their Anthropic API key (if not already set)
-
-## Step 1: Check Environment
-
-Run `bash .claude/skills/setup/scripts/01-check-env.sh` and parse the JSON
-output. For each component:
-
-- **ok**: Move on
-- **missing**: Install it (see Step 2)
-- **outdated**: Upgrade it
-
-Required components: Python 3.11+, uv, Node.js 22+, Docker, git
-
-## Step 2: Install Dependencies
-
-Run `bash .claude/skills/setup/scripts/02-install-deps.sh` for any missing
-components. If a package manager isn't available, install the appropriate one
-first.
-
-Then install Warren's Python dependencies:
-```
-cd server && uv sync
-```
-
-## Step 3: Configure
-
-Run `bash .claude/skills/setup/scripts/03-configure.sh` to check for existing
-configuration. Then:
-
-1. **config.yaml** — Create or update with repo paths. Ask the user which
-   repos they want Warren to manage.
-2. **.env** — Set `WARREN_ANTHROPIC_API_KEY`, `WARREN_ADMIN_TOKEN` (generate
-   if not set), `WARREN_BACKEND` (direct or nanoclaw).
-3. **Deployment** — Based on their deployment method:
-   - Dokploy: guide service creation, volume mounts, env vars
-   - systemd: create service unit file
-   - docker-compose: create compose.yaml
-
-## Step 4: Verify
-
-Run `bash .claude/skills/setup/scripts/04-verify.sh` and parse the JSON
-output. For each component:
-
-- **ok**: Report success
-- **error**: Diagnose and fix, then re-run verification
-
-Required checks: Warren starts, health endpoint responds, config loads, at
-least one repo is accessible.
-
-## Troubleshooting
-
-If any step fails:
-1. Read the error output carefully
-2. Check common causes (permissions, ports in use, missing env vars)
-3. Fix the issue
-4. Re-run the failing step
-5. If stuck after 2 attempts, ask the user for help
-```
-
-**Step 2: Write check scripts**
-
-Create `.claude/skills/setup/scripts/01-check-env.sh`:
-
-```bash
-#!/usr/bin/env bash
-set -euo pipefail
-
-check() {
-    local name="$1" cmd="$2" min_version="${3:-}"
-    if command -v "$cmd" &>/dev/null; then
-        version=$("$cmd" --version 2>&1 | head -1 || echo "unknown")
-        echo "{\"component\":\"$name\",\"status\":\"ok\",\"version\":\"$version\"}"
-    else
-        echo "{\"component\":\"$name\",\"status\":\"missing\"}"
-    fi
-}
-
-check "python" "python3" "3.11"
-check "uv" "uv"
-check "node" "node" "22"
-check "docker" "docker"
-check "git" "git"
-```
-
-Create `.claude/skills/setup/scripts/02-install-deps.sh`:
-
-```bash
-#!/usr/bin/env bash
-set -euo pipefail
-
-# Detect OS
-if [ -f /etc/os-release ]; then
-    . /etc/os-release
-    echo "{\"component\":\"os\",\"status\":\"ok\",\"distro\":\"$ID\",\"version\":\"$VERSION_ID\"}"
-else
-    echo "{\"component\":\"os\",\"status\":\"error\",\"detail\":\"Cannot detect OS\"}"
-    exit 1
-fi
-
-# Install uv if missing
-if ! command -v uv &>/dev/null; then
-    curl -LsSf https://astral.sh/uv/install.sh | sh
-    echo "{\"component\":\"uv\",\"status\":\"installed\"}"
-fi
-
-# Install project deps
-cd "$(dirname "$0")/../../.." || exit 1
-if [ -d server ]; then
-    cd server && uv sync
-    echo "{\"component\":\"warren_deps\",\"status\":\"ok\"}"
-fi
-```
-
-Create `.claude/skills/setup/scripts/03-configure.sh`:
-
-```bash
-#!/usr/bin/env bash
-set -euo pipefail
-
-PROJECT_ROOT="$(cd "$(dirname "$0")/../../.." && pwd)"
-
-# Check existing config
-if [ -f "$PROJECT_ROOT/server/config.yaml" ]; then
-    echo "{\"component\":\"config\",\"status\":\"ok\",\"path\":\"$PROJECT_ROOT/server/config.yaml\"}"
-else
-    echo "{\"component\":\"config\",\"status\":\"missing\",\"path\":\"$PROJECT_ROOT/server/config.yaml\"}"
-fi
-
-# Check .env
-if [ -f "$PROJECT_ROOT/server/.env" ]; then
-    echo "{\"component\":\"env\",\"status\":\"ok\"}"
-else
-    echo "{\"component\":\"env\",\"status\":\"missing\"}"
-fi
-```
-
-Create `.claude/skills/setup/scripts/04-verify.sh`:
-
-```bash
-#!/usr/bin/env bash
-set -euo pipefail
-
-# Check if Warren is running by hitting health endpoint
-WARREN_PORT="${WARREN_PORT:-4030}"
-
-if curl -sf "http://localhost:$WARREN_PORT/health" >/dev/null 2>&1; then
-    echo "{\"component\":\"warren\",\"status\":\"ok\"}"
-else
-    echo "{\"component\":\"warren\",\"status\":\"error\",\"detail\":\"Health endpoint not responding on port $WARREN_PORT\"}"
-fi
-```
-
-**Step 3: Make scripts executable and commit**
-
-```bash
-chmod +x .claude/skills/setup/scripts/*.sh
-git add .claude/skills/setup/
-git commit -m "feat: add /setup skill for AI-guided deployment"
-```
-
----
-
-### Task 7: Create /customize skill
-
-**Files:**
-- Create: `.claude/skills/customize/SKILL.md`
-
-**Step 1: Write SKILL.md**
-
-Create `.claude/skills/customize/SKILL.md`:
-
-```markdown
----
-name: customize
-description: Modify Warren to fit your needs. Add endpoints, change behavior, configure backends, extend the UI.
----
-
-# Customize Warren
-
-This skill handles any open-ended customization request. Understand the
-request, plan changes, implement, test.
-
-## Warren Architecture
-
-Before making changes, understand the codebase:
-
-- `server/src/warren/app.py` — FastAPI app factory, all endpoints, SSE streaming
-- `server/src/warren/adapters/` — Backend adapter interface and implementations
-  - `__init__.py` — BackendAdapter protocol, BackendMessage, BackendEvent types
-  - `direct.py` — DirectAdapter (bare Claude CLI via subprocess)
-  - `nanoclaw.py` — NanoClawAdapter stub (file-based IPC)
-- `server/src/warren/claude_runner.py` — Claude CLI subprocess management
-- `server/src/warren/config.py` — pydantic-settings (env) + YAML config
-- `server/src/warren/db.py` — SQLite via aiosqlite
-- `server/src/warren/lifecycle.py` — Background session lifecycle sweep
-- `server/src/warren/auth.py` — QR pairing, device tokens
-- `creation/` — R1 WebView UI (vanilla HTML/CSS/JS, 240x282 viewport)
-
-## Process
-
-1. **Understand** — Read the relevant source files. Check CLAUDE.md for
-   conventions and extension points.
-2. **Plan** — Describe what you'll change and why. List files to create/modify.
-3. **Test first** — Write a failing test for the new behavior.
-4. **Implement** — Write minimal code to make the test pass.
-5. **Verify** — Run `cd server && uv run pytest tests/ -v` and
-   `cd server && uv run ruff check src/ tests/`.
-6. **Commit** — Small, focused commits.
-
-## Common Customizations
-
-### Add an API endpoint
-1. Add route in `app.py` inside `create_app()`
-2. Add test in `tests/test_app.py`
-3. If it needs auth, use `Depends(require_auth)` or `Depends(require_admin)`
-
-### Add a backend adapter
-1. Create `server/src/warren/adapters/{name}.py` implementing `BackendAdapter`
-2. Add tests in `tests/test_adapters.py`
-3. Register in `create_adapter()` in `app.py`
-4. Add config section to `backend_config` in config.yaml
-
-### Modify R1 UI
-1. Edit files in `creation/` (vanilla HTML/CSS/JS)
-2. Viewport is 240x282 pixels — design for this exact size
-3. Test in browser at that resolution
-
-### Change config
-1. Add pydantic model fields in `config.py`
-2. Add tests in `tests/test_config.py`
-3. Update CLAUDE.md if it's an extension point
-```
-
-**Step 2: Commit**
-
-```bash
-git add .claude/skills/customize/
-git commit -m "feat: add /customize skill for AI-driven extensibility"
-```
-
----
-
-### Task 8: Create /debug skill
-
-**Files:**
-- Create: `.claude/skills/debug/SKILL.md`
-
-**Step 1: Write SKILL.md**
-
-Create `.claude/skills/debug/SKILL.md`:
-
-```markdown
----
-name: debug
-description: Diagnose and fix Warren issues. Adapter failures, session problems, networking, configuration.
----
-
-# Debug Warren
-
-Systematic troubleshooting for Warren. Run diagnostics, parse output, fix
-issues automatically.
-
-**Principle:** Fix it. Don't tell the user to go fix it themselves.
-
-## Diagnostic Steps
-
-### 1. Check service health
-```bash
-curl -sf http://localhost:${WARREN_PORT:-4030}/health
-```
-If this fails, Warren isn't running. Check:
-- Process running: `pgrep -f "warren"` or `docker ps`
-- Port in use: `ss -tlnp | grep ${WARREN_PORT:-4030}`
-- Logs: `journalctl -u warren --since "5 minutes ago"` or docker logs
-
-### 2. Check configuration
-```bash
-cd server && uv run python -c "
-from warren.config import Settings, load_config
-s = Settings()
-print(f'Backend: {s.backend}')
-print(f'DB: {s.db}')
-print(f'Config: {s.config}')
-c = load_config(s.config)
-print(f'Repos: {list(c.repos.keys())}')
-print(f'Backend config: {c.backend_config}')
-"
-```
-
-### 3. Check adapter connectivity
-For direct backend:
-- Verify `claude` binary exists: `which claude`
-- Test Claude CLI: `claude --version`
-- Check API key: verify `ANTHROPIC_API_KEY` is set
-
-For NanoClaw backend:
-- Check IPC directory exists and is writable
-- Check NanoClaw process is running
-- Check Docker socket is accessible
-
-### 4. Check database
-```bash
-cd server && uv run python -c "
-import asyncio, aiosqlite
-async def check():
-    db = await aiosqlite.connect('warren.db')
-    tables = await (await db.execute(
-        \"SELECT name FROM sqlite_master WHERE type='table'\"
-    )).fetchall()
-    print(f'Tables: {[t[0] for t in tables]}')
-    sessions = await (await db.execute(
-        'SELECT count(*), status FROM sessions GROUP BY status'
-    )).fetchall()
-    print(f'Sessions: {dict(sessions)}')
-    await db.close()
-asyncio.run(check())
-"
-```
-
-### 5. Check session streaming
-If SSE isn't working:
-- Verify event queue exists for the session
-- Check adapter is producing events
-- Test with curl: `curl -N -H "Authorization: Bearer TOKEN" http://localhost:4030/sessions/SID/stream`
-
-## Common Issues
-
-| Symptom | Likely Cause | Fix |
-|---------|-------------|-----|
-| 401 on all endpoints | Invalid device token | Re-pair via /admin |
-| Session stays "working" | Claude subprocess hung | Check adapter tasks, kill stale processes |
-| No SSE events | Event pump not running | Check logs for pump errors, restart |
-| Config not loading | YAML syntax error | Validate with `python -c "import yaml; yaml.safe_load(open('config.yaml'))"` |
-| Port already in use | Another instance running | `kill $(lsof -t -i:4030)` or change port |
-```
-
-**Step 2: Commit**
-
-```bash
-git add .claude/skills/debug/
-git commit -m "feat: add /debug skill for AI-guided troubleshooting"
-```
-
----
-
-### Task 9: Expand CLAUDE.md with extension points
-
-**Files:**
-- Modify: `CLAUDE.md`
-
-**Step 1: Add extension points section**
-
-Append to `CLAUDE.md`:
-
-```markdown
-
-## Extension Points
-
-### Add a backend adapter
-Create `server/src/warren/adapters/{name}.py` implementing the `BackendAdapter`
-protocol from `warren.adapters`. Register it in `create_adapter()` in `app.py`.
-Add backend-specific config under `backend_config` in config.yaml.
-
-### Add an API endpoint
-Add a route inside `create_app()` in `app.py`. Use `Depends(require_auth)` for
-device-token auth or `Depends(require_admin)` for admin-only access.
-
-### Add a skill
-Create `.claude/skills/{name}/SKILL.md`. The skill is automatically available
-to Claude Code when working in the Warren repo.
-
-### Key abstractions
-- `BackendAdapter` (`server/src/warren/adapters/__init__.py`) — protocol for AI backends
-- `BackendMessage` / `BackendEvent` — message types for adapter communication
-- `DirectAdapter` (`server/src/warren/adapters/direct.py`) — bare Claude CLI adapter
-- `Database` (`server/src/warren/db.py`) — SQLite via aiosqlite
-- `WarrenConfig` / `Settings` (`server/src/warren/config.py`) — YAML + env config
-```
-
-**Step 2: Commit**
-
-```bash
-git add CLAUDE.md
-git commit -m "docs: add extension points to CLAUDE.md"
-```
-
----
-
-## Dependency Graph
-
-```
-Task 1 (Adapter types)
-  |
-Task 2 (DirectAdapter)
-  |
-Task 3 (Config)
-  |
-Task 4 (Wire into app.py) ---- Tasks 6-9 can run in parallel
-  |                              (no code dependencies)
-Task 5 (NanoClaw stub)
-
-Tasks 6-9 (Skills + CLAUDE.md) -- independent of Phase 1
-```
-
-## Notes
-
-- **Phase 2 (skills) has no code dependencies** — can start immediately in
-  parallel with Phase 1.
-- `WARREN_BACKEND=direct` is default — existing deploys unaffected.
-- `ClaudeRunner` and `parse_stream_line` stay unchanged — DirectAdapter wraps them.
-- All existing tests must pass after every task.
-- The NanoClaw adapter is a stub. Full IPC implementation is a separate
-  future plan once NanoClaw is deployed alongside Warren.
diff --git a/docs/plans/2026-02-21-agent-progress-stream-design.md b/docs/plans/2026-02-21-agent-progress-stream-design.md
deleted file mode 100644
index 4c01e1d..0000000
--- a/docs/plans/2026-02-21-agent-progress-stream-design.md
+++ /dev/null
@@ -1,113 +0,0 @@
-# Agent Progress Stream & Auto-Memory Design
-
-## Goal
-
-Give the user real-time visibility into what the agent is doing (tool activity stream) and let the agent build persistent knowledge over time (auto-memory).
-
-## Problem
-
-1. **Silent work**: When the agent reads files, runs commands, or searches code, the user sees nothing until the final response arrives. On long tasks this feels like the agent is hung.
-2. **No memory**: Each session starts from scratch. The agent doesn't learn user preferences or project patterns over time.
-
-## Design
-
-### Tool Activity Stream
-
-**Approach:** Extend the existing result pipeline to also forward tool_use events. Every layer already exists — we widen what flows through it.
-
-**Data flow:**
-
-```
-SDK query() yields assistant message with tool_use content blocks
-  → agent-runner extracts tool name + input, formats compact summary
-  → writeOutput({ type: "progress", tool, summary }) via stdout markers
-  → container-runner parses markers in real-time, calls onOutput()
-  → NanoClaw onOutput dispatches on type: progress → channel.sendProgress()
-  → WarrenChannel POSTs { session_id, type: "tool", tool, summary }
-  → Warren callback puts { type: "tool", tool, summary } on SSE queue
-  → R1 UI handleEvent case "tool": renders compact one-liner
-```
-
-**Agent-runner changes:**
-
-The message loop adds handling for `assistant` messages containing `tool_use` blocks:
-
-```typescript
-if (message.type === 'assistant') {
-  for (const block of message.content) {
-    if (block.type === 'tool_use') {
-      const summary = summarizeTool(block.name, block.input);
-      writeOutput({ type: 'progress', tool: block.name, summary });
-    }
-  }
-}
-```
-
-Tool summary extraction per tool:
-- `Read` → file path
-- `Bash` → first ~40 chars of command
-- `Write`/`Edit` → file path
-- `Grep`/`Glob` → pattern
-- `WebSearch`/`WebFetch` → query/URL
-- Other → just the tool name
-
-**Output protocol change:**
-
-Currently `ContainerOutput` is `{ status, result, newSessionId }`. We introduce a discriminated union:
-
-```typescript
-type ContainerEvent =
-  | { type: 'result'; status: 'success' | 'error'; result: string | null; newSessionId?: string }
-  | { type: 'progress'; tool: string; summary: string };
-```
-
-The container-runner dispatches based on `type`. Progress events call a new callback; result events use the existing path.
-
-**NanoClaw changes:**
-
-The `onOutput` handler dispatches on event type:
-- `type: 'progress'` → `channel.sendProgress(jid, tool, summary)`
-- `type: 'result'` → existing `channel.sendMessage(jid, text)` path
-
-`WarrenChannel` gets `sendProgress()` that POSTs to the same callback URL:
-```json
-{ "session_id": "abc", "type": "tool", "tool": "Read", "summary": "server/app.py" }
-```
-
-**Warren changes:**
-
-The callback handler adds a case for `type: "tool"`:
-```python
-elif event_type == "tool":
-    await queue.put({"type": "tool", "tool": body.get("tool"), "summary": body.get("summary")})
-```
-
-**R1 UI changes:**
-
-Replace the tool-buffering approach with immediate one-liners:
-
-```
-Reading server/app.py
-Running npm test
-Searching for "handleEvent"
-```
-
-Styled as `msg-tool` — dimmer than assistant text, monospace, single line with `text-overflow: ellipsis` for the 240px R1 screen. No debounce, no batching. They scroll by like a build log.
-
-### Auto-Memory
-
-The agent container already has the prerequisites:
-- `CLAUDE_CODE_DISABLE_AUTO_MEMORY: '0'` in settings.json
-- Persistent `/home/node/.claude` mounted from the data volume
-- `settingSources: ['project', 'user']` in SDK query options
-
-The SDK should write to `~/.claude/projects/<path>/memory/MEMORY.md` and it persists across sessions via the volume mount. This is a verification task — confirm it works, fix path issues if not. No architectural changes needed.
-
-## What This Does NOT Include
-
-- Streaming text chunks (partial responses as they're written)
-- Token usage or context size display
-- Progress percentages or time estimates
-- Claude's internal thinking/reasoning steps
-
-These could be added later using the same pipeline, but are out of scope for this design.
diff --git a/docs/plans/2026-02-21-agent-progress-stream-plan.md b/docs/plans/2026-02-21-agent-progress-stream-plan.md
deleted file mode 100644
index 6e5e9ec..0000000
--- a/docs/plans/2026-02-21-agent-progress-stream-plan.md
+++ /dev/null
@@ -1,450 +0,0 @@
-# Agent Progress Stream Implementation Plan
-
-> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
-
-**Goal:** Stream tool activity from agent containers to the R1 UI as compact one-liners so the user sees real-time progress instead of silence.
-
-**Architecture:** Extend the existing stdout-markers → container-runner → callback → SSE pipeline. The agent-runner emits progress events for each tool_use block. NanoClaw forwards them via the Warren callback. Warren puts them on the SSE queue. The R1 UI renders them as compact one-liners.
-
-**Tech Stack:** TypeScript (agent-runner, NanoClaw), Python/FastAPI (Warren), vanilla JS (R1 UI)
-
----
-
-### Task 1: Add progress output to agent-runner
-
-The agent-runner message loop currently only emits `result` messages via `writeOutput()`. We add progress events for `tool_use` content blocks.
-
-**Files:**
-- Modify: `nanoclaw/container/agent-runner/src/index.ts:33-38` (ContainerOutput type)
-- Modify: `nanoclaw/container/agent-runner/src/index.ts:111-115` (writeOutput function)
-- Modify: `nanoclaw/container/agent-runner/src/index.ts:457-486` (message loop)
-
-**Step 1: Extend the output types**
-
-Change the `ContainerOutput` interface and add a `ProgressEvent` type. Update `writeOutput` to accept both:
-
-```typescript
-// Replace the existing ContainerOutput (line 33-38) with:
-interface ContainerResult {
-  type: 'result';
-  status: 'success' | 'error';
-  result: string | null;
-  newSessionId?: string;
-  error?: string;
-}
-
-interface ContainerProgress {
-  type: 'progress';
-  tool: string;
-  summary: string;
-}
-
-type ContainerOutput = ContainerResult | ContainerProgress;
-```
-
-Update `writeOutput` — it already works since it just `JSON.stringify`s, but update the existing result writes to include `type: 'result'`:
-
-```typescript
-// In the result handler (line 480-484), change to:
-writeOutput({
-  type: 'result',
-  status: 'success',
-  result: textResult || null,
-  newSessionId
-});
-```
-
-Also update the error output at the bottom of `main()` (search for `writeOutput` with `status: 'error'`):
-
-```typescript
-writeOutput({
-  type: 'result',
-  status: 'error',
-  result: null,
-  error: `...`
-});
-```
-
-**Step 2: Add tool summary helper**
-
-Add a `summarizeTool` function before the message loop:
-
-```typescript
-function summarizeTool(toolName: string, input: Record<string, unknown>): string {
-  switch (toolName) {
-    case 'Read':
-      return String(input.file_path || '').split('/').slice(-2).join('/');
-    case 'Write':
-    case 'Edit':
-      return String(input.file_path || '').split('/').slice(-2).join('/');
-    case 'Bash':
-      return ('$ ' + String(input.command || '')).slice(0, 60);
-    case 'Grep':
-      return String(input.pattern || '');
-    case 'Glob':
-      return String(input.pattern || '');
-    case 'WebSearch':
-      return String(input.query || '');
-    case 'WebFetch':
-      return String(input.url || '').slice(0, 60);
-    default:
-      return toolName;
-  }
-}
-```
-
-**Step 3: Emit progress events in the message loop**
-
-In the `for await` loop (after line 464, inside the `message.type === 'assistant'` block), add tool_use extraction:
-
-```typescript
-if (message.type === 'assistant' && 'uuid' in message) {
-  lastAssistantUuid = (message as { uuid: string }).uuid;
-  // Emit progress events for tool_use blocks
-  const content = (message as { content?: Array<{ type: string; name?: string; input?: Record<string, unknown> }> }).content;
-  if (content) {
-    for (const block of content) {
-      if (block.type === 'tool_use' && block.name) {
-        writeOutput({
-          type: 'progress',
-          tool: block.name,
-          summary: summarizeTool(block.name, block.input || {}),
-        });
-      }
-    }
-  }
-}
-```
-
-**Step 4: Build and verify**
-
-Run: `cd nanoclaw && npm run build`
-Expected: Clean compile, no errors.
-
-**Step 5: Commit**
-
-```bash
-cd nanoclaw
-git add container/agent-runner/src/index.ts
-git commit -m "feat: emit tool progress events from agent-runner"
-```
-
----
-
-### Task 2: Handle progress events in container-runner and NanoClaw
-
-The container-runner parses stdout markers and calls `onOutput`. Currently it types the parsed JSON as `ContainerOutput` (the old shape). We need to handle the new discriminated union. The NanoClaw `onOutput` handler in `index.ts` needs to dispatch progress events to the channel.
-
-**Files:**
-- Modify: `nanoclaw/src/container-runner.ts:83-99` (ContainerOutput type)
-- Modify: `nanoclaw/src/container-runner.ts:373-383` (streaming output parsing)
-- Modify: `nanoclaw/src/index.ts:189-207` (onOutput handler)
-- Modify: `nanoclaw/src/channels/warren.ts:41-51` (add sendProgress)
-- Modify: `nanoclaw/src/types.ts:81-90` (Channel interface)
-
-**Step 1: Update ContainerOutput type in container-runner.ts**
-
-Replace the existing `ContainerOutput` interface (around line 93):
-
-```typescript
-export interface ContainerResult {
-  type: 'result';
-  status: 'success' | 'error';
-  result: string | null;
-  newSessionId?: string;
-  error?: string;
-}
-
-export interface ContainerProgress {
-  type: 'progress';
-  tool: string;
-  summary: string;
-}
-
-export type ContainerOutput = ContainerResult | ContainerProgress;
-```
-
-Update the `runContainerAgent` return type and final output parsing to handle both types. In the streaming parser (line 374), `parsed` is already generic JSON — it will have `type: 'progress'` or `type: 'result'`. The `newSessionId` extraction (line 375-377) should only run for result events:
-
-```typescript
-const parsed: ContainerOutput = JSON.parse(jsonStr);
-if (parsed.type === 'result' && parsed.newSessionId) {
-  newSessionId = parsed.newSessionId;
-}
-```
-
-Note: if `type` is missing (backwards compat with old agent images), treat as result.
-
-**Step 2: Add sendProgress to Channel interface and WarrenChannel**
-
-In `nanoclaw/src/types.ts`, add to the Channel interface:
-
-```typescript
-export interface Channel {
-  // ... existing methods ...
-  sendProgress?(jid: string, tool: string, summary: string): Promise<void>;
-}
-```
-
-In `nanoclaw/src/channels/warren.ts`, add the method:
-
-```typescript
-async sendProgress(jid: string, tool: string, summary: string): Promise<void> {
-  const sessionId = jid.replace('warren:', '');
-  const resp = await fetch(this.callbackUrl, {
-    method: 'POST',
-    headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({ session_id: sessionId, type: 'tool', tool, summary }),
-  });
-  if (!resp.ok) {
-    throw new Error(`Warren callback failed: ${resp.status}`);
-  }
-}
-```
-
-**Step 3: Dispatch progress events in NanoClaw's onOutput handler**
-
-In `nanoclaw/src/index.ts`, update the `onOutput` callback (line 189-207):
-
-```typescript
-const output = await runAgent(group, prompt, chatJid, async (result) => {
-  if (result.type === 'progress') {
-    // Forward tool activity to the user in real-time
-    await channel.sendProgress?.(chatJid, result.tool, result.summary);
-    return;
-  }
-  // Existing result handling (unchanged)
-  if (result.result) {
-    const raw = typeof result.result === 'string' ? result.result : JSON.stringify(result.result);
-    const text = raw.replace(/<internal>[\s\S]*?<\/internal>/g, '').trim();
-    logger.info({ group: group.name }, `Agent output: ${raw.slice(0, 200)}`);
-    if (text) {
-      await channel.sendMessage(chatJid, text);
-      outputSentToUser = true;
-    }
-    resetIdleTimer();
-  }
-
-  if (result.status === 'error') {
-    hadError = true;
-  }
-});
-```
-
-Note: We need to handle `result.status` and `result.result` only for result-type events. Since the old `ContainerOutput` didn't have `type`, access `result.status` only when `result.type !== 'progress'` (or `result.type === 'result'` or `type` is undefined for backwards compat).
-
-**Step 4: Build and verify**
-
-Run: `cd nanoclaw && npm run build`
-Expected: Clean compile.
-
-**Step 5: Commit**
-
-```bash
-cd nanoclaw
-git add src/container-runner.ts src/index.ts src/channels/warren.ts src/types.ts
-git commit -m "feat: forward tool progress events through NanoClaw pipeline"
-```
-
----
-
-### Task 3: Handle tool events in Warren's callback endpoint
-
-Warren's callback handler needs to recognize `type: "tool"` and put it on the SSE queue.
-
-**Files:**
-- Modify: `warren/server/src/warren/app.py:620-638` (nanoclaw_callback)
-
-**Step 1: Add tool event handling**
-
-In the `nanoclaw_callback` function, add a case for tool events:
-
-```python
-@app.post("/internal/nanoclaw/callback")
-async def nanoclaw_callback(request: Request):
-    body = await request.json()
-    session_id = body.get("session_id")
-    text = body.get("text", "")
-    event_type = body.get("type", "text")
-
-    queues: dict[str, asyncio.Queue] = request.app.state.event_queues
-    queue = queues.get(session_id)
-    if queue:
-        if event_type == "result":
-            await queue.put({
-                "type": "result",
-                "summary": text,
-                "session_id": body.get("nanoclaw_session_id", ""),
-            })
-        elif event_type == "tool":
-            await queue.put({
-                "type": "tool",
-                "tool": body.get("tool", ""),
-                "summary": body.get("summary", ""),
-            })
-        else:
-            await queue.put({"type": "text", "content": text})
-    return {"status": "ok"}
-```
-
-**Step 2: Verify lint passes**
-
-Run: `cd server && uv run ruff check src/`
-Expected: No errors.
-
-**Step 3: Run tests**
-
-Run: `cd server && uv run pytest tests/ -v`
-Expected: All existing tests pass.
-
-**Step 4: Commit**
-
-```bash
-cd warren
-git add server/src/warren/app.py
-git commit -m "feat: handle tool progress events in nanoclaw callback"
-```
-
----
-
-### Task 4: Update R1 UI tool rendering
-
-The R1 UI already has a `case "tool"` handler, but it buffers tools and shows summaries like "3 tools: Read, Bash, Grep". We change it to render immediate compact one-liners using the `summary` field from the new event format.
-
-**Files:**
-- Modify: `warren/creation/js/app.js:264-279` (handleEvent tool case)
-
-**Step 1: Update the tool event handler**
-
-Replace the existing `case "tool"` block in `handleEvent`:
-
-```javascript
-case "tool": {
-  // Show compact one-liner for each tool use
-  const summary = event.summary
-    ? event.tool + " " + event.summary
-    : event.tool || event.action || "working";
-  appendRawMsg("msg-tool", summary);
-  break;
-}
-```
-
-This removes the debounce buffer — each tool event renders immediately. The existing `msg-tool` CSS already handles the styling (green, 10px, `[TOOL]` prefix).
-
-Also remove the now-unused `toolBuffer`, `toolFlushTimer`, and `flushToolBuffer` references. Search for all calls to `flushToolBuffer()` in other cases (`text`, `result`, `status`, `error`) and remove them since there's no longer a buffer to flush.
-
-Actually, keep `flushToolBuffer` as a no-op for safety — the DirectAdapter path (non-nanoclaw) might still use the old format:
-
-```javascript
-let toolBuffer = [];
-let toolFlushTimer;
-function flushToolBuffer() {
-  if (toolBuffer.length === 0) return;
-  if (toolBuffer.length === 1) {
-    appendRawMsg("msg-tool", toolBuffer[0]);
-  } else {
-    // ... existing group rendering ...
-  }
-  toolBuffer = [];
-}
-```
-
-Keep `flushToolBuffer` calls in other cases. The tool case itself just renders directly when `event.summary` exists (NanoClaw path), or falls back to the old buffer behavior otherwise:
-
-```javascript
-case "tool": {
-  if (event.summary !== undefined) {
-    // NanoClaw path: immediate one-liner
-    flushToolBuffer();
-    const label = event.tool + " " + event.summary;
-    appendRawMsg("msg-tool", label);
-  } else {
-    // Direct adapter path: buffer and batch
-    const label = event.file
-      ? event.action + " " + event.file
-      : event.cmd
-      ? "$ " + event.cmd
-      : event.action;
-    toolBuffer.push(label);
-    clearTimeout(toolFlushTimer);
-    toolFlushTimer = setTimeout(flushToolBuffer, 300);
-  }
-  break;
-}
-```
-
-**Step 2: Commit**
-
-```bash
-cd warren
-git add creation/js/app.js
-git commit -m "feat: render tool progress as immediate one-liners in R1 UI"
-```
-
----
-
-### Task 5: Verify auto-memory
-
-Verify that Claude Code's auto-memory works inside agent containers. The settings are already configured (`CLAUDE_CODE_DISABLE_AUTO_MEMORY: '0'`), the `.claude` directory persists on the volume, and `settingSources: ['project', 'user']` is set.
-
-**Files:**
-- Check: `nanoclaw/src/container-runner.ts:124-140` (settings.json written to .claude)
-- Check: Volume mount at `/home/node/.claude`
-
-**Step 1: Test auto-memory**
-
-Start a Warren session and ask the agent: "Remember that I prefer TypeScript over JavaScript for new code." End the session. Start a new session and ask: "What are my preferences?" If the agent remembers, auto-memory works.
-
-**Step 2: If auto-memory doesn't work, check paths**
-
-The SDK writes memory to `~/.claude/projects/<cwd-path>/memory/MEMORY.md`. Inside the container, cwd is `/workspace/group`, so the memory file would be at `/home/node/.claude/projects/-workspace-group/memory/MEMORY.md`. Verify this file exists on the volume after a session.
-
-If the file doesn't exist, the `query()` SDK might not support auto-memory. In that case, we'd need to investigate the SDK docs or add a manual memory mechanism. Document findings either way.
-
-**Step 3: Commit (if changes needed)**
-
-```bash
-cd nanoclaw
-git add <any changed files>
-git commit -m "fix: ensure auto-memory works in agent containers"
-```
-
----
-
-### Task 6: Integration test — end to end
-
-Deploy and verify the full pipeline works.
-
-**Step 1: Push both repos**
-
-```bash
-cd nanoclaw && git push origin main
-cd warren && git -C nanoclaw pull origin main && git add nanoclaw && git commit -m "chore: update nanoclaw submodule" && git push origin main
-```
-
-**Step 2: Redeploy via Dokploy**
-
-Trigger a redeploy. Watch the nanoclaw logs for the `DinD volume mapping detected` message to confirm startup.
-
-**Step 3: Test tool activity stream**
-
-Start a new Warren session. Send a prompt that triggers multiple tools, e.g.: "Read the README.md in this workspace and summarize it."
-
-Expected in R1 UI:
-```
-[TOOL] Read README.md
-[assistant response text]
-```
-
-**Step 4: Test rapid tool sequences**
-
-Send: "Search for all TypeScript files and count the lines in each."
-
-Expected: Multiple `[TOOL]` one-liners scroll by in real time, followed by the result.
-
-**Step 5: Verify no regressions**
-
-- Text messages still render correctly
-- Results still show with `msg-result` styling
-- Status dot still updates (working → waiting)
-- Sessions still persist and can be reopened
diff --git a/docs/plans/2026-02-21-r1-workflow-engine-design.md b/docs/plans/2026-02-21-r1-workflow-engine-design.md
deleted file mode 100644
index 0e3d692..0000000
--- a/docs/plans/2026-02-21-r1-workflow-engine-design.md
+++ /dev/null
@@ -1,148 +0,0 @@
-# R1 Workflow Engine Design
-
-## Problem
-
-Using the Rabbit R1 for development work is freeform — every interaction is an
-unstructured chat session. There's no way to trigger predefined multi-step
-workflows, no visibility into step-by-step progress on the 240x282 screen, and
-no structured failure handling. The R1 should be a control surface for
-steering autonomous dev work, not a code editor.
-
-## Goals
-
-- One-tap multi-step workflows from the R1 command palette (deploy, PR review,
-  run tests)
-- Real-time step progress visible on the small screen
-- Agent-planned workflows get the same structured UI as templates
-- Auto-continue by default; pause only on failure
-- No new backend infrastructure — leverage existing session/message/SSE flow
-
-## Design
-
-### Workflow Templates
-
-Templates live in Warren's `config.yaml` under a new `workflows` key:
-
-```yaml
-workflows:
-  deploy:
-    label: "Deploy"
-    steps:
-      - "Pull latest from main"
-      - "Run the test suite"
-      - "If tests pass, push to production"
-      - "Report deploy status"
-
-  pr-review:
-    label: "Review PR"
-    input: "pr_url"
-    steps:
-      - "Fetch the PR diff and description"
-      - "Review for bugs, security issues, and style"
-      - "Post a summary of findings"
-
-  run-tests:
-    label: "Run Tests"
-    steps:
-      - "Run the full test suite"
-      - "Report pass/fail counts and any failures"
-```
-
-Each step is a natural language instruction the agent interprets and executes.
-The `input` field means the R1 prompts for a value before starting.
-
-When triggered, Warren composes a single prompt from the steps and sends it to
-the agent with a preamble instructing it to report progress after each step and
-stop on failure.
-
-Templates are deliberately simple — no DAGs, no conditionals. The agent handles
-sequencing and error decisions. A template is a structured prompt.
-
-### Step Progress Protocol
-
-The agent reports progress using a text convention in `send_message`:
-
-```
-[STEP:1/4:done] Created auth module
-[STEP:2/4:running] Writing tests
-[STEP:3/4:pending] Create PR
-[STEP:4/4:pending] Run CI
-```
-
-Statuses: `pending`, `running`, `done`, `failed`.
-
-This is a convention, not a new IPC type. The R1 UI parses these from regular
-chat messages. Messages without the prefix render as normal chat. The agent
-learns this convention from its system prompt / CLAUDE.md.
-
-Agent-planned workflows use the same protocol. When the agent generates a plan
-(via brainstorming/writing-plans skills), it reports steps in the same format,
-giving the R1 identical rendering for both template and freeform workflows.
-
-### R1 UI Changes
-
-**Command palette** gains a Workflows category alongside existing Quick Actions
-and Templates. Tapping a workflow starts a session with the composed prompt.
-
-**Active workflow view** — chat view gets a compact step progress header:
-
-```
-+--------------------------+
-| <- DEPLOY         [STOP] |
-| +----------------------+ |
-| | * Pull latest        | |
-| | > Running tests...   | |
-| | o Push to prod       | |
-| | o Report status      | |
-| +----------------------+ |
-|                          |
-| Running test suite...    |
-| 142 passed, 0 failed    |
-|                          |
-| > _               [CMD] |
-+--------------------------+
-```
-
-Step list is compact, fitting the 240x282 screen. Agent messages appear below.
-The user can type to intervene at any point.
-
-No polling needed — the existing SSE stream pushes events in real-time.
-
-### Failure Handling
-
-When a step fails, the agent sends `[STEP:2/4:failed] Tests failed: 3 errors`
-and stops. The R1 shows the failure in red. The user can:
-
-- **Type a response** — "skip tests and deploy anyway" or "fix the failing tests"
-- **Tap [STOP]** — abort the workflow
-
-The agent resumes based on the user's response. No retry machinery — it's a
-conversation with structure.
-
-## Changes by Component
-
-| Component | Change |
-|---|---|
-| `config.yaml` / `config.py` | New `workflows` section with named templates |
-| `app.py` | Workflows exposed via existing internal config endpoints |
-| `creation/js/app.js` | Parse `[STEP:...]` messages, render step list |
-| `creation/index.html` | Workflow progress header in chat view |
-| `creation/css/styles.css` | Step list styling (done/running/pending/failed) |
-| Agent CLAUDE.md | Step reporting convention instructions |
-| NanoClaw container skills | Optional: workflow execution skill |
-| Command palette | New "Workflows" category from config |
-
-## What Doesn't Change
-
-- No new backend endpoints — workflows are structured prompts via existing
-  session/message flow
-- No new IPC types — step progress goes through existing `send_message`
-- No new MCP tools — agent already has everything it needs
-- SSE streaming already works — no polling
-- NanoClaw container-runner, IPC watcher unchanged
-
-## Key Insight
-
-A workflow is a well-structured prompt plus a UI convention for rendering
-progress. The agent is already capable of multi-step execution. This design
-adds structure and visibility, not new capabilities.
diff --git a/docs/plans/2026-02-21-r1-workflow-engine-plan.md b/docs/plans/2026-02-21-r1-workflow-engine-plan.md
deleted file mode 100644
index 93a69a5..0000000
--- a/docs/plans/2026-02-21-r1-workflow-engine-plan.md
+++ /dev/null
@@ -1,676 +0,0 @@
-# R1 Workflow Engine Implementation Plan
-
-> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
-
-**Goal:** Add structured multi-step workflows to the R1 Creation UI — template-based or agent-planned — with real-time step progress rendering on the 240x282 screen.
-
-**Architecture:** Workflows are structured prompts composed from template steps in `config.yaml`. Step progress is a text convention (`[STEP:n/total:status]`) embedded in regular `send_message` calls, parsed client-side. No new backend infrastructure, IPC types, or MCP tools needed.
-
-**Tech Stack:** Python (FastAPI/Pydantic), vanilla JS (R1 Creation), CSS, YAML config
-
----
-
-### Task 1: Add Workflow Config Model
-
-**Files:**
-- Modify: `server/src/warren/config.py:29-43`
-- Test: `server/tests/test_config.py` (create)
-
-**Step 1: Write the test**
-
-Create `server/tests/test_config.py`:
-
-```python
-import pytest
-from warren.config import WarrenConfig, WorkflowConfig
-
-
-def test_workflow_config_from_yaml():
-    data = {
-        "workflows": {
-            "deploy": {
-                "label": "Deploy",
-                "steps": [
-                    "Pull latest from main",
-                    "Run the test suite",
-                    "Push to production",
-                ],
-            },
-            "pr-review": {
-                "label": "Review PR",
-                "input": "pr_url",
-                "steps": [
-                    "Fetch the PR diff",
-                    "Review for bugs and style",
-                ],
-            },
-        }
-    }
-    config = WarrenConfig(**data)
-    assert len(config.workflows) == 2
-    assert config.workflows["deploy"].label == "Deploy"
-    assert len(config.workflows["deploy"].steps) == 3
-    assert config.workflows["deploy"].input is None
-    assert config.workflows["pr-review"].input == "pr_url"
-
-
-def test_workflow_config_defaults_empty():
-    config = WarrenConfig()
-    assert config.workflows == {}
-```
-
-**Step 2: Run test to verify it fails**
-
-Run: `cd server && uv run pytest tests/test_config.py -v`
-Expected: FAIL — `WorkflowConfig` not defined
-
-**Step 3: Write minimal implementation**
-
-In `server/src/warren/config.py`, add before `CommandsConfig`:
-
-```python
-class WorkflowConfig(BaseModel):
-    label: str
-    steps: list[str]
-    input: str | None = None
-```
-
-Add to `WarrenConfig`:
-
-```python
-workflows: dict[str, WorkflowConfig] = {}
-```
-
-**Step 4: Run test to verify it passes**
-
-Run: `cd server && uv run pytest tests/test_config.py -v`
-Expected: PASS
-
-**Step 5: Lint**
-
-Run: `cd server && uv run ruff check src/ tests/`
-Expected: Clean
-
-**Step 6: Commit**
-
-```bash
-git add server/src/warren/config.py server/tests/test_config.py
-git commit -m "feat: add WorkflowConfig model to config"
-```
-
----
-
-### Task 2: Expose Workflows in Commands Endpoint
-
-**Files:**
-- Modify: `server/src/warren/app.py:313-321` (the `list_commands` endpoint)
-- Modify: `server/tests/test_app.py`
-
-**Step 1: Write the test**
-
-Add to `server/tests/test_app.py` in the appropriate test class (or create one):
-
-```python
-@pytest.mark.asyncio
-async def test_commands_include_workflows(client, app):
-    app.state.config.workflows = {
-        "deploy": WorkflowConfig(
-            label="Deploy",
-            steps=["Pull latest", "Run tests", "Push"],
-        ),
-    }
-    resp = await client.get("/commands", headers=auth_headers())
-    assert resp.status_code == 200
-    data = resp.json()
-    assert "workflows" in data
-    assert len(data["workflows"]) == 1
-    wf = data["workflows"][0]
-    assert wf["name"] == "deploy"
-    assert wf["label"] == "Deploy"
-    assert wf["steps"] == ["Pull latest", "Run tests", "Push"]
-    assert wf["input"] is None
-```
-
-Import `WorkflowConfig` at the top of the test file.
-
-**Step 2: Run test to verify it fails**
-
-Run: `cd server && uv run pytest tests/test_app.py::TestClassName::test_commands_include_workflows -v`
-Expected: FAIL — no `workflows` key in response
-
-**Step 3: Modify the endpoint**
-
-In `app.py`, modify the `list_commands` endpoint to include workflows:
-
-```python
-@app.get("/commands")
-async def list_commands(request: Request, _token: str = Depends(require_auth)):
-    config = request.app.state.config
-    return {
-        "actions": [{"label": a.label, "prompt": a.prompt} for a in config.commands.actions],
-        "templates": [
-            {"label": t.label, "prompt": t.prompt} for t in config.commands.templates
-        ],
-        "workflows": [
-            {
-                "name": name,
-                "label": wf.label,
-                "steps": wf.steps,
-                "input": wf.input,
-            }
-            for name, wf in config.workflows.items()
-        ],
-    }
-```
-
-**Step 4: Run test to verify it passes**
-
-Run: `cd server && uv run pytest tests/test_app.py -v`
-Expected: PASS (all existing tests still pass too)
-
-**Step 5: Lint**
-
-Run: `cd server && uv run ruff check src/ tests/`
-
-**Step 6: Commit**
-
-```bash
-git add server/src/warren/app.py server/tests/test_app.py
-git commit -m "feat: expose workflows in /commands endpoint"
-```
-
----
-
-### Task 3: Parse Step Progress in Chat Messages
-
-**Files:**
-- Modify: `creation/js/app.js:279-325` (the `handleEvent` function)
-
-This is the core UI change. When a `text` event contains `[STEP:n/total:status]` lines, we extract them and update a step tracker instead of rendering them as regular chat messages.
-
-**Step 1: Add step state and parser**
-
-At the top of the IIFE in `app.js` (after the existing state declarations around line 14), add:
-
-```javascript
-let workflowSteps = []; // [{index, total, status, text}]
-```
-
-Add a parser function after the `esc()` helper (around line 45):
-
-```javascript
-/** Parse [STEP:n/total:status] lines from a message, return {steps, rest}. */
-function parseSteps(text) {
-  const stepRegex = /^\[STEP:(\d+)\/(\d+):(pending|running|done|failed)\]\s*(.*)$/;
-  const lines = text.split('\n');
-  const steps = [];
-  const rest = [];
-  for (const line of lines) {
-    const m = line.match(stepRegex);
-    if (m) {
-      steps.push({ index: parseInt(m[1]), total: parseInt(m[2]), status: m[3], text: m[4] });
-    } else if (line.trim()) {
-      rest.push(line);
-    }
-  }
-  return { steps, rest: rest.join('\n') };
-}
-```
-
-**Step 2: Modify handleEvent to detect steps**
-
-In `handleEvent`, modify the `"text"` case (line 281-283):
-
-```javascript
-case "text": {
-  flushToolBuffer();
-  const { steps, rest } = parseSteps(event.content);
-  if (steps.length > 0) {
-    workflowSteps = steps;
-    renderStepProgress();
-  }
-  if (rest) {
-    appendTruncatedMsg("msg-text", rest);
-  }
-  break;
-}
-```
-
-**Step 3: Add renderStepProgress function**
-
-After the `handleEvent` function, add:
-
-```javascript
-function renderStepProgress() {
-  let container = document.getElementById("step-progress");
-  if (!container) return;
-  container.textContent = "";
-  if (workflowSteps.length === 0) {
-    container.classList.add("hidden");
-    return;
-  }
-  container.classList.remove("hidden");
-  workflowSteps.forEach((step) => {
-    const el = document.createElement("div");
-    el.className = "step-item step-" + step.status;
-    const icon = step.status === "done" ? "*"
-      : step.status === "running" ? ">"
-      : step.status === "failed" ? "!"
-      : "o";
-    el.textContent = icon + " " + step.text;
-    container.appendChild(el);
-  });
-}
-```
-
-**Step 4: Clear steps when opening a new session**
-
-In `openSession` (line 192), add `workflowSteps = [];` after clearing chat:
-
-```javascript
-async function openSession(session) {
-  currentSessionId = session.id;
-  chatTitle.textContent = session.name;
-  chatMessages.textContent = "";
-  workflowSteps = [];
-  // ...
-```
-
-Also call `renderStepProgress()` after loading message history (line 212), so steps from history are displayed.
-
-**Step 5: Test manually**
-
-No automated test for UI parsing, but verify by:
-- Deploy and send a message containing `[STEP:1/3:done] Pulled latest\n[STEP:2/3:running] Running tests\n[STEP:3/3:pending] Push to prod`
-- The step header should appear with done/running/pending indicators
-
-**Step 6: Commit**
-
-```bash
-git add creation/js/app.js
-git commit -m "feat: parse [STEP:] progress from chat messages"
-```
-
----
-
-### Task 4: Step Progress UI (HTML + CSS)
-
-**Files:**
-- Modify: `creation/index.html:27-42` (chat view)
-- Modify: `creation/css/styles.css` (add step styles)
-
-**Step 1: Add step progress container to chat view**
-
-In `creation/index.html`, add a `step-progress` div between the chat header and chat messages. Replace the chat view section (lines 27-42):
-
-```html
-<!-- Chat View -->
-<div id="view-chat" class="view">
-  <div class="header">
-    <button id="btn-back" class="btn-back">&lt;-</button>
-    <span id="chat-title" class="title"></span>
-    <button id="btn-stop" class="btn-stop hidden">[STOP]</button>
-    <span id="chat-status" class="dot dot-green"></span>
-  </div>
-  <div id="step-progress" class="step-progress hidden"></div>
-  <div id="chat-messages" class="scroll-area"></div>
-  <div class="footer">
-    <div id="ptt-indicator" class="ptt hidden">Listening...</div>
-    <div class="chat-input-row">
-      <input id="msg-input" type="text" placeholder="> _" class="msg-input">
-      <button id="btn-cmd" class="btn-cmd">[CMD]</button>
-    </div>
-  </div>
-</div>
-```
-
-**Step 2: Add CSS for step progress**
-
-Append to `creation/css/styles.css`:
-
-```css
-/* Workflow step progress */
-.step-progress {
-  padding: 4px 8px;
-  border-bottom: 1px solid #0a6e0a;
-  max-height: 80px;
-  overflow-y: auto;
-}
-
-.step-progress.hidden {
-  display: none;
-}
-
-.step-progress::-webkit-scrollbar {
-  display: none;
-}
-
-.step-item {
-  font-size: 10px;
-  padding: 1px 0;
-  white-space: nowrap;
-  overflow: hidden;
-  text-overflow: ellipsis;
-}
-
-.step-done {
-  color: #0a6e0a;
-}
-
-.step-running {
-  color: #33ff33;
-  text-shadow: 0 0 6px rgba(51, 255, 51, 0.5);
-}
-
-.step-pending {
-  color: #0a4e0a;
-}
-
-.step-failed {
-  color: #ff3333;
-  text-shadow: 0 0 4px rgba(255, 51, 51, 0.3);
-}
-```
-
-**Step 3: Verify HTML renders**
-
-Load the creation in a browser. The step-progress div should be hidden by default. When steps are received via SSE, they should appear between the header and messages.
-
-**Step 4: Commit**
-
-```bash
-git add creation/index.html creation/css/styles.css
-git commit -m "feat: step progress header UI in chat view"
-```
-
----
-
-### Task 5: Workflow Category in Command Palette
-
-**Files:**
-- Modify: `creation/js/app.js:475-548` (the `openPalette` function)
-
-**Step 1: Add workflow items to palette**
-
-In `openPalette()`, after the Templates section (line 520) and before the Recents section (line 522), add a Workflows section:
-
-```javascript
-// Workflows
-if (cmds.workflows && cmds.workflows.length > 0) {
-  const header = document.createElement("div");
-  header.className = "palette-section";
-  header.textContent = "WORKFLOWS";
-  paletteList.appendChild(header);
-
-  cmds.workflows.forEach((w) => {
-    paletteItems.push({
-      type: "workflow",
-      label: w.label,
-      steps: w.steps,
-      input: w.input,
-      name: w.name,
-    });
-    const el = document.createElement("div");
-    el.className = "palette-item";
-    el.textContent = w.label;
-    el.addEventListener("click", () => selectPaletteItem(paletteItems.length - 1));
-    paletteList.appendChild(el);
-  });
-}
-```
-
-**Step 2: Handle workflow selection in selectPaletteItem**
-
-In `selectPaletteItem()` (line 566), add handling for the `workflow` type. Before the existing `if (item.type === "template")` block (line 576):
-
-```javascript
-if (item.type === "workflow") {
-  if (item.input) {
-    // Workflow needs user input — prefill msg input with placeholder
-    msgInput.value = "";
-    msgInput.placeholder = item.input + "...";
-    msgInput.dataset.workflowSteps = JSON.stringify(item.steps);
-    msgInput.dataset.workflowName = item.name;
-    msgInput.focus();
-  } else {
-    // No input needed — compose and send immediately
-    const prompt = composeWorkflowPrompt(item.steps);
-    msgInput.value = prompt;
-    sendMsg();
-  }
-  return;
-}
-```
-
-**Step 3: Handle workflow selection from sessions view**
-
-In `createSessionFromPalette()` (line 595), add handling for workflows before the existing try block:
-
-```javascript
-if (item.type === "workflow") {
-  if (item.input) {
-    // Go to new session view with workflow context
-    showView("new");
-    initialMsg.value = "";
-    initialMsg.placeholder = item.input + "...";
-    initialMsg.dataset.workflowSteps = JSON.stringify(item.steps);
-    initialMsg.dataset.workflowName = item.name;
-    initialMsg.focus();
-  } else {
-    // Create session with composed workflow prompt
-    try {
-      let repoName = "";
-      if (backendType !== "nanoclaw") {
-        const repos = await Warren.fetchRepos();
-        repoName = Object.keys(repos)[0];
-        if (!repoName) { appendRawMsg("msg-error", "No repos configured"); return; }
-      }
-      const prompt = composeWorkflowPrompt(item.steps);
-      const data = await Warren.createSession(repoName, prompt);
-      openSession({ id: data.session_id, name: data.name, repo: data.repo });
-    } catch (err) {
-      appendRawMsg("msg-error", "Failed to create session");
-    }
-  }
-  return;
-}
-```
-
-**Step 4: Add composeWorkflowPrompt helper**
-
-Add this helper function near the other helpers (after `parseSteps`):
-
-```javascript
-function composeWorkflowPrompt(steps) {
-  const preamble = "Execute this workflow step by step. After completing each step, report progress using the format: [STEP:n/total:status] description (status: pending, running, done, failed). Stop on failure and report the error.\n\nSteps:";
-  const numbered = steps.map((s, i) => (i + 1) + ". " + s).join("\n");
-  return preamble + "\n" + numbered;
-}
-```
-
-**Step 5: Handle sendMsg for workflows with input**
-
-In `sendMsg()` (line 625), check if the input has workflow context and compose the prompt:
-
-After `const text = msgInput.value.trim();` and before `if (!text || !currentSessionId) return;`, add:
-
-```javascript
-const wfSteps = msgInput.dataset.workflowSteps;
-if (wfSteps) {
-  const steps = JSON.parse(wfSteps);
-  const prompt = composeWorkflowPrompt(steps).replace('{' + (msgInput.dataset.workflowInput || 'input') + '}', text);
-  msgInput.dataset.workflowSteps = "";
-  msgInput.dataset.workflowName = "";
-  msgInput.placeholder = "> _";
-  msgInput.value = prompt + "\n\nInput: " + text;
-}
-```
-
-Actually, simpler approach — when the user provides input, prepend it to the workflow prompt:
-
-```javascript
-const wfSteps = msgInput.dataset.workflowSteps;
-if (wfSteps) {
-  const steps = JSON.parse(wfSteps);
-  msgInput.value = composeWorkflowPrompt(steps) + "\n\nInput: " + text;
-  delete msgInput.dataset.workflowSteps;
-  delete msgInput.dataset.workflowName;
-  msgInput.placeholder = "> _";
-}
-```
-
-**Step 6: Commit**
-
-```bash
-git add creation/js/app.js
-git commit -m "feat: workflow category in command palette + launcher"
-```
-
----
-
-### Task 6: Agent Step Reporting Convention
-
-**Files:**
-- Create: `nanoclaw/container/skills/workflow-steps/SKILL.md`
-
-This skill teaches the agent the step reporting convention. It's loaded when the agent is running inside NanoClaw containers.
-
-**Step 1: Create the skill file**
-
-```markdown
----
-name: workflow-steps
-description: Report step-by-step progress for multi-step workflows. Use when executing a workflow with numbered steps, or when you plan and execute a multi-step task.
----
-
-# Workflow Step Progress
-
-When executing multi-step work, report progress after each step using the `send_message` tool with this format:
-
-```
-[STEP:1/4:done] Pulled latest from main
-[STEP:2/4:running] Running test suite
-[STEP:3/4:pending] Push to production
-[STEP:4/4:pending] Report deploy status
-```
-
-## Format
-
-`[STEP:{current}/{total}:{status}] {description}`
-
-**Statuses:** `pending`, `running`, `done`, `failed`
-
-## Rules
-
-- Send ALL steps in every progress update (not just the current one)
-- Mark completed steps as `done`, current step as `running`, future steps as `pending`
-- On failure: mark the step as `failed`, stop execution, describe the error
-- Send progress via `send_message` after each step completes
-- You can include additional text after the step block — it renders as normal chat
-
-## Example: Success
-
-```
-[STEP:1/3:done] Cloned repository
-[STEP:2/3:done] Tests passed (142/142)
-[STEP:3/3:running] Creating pull request
-```
-
-## Example: Failure
-
-```
-[STEP:1/3:done] Cloned repository
-[STEP:2/3:failed] Tests failed: 3 errors in auth module
-[STEP:3/3:pending] Create pull request
-
-Test failures:
-- test_login_redirect: AssertionError
-- test_token_refresh: TimeoutError
-- test_logout: 404 response
-```
-
-## Agent-Planned Workflows
-
-When a user gives you a complex task without predefined steps, you can plan your own steps and report them the same way. Create your step list, then report progress as you execute.
-```
-
-**Step 2: Verify skill is loadable**
-
-Run: `ls -la nanoclaw/container/skills/workflow-steps/SKILL.md`
-Expected: File exists with correct content
-
-**Step 3: Commit**
-
-```bash
-git add nanoclaw/container/skills/workflow-steps/SKILL.md
-git commit -m "feat: add workflow-steps skill for agent step reporting"
-```
-
----
-
-### Task 7: Add Sample Workflows to Config
-
-**Files:**
-- Modify: `server/src/warren/config.py` (no changes needed — model already supports it)
-- Note: No `config.yaml` exists in repo (it's deployment-specific)
-
-This task documents the sample config and verifies the end-to-end flow works.
-
-**Step 1: Add sample workflows to config documentation**
-
-Update the `CLAUDE.md` deploy section or the design doc with example config:
-
-```yaml
-workflows:
-  deploy:
-    label: "Deploy"
-    steps:
-      - "Pull latest from main"
-      - "Run the test suite"
-      - "If tests pass, push to production"
-      - "Report deploy status"
-  pr-review:
-    label: "Review PR"
-    input: "pr_url"
-    steps:
-      - "Fetch the PR diff and description"
-      - "Review for bugs, security issues, and style"
-      - "Post a summary of findings"
-  run-tests:
-    label: "Run Tests"
-    steps:
-      - "Run the full test suite"
-      - "Report pass/fail counts and any failures"
-```
-
-**Step 2: Run full test suite**
-
-Run: `cd server && uv run pytest tests/ -v`
-Expected: All tests pass
-
-**Step 3: Build nanoclaw**
-
-Run: `cd nanoclaw && npm run build`
-Expected: Clean compilation
-
-**Step 4: Commit**
-
-```bash
-git add -A
-git commit -m "feat: R1 workflow engine — complete implementation"
-```
-
----
-
-## Verification Checklist
-
-1. `cd server && uv run ruff check src/ tests/` — lint clean
-2. `cd server && uv run pytest tests/ -v` — all tests pass
-3. `cd nanoclaw && npm run build` — compiles clean
-4. Deploy and open R1:
-   - Command palette shows WORKFLOWS section
-   - Tap a workflow → creates session with composed prompt
-   - Agent sends `[STEP:...]` messages → step header appears
-   - Failed step shows red, agent stops
-   - User can type to intervene or tap [STOP]
-5. Agent-planned workflows: ask agent a complex task → it uses workflow-steps skill to report steps in same format
diff --git a/docs/plans/2026-02-22-code-reduction-design.md b/docs/plans/2026-02-22-code-reduction-design.md
deleted file mode 100644
index 9f0b17c..0000000
--- a/docs/plans/2026-02-22-code-reduction-design.md
+++ /dev/null
@@ -1,139 +0,0 @@
-# Code Reduction + R1 Test Harness Design
-
-## Context
-
-Now that PTT voice is the primary interaction method and NanoClaw is the only
-production backend, Warren carries ~860 lines of dead code: the DirectAdapter
-subprocess system, unused Creation JS modules, and backend-switching logic.
-
-Additionally, there's no way to test the full R1 pipeline without a physical
-device. An E2E test harness that emulates both R1 interfaces (voice PTT and
-Creation WebView) against a local docker-compose stack would catch regressions
-automatically.
-
-## Part 1: Surgical Code Removal
-
-### 1A. Remove DirectAdapter + ClaudeRunner
-
-**Delete entirely:**
-- `server/src/warren/adapters/direct.py` (118 lines)
-- `server/src/warren/claude_runner.py` (148 lines)
-- `server/tests/test_claude_runner.py` (119 lines)
-
-**Modify `server/tests/test_adapters.py`:**
-- Delete `TestDirectAdapter` class (lines 37-159)
-- Keep `TestNanoClawAdapter` and message/event unit tests
-
-**Simplify `server/src/warren/app.py`:**
-- Remove `from warren.adapters.direct import DirectAdapter` import
-- Simplify `create_adapter()` to always create NanoClawAdapter (no switch)
-- Remove `backend != "nanoclaw"` special case in `create_session` (line 480-481)
-
-**Fix test fixtures** in 6 files that use DirectAdapter as a mock target:
-- `test_app.py`, `test_streaming.py`, `test_integration.py`
-- `test_voice.py`, `test_voice_integration.py`, `test_voice_pairing.py`
-- Replace `DirectAdapter(...)` with `NanoClawAdapter(ipc_dir=...)` in fixtures
-
-### 1B. Remove Dead Creation JS
-
-**Delete entirely:**
-- `creation/js/probe.js` (234 lines) -- network diagnostics, triple-tap activated
-- `creation/js/hardware.js` (45 lines) -- unimplemented R1 STT button mapping
-
-**Modify `creation/index.html`:**
-- Remove `<script>` tags for probe.js and hardware.js
-
-**Modify `creation/js/app.js`:**
-- Remove Probe attachment code (lines 999-1002)
-
-### 1C. Clean Up Backend Config
-
-**Modify `server/src/warren/config.py`:**
-- Remove `Settings.backend` field (always nanoclaw)
-- Remove `Settings.claude_bin` field (only used by DirectAdapter)
-
-**Modify `server/src/warren/app.py`:**
-- `create_adapter()` always creates NanoClawAdapter, no switch statement
-- Remove `/backend` endpoint (line 314-316)
-- Remove `backend != "nanoclaw"` guards in monitor endpoints
-
-**Modify `creation/js/api.js`:**
-- Remove `Warren.fetchBackend()` function
-
-**Modify `creation/js/app.js`:**
-- Remove `backendType` variable and all branches that check it
-- Take the nanoclaw code path unconditionally (repo picker hidden, etc.)
-
-### Estimated Savings
-
-| Component | Lines removed |
-|-----------|--------------|
-| DirectAdapter + ClaudeRunner + tests | ~500 |
-| probe.js + hardware.js | ~280 |
-| Backend config cleanup | ~80 |
-| **Total** | **~860** |
-
-## Part 2: R1 E2E Test Harness
-
-### Overview
-
-A pytest-based E2E suite that emulates the R1's two interfaces against a local
-docker-compose stack with a mock Claude CLI (no API key needed).
-
-### Architecture
-
-```
-tests/e2e/
-  conftest.py           -- docker-compose up/down, health wait, auth helpers
-  test_voice_e2e.py     -- OpenClaw WebSocket protocol emulation
-  test_creation_e2e.py  -- Playwright headless browser at 240x282
-  test_pipeline_e2e.py  -- Full NanoClaw round-trip
-  mock-claude            -- Fake claude binary (canned stream-json output)
-  docker-compose.test.yml -- Override: mounts mock-claude, disables real API
-```
-
-### Test Coverage
-
-**Voice PTT (`test_voice_e2e.py`):**
-- WebSocket connect, challenge, auth handshake
-- `chat.send` request, verify `chat-ack`
-- Verify `chat-event` deltas and `final` arrive
-- Keepalive tick verification
-
-**Creation UI (`test_creation_e2e.py`):**
-- Playwright loads `/creation/?token=...` at 240x282 viewport
-- Session list renders
-- Open session, SSE events arrive and render as chat messages
-- Monitor view loads and shows agent status
-
-**Full pipeline (`test_pipeline_e2e.py`):**
-- Voice `chat.send` triggers NanoClaw IPC pickup
-- Mock agent runs, posts callback to Warren
-- SSE and WebSocket both receive the response
-- Message persists in database
-
-### Mock Claude CLI
-
-A shell script or Python script placed in agent containers that:
-- Reads the `-p` argument (the user message)
-- Outputs canned `stream-json` lines (status, text, result)
-- Exits 0
-
-This makes tests fast (~seconds), deterministic, and free.
-
-### Docker Compose Test Override
-
-`docker-compose.test.yml` overrides the base compose to:
-- Mount `mock-claude` as `/usr/local/bin/claude` in agent containers
-- Set `ANTHROPIC_API_KEY=sk-test` (mock doesn't use it)
-- Set `WARREN_INTERNAL_SECRET=test-secret`
-- Expose ports for test client access
-
-## What Stays Untouched
-
-- Voice channel (PTT + Creation UI)
-- Monitor view and endpoints
-- SessionBus, lifecycle, auth, database
-- NanoClawAdapter, IPC, nanoclaw_db
-- All NanoClaw submodule code
-- Notify overlay, command palette
diff --git a/docs/plans/2026-02-22-code-reduction-plan.md b/docs/plans/2026-02-22-code-reduction-plan.md
deleted file mode 100644
index 9229550..0000000
--- a/docs/plans/2026-02-22-code-reduction-plan.md
+++ /dev/null
@@ -1,1219 +0,0 @@
-# Code Reduction + R1 Test Harness Implementation Plan
-
-> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
-
-**Goal:** Remove ~860 lines of dead code (DirectAdapter, unused JS, backend-switching logic) and add an E2E test harness that emulates the R1.
-
-**Architecture:** Part 1 removes DirectAdapter/ClaudeRunner, dead JS files, and backend config switching. Part 2 adds pytest E2E tests using WebSocket client + Playwright against a docker-compose test stack with a mock Claude CLI.
-
-**Tech Stack:** Python/FastAPI, pytest, websockets, Playwright, Docker Compose, bash
-
----
-
-### Task 1: Delete DirectAdapter, ClaudeRunner, and their tests
-
-**Files:**
-- Delete: `server/src/warren/adapters/direct.py`
-- Delete: `server/src/warren/claude_runner.py`
-- Delete: `server/tests/test_claude_runner.py`
-- Modify: `server/tests/test_adapters.py`
-
-**Step 1: Delete the three files**
-
-```bash
-rm server/src/warren/adapters/direct.py
-rm server/src/warren/claude_runner.py
-rm server/tests/test_claude_runner.py
-```
-
-**Step 2: Remove TestDirectAdapter from test_adapters.py**
-
-In `server/tests/test_adapters.py`, delete the entire `TestDirectAdapter` class (lines 37-159) and its imports. Keep only:
-- `test_backend_message_defaults` (line 7)
-- `test_backend_event_creation` (line 14)
-- `test_backend_message_with_payload` (line 23)
-- `TestNanoClawAdapter` class (line 162-237)
-
-The remaining file should have these imports only:
-
-```python
-"""Tests for backend adapter types."""
-
-
-def test_backend_message_defaults():
-    from warren.adapters import BackendMessage
-
-    msg = BackendMessage(kind="new_session")
-    assert msg.kind == "new_session"
-    assert msg.session_id is None
-    assert msg.payload == {}
-
-
-def test_backend_event_creation():
-    from warren.adapters import BackendEvent
-
-    event = BackendEvent(kind="text", session_id="abc", data={"content": "hello"})
-    assert event.kind == "text"
-    assert event.session_id == "abc"
-    assert event.data["content"] == "hello"
-
-
-def test_backend_message_with_payload():
-    from warren.adapters import BackendMessage
-
-    msg = BackendMessage(
-        kind="new_session",
-        session_id="abc",
-        payload={"message": "hello", "repo": "test"},
-    )
-    assert msg.session_id == "abc"
-    assert msg.payload["message"] == "hello"
-
-
-class TestNanoClawAdapter:
-    # ... keep all existing NanoClaw tests unchanged ...
-```
-
-Remove the `import asyncio` and `from unittest.mock import patch` imports (only used by TestDirectAdapter).
-
-**Step 3: Run tests to verify nothing breaks**
-
-```bash
-cd server && uv run pytest tests/test_adapters.py -v
-```
-
-Expected: All NanoClaw adapter tests pass. No TestDirectAdapter tests (they were deleted).
-
-**Step 4: Commit**
-
-```bash
-git add -A && git commit -m "refactor: remove DirectAdapter + ClaudeRunner (nanoclaw-only)"
-```
-
----
-
-### Task 2: Update test fixtures from DirectAdapter to NanoClawAdapter
-
-Six test files create a `DirectAdapter` in their fixtures but only use it as a mock target (they patch `adapter.send`). Switch them to `NanoClawAdapter`.
-
-**Files:**
-- Modify: `server/tests/test_app.py`
-- Modify: `server/tests/test_streaming.py`
-- Modify: `server/tests/test_integration.py`
-- Modify: `server/tests/test_voice.py`
-- Modify: `server/tests/test_voice_integration.py`
-- Modify: `server/tests/test_voice_pairing.py`
-
-**Step 1: Update each fixture**
-
-In each file, apply these changes:
-
-**Import change:**
-```python
-# BEFORE:
-from warren.adapters.direct import DirectAdapter
-
-# AFTER:
-from warren.adapters.nanoclaw import NanoClawAdapter
-```
-
-**Constructor change:**
-```python
-# BEFORE:
-adapter = DirectAdapter(claude_bin="claude", max_concurrent=2)
-
-# AFTER:
-adapter = NanoClawAdapter(ipc_dir=str(tmp_path / "ipc"))
-```
-
-**Settings change** (where `backend="direct"` is set):
-```python
-# BEFORE:
-backend="direct",
-
-# AFTER:
-# Delete this line entirely — 'backend' setting will be removed in Task 4
-```
-
-Here are the exact changes per file:
-
-**`test_app.py` (lines 15-16, 53-55):**
-- Line 15: `from warren.adapters.direct import DirectAdapter` → `from warren.adapters.nanoclaw import NanoClawAdapter`
-- Line 53: `adapter = DirectAdapter(claude_bin="claude", max_concurrent=2)` → `adapter = NanoClawAdapter(ipc_dir=str(tmp_path / "ipc"))`
-
-**`test_streaming.py` (lines 13-14, 37-38):**
-- Line 13: `from warren.adapters.direct import DirectAdapter` → `from warren.adapters.nanoclaw import NanoClawAdapter`
-- Line 37: `adapter = DirectAdapter(claude_bin="claude", max_concurrent=2)` → `adapter = NanoClawAdapter(ipc_dir=str(tmp_path / "ipc"))`
-
-**`test_integration.py` (lines 16-17, 40-41):**
-- Line 16: `from warren.adapters.direct import DirectAdapter` → `from warren.adapters.nanoclaw import NanoClawAdapter`
-- Line 40: `adapter = DirectAdapter(claude_bin="claude", max_concurrent=2)` → `adapter = NanoClawAdapter(ipc_dir=str(tmp_path / "ipc"))`
-
-**`test_voice.py` (lines 9, 25-26, 43-44, 64-65, 81-82):**
-- Line 9: `from warren.adapters.direct import DirectAdapter` → `from warren.adapters.nanoclaw import NanoClawAdapter`
-- Line 25-26: Remove `backend="direct",`
-- Line 43-44: `adapter = DirectAdapter(claude_bin="claude", max_concurrent=2)` → `adapter = NanoClawAdapter(ipc_dir=str(tmp_path / "ipc"))`
-- Line 64-65: Remove `backend="direct",`
-- Line 81-82: `adapter = DirectAdapter(claude_bin="claude", max_concurrent=2)` → `adapter = NanoClawAdapter(ipc_dir=str(tmp_path / "ipc"))`
-
-**`test_voice_integration.py` (lines 12, 30, 49-50):**
-- Line 12: `from warren.adapters.direct import DirectAdapter` → `from warren.adapters.nanoclaw import NanoClawAdapter`
-- Line 30: Remove `backend="direct",`
-- Line 49: `adapter = DirectAdapter(claude_bin="claude", max_concurrent=2)` → `adapter = NanoClawAdapter(ipc_dir=str(tmp_path / "ipc"))`
-
-**`test_voice_pairing.py` (lines 8, 22, 33-34):**
-- Line 8: `from warren.adapters.direct import DirectAdapter` → `from warren.adapters.nanoclaw import NanoClawAdapter`
-- Line 22: Remove `backend="direct",`
-- Line 33: `adapter = DirectAdapter(claude_bin="claude", max_concurrent=2)` → `adapter = NanoClawAdapter(ipc_dir=str(tmp_path / "ipc"))`
-
-**Step 2: Run full test suite**
-
-```bash
-cd server && uv run pytest tests/ -v
-```
-
-Expected: All tests pass. No remaining imports of `DirectAdapter` or `ClaudeRunner`.
-
-**Step 3: Verify no remaining references**
-
-```bash
-cd server && grep -r "DirectAdapter\|claude_runner\|ClaudeRunner" src/ tests/
-```
-
-Expected: No output.
-
-**Step 4: Commit**
-
-```bash
-git add -A && git commit -m "refactor: switch test fixtures from DirectAdapter to NanoClawAdapter"
-```
-
----
-
-### Task 3: Remove dead Creation JS files
-
-**Files:**
-- Delete: `creation/js/probe.js`
-- Delete: `creation/js/hardware.js`
-- Modify: `creation/index.html`
-- Modify: `creation/js/app.js`
-
-**Step 1: Delete the JS files**
-
-```bash
-rm creation/js/probe.js
-rm creation/js/hardware.js
-```
-
-**Step 2: Remove script tags from index.html**
-
-In `creation/index.html`, delete these two lines:
-
-```html
-  <script src="js/hardware.js?v=20"></script>
-```
-
-```html
-  <script src="js/probe.js?v=20"></script>
-```
-
-The scripts section should become:
-
-```html
-  <script src="js/api.js?v=20"></script>
-  <script src="js/notify.js?v=20"></script>
-  <script src="js/app.js?v=20"></script>
-```
-
-**Step 3: Remove Probe attachment from app.js**
-
-In `creation/js/app.js`, delete lines 999-1003:
-
-```javascript
-  // Attach API probe (triple-tap "WARREN://" title to activate)
-  if (typeof Probe !== "undefined") {
-    const probeTarget = document.querySelector("#view-sessions .title");
-    if (probeTarget) Probe.attach(probeTarget, sessionList);
-  }
-```
-
-**Step 4: Remove Hardware references from app.js**
-
-In `creation/js/app.js`, the `Hardware.bind(...)` calls (lines 833-959) reference the deleted `Hardware` global. Since `hardware.js` defined the `Hardware` object, these calls will throw `ReferenceError`.
-
-However — looking more carefully, `Hardware.bind` is used for scroll wheel, side button, and long-press PTT. These are still valuable for the R1 physical buttons. The `hardware.js` file defines the Hardware object that translates raw DOM events (wheel, pointerdown, etc.) into named actions.
-
-**Actually — re-read hardware.js before deleting.** Let me check what it does:
-
-The `Hardware` object maps:
-- `scrollUp` / `scrollDown` — scroll wheel events → navigate session list, palette, chat
-- `sideClick` — R1 side button → select session, send message, etc.
-- `longPressStart` / `longPressEnd` — long-press side button → STT recording
-
-**The scroll wheel and side button bindings are actively used on the R1.** Only the STT (longPress) part is dead code.
-
-**Revised plan:** Keep `hardware.js`. Only delete `probe.js`.
-
-Updated step 1:
-```bash
-rm creation/js/probe.js
-```
-
-Updated step 2 — only remove probe.js script tag from index.html:
-```html
-<!-- DELETE THIS LINE: -->
-  <script src="js/probe.js?v=20"></script>
-```
-
-Step 3 stays the same (remove Probe attachment from app.js).
-
-Step 4 is no longer needed — hardware.js stays.
-
-**Step 5: Verify the creation page loads**
-
-Open `creation/index.html` in a browser or check no JS errors:
-```bash
-cd server && uv run python -c "
-from pathlib import Path
-html = Path('../creation/index.html').read_text()
-assert 'probe.js' not in html
-assert 'hardware.js' in html
-print('OK: probe.js removed, hardware.js kept')
-"
-```
-
-**Step 6: Commit**
-
-```bash
-git add -A && git commit -m "refactor: remove dead probe.js from Creation UI"
-```
-
----
-
-### Task 4: Clean up backend config and switching logic
-
-**Files:**
-- Modify: `server/src/warren/config.py`
-- Modify: `server/src/warren/app.py`
-- Modify: `creation/js/api.js`
-- Modify: `creation/js/app.js`
-
-**Step 1: Remove `backend` and `claude_bin` from Settings**
-
-In `server/src/warren/config.py`, delete these two lines from the `Settings` class:
-
-```python
-    claude_bin: str = "claude"
-    backend: str = "direct"
-```
-
-The Settings class should have these fields remaining:
-```python
-class Settings(BaseSettings):
-    model_config = {"env_prefix": "WARREN_"}
-
-    host: str = "0.0.0.0"
-    port: int = 4030
-    db: str = "warren.db"
-    config: str = "config.yaml"
-    anthropic_api_key: str = ""
-    admin_token: str = ""
-    data_dir: str = "data"
-    nanoclaw_db: str = ""
-    nanoclaw_ipc_dir: str = ""
-    voice_enabled: bool = False
-    voice_port: int = 443
-    internal_secret: str = ""
-```
-
-**Step 2: Simplify `create_adapter()` in app.py**
-
-Replace the entire `create_adapter` function (lines 72-85) with:
-
-```python
-def create_adapter(settings: Settings, config):
-    """Create the NanoClaw backend adapter."""
-    from warren.adapters.nanoclaw import NanoClawAdapter
-
-    db_path, ipc_dir = _nanoclaw_paths(settings, config)
-    return NanoClawAdapter(ipc_dir=ipc_dir, db_path=db_path)
-```
-
-Also remove the DirectAdapter import at the top of app.py (line 21):
-```python
-from warren.adapters.direct import DirectAdapter
-```
-
-**Step 3: Remove `/backend` endpoint from app.py**
-
-Delete the `/backend` endpoint (lines 314-316):
-
-```python
-    @app.get("/backend")
-    async def get_backend(request: Request, _token: str = Depends(require_auth)):
-        return {"backend": request.app.state.settings.backend}
-```
-
-**Step 4: Remove `backend != "nanoclaw"` guards in app.py**
-
-In `create_session` (around line 480), delete this line:
-```python
-        elif repo_name and settings.backend != "nanoclaw":
-            raise HTTPException(status_code=404, detail=f"Unknown repo: {repo_name}")
-```
-
-In `monitor_health` (around line 637), simplify to always check nanoclaw:
-```python
-    @app.get("/monitor/health")
-    async def monitor_health(
-        request: Request,
-        _token: str = Depends(require_auth),
-    ):
-        import os
-
-        settings = request.app.state.settings
-        result = {"warren": {"status": "ok"}}
-
-        db_path, _ = _nanoclaw_paths(settings, request.app.state.config)
-        if os.path.exists(db_path):
-            result["nanoclaw"] = {"status": "ok"}
-        else:
-            result["nanoclaw"] = {
-                "status": "unavailable",
-                "detail": "DB not found",
-            }
-
-        return result
-```
-
-In `monitor_tasks` (around line 664), remove the `backend != "nanoclaw"` guard:
-```python
-    @app.get("/monitor/tasks")
-    async def monitor_tasks(
-        request: Request,
-        _token: str = Depends(require_auth),
-    ):
-        settings = request.app.state.settings
-        db_path, _ = _nanoclaw_paths(settings, request.app.state.config)
-        from warren.nanoclaw_db import NanoClawDbReader
-
-        reader = NanoClawDbReader(db_path)
-        tasks = await reader.list_tasks()
-        return {"tasks": tasks}
-```
-
-In `pause_task` (around line 680), remove the guard:
-```python
-    @app.post("/monitor/tasks/{task_id}/pause")
-    async def pause_task(
-        task_id: str,
-        request: Request,
-        _token: str = Depends(require_auth),
-    ):
-        settings = request.app.state.settings
-        _, ipc_dir = _nanoclaw_paths(settings, request.app.state.config)
-        from warren.ipc import write_ipc_task
-
-        write_ipc_task(
-            ipc_dir=ipc_dir,
-            group="warren",
-            task_id=f"pause-{task_id}",
-            payload={
-                "type": "pause_task",
-                "task_id": task_id,
-                "owner_group": "warren",
-            },
-        )
-        return {"status": "paused", "task_id": task_id}
-```
-
-Same for `resume_task` — remove the guard.
-
-**Step 5: Remove `fetchBackend` from api.js**
-
-In `creation/js/api.js`, delete the function (lines 49-51):
-```javascript
-  async function fetchBackend() {
-    return _fetch("/backend");
-  }
-```
-
-And remove `fetchBackend` from the return object (line 153):
-```javascript
-    fetchBackend,
-```
-
-**Step 6: Remove `backendType` from app.js**
-
-In `creation/js/app.js`, make these changes:
-
-1. Delete line 13: `let backendType = "direct";`
-
-2. Replace the `initialize` function (lines 1005-1027) — remove the fetchBackend call:
-```javascript
-  async function initialize() {
-    const result = Warren.initFromUrl();
-    if (result.ok) {
-      showView("sessions");
-      return;
-    }
-    sessionList.textContent = "";
-    const empty = document.createElement("div");
-    empty.className = "empty-state";
-    empty.textContent = "ERR: " + result.reason;
-    sessionList.appendChild(empty);
-
-    const debug = document.createElement("div");
-    debug.className = "empty-state";
-    debug.style.fontSize = "10px";
-    debug.style.wordBreak = "break-all";
-    debug.textContent = window.location.href;
-    sessionList.appendChild(debug);
-  }
-```
-
-3. In `createSessionFromPalette` (around line 690), replace the workflow branch that checks backendType:
-
-```javascript
-    // BEFORE (lines 699-711):
-    } else {
-        try {
-          let repoName = "";
-          if (backendType !== "nanoclaw") {
-            const repos = await Warren.fetchRepos();
-            repoName = Object.keys(repos)[0];
-            if (!repoName) { appendRawMsg("msg-error", "No repos configured"); return; }
-          }
-          const prompt = composeWorkflowPrompt(item.steps);
-          const data = await Warren.createSession(repoName, prompt);
-          openSession({ id: data.session_id, name: data.name, repo: data.repo });
-        } catch (err) {
-          appendRawMsg("msg-error", "Failed to create session");
-        }
-    }
-
-    // AFTER:
-    } else {
-        try {
-          const prompt = composeWorkflowPrompt(item.steps);
-          const data = await Warren.createSession("", prompt);
-          openSession({ id: data.session_id, name: data.name, repo: data.repo });
-        } catch (err) {
-          appendRawMsg("msg-error", "Failed to create session");
-        }
-    }
-```
-
-4. Same pattern for the action/recent branch (around lines 728-742):
-```javascript
-    // BEFORE:
-    try {
-      let repoName = "";
-      if (backendType !== "nanoclaw") {
-        const repos = await Warren.fetchRepos();
-        repoName = Object.keys(repos)[0];
-        if (!repoName) {
-          appendRawMsg("msg-error", "No repos configured");
-          return;
-        }
-      }
-      const data = await Warren.createSession(repoName, item.prompt);
-
-    // AFTER:
-    try {
-      const data = await Warren.createSession("", item.prompt);
-```
-
-5. Simplify `loadRepos` (around line 772) — always hide repo picker:
-```javascript
-  async function loadRepos() {
-    const repoLabel = repoSelect.previousElementSibling;
-    repoSelect.style.display = "none";
-    if (repoLabel) repoLabel.style.display = "none";
-    initialMsg.focus();
-  }
-```
-
-6. In `startSession` (around line 811), remove the backendType guard:
-```javascript
-    // BEFORE:
-    if (!repo && backendType !== "nanoclaw") return;
-
-    // AFTER:
-    // (delete this line entirely — repo is always optional with nanoclaw)
-```
-
-7. In `handleEvent` case `"tool"` (around lines 313-329), remove the DirectAdapter branch:
-```javascript
-      case "tool": {
-        flushToolBuffer();
-        const label = event.tool
-          ? event.tool + " " + (event.summary || "")
-          : event.action || "tool";
-        appendRawMsg("msg-tool", label);
-        break;
-      }
-```
-
-**Step 7: Run lint and tests**
-
-```bash
-cd server && uv run ruff check src/ tests/
-cd server && uv run pytest tests/ -v
-```
-
-Expected: Lint clean, all tests pass.
-
-**Step 8: Verify no remaining references to removed code**
-
-```bash
-grep -r "backendType\|fetchBackend\|backend.*direct\|DirectAdapter\|claude_runner\|ClaudeRunner\|claude_bin" server/src/ server/tests/ creation/
-```
-
-Expected: No output (except possibly CLAUDE.md docs, which is fine).
-
-**Step 9: Commit**
-
-```bash
-git add -A && git commit -m "refactor: remove backend switching logic (nanoclaw-only)"
-```
-
----
-
-### Task 5: Update tests for removed /backend endpoint and Settings fields
-
-**Files:**
-- Modify: `server/tests/test_app.py` (if it tests `/backend`)
-- Modify: `server/tests/test_monitor.py` (monitor tests may assume direct backend)
-- Modify: `server/tests/test_config.py` (if it tests `backend` or `claude_bin` fields)
-
-**Step 1: Check and fix test_config.py**
-
-Search for tests that reference `backend` or `claude_bin` in Settings and remove/update them.
-
-```bash
-cd server && grep -n "backend\|claude_bin" tests/test_config.py
-```
-
-Remove any tests that assert on `settings.backend` or `settings.claude_bin`.
-
-**Step 2: Check and fix test_monitor.py**
-
-The `monitor_app` fixture (line 19) doesn't set `backend`, so it defaults. Since we removed the `backend` field, the default is now always nanoclaw. The test `test_returns_empty_when_no_nanoclaw` (line 100) may need updating — it currently expects `{"tasks": []}` when `backend != "nanoclaw"`, but now monitor_tasks always tries to read the NanoClaw DB. If the DB doesn't exist, it should still return empty or handle gracefully.
-
-Check the monitor fixture uses the lifespan (which calls `create_adapter`), so it will now try to create a NanoClawAdapter. The fixture already has `tmp_path`, so the IPC dir will be created under tmp_path. But the NanoClaw DB won't exist, so `monitor_tasks` will try to read a non-existent file.
-
-Fix: ensure `NanoClawDbReader.list_tasks()` handles missing DB gracefully, OR update the test to expect the actual behavior.
-
-**Step 3: Run full test suite**
-
-```bash
-cd server && uv run pytest tests/ -v
-```
-
-Fix any failures.
-
-**Step 4: Commit**
-
-```bash
-git add -A && git commit -m "test: update tests for nanoclaw-only backend"
-```
-
----
-
-### Task 6: Create mock Claude CLI
-
-**Files:**
-- Create: `tests/e2e/mock-claude`
-
-**Step 1: Write the mock claude script**
-
-Create `tests/e2e/mock-claude` — a shell script that mimics `claude -p "..." --output-format stream-json`:
-
-```bash
-#!/bin/bash
-# Mock Claude CLI for E2E testing.
-# Reads -p argument and returns canned stream-json output.
-
-MESSAGE=""
-while [[ $# -gt 0 ]]; do
-  case $1 in
-    -p) MESSAGE="$2"; shift 2 ;;
-    *) shift ;;
-  esac
-done
-
-# Simulate short processing delay
-sleep 0.5
-
-# Emit stream-json lines matching real Claude CLI output format
-echo '{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"I received your message: '"${MESSAGE}"'"}]}}'
-sleep 0.2
-echo '{"type":"result","result":"Mock response complete.","session_id":"mock-session-001"}'
-```
-
-**Step 2: Make it executable**
-
-```bash
-chmod +x tests/e2e/mock-claude
-```
-
-**Step 3: Test it locally**
-
-```bash
-./tests/e2e/mock-claude -p "hello world" --output-format stream-json
-```
-
-Expected output:
-```
-{"type":"assistant","message":{"role":"assistant","content":[{"type":"text","text":"I received your message: hello world"}]}}
-{"type":"result","result":"Mock response complete.","session_id":"mock-session-001"}
-```
-
-**Step 4: Commit**
-
-```bash
-git add tests/e2e/mock-claude && git commit -m "test: add mock Claude CLI for E2E testing"
-```
-
----
-
-### Task 7: Create docker-compose test override
-
-**Files:**
-- Create: `tests/e2e/docker-compose.test.yml`
-
-**Step 1: Write the test compose override**
-
-This overlays the base `docker-compose.yml` to:
-- Mount mock-claude into the NanoClaw agent container image build
-- Set test secrets
-- Expose Warren on a fixed test port
-
-```yaml
-# tests/e2e/docker-compose.test.yml
-# Override for E2E testing — uses mock Claude CLI, no real API key needed.
-# Usage: docker compose -f docker-compose.yml -f tests/e2e/docker-compose.test.yml up -d
-
-services:
-  warren:
-    environment:
-      - ANTHROPIC_API_KEY=sk-test-mock
-      - WARREN_ADMIN_TOKEN=test-admin-token
-      - WARREN_INTERNAL_SECRET=test-internal-secret
-      - WARREN_VOICE_ENABLED=true
-      - WARREN_VOICE_PORT=443
-    ports:
-      - "14030:4030"
-
-  nanoclaw:
-    environment:
-      - ANTHROPIC_API_KEY=sk-test-mock
-      - WARREN_INTERNAL_SECRET=test-internal-secret
-    volumes:
-      - ./tests/e2e/mock-claude:/usr/local/bin/claude:ro
-```
-
-**Step 2: Commit**
-
-```bash
-git add tests/e2e/docker-compose.test.yml && git commit -m "test: add docker-compose test override for E2E"
-```
-
----
-
-### Task 8: Create E2E test conftest with docker-compose lifecycle
-
-**Files:**
-- Create: `tests/e2e/__init__.py`
-- Create: `tests/e2e/conftest.py`
-
-**Step 1: Write conftest.py**
-
-```python
-"""E2E test fixtures — manages docker-compose lifecycle."""
-
-import os
-import secrets
-import subprocess
-import time
-
-import httpx
-import pytest
-
-WARREN_URL = os.environ.get("WARREN_E2E_URL", "http://localhost:14030")
-COMPOSE_FILES = [
-    "docker-compose.yml",
-    "tests/e2e/docker-compose.test.yml",
-]
-PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
-
-
-def _compose_cmd(*args: str) -> list[str]:
-    """Build a docker compose command with the right -f flags."""
-    cmd = ["docker", "compose"]
-    for f in COMPOSE_FILES:
-        cmd.extend(["-f", os.path.join(PROJECT_ROOT, f)])
-    cmd.extend(args)
-    return cmd
-
-
-@pytest.fixture(scope="session")
-def compose_stack():
-    """Start docker-compose stack for the test session, tear down after."""
-    # Build and start
-    subprocess.run(_compose_cmd("build"), check=True, cwd=PROJECT_ROOT)
-    subprocess.run(
-        _compose_cmd("up", "-d", "--wait"),
-        check=True,
-        cwd=PROJECT_ROOT,
-    )
-
-    # Wait for Warren health endpoint
-    deadline = time.time() + 60
-    while time.time() < deadline:
-        try:
-            resp = httpx.get(f"{WARREN_URL}/health", timeout=2)
-            if resp.status_code == 200:
-                break
-        except httpx.ConnectError:
-            pass
-        time.sleep(1)
-    else:
-        # Dump logs on failure
-        subprocess.run(_compose_cmd("logs"), cwd=PROJECT_ROOT)
-        pytest.fail("Warren did not become healthy within 60 seconds")
-
-    yield WARREN_URL
-
-    # Tear down
-    subprocess.run(_compose_cmd("down", "-v"), cwd=PROJECT_ROOT)
-
-
-@pytest.fixture
-def warren_url(compose_stack) -> str:
-    """Return the Warren base URL."""
-    return compose_stack
-
-
-@pytest.fixture
-def admin_headers() -> dict:
-    """Admin auth headers for test endpoints."""
-    return {"X-Admin-Token": "test-admin-token"}
-
-
-@pytest.fixture
-def device_token(warren_url, admin_headers) -> str:
-    """Create a device token via pairing flow."""
-    resp = httpx.get(f"{warren_url}/pair", headers=admin_headers)
-    assert resp.status_code == 200
-    pairing_token = resp.json()["pairing_token"]
-
-    resp = httpx.post(
-        f"{warren_url}/pair/confirm",
-        json={"pairing_token": pairing_token, "device_name": "e2e-test"},
-    )
-    assert resp.status_code == 200
-    return resp.json()["device_token"]
-
-
-@pytest.fixture
-def auth_headers(device_token) -> dict:
-    """Device auth headers."""
-    return {"Authorization": f"Bearer {device_token}"}
-```
-
-Also create the empty `__init__.py`:
-
-```python
-```
-
-**Step 2: Commit**
-
-```bash
-git add tests/e2e/ && git commit -m "test: add E2E conftest with docker-compose lifecycle"
-```
-
----
-
-### Task 9: Create voice PTT E2E tests
-
-**Files:**
-- Create: `tests/e2e/test_voice_e2e.py`
-
-**Step 1: Write voice E2E tests**
-
-```python
-"""E2E tests for R1 voice PTT interface (OpenClaw WebSocket protocol)."""
-
-import json
-
-import pytest
-import websockets.sync.client as ws_client
-
-
-def _ws_url(warren_url: str) -> str:
-    """Convert http://host:port to ws://host:port/."""
-    return warren_url.replace("http://", "ws://") + "/"
-
-
-class TestVoiceHandshake:
-    def test_challenge_on_connect(self, warren_url):
-        """Server sends connect.challenge immediately after WebSocket open."""
-        with ws_client.connect(_ws_url(warren_url)) as ws:
-            msg = json.loads(ws.recv(timeout=5))
-            assert msg["type"] == "event"
-            assert msg["event"] == "connect.challenge"
-            assert "nonce" in msg["payload"]
-
-    def test_full_handshake(self, warren_url, device_token):
-        """Challenge -> connect -> hello-ok with valid token."""
-        with ws_client.connect(_ws_url(warren_url)) as ws:
-            challenge = json.loads(ws.recv(timeout=5))
-            nonce = challenge["payload"]["nonce"]
-
-            ws.send(json.dumps({
-                "type": "req",
-                "id": "connect-1",
-                "method": "connect",
-                "params": {
-                    "minProtocol": 3,
-                    "maxProtocol": 3,
-                    "auth": {"token": device_token},
-                    "client": {"id": "e2e-test", "version": "1.0", "platform": "test"},
-                    "role": "node",
-                    "device": {"id": "e2e-device", "nonce": nonce},
-                },
-            }))
-
-            hello = json.loads(ws.recv(timeout=5))
-            assert hello["type"] == "res"
-            assert hello["ok"] is True
-            assert hello["payload"]["type"] == "hello-ok"
-
-    def test_invalid_token_rejected(self, warren_url):
-        """Connect with bad token returns hello-error."""
-        with ws_client.connect(_ws_url(warren_url)) as ws:
-            challenge = json.loads(ws.recv(timeout=5))
-            nonce = challenge["payload"]["nonce"]
-
-            ws.send(json.dumps({
-                "type": "req",
-                "id": "connect-1",
-                "method": "connect",
-                "params": {
-                    "minProtocol": 3,
-                    "maxProtocol": 3,
-                    "auth": {"token": "bad-token"},
-                    "client": {"id": "e2e-test", "version": "1.0", "platform": "test"},
-                    "role": "node",
-                    "device": {"id": "e2e-device", "nonce": nonce},
-                },
-            }))
-
-            hello = json.loads(ws.recv(timeout=5))
-            assert hello["type"] == "res"
-            assert hello["ok"] is False
-
-
-class TestVoiceChatSend:
-    def _handshake(self, ws, device_token: str) -> None:
-        """Helper: complete handshake."""
-        challenge = json.loads(ws.recv(timeout=5))
-        nonce = challenge["payload"]["nonce"]
-        ws.send(json.dumps({
-            "type": "req",
-            "id": "connect-1",
-            "method": "connect",
-            "params": {
-                "minProtocol": 3,
-                "maxProtocol": 3,
-                "auth": {"token": device_token},
-                "client": {"id": "e2e-test", "version": "1.0", "platform": "test"},
-                "role": "node",
-                "device": {"id": "e2e-device", "nonce": "test"},
-            },
-        }))
-        hello = json.loads(ws.recv(timeout=5))
-        assert hello["ok"] is True
-
-    def test_chat_send_gets_ack(self, warren_url, device_token):
-        """Sending chat.send returns a chat-ack response."""
-        with ws_client.connect(_ws_url(warren_url)) as ws:
-            self._handshake(ws, device_token)
-
-            ws.send(json.dumps({
-                "type": "req",
-                "id": "chat-1",
-                "method": "chat.send",
-                "params": {
-                    "sessionKey": "main",
-                    "message": "E2E test message",
-                    "idempotencyKey": "e2e-idem-1",
-                },
-            }))
-
-            ack = json.loads(ws.recv(timeout=10))
-            assert ack["type"] == "res"
-            assert ack["ok"] is True
-            assert ack["id"] == "chat-1"
-```
-
-**Step 2: Commit**
-
-```bash
-git add tests/e2e/test_voice_e2e.py && git commit -m "test: add voice PTT E2E tests"
-```
-
----
-
-### Task 10: Create Creation UI E2E tests with Playwright
-
-**Files:**
-- Create: `tests/e2e/test_creation_e2e.py`
-
-**Step 1: Install Playwright test dependency**
-
-```bash
-pip install playwright pytest-playwright
-playwright install chromium
-```
-
-**Step 2: Write Creation E2E tests**
-
-```python
-"""E2E tests for R1 Creation UI at 240x282 viewport."""
-
-import re
-
-import pytest
-from playwright.sync_api import Page, expect
-
-
-@pytest.fixture
-def r1_page(page: Page, warren_url, device_token) -> Page:
-    """Navigate to Creation UI at R1 viewport size."""
-    page.set_viewport_size({"width": 240, "height": 282})
-    page.goto(f"{warren_url}/creation/?token={device_token}")
-    return page
-
-
-class TestCreationLoads:
-    def test_session_list_renders(self, r1_page):
-        """Creation UI loads and shows session list or empty state."""
-        expect(r1_page.locator("#view-sessions")).to_be_visible()
-        # Either sessions or "NO SESSIONS" text
-        r1_page.wait_for_selector("#session-list", timeout=5000)
-
-    def test_connection_dot_green(self, r1_page):
-        """Connection indicator turns green when server is reachable."""
-        dot = r1_page.locator("#connection-dot")
-        expect(dot).to_have_class(re.compile(r"dot-green"))
-
-    def test_viewport_is_240x282(self, r1_page):
-        """Verify viewport matches R1 dimensions."""
-        size = r1_page.viewport_size
-        assert size["width"] == 240
-        assert size["height"] == 282
-
-
-class TestMonitorView:
-    def test_monitor_loads(self, r1_page):
-        """Clicking MON button shows the monitor view."""
-        r1_page.click("#btn-monitor")
-        expect(r1_page.locator("#view-monitor")).to_have_class(re.compile(r"active"))
-        expect(r1_page.locator("#monitor-health")).to_be_visible()
-
-    def test_monitor_shows_health(self, r1_page):
-        """Monitor view shows health status after refresh."""
-        r1_page.click("#btn-monitor")
-        r1_page.click("#btn-monitor-refresh")
-        r1_page.wait_for_timeout(1000)
-        # Should show warren health status
-        health = r1_page.locator("#monitor-health")
-        expect(health).to_contain_text("warren")
-```
-
-**Step 3: Commit**
-
-```bash
-git add tests/e2e/test_creation_e2e.py && git commit -m "test: add Creation UI E2E tests with Playwright"
-```
-
----
-
-### Task 11: Create full pipeline E2E test
-
-**Files:**
-- Create: `tests/e2e/test_pipeline_e2e.py`
-
-**Step 1: Write pipeline E2E test**
-
-This test verifies the full round-trip: voice message → NanoClaw → mock agent → callback → Warren → SSE/WebSocket.
-
-```python
-"""E2E test for full NanoClaw pipeline: voice → agent → callback → response."""
-
-import json
-import threading
-import time
-
-import httpx
-import websockets.sync.client as ws_client
-
-
-def _ws_url(warren_url: str) -> str:
-    return warren_url.replace("http://", "ws://") + "/"
-
-
-def _handshake(ws, device_token: str) -> None:
-    """Complete OpenClaw handshake."""
-    challenge = json.loads(ws.recv(timeout=5))
-    nonce = challenge["payload"]["nonce"]
-    ws.send(json.dumps({
-        "type": "req",
-        "id": "connect-1",
-        "method": "connect",
-        "params": {
-            "minProtocol": 3,
-            "maxProtocol": 3,
-            "auth": {"token": device_token},
-            "client": {"id": "e2e-test", "version": "1.0", "platform": "test"},
-            "role": "node",
-            "device": {"id": "e2e-device", "nonce": nonce},
-        },
-    }))
-    hello = json.loads(ws.recv(timeout=5))
-    assert hello["ok"] is True
-
-
-class TestFullPipeline:
-    def test_voice_message_round_trip(self, warren_url, device_token, auth_headers):
-        """Voice chat.send → NanoClaw processes → response arrives via WebSocket."""
-        with ws_client.connect(_ws_url(warren_url), close_timeout=5) as ws:
-            _handshake(ws, device_token)
-
-            # Send a message via voice
-            ws.send(json.dumps({
-                "type": "req",
-                "id": "chat-1",
-                "method": "chat.send",
-                "params": {
-                    "sessionKey": "main",
-                    "message": "Full pipeline test",
-                    "idempotencyKey": "pipeline-1",
-                },
-            }))
-
-            # Collect events (ack + chat events) with timeout
-            events = []
-            deadline = time.time() + 120  # 2 min for agent to complete
-            got_final = False
-
-            while time.time() < deadline and not got_final:
-                try:
-                    raw = ws.recv(timeout=10)
-                    msg = json.loads(raw)
-                    events.append(msg)
-
-                    # Check for final chat event
-                    if (
-                        msg.get("type") == "event"
-                        and msg.get("event") == "chat"
-                        and msg.get("payload", {}).get("state") == "final"
-                    ):
-                        got_final = True
-                except TimeoutError:
-                    continue
-
-            # Verify we got an ack
-            acks = [e for e in events if e.get("type") == "res" and e.get("id") == "chat-1"]
-            assert len(acks) == 1, f"Expected 1 ack, got {len(acks)}: {acks}"
-
-            # Verify we got at least one chat event
-            chat_events = [e for e in events if e.get("event") == "chat"]
-            assert len(chat_events) >= 1, f"Expected chat events, got: {events}"
-
-            # Verify final event has content
-            if got_final:
-                final = [e for e in chat_events if e["payload"].get("state") == "final"]
-                assert len(final) == 1
-                content = final[0]["payload"]["message"]
-                assert isinstance(content, dict)
-
-    def test_voice_message_visible_in_session_list(
-        self, warren_url, device_token, auth_headers,
-    ):
-        """After voice chat.send, session appears in HTTP session list."""
-        with ws_client.connect(_ws_url(warren_url), close_timeout=5) as ws:
-            _handshake(ws, device_token)
-
-            ws.send(json.dumps({
-                "type": "req",
-                "id": "chat-1",
-                "method": "chat.send",
-                "params": {
-                    "sessionKey": "main",
-                    "message": "Session visibility test",
-                    "idempotencyKey": "visibility-1",
-                },
-            }))
-
-            # Wait a moment for session creation
-            time.sleep(2)
-
-        # Check via HTTP API
-        resp = httpx.get(f"{warren_url}/sessions", headers=auth_headers)
-        assert resp.status_code == 200
-        sessions = resp.json()
-        assert len(sessions) >= 1
-```
-
-**Step 2: Commit**
-
-```bash
-git add tests/e2e/test_pipeline_e2e.py && git commit -m "test: add full pipeline E2E test"
-```
-
----
-
-### Task 12: Run full verification
-
-**Step 1: Lint**
-
-```bash
-cd server && uv run ruff check src/ tests/
-```
-
-Expected: Clean.
-
-**Step 2: Unit tests**
-
-```bash
-cd server && uv run pytest tests/ -v --ignore=tests/e2e
-```
-
-Expected: All pass.
-
-**Step 3: Verify no dead references**
-
-```bash
-grep -r "DirectAdapter\|ClaudeRunner\|claude_runner\|claude_bin\|fetchBackend\|backendType\|probe\.js" server/src/ server/tests/ creation/ --include="*.py" --include="*.js" --include="*.html" --include="*.ts"
-```
-
-Expected: No output (except CLAUDE.md or design docs).
-
-**Step 4: E2E tests (if docker available)**
-
-```bash
-cd /home/dakota/Work/github/dsr-restyn/warren && docker compose -f docker-compose.yml -f tests/e2e/docker-compose.test.yml up -d --build --wait
-cd server && uv run pytest tests/e2e/ -v
-docker compose -f docker-compose.yml -f tests/e2e/docker-compose.test.yml down -v
-```
-
-**Step 5: Final commit and push**
-
-```bash
-git add -A && git commit -m "chore: final cleanup and verification"
-git push
-```
diff --git a/docs/plans/2026-02-22-voice-channel-design.md b/docs/plans/2026-02-22-voice-channel-design.md
deleted file mode 100644
index 9de402d..0000000
--- a/docs/plans/2026-02-22-voice-channel-design.md
+++ /dev/null
@@ -1,146 +0,0 @@
-# Voice Channel Design
-
-## Problem
-
-The R1's `webkitSpeechRecognition` doesn't work in the Android WebView. The R1
-natively supports OpenClaw-compatible voice gateways: hardware PTT button
-triggers cloud STT, transcripts arrive over WebSocket as plain text, responses
-go back over the same WS for native TTS. No audio processing needed on our end.
-
-Currently, Warren only has one input path (HTTP from the Creation browser).
-Voice adds a second path that shares the same sessions.
-
-## Architecture
-
-Warren becomes a protocol translator. Both input paths feed the same sessions
-and adapter layer. NanoClaw doesn't know or care how the user is talking.
-
-```
-R1 Browser (Creation) ──HTTP/SSE──┐
-                                   ├── Warren ──adapter──▶ NanoClaw
-R1 PTT (Voice) ───────WebSocket───┘
-```
-
-Both paths read from the same event stream via a new pub/sub `SessionBus` that
-replaces the current single-consumer `event_queues` dict.
-
-## Components
-
-### 1. SessionBus
-
-**Problem:** `event_queues` is `dict[str, asyncio.Queue]` — one queue per
-session, one consumer. Voice needs a second consumer on the same session.
-
-**Solution:** `SessionBus` — publish/subscribe, multiple queues per session.
-
-```
-callback ──publish──▶ SessionBus ──▶ SSE subscriber queue
-                                  ──▶ Voice WS subscriber queue
-```
-
-- SSE handler calls `bus.subscribe(session_id)`, gets its own queue
-- Voice WS handler does the same
-- Callback endpoint calls `bus.publish(session_id, event)` — all subscribers get it
-- On disconnect, each handler calls `bus.unsubscribe()`
-
-**File:** `server/src/warren/bus.py` (~40 lines)
-
-### 2. Voice WebSocket Endpoint
-
-**Path:** `/ws/voice` on Warren's existing port (4030).
-
-Speaks a minimal subset of the OpenClaw protocol:
-
-| Step | Direction | Message |
-|------|-----------|---------|
-| 1 | S→C | `connect.challenge` with random nonce |
-| 2 | C→S | `connect` with `authToken` (device token) + nonce |
-| 3 | S→C | `hello-ok` or `hello-error` |
-| 4 | C→S | `chat.send` with `{message}` |
-| 5 | S→C | `chat` with `state: "delta"` (streaming) |
-| 6 | S→C | `chat` with `state: "final"` (done) |
-| - | S→C | `tick` keepalive every 15s |
-
-**Session binding:** Voice targets the device's most recently active
-non-terminated session. If none exists, creates one with the first configured
-repo when the first `chat.send` arrives.
-
-**Event filtering:** Voice only sends `text` and `result` events as TTS.
-Tool events and status updates are skipped (the user doesn't need to hear
-"Running Bash: git status").
-
-**File:** `server/src/warren/voice.py` (~200 lines)
-
-### 3. Voice Pairing
-
-Admin endpoint `GET /pair/voice/qr` generates a QR code in the format the R1
-expects for gateway pairing:
-
-```json
-{
-  "type": "clawdbot-gateway",
-  "version": 1,
-  "ips": ["<server_host>"],
-  "port": 4030,
-  "token": "<device_token>"
-}
-```
-
-Reuses Warren's existing device token system. The token in the QR is a real
-device token stored in the DB — same token used by Creation, same auth
-middleware.
-
-**Added to:** `server/src/warren/auth.py` (one function) and `app.py` (one route)
-
-### 4. Config
-
-Single new env var: `WARREN_VOICE_ENABLED` (default: `false`).
-
-When disabled, the WS endpoint rejects connections with 403. When enabled,
-accepts and runs the protocol.
-
-No new ports — voice runs on Warren's existing port 4030. FastAPI handles
-WebSocket upgrades natively.
-
-## What Doesn't Change
-
-- **NanoClaw:** Zero modifications. Doesn't know voice exists.
-- **Creation UI:** Unchanged. Voice messages appear in shared sessions.
-- **Adapter layer:** Voice calls the same `adapter.send()` as HTTP.
-- **Auth:** Reuses device tokens. No new auth mechanism.
-- **Docker networking:** No new ports. WS upgrade on existing port 4030.
-
-## Event Flow
-
-```
-Voice (PTT press):
-  R1 hardware STT ──▶ transcript text
-  R1 WS client ──────▶ chat.send {message: "check the tests"}
-  Warren voice.py ───▶ adapter.send(BackendMessage(kind="user_message",...))
-  NanoClaw ──────────▶ runs agent, streams callbacks
-  Warren callback ───▶ bus.publish(session_id, {type:"text", content:"..."})
-  bus ───────────────▶ SSE queue (Creation sees it)
-                     ▶ Voice WS queue (voice.py sends chat delta)
-  R1 native TTS ────▶ speaks the response
-
-Creation (typed):
-  R1 browser ────────▶ POST /sessions/{id}/msg
-  Warren app.py ─────▶ adapter.send(BackendMessage(kind="user_message",...))
-  NanoClaw ──────────▶ runs agent, streams callbacks
-  Warren callback ───▶ bus.publish(session_id, {type:"text", content:"..."})
-  bus ───────────────▶ SSE queue (Creation sees it)
-                     ▶ Voice WS queue (if connected, speaks it)
-```
-
-Both paths are symmetric. A voice message shows up in the Creation chat.
-A typed message plays through TTS if the voice WS is connected.
-
-## Open Questions
-
-1. **WS path:** The R1 firmware may expect the WebSocket at `/` rather than
-   `/ws/voice`. Need to test with a real device and adjust if needed.
-2. **TLS:** Traefik handles TLS termination for HTTPS. Need to confirm it also
-   handles `wss://` upgrade correctly (it should by default).
-3. **Text accumulation:** Should we accumulate text events into one final TTS
-   utterance, or stream deltas for real-time speech? Start with accumulating
-   (simpler, avoids choppy TTS), optimize later if latency is bad.
diff --git a/docs/plans/2026-02-22-voice-channel-plan.md b/docs/plans/2026-02-22-voice-channel-plan.md
deleted file mode 100644
index f7d8d1a..0000000
--- a/docs/plans/2026-02-22-voice-channel-plan.md
+++ /dev/null
@@ -1,1164 +0,0 @@
-# Voice Channel Implementation Plan
-
-> **For Claude:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
-
-**Goal:** Add a WebSocket voice endpoint to Warren that speaks the R1's OpenClaw protocol, sharing sessions with the existing Creation UI.
-
-**Architecture:** Warren gets a `SessionBus` (pub/sub replacing single-consumer event queues), a voice WebSocket handler speaking the OpenClaw protocol, and a pairing endpoint. NanoClaw is untouched — Warren translates the voice protocol into the same adapter calls the HTTP path uses.
-
-**Tech Stack:** Python 3.11+, FastAPI WebSocket, asyncio, pydantic, pytest, `websockets` (test client)
-
----
-
-### Task 1: SessionBus
-
-Replace the single-consumer `event_queues: dict[str, asyncio.Queue]` with a
-pub/sub bus that supports multiple subscribers per session.
-
-**Files:**
-- Create: `server/src/warren/bus.py`
-- Create: `server/tests/test_bus.py`
-
-**Step 1: Write the failing tests**
-
-```python
-# server/tests/test_bus.py
-"""Tests for SessionBus pub/sub."""
-
-import asyncio
-import pytest
-from warren.bus import SessionBus
-
-
-@pytest.fixture
-def bus():
-    return SessionBus()
-
-
-@pytest.mark.asyncio
-async def test_single_subscriber_receives_event(bus):
-    q = bus.subscribe("s1")
-    await bus.publish("s1", {"type": "text", "content": "hello"})
-    event = q.get_nowait()
-    assert event == {"type": "text", "content": "hello"}
-
-
-@pytest.mark.asyncio
-async def test_multiple_subscribers_receive_same_event(bus):
-    q1 = bus.subscribe("s1")
-    q2 = bus.subscribe("s1")
-    await bus.publish("s1", {"type": "text", "content": "hello"})
-    assert q1.get_nowait() == {"type": "text", "content": "hello"}
-    assert q2.get_nowait() == {"type": "text", "content": "hello"}
-
-
-@pytest.mark.asyncio
-async def test_unsubscribe_stops_delivery(bus):
-    q = bus.subscribe("s1")
-    bus.unsubscribe("s1", q)
-    await bus.publish("s1", {"type": "text", "content": "hello"})
-    assert q.empty()
-
-
-@pytest.mark.asyncio
-async def test_publish_to_nonexistent_session_is_noop(bus):
-    # Should not raise
-    await bus.publish("nonexistent", {"type": "text"})
-
-
-@pytest.mark.asyncio
-async def test_different_sessions_are_isolated(bus):
-    q1 = bus.subscribe("s1")
-    q2 = bus.subscribe("s2")
-    await bus.publish("s1", {"type": "text", "content": "for s1"})
-    assert not q1.empty()
-    assert q2.empty()
-
-
-@pytest.mark.asyncio
-async def test_has_subscribers(bus):
-    assert not bus.has_subscribers("s1")
-    q = bus.subscribe("s1")
-    assert bus.has_subscribers("s1")
-    bus.unsubscribe("s1", q)
-    assert not bus.has_subscribers("s1")
-
-
-@pytest.mark.asyncio
-async def test_remove_session(bus):
-    q = bus.subscribe("s1")
-    bus.remove_session("s1")
-    assert not bus.has_subscribers("s1")
-    # Queue still exists but won't receive new events
-    await bus.publish("s1", {"type": "text"})
-    assert q.empty()
-```
-
-**Step 2: Run tests to verify they fail**
-
-Run: `cd server && uv run pytest tests/test_bus.py -v`
-Expected: FAIL — `ModuleNotFoundError: No module named 'warren.bus'`
-
-**Step 3: Implement SessionBus**
-
-```python
-# server/src/warren/bus.py
-"""Pub/sub event bus for session events.
-
-Multiple consumers (SSE, voice WS) can subscribe to the same session.
-Each subscriber gets its own asyncio.Queue. Publishing broadcasts to all.
-"""
-
-from __future__ import annotations
-
-import asyncio
-
-
-class SessionBus:
-    def __init__(self) -> None:
-        self._subs: dict[str, list[asyncio.Queue]] = {}
-
-    def subscribe(self, session_id: str) -> asyncio.Queue:
-        """Create and return a new subscriber queue for a session."""
-        q: asyncio.Queue = asyncio.Queue()
-        self._subs.setdefault(session_id, []).append(q)
-        return q
-
-    def unsubscribe(self, session_id: str, q: asyncio.Queue) -> None:
-        """Remove a subscriber queue. Safe to call if already removed."""
-        subs = self._subs.get(session_id, [])
-        if q in subs:
-            subs.remove(q)
-        if not subs:
-            self._subs.pop(session_id, None)
-
-    async def publish(self, session_id: str, event: dict) -> None:
-        """Send an event to all subscribers of a session."""
-        for q in self._subs.get(session_id, []):
-            await q.put(event)
-
-    def has_subscribers(self, session_id: str) -> bool:
-        """Check if any queues are subscribed to a session."""
-        return bool(self._subs.get(session_id))
-
-    def remove_session(self, session_id: str) -> None:
-        """Remove all subscribers for a session."""
-        self._subs.pop(session_id, None)
-```
-
-**Step 4: Run tests to verify they pass**
-
-Run: `cd server && uv run pytest tests/test_bus.py -v`
-Expected: 7 passed
-
-**Step 5: Commit**
-
-```bash
-git add server/src/warren/bus.py server/tests/test_bus.py
-git commit -m "feat: add SessionBus pub/sub for multi-consumer event delivery"
-```
-
----
-
-### Task 2: Migrate app.py from event_queues to SessionBus
-
-Replace all `event_queues` usage in `app.py` with `SessionBus`. This is a
-refactor — behavior is identical from the outside, but now supports multiple
-subscribers.
-
-**Files:**
-- Modify: `server/src/warren/app.py`
-
-**Changes summary:**
-
-1. **Lifespan** (line 148-149): Replace `app.state.event_queues: dict[str, asyncio.Queue] = {}`
-   with `app.state.bus = SessionBus()`
-
-2. **event_pump** (line 85-116): Change `queue = event_queues.get(event.session_id)` /
-   `await queue.put(event.data)` to `await bus.publish(event.session_id, event.data)`.
-   The DB persistence logic stays the same but uses `bus.has_subscribers()` instead of
-   checking dict membership.
-
-3. **create_session** (line 477): Replace `queues[session_id] = asyncio.Queue()` with
-   `bus.subscribe(session_id)` — but we don't need the queue here yet. Actually, we just
-   need the session to exist in the bus. The SSE handler will subscribe when it connects.
-   Remove the queue creation entirely — SSE `stream_session` already creates/gets a queue.
-
-4. **send_message** (lines 521-522): Remove `if session_id not in queues: queues[session_id] = asyncio.Queue()`.
-   The bus handles this automatically when SSE subscribes.
-
-5. **stream_session** (lines 557-585): Replace queue lookup with `bus.subscribe()`. Add
-   cleanup via `bus.unsubscribe()` when the SSE generator exits.
-
-6. **stop_session** (lines 600-603): Replace `queue.put()` with `bus.publish()`.
-
-7. **delete_session** (lines 617-618): Replace `queues.pop()` with `bus.remove_session()`.
-
-8. **nanoclaw_callback** (lines 775-819): Replace all `queue = queues.get(session_id)` /
-   `await queue.put(...)` with `await bus.publish(session_id, ...)`. Remove queue existence
-   checks — publishing to a session with no subscribers is a safe no-op.
-
-9. **nanoclaw_typing** (lines 827-831): Same pattern — replace queue.put with bus.publish.
-
-**Step 1: Apply the refactor**
-
-In `app.py`, add import:
-```python
-from warren.bus import SessionBus
-```
-
-Change `event_pump` signature and body:
-```python
-async def event_pump(adapter, bus: SessionBus, db):
-    """Route backend events to session subscribers."""
-    async for event in adapter.events():
-        await bus.publish(event.session_id, event.data)
-        if event.kind == "result" and event.data.get("session_id"):
-            await db.update_claude_session_id(event.session_id, event.data["session_id"])
-
-        # Save assistant messages to database
-        event_type = event.data.get("type")
-        if event_type in ["text", "result", "tool"]:
-            content = ""
-            if event_type == "text":
-                content = event.data.get("content", "")
-            elif event_type == "result":
-                content = event.data.get("summary", "")
-            elif event_type == "tool":
-                tool = event.data.get("action", event.data.get("tool", ""))
-                detail = event.data.get(
-                    "file", event.data.get("cmd", event.data.get("summary", ""))
-                )
-                content = f"{tool} {detail}".strip()
-
-            if content:
-                await db.save_message(
-                    event.session_id,
-                    content,
-                    role="assistant",
-                    event_type=event_type,
-                    event_data=json.dumps(event.data),
-                )
-```
-
-Change lifespan:
-```python
-app.state.bus = SessionBus()
-
-pump_task = asyncio.create_task(
-    event_pump(adapter, app.state.bus, db)
-)
-```
-
-Change `create_session` — remove queue creation line (`queues[session_id] = asyncio.Queue()`).
-The bus doesn't need pre-registration.
-
-Change `send_message` — remove queue existence check block.
-
-Change `stream_session`:
-```python
-bus: SessionBus = app.state.bus
-queue = bus.subscribe(session_id)
-
-# Drain stale events (for backward compat with `fresh` param)
-if fresh:
-    while not queue.empty():
-        try:
-            queue.get_nowait()
-        except asyncio.QueueEmpty:
-            break
-
-async def event_generator():
-    try:
-        while True:
-            try:
-                event = await asyncio.wait_for(queue.get(), timeout=30.0)
-                yield {
-                    "event": event.get("type", "message"),
-                    "data": json.dumps(event),
-                }
-                if event.get("type") == "status" and event.get("state") == "waiting":
-                    break
-            except asyncio.TimeoutError:
-                yield {"event": "ping", "data": ""}
-    finally:
-        bus.unsubscribe(session_id, queue)
-
-return EventSourceResponse(event_generator())
-```
-
-Change `stop_session`:
-```python
-bus: SessionBus = app.state.bus
-await bus.publish(session_id, {"type": "status", "state": "waiting"})
-```
-
-Change `delete_session`:
-```python
-bus: SessionBus = app.state.bus
-bus.remove_session(session_id)
-```
-
-Change `nanoclaw_callback`:
-```python
-bus: SessionBus = request.app.state.bus
-# Replace all `queue = queues.get(session_id)` / `await queue.put(event_data)`
-# with `await bus.publish(session_id, event_data)`
-# Remove if-queue checks — publish to no subscribers is a no-op
-```
-
-Change `nanoclaw_typing`:
-```python
-bus: SessionBus = request.app.state.bus
-state = "working" if is_typing else "waiting"
-await bus.publish(session_id, {"type": "status", "state": state})
-```
-
-Change `monitor_agents`:
-```python
-# This returned list(queues.keys()). With SessionBus, expose _subs keys.
-# Add a method to SessionBus:
-def active_sessions(self) -> list[str]:
-    return list(self._subs.keys())
-```
-
-Then in the endpoint:
-```python
-bus: SessionBus = request.app.state.bus
-return {"active_sessions": bus.active_sessions()}
-```
-
-**Step 2: Run existing tests**
-
-Run: `cd server && uv run pytest tests/ -v`
-Expected: All existing tests still pass.
-
-Run: `cd server && uv run ruff check src/ tests/`
-Expected: Clean.
-
-**Step 3: Commit**
-
-```bash
-git add server/src/warren/app.py server/src/warren/bus.py
-git commit -m "refactor: replace event_queues with SessionBus pub/sub"
-```
-
----
-
-### Task 3: Voice Protocol Module
-
-Message types and builders for the OpenClaw voice protocol. Pure functions,
-no I/O — easy to test.
-
-**Files:**
-- Create: `server/src/warren/voice_protocol.py`
-- Create: `server/tests/test_voice_protocol.py`
-
-**Step 1: Write the failing tests**
-
-```python
-# server/tests/test_voice_protocol.py
-"""Tests for OpenClaw voice protocol message builders/parsers."""
-
-import json
-import pytest
-from warren.voice_protocol import (
-    build_challenge,
-    build_hello_ok,
-    build_hello_error,
-    build_chat_delta,
-    build_chat_final,
-    build_tick,
-    parse_client_message,
-)
-
-
-def test_build_challenge():
-    msg = build_challenge("abc123")
-    assert msg["type"] == "connect.challenge"
-    assert msg["nonce"] == "abc123"
-
-
-def test_build_hello_ok():
-    msg = build_hello_ok("device-1")
-    assert msg["type"] == "hello-ok"
-    assert msg["deviceId"] == "device-1"
-
-
-def test_build_hello_error():
-    msg = build_hello_error("bad token")
-    assert msg["type"] == "hello-error"
-    assert msg["reason"] == "bad token"
-
-
-def test_build_chat_delta():
-    msg = build_chat_delta("partial text")
-    assert msg["type"] == "chat"
-    assert msg["data"]["state"] == "delta"
-    assert msg["data"]["text"] == "partial text"
-
-
-def test_build_chat_final():
-    msg = build_chat_final("full response")
-    assert msg["type"] == "chat"
-    assert msg["data"]["state"] == "final"
-    assert msg["data"]["text"] == "full response"
-
-
-def test_build_tick():
-    msg = build_tick()
-    assert msg["type"] == "tick"
-    assert "timestamp" in msg
-
-
-def test_parse_connect_message():
-    raw = json.dumps({
-        "type": "connect",
-        "authToken": "tok123",
-        "nonce": "abc",
-        "protocol": "1.0",
-    })
-    parsed = parse_client_message(raw)
-    assert parsed["type"] == "connect"
-    assert parsed["authToken"] == "tok123"
-    assert parsed["nonce"] == "abc"
-
-
-def test_parse_chat_send():
-    raw = json.dumps({
-        "type": "chat.send",
-        "data": {"message": "hello world"},
-    })
-    parsed = parse_client_message(raw)
-    assert parsed["type"] == "chat.send"
-    assert parsed["data"]["message"] == "hello world"
-
-
-def test_parse_invalid_json():
-    with pytest.raises(ValueError, match="Invalid message"):
-        parse_client_message("not json {{{")
-
-
-def test_parse_missing_type():
-    with pytest.raises(ValueError, match="Missing 'type'"):
-        parse_client_message(json.dumps({"data": "no type"}))
-```
-
-**Step 2: Run tests to verify they fail**
-
-Run: `cd server && uv run pytest tests/test_voice_protocol.py -v`
-Expected: FAIL — `ModuleNotFoundError`
-
-**Step 3: Implement the protocol module**
-
-```python
-# server/src/warren/voice_protocol.py
-"""OpenClaw voice protocol message builders and parsers.
-
-Implements the minimal subset of the OpenClaw WebSocket protocol needed
-for R1 PTT voice input/output. Pure functions, no I/O.
-
-Protocol flow:
-  S→C: connect.challenge {nonce}
-  C→S: connect {authToken, nonce, protocol}
-  S→C: hello-ok {deviceId} | hello-error {reason}
-  C→S: chat.send {data: {message}}
-  S→C: chat {data: {state: "delta"|"final", text}}
-  S→C: tick {timestamp}
-"""
-
-from __future__ import annotations
-
-import json
-import time
-
-
-def build_challenge(nonce: str) -> dict:
-    return {"type": "connect.challenge", "nonce": nonce}
-
-
-def build_hello_ok(device_id: str) -> dict:
-    return {"type": "hello-ok", "deviceId": device_id}
-
-
-def build_hello_error(reason: str) -> dict:
-    return {"type": "hello-error", "reason": reason}
-
-
-def build_chat_delta(text: str) -> dict:
-    return {"type": "chat", "data": {"state": "delta", "text": text}}
-
-
-def build_chat_final(text: str) -> dict:
-    return {"type": "chat", "data": {"state": "final", "text": text}}
-
-
-def build_tick() -> dict:
-    return {"type": "tick", "timestamp": int(time.time() * 1000)}
-
-
-def parse_client_message(raw: str) -> dict:
-    """Parse a client WebSocket message. Raises ValueError on bad input."""
-    try:
-        msg = json.loads(raw)
-    except json.JSONDecodeError as e:
-        raise ValueError(f"Invalid message: {e}") from e
-    if "type" not in msg:
-        raise ValueError("Missing 'type' field in message")
-    return msg
-```
-
-**Step 4: Run tests to verify they pass**
-
-Run: `cd server && uv run pytest tests/test_voice_protocol.py -v`
-Expected: 10 passed
-
-**Step 5: Commit**
-
-```bash
-git add server/src/warren/voice_protocol.py server/tests/test_voice_protocol.py
-git commit -m "feat: add OpenClaw voice protocol message builders/parsers"
-```
-
----
-
-### Task 4: Voice WebSocket Handler
-
-The core voice endpoint. Handles the OpenClaw handshake, routes `chat.send`
-messages to sessions via the adapter, and streams responses back as `chat`
-events for TTS.
-
-**Files:**
-- Create: `server/src/warren/voice.py`
-- Create: `server/tests/test_voice.py`
-
-**Step 1: Write the failing tests**
-
-```python
-# server/tests/test_voice.py
-"""Tests for voice WebSocket handler."""
-
-import asyncio
-import json
-import pytest
-from unittest.mock import AsyncMock, MagicMock, patch
-from fastapi.testclient import TestClient
-from starlette.testclient import TestClient as StarletteTestClient
-
-from warren.app import create_app
-from warren.config import Settings
-
-
-@pytest.fixture
-def settings(tmp_path):
-    db_path = str(tmp_path / "test.db")
-    config_path = str(tmp_path / "config.yaml")
-    (tmp_path / "config.yaml").write_text("repos: {}")
-    return Settings(
-        db=db_path,
-        config=config_path,
-        backend="direct",
-        admin_token="admin123",
-        voice_enabled=True,
-    )
-
-
-@pytest.fixture
-def app(settings):
-    return create_app(settings)
-
-
-@pytest.fixture
-def client(app):
-    return TestClient(app)
-
-
-def test_voice_disabled_rejects_connection(tmp_path):
-    """When voice_enabled=False, WS connections are rejected."""
-    db_path = str(tmp_path / "test.db")
-    config_path = str(tmp_path / "config.yaml")
-    (tmp_path / "config.yaml").write_text("repos: {}")
-    s = Settings(
-        db=db_path,
-        config=config_path,
-        backend="direct",
-        admin_token="admin123",
-        voice_enabled=False,
-    )
-    app = create_app(s)
-    client = TestClient(app)
-    with pytest.raises(Exception):
-        with client.websocket_connect("/ws/voice"):
-            pass
-
-
-def test_handshake_sends_challenge(client):
-    """Server sends connect.challenge immediately on connection."""
-    with client.websocket_connect("/ws/voice") as ws:
-        msg = json.loads(ws.receive_text())
-        assert msg["type"] == "connect.challenge"
-        assert "nonce" in msg
-
-
-def test_handshake_valid_token(client):
-    """Valid device token completes handshake with hello-ok."""
-    # Create a device token via the admin API
-    resp = client.get("/pair", headers={"x-admin-token": "admin123"})
-    pairing_token = resp.json()["pairing_token"]
-    resp = client.post("/pair/confirm", json={
-        "pairing_token": pairing_token,
-        "device_name": "test-voice",
-    })
-    device_token = resp.json()["device_token"]
-
-    with client.websocket_connect("/ws/voice") as ws:
-        challenge = json.loads(ws.receive_text())
-        nonce = challenge["nonce"]
-        ws.send_text(json.dumps({
-            "type": "connect",
-            "authToken": device_token,
-            "nonce": nonce,
-            "protocol": "1.0",
-        }))
-        hello = json.loads(ws.receive_text())
-        assert hello["type"] == "hello-ok"
-
-
-def test_handshake_invalid_token(client):
-    """Invalid token gets hello-error and connection closes."""
-    with client.websocket_connect("/ws/voice") as ws:
-        challenge = json.loads(ws.receive_text())
-        ws.send_text(json.dumps({
-            "type": "connect",
-            "authToken": "bad-token",
-            "nonce": challenge["nonce"],
-            "protocol": "1.0",
-        }))
-        hello = json.loads(ws.receive_text())
-        assert hello["type"] == "hello-error"
-```
-
-**Step 2: Run tests to verify they fail**
-
-Run: `cd server && uv run pytest tests/test_voice.py -v`
-Expected: FAIL — voice_enabled not a valid Setting field / no /ws/voice route
-
-**Step 3: Add `voice_enabled` to Settings**
-
-In `server/src/warren/config.py`, add to `Settings`:
-
-```python
-voice_enabled: bool = False
-```
-
-**Step 4: Implement the voice handler**
-
-```python
-# server/src/warren/voice.py
-"""Voice WebSocket handler — OpenClaw protocol for R1 PTT.
-
-Handles the handshake, routes chat.send messages to the active session
-via the adapter, and streams responses back as chat events for TTS.
-"""
-
-from __future__ import annotations
-
-import asyncio
-import json
-import logging
-import secrets
-
-from fastapi import WebSocket, WebSocketDisconnect
-
-from warren.bus import SessionBus
-from warren.db import Database
-from warren.voice_protocol import (
-    build_challenge,
-    build_chat_delta,
-    build_chat_final,
-    build_hello_error,
-    build_hello_ok,
-    build_tick,
-    parse_client_message,
-)
-
-logger = logging.getLogger(__name__)
-
-# How often to send keepalive ticks (seconds)
-TICK_INTERVAL = 15
-
-
-async def voice_handler(
-    ws: WebSocket,
-    bus: SessionBus,
-    db: Database,
-    adapter,
-    config,
-    voice_enabled: bool,
-) -> None:
-    """Main voice WebSocket handler. Called from the app route."""
-    if not voice_enabled:
-        await ws.close(code=4003, reason="Voice disabled")
-        return
-
-    await ws.accept()
-
-    # --- Handshake ---
-    nonce = secrets.token_hex(16)
-    await ws.send_text(json.dumps(build_challenge(nonce)))
-
-    try:
-        raw = await asyncio.wait_for(ws.receive_text(), timeout=10.0)
-    except (asyncio.TimeoutError, WebSocketDisconnect):
-        return
-
-    try:
-        msg = parse_client_message(raw)
-    except ValueError:
-        await ws.send_text(json.dumps(build_hello_error("invalid message")))
-        await ws.close()
-        return
-
-    if msg.get("type") != "connect":
-        await ws.send_text(json.dumps(build_hello_error("expected connect")))
-        await ws.close()
-        return
-
-    token = msg.get("authToken", "")
-    if not await db.validate_device_token(token):
-        await ws.send_text(json.dumps(build_hello_error("invalid token")))
-        await ws.close()
-        return
-
-    device_id = token[:8]
-    await ws.send_text(json.dumps(build_hello_ok(device_id)))
-    logger.info("Voice client authenticated: %s", device_id)
-
-    # --- Connected session ---
-    # Track the session this voice connection is bound to
-    active_session_id: str | None = None
-    event_queue: asyncio.Queue | None = None
-
-    async def find_or_create_session(message: str) -> str:
-        """Find the most recent active session or create one."""
-        nonlocal active_session_id, event_queue
-
-        # If already bound to a session, reuse it
-        if active_session_id:
-            session = await db.get_session(active_session_id)
-            if session and session.get("status") != "terminated":
-                return active_session_id
-
-        # Find most recent non-terminated session
-        sessions = await db.list_sessions()
-        if sessions:
-            active_session_id = sessions[0]["id"]
-        else:
-            # Create a new session with the first configured repo
-            import uuid
-
-            session_id = uuid.uuid4().hex
-            repo_name = ""
-            if config.repos:
-                repo_name = next(iter(config.repos))
-            await db.create_session(session_id, "voice session", repo_name)
-            active_session_id = session_id
-
-        # Subscribe to events for this session
-        if event_queue:
-            bus.unsubscribe(active_session_id, event_queue)
-        event_queue = bus.subscribe(active_session_id)
-        return active_session_id
-
-    async def handle_chat_send(msg: dict) -> None:
-        """Route a voice message to the active session."""
-        data = msg.get("data", {})
-        message = data.get("message", "")
-        if not message:
-            return
-
-        session_id = await find_or_create_session(message)
-        session = await db.get_session(session_id)
-
-        repo_path = ""
-        model = ""
-        append_system_prompt = ""
-        repo_name = session.get("repo", "") if session else ""
-
-        if repo_name and repo_name in config.repos:
-            repo_config = config.repos[repo_name]
-            repo_path = repo_config.path
-            model = repo_config.model
-            append_system_prompt = repo_config.append_system_prompt
-
-        await db.touch(session_id)
-        await db.save_message(session_id, message)
-        await db.update_status(session_id, "active")
-
-        claude_sid = session.get("claude_session_id") if session else None
-
-        from warren.adapters import BackendMessage
-
-        await adapter.send(
-            BackendMessage(
-                kind="user_message" if claude_sid else "new_session",
-                session_id=session_id,
-                payload={
-                    "message": message,
-                    "repo_path": repo_path,
-                    "resume_session_id": claude_sid,
-                    "model": model,
-                    "append_system_prompt": append_system_prompt,
-                },
-            )
-        )
-
-    async def tick_loop():
-        """Send keepalive ticks."""
-        try:
-            while True:
-                await asyncio.sleep(TICK_INTERVAL)
-                await ws.send_text(json.dumps(build_tick()))
-        except (WebSocketDisconnect, Exception):
-            pass
-
-    async def event_relay():
-        """Relay session events to voice client as chat messages."""
-        nonlocal event_queue
-        accumulated_text = ""
-        try:
-            while True:
-                if not event_queue:
-                    await asyncio.sleep(0.1)
-                    continue
-                try:
-                    event = await asyncio.wait_for(event_queue.get(), timeout=1.0)
-                except asyncio.TimeoutError:
-                    continue
-
-                event_type = event.get("type")
-                if event_type == "text":
-                    content = event.get("content", "")
-                    if content:
-                        accumulated_text += content
-                        await ws.send_text(json.dumps(build_chat_delta(content)))
-                elif event_type == "result":
-                    summary = event.get("summary", accumulated_text)
-                    await ws.send_text(json.dumps(build_chat_final(summary or accumulated_text)))
-                    accumulated_text = ""
-                # Skip tool and status events — not useful for TTS
-        except (WebSocketDisconnect, Exception):
-            pass
-
-    # --- Main loop ---
-    tick_task = asyncio.create_task(tick_loop())
-    relay_task = asyncio.create_task(event_relay())
-
-    try:
-        while True:
-            raw = await ws.receive_text()
-            try:
-                msg = parse_client_message(raw)
-            except ValueError:
-                continue
-
-            if msg["type"] == "chat.send":
-                await handle_chat_send(msg)
-    except WebSocketDisconnect:
-        logger.info("Voice client disconnected: %s", device_id)
-    finally:
-        tick_task.cancel()
-        relay_task.cancel()
-        if event_queue and active_session_id:
-            bus.unsubscribe(active_session_id, event_queue)
-```
-
-**Step 5: Run tests**
-
-Run: `cd server && uv run pytest tests/test_voice.py -v`
-Expected: All pass (note: test_voice_disabled may need adjustment based on how
-FastAPI handles WebSocket rejection — may raise a different exception type)
-
-**Step 6: Run all tests + lint**
-
-Run: `cd server && uv run pytest tests/ -v && uv run ruff check src/ tests/`
-Expected: All pass, clean.
-
-**Step 7: Commit**
-
-```bash
-git add server/src/warren/voice.py server/tests/test_voice.py server/src/warren/config.py
-git commit -m "feat: add voice WebSocket handler with OpenClaw protocol"
-```
-
----
-
-### Task 5: Voice Pairing + Wire into app.py
-
-Add the voice pairing QR endpoint and mount the voice WebSocket route.
-
-**Files:**
-- Modify: `server/src/warren/auth.py` — add `generate_voice_qr_png()`
-- Modify: `server/src/warren/app.py` — add `/ws/voice` route + `/pair/voice/qr` endpoint
-
-**Step 1: Add voice QR generator to auth.py**
-
-```python
-# Add to server/src/warren/auth.py
-
-def generate_voice_qr_png(server_host: str, port: int, device_token: str) -> bytes:
-    """Generate R1 voice gateway pairing QR code.
-
-    Encodes connection info in the format the R1 expects for
-    OpenClaw-compatible gateway pairing.
-    """
-    payload = json.dumps(
-        {
-            "type": "clawdbot-gateway",
-            "version": 1,
-            "ips": [server_host],
-            "port": port,
-            "token": device_token,
-        }
-    )
-    img = qrcode.make(payload)
-    buf = io.BytesIO()
-    img.save(buf, format="PNG")
-    return buf.getvalue()
-```
-
-**Step 2: Add routes to app.py**
-
-Add import:
-```python
-from fastapi import WebSocket
-from warren.voice import voice_handler
-from warren.auth import generate_voice_qr_png
-```
-
-Add WebSocket route (inside `create_app`, after the other routes):
-```python
-@app.websocket("/ws/voice")
-async def voice_ws(websocket: WebSocket):
-    await voice_handler(
-        ws=websocket,
-        bus=app.state.bus,
-        db=app.state.db,
-        adapter=app.state.adapter,
-        config=app.state.config,
-        voice_enabled=settings.voice_enabled,
-    )
-```
-
-Add pairing endpoint (near the other `/pair` routes):
-```python
-@app.get("/pair/voice/qr", dependencies=[Depends(require_admin)])
-async def pair_voice_qr(request: Request, db: Database = Depends(get_db)):
-    """Generate QR code for R1 voice gateway pairing."""
-    if not settings.voice_enabled:
-        raise HTTPException(status_code=400, detail="Voice not enabled")
-    device_token = secrets.token_hex(32)
-    await db.store_device_token(device_token, "R1 Voice")
-    server_host = request.base_url.hostname
-    png = generate_voice_qr_png(server_host, settings.port, device_token)
-    return Response(content=png, media_type="image/png")
-```
-
-**Step 3: Run all tests + lint**
-
-Run: `cd server && uv run pytest tests/ -v && uv run ruff check src/ tests/`
-Expected: All pass, clean.
-
-**Step 4: Commit**
-
-```bash
-git add server/src/warren/auth.py server/src/warren/app.py
-git commit -m "feat: add voice pairing QR + mount /ws/voice route"
-```
-
----
-
-### Task 6: Config + Docker
-
-Add the `WARREN_VOICE_ENABLED` environment variable to docker-compose and
-update CLAUDE.md.
-
-**Files:**
-- Modify: `docker-compose.yml`
-- Modify: `CLAUDE.md`
-
-**Step 1: Add env var to docker-compose.yml**
-
-In the `warren` service `environment` section, add:
-```yaml
-- WARREN_VOICE_ENABLED=${WARREN_VOICE_ENABLED:-false}
-```
-
-**Step 2: Update CLAUDE.md**
-
-Add to the SSE Event Types section or a new "Voice" section:
-```markdown
-## Voice Channel
-
-Warren supports an optional WebSocket voice endpoint for R1 PTT interaction.
-Enable with `WARREN_VOICE_ENABLED=true`. Speaks the OpenClaw protocol at
-`/ws/voice`. Voice messages route to the same sessions as the Creation UI.
-
-Pair via `GET /pair/voice/qr` (admin). The QR encodes gateway connection info
-for the R1's native voice assistant pairing.
-```
-
-**Step 3: Commit**
-
-```bash
-git add docker-compose.yml CLAUDE.md
-git commit -m "feat: add voice channel config and documentation"
-```
-
----
-
-### Task 7: Integration Test
-
-WebSocket integration test using `websockets` library to verify the full
-handshake and chat flow.
-
-**Files:**
-- Create: `server/tests/test_voice_integration.py`
-
-**Step 1: Write integration test**
-
-```python
-# server/tests/test_voice_integration.py
-"""Integration test for voice WebSocket — full handshake + chat flow."""
-
-import asyncio
-import json
-import pytest
-from fastapi.testclient import TestClient
-
-from warren.app import create_app
-from warren.config import Settings
-
-
-@pytest.fixture
-def settings(tmp_path):
-    db_path = str(tmp_path / "test.db")
-    config_path = str(tmp_path / "config.yaml")
-    (tmp_path / "config.yaml").write_text("repos:\n  test:\n    path: /tmp/test")
-    return Settings(
-        db=db_path,
-        config=config_path,
-        backend="direct",
-        admin_token="admin123",
-        voice_enabled=True,
-    )
-
-
-@pytest.fixture
-def app(settings):
-    return create_app(settings)
-
-
-@pytest.fixture
-def client(app):
-    return TestClient(app)
-
-
-def _get_device_token(client) -> str:
-    """Helper: create a device token via pairing flow."""
-    resp = client.get("/pair", headers={"x-admin-token": "admin123"})
-    pairing_token = resp.json()["pairing_token"]
-    resp = client.post("/pair/confirm", json={
-        "pairing_token": pairing_token,
-        "device_name": "test-voice",
-    })
-    return resp.json()["device_token"]
-
-
-def _do_handshake(ws, device_token: str) -> dict:
-    """Helper: complete the OpenClaw handshake."""
-    challenge = json.loads(ws.receive_text())
-    assert challenge["type"] == "connect.challenge"
-    ws.send_text(json.dumps({
-        "type": "connect",
-        "authToken": device_token,
-        "nonce": challenge["nonce"],
-        "protocol": "1.0",
-    }))
-    hello = json.loads(ws.receive_text())
-    assert hello["type"] == "hello-ok"
-    return hello
-
-
-def test_full_handshake_flow(client):
-    """Complete handshake: challenge → connect → hello-ok."""
-    token = _get_device_token(client)
-    with client.websocket_connect("/ws/voice") as ws:
-        _do_handshake(ws, token)
-
-
-def test_chat_send_creates_session(client):
-    """Sending chat.send when no sessions exist creates one."""
-    token = _get_device_token(client)
-
-    # Verify no sessions exist
-    resp = client.get("/sessions", headers={"Authorization": f"Bearer {token}"})
-    assert resp.json() == []
-
-    with client.websocket_connect("/ws/voice") as ws:
-        _do_handshake(ws, token)
-        ws.send_text(json.dumps({
-            "type": "chat.send",
-            "data": {"message": "hello from voice"},
-        }))
-        # Give the handler a moment to process
-        # (In test, DirectAdapter may not produce output, but the session
-        # should be created)
-
-    # Session should now exist
-    resp = client.get("/sessions", headers={"Authorization": f"Bearer {token}"})
-    sessions = resp.json()
-    assert len(sessions) >= 1
-
-
-def test_voice_pairing_qr(client):
-    """GET /pair/voice/qr returns a PNG image."""
-    resp = client.get("/pair/voice/qr", headers={"x-admin-token": "admin123"})
-    assert resp.status_code == 200
-    assert resp.headers["content-type"] == "image/png"
-    # PNG magic bytes
-    assert resp.content[:4] == b"\x89PNG"
-```
-
-**Step 2: Run integration tests**
-
-Run: `cd server && uv run pytest tests/test_voice_integration.py -v`
-Expected: All pass.
-
-**Step 3: Run full test suite + lint**
-
-Run: `cd server && uv run pytest tests/ -v && uv run ruff check src/ tests/`
-Expected: All pass, clean.
-
-**Step 4: Commit**
-
-```bash
-git add server/tests/test_voice_integration.py
-git commit -m "test: add voice WebSocket integration tests"
-```
-
----
-
-## Verification Checklist
-
-After all tasks:
-
-1. `cd server && uv run ruff check src/ tests/` — lint clean
-2. `cd server && uv run pytest tests/ -v` — all tests pass
-3. `docker compose build warren` — builds clean
-4. Deploy with `WARREN_VOICE_ENABLED=true`
-5. `GET /pair/voice/qr` — generates QR PNG
-6. Scan QR with R1 — confirms pairing
-7. Press PTT on R1 — voice transcript appears in Warren session
-8. Response plays back via R1 TTS
-9. Same session visible in Creation browser
diff --git a/docs/superpowers/plans/2026-06-18-remove-nanoclaw.md b/docs/superpowers/plans/2026-06-18-remove-nanoclaw.md
deleted file mode 100644
index cfc12a5..0000000
--- a/docs/superpowers/plans/2026-06-18-remove-nanoclaw.md
+++ /dev/null
@@ -1,413 +0,0 @@
-# Remove NanoClaw Implementation Plan
-
-> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
-
-**Goal:** Remove all NanoClaw-specific code, config, tests, the git submodule, and docs from Warren, leaving a clean WebSocket-proxy + Orca backend architecture.
-
-**Architecture:** Warren keeps two backends selected by `settings.backend`: `websocket` (default, `AgentWebSocketAdapter`) and `orca` (`OrcaAdapter`). The `nanoclaw` branch, file-IPC transport, read-only SQLite monitoring, and HTTP callback/typing endpoints are deleted. Endpoints the R1 frontend still calls (`/monitor/tasks`, pause/resume) are kept as safe no-ops so the UI does not 404.
-
-**Tech Stack:** Python 3, FastAPI, pytest, ruff, uv. Work happens in the worktree `/home/dakota/Work/github/dsr-restyn/warren` on branch `feature/decouple-and-orca-proxy`.
-
-## Global Constraints
-
-- All commands run from `/home/dakota/Work/github/dsr-restyn/warren` (the PR-branch worktree), not the refactor worktree.
-- Lint: `cd server && uv run ruff check src/ tests/` must pass before every commit.
-- Tests: `cd server && uv run pytest -q` must pass (green) at the end of every task — the suite never goes red between commits.
-- Keep `settings.internal_secret` (used by `AgentWebSocketAdapter` auth) and `settings.backend`. Remove only nanoclaw-specific fields.
-- Keep the three frontend-facing monitor endpoints reachable (return empty/no-op), since `creation/js/api.js:135-143` calls them.
-- One commit per task. Commit message trailer: `Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>`.
-- Do NOT touch the Orca adapter or the WebSocket adapter logic (out of scope; separate review).
-
----
-
-### Task 1: Remove NanoClaw internal HTTP endpoints + `require_internal`
-
-Removes the inbound callback/typing/config endpoints NanoClaw used, and the auth dependency that only guarded them.
-
-**Files:**
-- Modify: `server/src/warren/app.py` — delete `require_internal` (~740-749), `/internal/nanoclaw/config` GET+PUT (~751-781), `/internal/nanoclaw/callback` (~783-839), `/internal/nanoclaw/typing` (~841-850)
-- Modify: `server/tests/test_app.py` — delete `class TestInternalConfigEndpoints` (~409-472)
-- Delete: `server/tests/test_nanoclaw_callback.py`
-
-**Interfaces:**
-- Consumes: nothing from other tasks.
-- Produces: removes `require_internal` symbol — no later task may reference it.
-
-- [ ] **Step 1: Delete the four nanoclaw internal endpoints and `require_internal` from `app.py`**
-
-Remove the `async def require_internal(...)` dependency and all four route handlers decorated with `dependencies=[Depends(require_internal)]` (`/internal/nanoclaw/config` GET, `/internal/nanoclaw/config` PUT, `/internal/nanoclaw/callback`, `/internal/nanoclaw/typing`). Leave the `/api/agent/ws` and voice WebSocket routes that follow intact.
-
-- [ ] **Step 2: Delete the callback test file**
-
-```bash
-git rm server/tests/test_nanoclaw_callback.py
-```
-
-- [ ] **Step 3: Remove `TestInternalConfigEndpoints` from `test_app.py`**
-
-Delete the whole class (docstring `"Tests for /internal/nanoclaw/config (requires x-internal-secret)."`).
-
-- [ ] **Step 4: Verify the import for `secrets` is still needed**
-
-Run: `cd server && grep -n "secrets\." src/warren/app.py`
-Expected: still used by `require_admin`/token compare. If zero hits, remove `import secrets`.
-
-- [ ] **Step 5: Lint + test**
-
-Run: `cd server && uv run ruff check src/ tests/ && uv run pytest -q`
-Expected: PASS, no collection errors.
-
-- [ ] **Step 6: Commit**
-
-```bash
-git add -A && git commit -m "refactor: remove nanoclaw internal callback/config/typing endpoints"
-```
-
----
-
-### Task 2: Remove the NanoClaw adapter and swap test fixtures to WebSocket
-
-The largest task: deletes the adapter, its `create_adapter` branch and lifespan warning, and repoints every test fixture that instantiated `NanoClawAdapter` to `AgentWebSocketAdapter`.
-
-**Files:**
-- Modify: `server/src/warren/app.py` — remove nanoclaw branch in `create_adapter` (~73-77) and the `if settings.backend == "nanoclaw":` API-key warning block in lifespan (~136-145)
-- Delete: `server/src/warren/adapters/nanoclaw.py`
-- Modify: `server/tests/test_adapters.py` — delete `class TestNanoClawAdapter` (~34-110)
-- Modify: `server/tests/test_app.py`, `test_integration.py`, `test_streaming.py`, `test_voice.py`, `test_voice_integration.py`, `test_voice_pairing.py` — replace `NanoClawAdapter` fixtures with `AgentWebSocketAdapter`
-
-**Interfaces:**
-- Consumes: `AgentWebSocketAdapter(internal_secret="", admin_token="")` from `warren.adapters.websocket` — constructor takes no IPC dir, `start()`/`stop()` are no-ops, `send()` queues an error event if no agent connected (safe for tests that only check session/DB behavior).
-- Produces: `create_adapter` now returns only `OrcaAdapter` or `AgentWebSocketAdapter`.
-
-- [ ] **Step 1: Remove the nanoclaw branch from `create_adapter` in `app.py`**
-
-Delete the `if settings.backend == "nanoclaw": ...` block so the function starts at `if settings.backend == "orca":` then `else:` → websocket. (Leave the orca/else branches unchanged.)
-
-- [ ] **Step 2: Remove the nanoclaw API-key warning in lifespan**
-
-Delete the `if settings.backend == "nanoclaw":` block (lines ~136-145) that warns about `ANTHROPIC_API_KEY`/`CLAUDE_CODE_OAUTH_TOKEN`.
-
-- [ ] **Step 3: Delete the adapter file**
-
-```bash
-git rm server/src/warren/adapters/nanoclaw.py
-```
-
-- [ ] **Step 4: Delete `TestNanoClawAdapter` from `test_adapters.py`**
-
-Remove the entire class; keep the generic `BackendMessage`/`BackendEvent` tests.
-
-- [ ] **Step 5: Repoint test fixtures**
-
-In each of `test_app.py`, `test_integration.py`, `test_streaming.py`, `test_voice.py` (two fixtures), `test_voice_integration.py`, `test_voice_pairing.py`, replace the pattern:
-
-```python
-from warren.adapters.nanoclaw import NanoClawAdapter
-...
-adapter = NanoClawAdapter(ipc_dir=str(tmp_path / "ipc"))
-await adapter.start()
-```
-
-with:
-
-```python
-from warren.adapters.websocket import AgentWebSocketAdapter
-...
-adapter = AgentWebSocketAdapter()
-await adapter.start()
-```
-
-For `test_voice_integration.py`, keep the `adapter.send = AsyncMock()` line (still valid). Update any docstring/comment mentioning "NanoClaw backend".
-
-- [ ] **Step 6: Lint + test**
-
-Run: `cd server && uv run ruff check src/ tests/ && uv run pytest -q`
-Expected: PASS. Watch for any test asserting IPC-file side effects — those belonged to the deleted `TestNanoClawAdapter` and should already be gone.
-
-- [ ] **Step 7: Commit**
-
-```bash
-git add -A && git commit -m "refactor: remove NanoClawAdapter, default to websocket backend in tests"
-```
-
----
-
-### Task 3: Strip NanoClaw monitoring; keep frontend endpoints as no-ops
-
-Removes the SQLite reader and the nanoclaw health/task internals, but preserves the three endpoints the R1 UI calls.
-
-**Files:**
-- Modify: `server/src/warren/app.py` — remove nanoclaw block in `/monitor/health` (~632-641); simplify `/monitor/tasks` to return `{"tasks": []}`; make pause/resume no-ops; remove `_nanoclaw_paths` (~56-68) once its last caller is gone
-- Delete: `server/src/warren/nanoclaw_db.py`
-- Delete: `server/tests/test_nanoclaw_db.py`
-- Modify: `server/tests/test_monitor.py` — rename `test_returns_empty_when_no_nanoclaw` → `test_returns_empty_tasks` (~100-113)
-
-**Interfaces:**
-- Consumes: nothing new.
-- Produces: `_nanoclaw_paths` no longer exists — no later task may reference it.
-
-- [ ] **Step 1: Remove the nanoclaw branch in `/monitor/health`**
-
-Delete the `if settings.backend == "nanoclaw":` block that checks `os.path.exists(db_path)`. Keep the orca and websocket health branches.
-
-- [ ] **Step 2: Simplify `/monitor/tasks`**
-
-Replace the body with a backend-agnostic stub (frontend `getTasks()` still calls this):
-
-```python
-    @app.get("/monitor/tasks")
-    async def monitor_tasks(
-        request: Request,
-        _token: str = Depends(require_auth),
-    ):
-        # Live task monitoring was nanoclaw-specific; no tasks on websocket/orca.
-        return {"tasks": []}
-```
-
-- [ ] **Step 3: Make pause/resume no-ops (keep the routes — frontend calls them)**
-
-```python
-    @app.post("/monitor/tasks/{task_id}/pause")
-    async def pause_task(task_id: str, _token: str = Depends(require_auth)):
-        return {"status": "ignored", "reason": "Task control not supported on this backend"}
-
-    @app.post("/monitor/tasks/{task_id}/resume")
-    async def resume_task(task_id: str, _token: str = Depends(require_auth)):
-        return {"status": "ignored", "reason": "Task control not supported on this backend"}
-```
-
-- [ ] **Step 4: Remove `_nanoclaw_paths` from `app.py`**
-
-Confirm no remaining callers, then delete the function:
-
-Run: `cd server && grep -n "_nanoclaw_paths" src/warren/app.py`
-Expected: zero hits after Steps 1-3 → delete the `def _nanoclaw_paths(...)` definition.
-
-- [ ] **Step 5: Delete the SQLite reader and its test**
-
-```bash
-git rm server/src/warren/nanoclaw_db.py server/tests/test_nanoclaw_db.py
-```
-
-- [ ] **Step 6: Rename the monitor test**
-
-In `test_monitor.py`, rename `test_returns_empty_when_no_nanoclaw` to `test_returns_empty_tasks` and drop any nanoclaw wording. The assertion `data["tasks"] == []` stays valid.
-
-- [ ] **Step 7: Lint + test**
-
-Run: `cd server && uv run ruff check src/ tests/ && uv run pytest -q`
-Expected: PASS.
-
-- [ ] **Step 8: Commit**
-
-```bash
-git add -A && git commit -m "refactor: drop nanoclaw monitoring, keep frontend task endpoints as no-ops"
-```
-
----
-
-### Task 4: Delete the IPC module
-
-By now `ipc.py` has no importers (adapter gone in Task 2, pause/resume no-ops in Task 3).
-
-**Files:**
-- Delete: `server/src/warren/ipc.py`
-- Delete: `server/tests/test_ipc.py`
-
-- [ ] **Step 1: Confirm no importers remain**
-
-Run: `cd server && grep -rn "from warren.ipc\|import ipc\|write_ipc" src/ tests/`
-Expected: zero hits. If any remain, stop and fix the referencing task first.
-
-- [ ] **Step 2: Delete the files**
-
-```bash
-git rm server/src/warren/ipc.py server/tests/test_ipc.py
-```
-
-- [ ] **Step 3: Lint + test**
-
-Run: `cd server && uv run ruff check src/ tests/ && uv run pytest -q`
-Expected: PASS.
-
-- [ ] **Step 4: Commit**
-
-```bash
-git add -A && git commit -m "refactor: remove file-IPC module (nanoclaw-only)"
-```
-
----
-
-### Task 5: Clean up config fields
-
-Removes nanoclaw-only settings; keeps `internal_secret` and `backend`.
-
-**Files:**
-- Modify: `server/src/warren/config.py` — remove `backend_config` (~50), `nanoclaw_db` (~65), `nanoclaw_ipc_dir` (~66)
-- Modify: `server/tests/test_config.py` — delete `test_config_with_backend_config` and `test_config_without_backend_config` (~194-220)
-
-**Interfaces:**
-- Consumes: nothing.
-- Produces: `Settings` no longer has `nanoclaw_db`/`nanoclaw_ipc_dir`; `WarrenConfig` no longer has `backend_config`.
-
-- [ ] **Step 1: Remove the three fields from `config.py`**
-
-Delete `backend_config: dict[str, dict] = {}` from `WarrenConfig`, and `nanoclaw_db: str = ""` + `nanoclaw_ipc_dir: str = ""` from `Settings`. Keep `internal_secret` and `backend`.
-
-- [ ] **Step 2: Confirm no remaining references**
-
-Run: `cd server && grep -rn "backend_config\|nanoclaw_db\|nanoclaw_ipc_dir" src/ tests/`
-Expected: zero hits.
-
-- [ ] **Step 3: Delete the two backend_config tests in `test_config.py`**
-
-Remove `test_config_with_backend_config` and `test_config_without_backend_config`.
-
-- [ ] **Step 4: Lint + test**
-
-Run: `cd server && uv run ruff check src/ tests/ && uv run pytest -q`
-Expected: PASS.
-
-- [ ] **Step 5: Commit**
-
-```bash
-git add -A && git commit -m "refactor: remove nanoclaw config fields"
-```
-
----
-
-### Task 6: Clean Docker entrypoint
-
-**Files:**
-- Modify: `server/entrypoint.sh` — remove the nanoclaw IPC dir setup (~8-10)
-
-- [ ] **Step 1: Remove the nanoclaw lines**
-
-Delete:
-```bash
-    # Ensure nanoclaw IPC dirs are writable by appuser (nanoclaw runs as root)
-    mkdir -p /data/nanoclaw/ipc/warren/messages
-    chown -R appuser:appuser /data/nanoclaw/ipc/warren
-```
-
-- [ ] **Step 2: Sanity-check the script parses**
-
-Run: `bash -n server/entrypoint.sh`
-Expected: no output (valid syntax).
-
-- [ ] **Step 3: Commit**
-
-```bash
-git add -A && git commit -m "chore: drop nanoclaw IPC dir setup from entrypoint"
-```
-
----
-
-### Task 7: Remove the NanoClaw git submodule
-
-**Files:**
-- Delete: `.gitmodules` (only entry is nanoclaw)
-- Delete: `nanoclaw/` (submodule working tree)
-- Modify: `.git/config` (remove submodule section)
-
-- [ ] **Step 1: Deinit and remove the submodule**
-
-```bash
-git submodule deinit -f nanoclaw
-git rm -f nanoclaw
-rm -rf .git/modules/nanoclaw
-```
-
-- [ ] **Step 2: Remove the now-empty `.gitmodules`**
-
-Run: `cat .gitmodules`
-Expected: empty or only the nanoclaw section. If empty/only-nanoclaw:
-```bash
-git rm -f .gitmodules
-```
-
-- [ ] **Step 3: Verify clean status**
-
-Run: `git status && git config --file .git/config --get-regexp submodule || true`
-Expected: no `submodule.nanoclaw` section remains; `nanoclaw/` gone.
-
-- [ ] **Step 4: Commit**
-
-```bash
-git add -A && git commit -m "chore: remove nanoclaw git submodule"
-```
-
----
-
-### Task 8: Update docs and skills; delete nanoclaw design docs
-
-**Files:**
-- Modify: `CLAUDE.md` — rewrite intro, architecture list, key abstractions, deploy section, callback-types section to drop nanoclaw
-- Modify: `README.md` — description, ASCII diagram, prerequisites, clone command (drop `--recurse-submodules`), `WARREN_NANOCLAW_*` env rows
-- Modify: `.claude/skills/setup/SKILL.md` (~51), `.claude/skills/customize/SKILL.md` (~18), `.claude/skills/debug/SKILL.md` (nanoclaw references)
-- Delete: `docs/plans/2026-02-20-nanoclaw-integration-design.md`, `-integration-plan.md`, `-service-design.md`, `-service-plan.md`
-
-- [ ] **Step 1: Delete the four nanoclaw design docs**
-
-```bash
-git rm docs/plans/2026-02-20-nanoclaw-integration-design.md \
-       docs/plans/2026-02-20-nanoclaw-integration-plan.md \
-       docs/plans/2026-02-20-nanoclaw-service-design.md \
-       docs/plans/2026-02-20-nanoclaw-service-plan.md
-```
-(Leave other dated historical plans untouched — they are records, not live docs.)
-
-- [ ] **Step 2: Rewrite `CLAUDE.md`**
-
-Replace nanoclaw framing with the WebSocket-proxy/Orca model: intro sentence, the `adapters/` file list (drop `nanoclaw.py`, `ipc.py`, `nanoclaw_db.py`), key-abstractions bullets, the "Deploy with NanoClaw" section, and the "NanoClaw Callback Types" section. Reference `PROTOCOL.md` for the WebSocket protocol.
-
-- [ ] **Step 3: Rewrite `README.md`**
-
-Update the one-line description, ASCII flow diagram (Warren → WebSocket → Agent), prerequisites, `git clone` (remove `--recurse-submodules`), and remove the `WARREN_NANOCLAW_IPC_DIR` / `WARREN_NANOCLAW_DB` env-var rows. Keep `WARREN_INTERNAL_SECRET` (websocket auth).
-
-- [ ] **Step 4: Update the three skills**
-
-In `setup`, `customize`, and `debug` SKILL.md files, replace nanoclaw references with websocket/orca equivalents (backend options, adapter list, debug steps).
-
-- [ ] **Step 5: Final repo-wide nanoclaw sweep**
-
-Run: `grep -rin nanoclaw . --exclude-dir=.git`
-Expected: only acceptable residue — incidental mentions inside *other* historical `docs/plans/*` files. Zero hits under `server/src`, `server/tests`, `creation/`, `CLAUDE.md`, `README.md`, `.claude/skills/`. If any code/doc-in-scope hit remains, fix it.
-
-- [ ] **Step 6: Commit**
-
-```bash
-git add -A && git commit -m "docs: remove nanoclaw references from docs and skills"
-```
-
----
-
-### Task 9: Full verification
-
-- [ ] **Step 1: Lint**
-
-Run: `cd server && uv run ruff check src/ tests/`
-Expected: PASS.
-
-- [ ] **Step 2: Full test suite**
-
-Run: `cd server && uv run pytest -q`
-Expected: PASS, no skipped-due-to-import-error, no collection errors.
-
-- [ ] **Step 3: App boots and routes resolve**
-
-Run: `cd server && uv run python -c "from warren.app import create_app; app = create_app(); print(sorted({r.path for r in app.routes}))"`
-Expected: prints routes; includes `/api/agent/ws`, `/monitor/tasks`; excludes any `/internal/nanoclaw/*`.
-
-- [ ] **Step 4: Confirm no nanoclaw imports survive**
-
-Run: `cd server && grep -rin "nanoclaw\|from warren.ipc\|write_ipc" src/ tests/`
-Expected: zero hits.
-
----
-
-## Open questions / out of scope
-
-1. **Frontend monitor view.** The R1 `creation/` monitor view still renders a task list and (likely) pause/resume controls that are now permanently empty/no-op. Cleaning that UI is a separate frontend task — not included here to keep changes minimal. Flag for follow-up if the empty monitor view is undesirable.
-2. **Orca adapter.** Left untouched. The earlier architectural sketch suggested Orca could become an external WS client too, but that is a separate decision beyond "drop nanoclaw."
-3. **`PROTOCOL.md` / `examples/agent.py`** already describe the websocket protocol and need no nanoclaw removal.
diff --git a/server/Dockerfile b/server/Dockerfile
index 64d4f02..b7384b7 100644
--- a/server/Dockerfile
+++ b/server/Dockerfile
@@ -14,16 +14,6 @@ RUN apt-get update && apt-get install -y --no-install-recommends gosu \
 RUN curl -LsSf https://astral.sh/uv/install.sh | sh \
     && cp /root/.local/bin/uv /usr/local/bin/uv
 
-# Install Claude Code standalone binary (installs to /root/.local/bin)
-# then copy to /usr/local/bin so non-root appuser can run it
-RUN curl -fsSL https://claude.ai/install.sh | bash \
-    && cp /root/.local/bin/claude /usr/local/bin/claude
-
-# Install Entire CLI (installs to /root/.local/bin)
-# then copy to /usr/local/bin so non-root appuser can run it
-RUN curl -fsSL https://entire.io/install.sh | bash \
-    && cp /root/.local/bin/entire /usr/local/bin/entire
-
 # Add GitHub SSH host key
 RUN mkdir -p /root/.ssh && ssh-keyscan github.com >> /root/.ssh/known_hosts
 
@@ -47,4 +37,7 @@ RUN uv sync --no-dev --frozen \
 
 EXPOSE 4030
 
+HEALTHCHECK --interval=30s --timeout=5s --start-period=10s --retries=3 \
+    CMD python -c "import urllib.request; urllib.request.urlopen('http://127.0.0.1:4030/health')" || exit 1
+
 ENTRYPOINT ["./entrypoint.sh"]
diff --git a/server/src/warren/__main__.py b/server/src/warren/__main__.py
index 2927910..a9699d6 100644
--- a/server/src/warren/__main__.py
+++ b/server/src/warren/__main__.py
@@ -6,6 +6,7 @@
 import json
 import logging
 import os
+import secrets
 import socket
 
 import qrcode
@@ -35,9 +36,11 @@ def print_startup_qr_codes(settings: Settings) -> str:
     """Pre-register device token and print ASCII QR codes for the R1."""
     local_ip = get_local_ip()
     server_url = f"http://{local_ip}:{settings.port}"
-    token = "local-r1-token"
+    # Generate a fresh, unguessable pairing token each startup. Set
+    # WARREN_LOCAL_TOKEN to keep a stable token across restarts during dev.
+    token = os.environ.get("WARREN_LOCAL_TOKEN") or secrets.token_urlsafe(24)
 
-    # Pre-register a static token in the database
+    # Register the pairing token in the database
     async def save_token():
         db = Database(settings.db)
         await db.initialize()
@@ -50,10 +53,10 @@ async def save_token():
         logger.warning("Database startup token setup failed: %s", e)
 
     print("\n" + "─" * 60)
-    print("🌅 WARREN RUNTIME SERVER STARTED 🌅")
-    print(f"Local Host IP:    {local_ip}")
-    print(f"Backend Adapter:  {settings.backend.upper()}")
-    print(f"Database Path:    {settings.db}")
+    print("Warren server started")
+    print(f"  Local host IP:   {local_ip}")
+    print(f"  Backend adapter: {settings.backend.upper()}")
+    print(f"  Database path:   {settings.db}")
     print("─" * 60)
 
     # 1. R1 Creation / WebView QR Code
@@ -68,7 +71,7 @@ async def save_token():
         }
     )
 
-    print("\n📱 SCAN TO PAIR R1 WEBVIEW / CREATION UI 📱")
+    print("\nPair R1 WebView / Creation UI")
     print("Scan this QR code with your Rabbit R1 to open the Warren WebView:")
     qr1 = qrcode.QRCode()
     qr1.add_data(creation_payload)
@@ -87,7 +90,7 @@ async def save_token():
                 "protocol": "ws",
             }
         )
-        print("\n🎙️ SCAN TO PAIR R1 VOICE PTT GATEWAY 🎙️")
+        print("\nPair R1 Voice PTT Gateway")
         print("Scan this QR code with your R1 to route native voice prompts:")
         qr2 = qrcode.QRCode()
         qr2.add_data(voice_payload)
diff --git a/server/src/warren/adapters/orca.py b/server/src/warren/adapters/orca.py
index e6b5649..51b8a9b 100644
--- a/server/src/warren/adapters/orca.py
+++ b/server/src/warren/adapters/orca.py
@@ -2,6 +2,10 @@
 
 Enables Warren to act as an R1 interface for Orca-managed worktrees,
 active terminal sessions, and running agents.
+
+Experimental: this adapter drives a single local Orca terminal and is not
+multi-session. The WebSocket adapter is the supported default backend; use
+Orca only for local, single-terminal experimentation.
 """
 
 from __future__ import annotations
diff --git a/server/src/warren/adapters/websocket.py b/server/src/warren/adapters/websocket.py
index 21b5da3..cd13d4d 100644
--- a/server/src/warren/adapters/websocket.py
+++ b/server/src/warren/adapters/websocket.py
@@ -8,6 +8,7 @@
 
 import asyncio
 import logging
+import secrets
 from collections.abc import AsyncIterator
 
 from fastapi import WebSocket
@@ -51,15 +52,22 @@ async def handle_websocket(self, websocket: WebSocket) -> None:
             if auth_header and auth_header.lower().startswith("bearer "):
                 token = auth_header[7:]
 
-        # Validate token if configured
-        is_authenticated = True
-        if self._internal_secret or self._admin_token:
-            is_authenticated = False
-            if token:
-                if self._internal_secret and token == self._internal_secret:
-                    is_authenticated = True
-                elif self._admin_token and token == self._admin_token:
-                    is_authenticated = True
+        # Fail closed: the agent gateway must have a secret configured. Without
+        # one, any anonymous client could connect and drive sessions.
+        if not (self._internal_secret or self._admin_token):
+            logger.error(
+                "Refusing agent connection: no WARREN_INTERNAL_SECRET (or admin "
+                "token) is configured. Set WARREN_INTERNAL_SECRET to enable the "
+                "agent WebSocket gateway."
+            )
+            await websocket.close(code=4001, reason="Server not configured for agent auth")
+            return
+
+        # Constant-time comparison against each configured secret.
+        is_authenticated = bool(token) and (
+            (self._internal_secret and secrets.compare_digest(token, self._internal_secret))
+            or (self._admin_token and secrets.compare_digest(token, self._admin_token))
+        )
 
         if not is_authenticated:
             logger.warning("Unauthenticated agent connection attempt from %s", websocket.client)
diff --git a/server/src/warren/app.py b/server/src/warren/app.py
index d826f29..eb942ec 100644
--- a/server/src/warren/app.py
+++ b/server/src/warren/app.py
@@ -627,22 +627,6 @@ async def monitor_agents(
         bus: SessionBus = request.app.state.bus
         return {"active_sessions": bus.active_sessions()}
 
-    @app.get("/monitor/tasks")
-    async def monitor_tasks(
-        request: Request,
-        _token: str = Depends(require_auth),
-    ):
-        # Live task monitoring is not supported on the websocket/orca backends.
-        return {"tasks": []}
-
-    @app.post("/monitor/tasks/{task_id}/pause")
-    async def pause_task(task_id: str, _token: str = Depends(require_auth)):
-        return {"status": "ignored", "reason": "Task control not supported on this backend"}
-
-    @app.post("/monitor/tasks/{task_id}/resume")
-    async def resume_task(task_id: str, _token: str = Depends(require_auth)):
-        return {"status": "ignored", "reason": "Task control not supported on this backend"}
-
     # -- Agent WebSocket --
 
     @app.websocket("/api/agent/ws")
diff --git a/server/src/warren/config.py b/server/src/warren/config.py
index 5340c60..9784f4e 100644
--- a/server/src/warren/config.py
+++ b/server/src/warren/config.py
@@ -58,7 +58,6 @@ class Settings(BaseSettings):
     port: int = 4030
     db: str = "warren.db"
     config: str = "config.yaml"
-    anthropic_api_key: str = ""
     admin_token: str = ""
     data_dir: str = "data"
     voice_enabled: bool = False
diff --git a/server/src/warren/voice.py b/server/src/warren/voice.py
index b1e2890..4d24c5b 100644
--- a/server/src/warren/voice.py
+++ b/server/src/warren/voice.py
@@ -8,9 +8,11 @@
 from __future__ import annotations
 
 import asyncio
+import hashlib
 import json
 import logging
 import secrets
+import time
 import uuid
 
 from fastapi import WebSocket, WebSocketDisconnect
@@ -56,7 +58,7 @@ async def voice_handler(
     try:
         data = await asyncio.wait_for(ws.receive(), timeout=10.0)
     except asyncio.TimeoutError:
-        logger.warning("Voice handshake: timeout waiting for connect")
+        logger.debug("Voice handshake: timeout waiting for connect")
         return
 
     if data.get("type") == "websocket.disconnect":
@@ -68,7 +70,7 @@ async def voice_handler(
     if not raw:
         return
 
-    logger.warning("Voice connect: %s", raw[:300])
+    logger.debug("Voice connect frame received (%d bytes)", len(raw))
 
     try:
         msg = parse_frame(raw)
@@ -86,15 +88,15 @@ async def voice_handler(
     token = auth.get("token", "") or auth.get("deviceToken", "")
 
     if not await db.validate_device_token(token):
-        tok_preview = token[:8] if token else "(empty)"
-        logger.warning("Voice handshake: invalid token: %s", tok_preview)
+        logger.warning("Voice handshake: invalid token")
         await ws.send_text(json.dumps(build_hello_error(request_id, "invalid token")))
         await ws.close()
         return
 
-    device_id = token[:8]
+    # Non-reversible correlation id for logs (never log token material).
+    device_id = hashlib.sha256(token.encode()).hexdigest()[:8] if token else "anon"
     await ws.send_text(json.dumps(build_hello_ok(request_id, token)))
-    logger.warning("Voice authenticated: %s", device_id)
+    logger.info("Voice authenticated: %s", device_id)
 
     # --- Session state ---
     active_session_id: str | None = None
@@ -135,7 +137,7 @@ async def handle_chat_send(msg: dict) -> None:
         if not message:
             return
 
-        logger.warning("Voice chat.send: %s", message[:100])
+        logger.debug("Voice chat.send (%d chars)", len(message))
 
         run_id = uuid.uuid4().hex[:12]
         chat_seq = 0
@@ -185,7 +187,7 @@ async def tick_loop():
             while True:
                 await asyncio.sleep(TICK_INTERVAL)
                 await ws.send_text(json.dumps(build_tick()))
-        except (WebSocketDisconnect, Exception):
+        except Exception:
             pass
 
     async def event_relay():
@@ -206,7 +208,6 @@ async def event_relay():
                     # If we have accumulated text and no events for a while,
                     # send final so the R1 can do TTS
                     if accumulated_text and idle_since is not None:
-                        import time
                         elapsed = time.monotonic() - idle_since
                         if elapsed >= IDLE_FINAL_TIMEOUT:
                             chat_seq += 1
@@ -216,21 +217,20 @@ async def event_relay():
                                 "final", accumulated_text,
                                 stop_reason="end_turn",
                             )
-                            logger.warning(
-                                "Voice relay: idle final: %s",
-                                accumulated_text[:80],
+                            logger.debug(
+                                "Voice relay: idle final (%d chars)",
+                                len(accumulated_text),
                             )
                             await ws.send_text(json.dumps(chat_msg))
                             accumulated_text = ""
                             idle_since = None
                     continue
 
-                import time
                 idle_since = time.monotonic()
 
                 session_key = client_session_key
                 event_type = event.get("type")
-                logger.warning("Voice relay event: %s", event_type)
+                logger.debug("Voice relay event: %s", event_type)
 
                 if event_type == "text":
                     content = event.get("content", "")
@@ -249,17 +249,15 @@ async def event_relay():
                         session_key, run_id, chat_seq, "final", summary,
                         stop_reason="end_turn",
                     )
-                    logger.warning(
-                        "Voice relay final: %s", summary[:80],
-                    )
+                    logger.debug("Voice relay final (%d chars)", len(summary))
                     await ws.send_text(json.dumps(chat_msg))
                     accumulated_text = ""
                     idle_since = None
 
         except WebSocketDisconnect:
-            logger.warning("Voice relay: WS disconnected")
-        except Exception as e:
-            logger.warning("Voice relay error: %s", e)
+            logger.debug("Voice relay: WS disconnected")
+        except Exception:
+            logger.exception("Voice relay error")
 
     # --- Main loop ---
     tick_task = asyncio.create_task(tick_loop())
@@ -268,7 +266,7 @@ async def event_relay():
     try:
         while True:
             raw = await ws.receive_text()
-            logger.warning("Voice recv: %s", raw[:300])
+            logger.debug("Voice recv (%d bytes)", len(raw))
             try:
                 msg = parse_frame(raw)
             except ValueError:
@@ -278,9 +276,9 @@ async def event_relay():
             if msg.get("type") == "req" and method == "chat.send":
                 await handle_chat_send(msg)
             elif msg.get("type") == "req":
-                logger.warning("Voice: unhandled method: %s", method)
+                logger.debug("Voice: unhandled method: %s", method)
     except WebSocketDisconnect:
-        logger.warning("Voice client disconnected: %s", device_id)
+        logger.info("Voice client disconnected: %s", device_id)
     finally:
         tick_task.cancel()
         relay_task.cancel()
diff --git a/server/tests/test_app.py b/server/tests/test_app.py
index f9a884e..159e5ac 100644
--- a/server/tests/test_app.py
+++ b/server/tests/test_app.py
@@ -36,7 +36,6 @@ async def client(tmp_path):
     }))
 
     settings = Settings(
-        anthropic_api_key="sk-test",
         db=str(tmp_path / "test.db"),
         config=str(config_file),
         admin_token="test-admin",
diff --git a/server/tests/test_config.py b/server/tests/test_config.py
index f78fc46..d31b4d6 100644
--- a/server/tests/test_config.py
+++ b/server/tests/test_config.py
@@ -7,7 +7,7 @@ def test_default_settings():
     """Settings loads with defaults when no env vars set."""
     from warren.config import Settings
 
-    s = Settings(anthropic_api_key="sk-test")
+    s = Settings()
     assert s.host == "0.0.0.0"
     assert s.port == 4030
     assert s.db == "warren.db"
diff --git a/server/tests/test_integration.py b/server/tests/test_integration.py
index 95ebdbf..867e972 100644
--- a/server/tests/test_integration.py
+++ b/server/tests/test_integration.py
@@ -25,7 +25,6 @@ async def client(tmp_path):
     }))
 
     settings = Settings(
-        anthropic_api_key="sk-test",
         db=str(tmp_path / "test.db"),
         config=str(config_file),
         admin_token="test-admin",
diff --git a/server/tests/test_monitor.py b/server/tests/test_monitor.py
index e048128..f739682 100644
--- a/server/tests/test_monitor.py
+++ b/server/tests/test_monitor.py
@@ -86,27 +86,3 @@ async def test_returns_active_sessions(
             data = resp.json()
             assert "active_sessions" in data
             assert "session-1" in data["active_sessions"]
-
-
-class TestMonitorTasks:
-    async def test_requires_auth(self, monitor_app):
-        async with AsyncClient(
-            transport=ASGITransport(app=monitor_app),
-            base_url="http://test",
-        ) as client:
-            resp = await client.get("/monitor/tasks")
-            assert resp.status_code == 401
-
-    async def test_returns_empty_tasks(
-        self, monitor_app, auth_headers,
-    ):
-        async with AsyncClient(
-            transport=ASGITransport(app=monitor_app),
-            base_url="http://test",
-        ) as client:
-            resp = await client.get(
-                "/monitor/tasks", headers=auth_headers,
-            )
-            assert resp.status_code == 200
-            data = resp.json()
-            assert data["tasks"] == []
diff --git a/server/tests/test_streaming.py b/server/tests/test_streaming.py
index e3904ea..abf0137 100644
--- a/server/tests/test_streaming.py
+++ b/server/tests/test_streaming.py
@@ -21,7 +21,6 @@ async def authed_client(tmp_path):
     }))
 
     settings = Settings(
-        anthropic_api_key="sk-test",
         db=str(tmp_path / "test.db"),
         config=str(config_file),
         admin_token="test-admin",
diff --git a/server/tests/test_websocket_adapter.py b/server/tests/test_websocket_adapter.py
index 4ba46b3..97b840d 100644
--- a/server/tests/test_websocket_adapter.py
+++ b/server/tests/test_websocket_adapter.py
@@ -49,6 +49,44 @@ async def ws_endpoint(websocket: WebSocket):
             websocket.receive_json()
     assert excinfo.value.code == 4001
 
+def test_websocket_no_secret_configured_fails_closed():
+    """With no secret configured the gateway must reject all connections."""
+    from fastapi.websockets import WebSocketDisconnect
+
+    app = FastAPI()
+    adapter = AgentWebSocketAdapter()  # no internal_secret, no admin_token
+
+    @app.websocket("/api/agent/ws")
+    async def ws_endpoint(websocket: WebSocket):
+        await adapter.handle_websocket(websocket)
+
+    client = TestClient(app)
+
+    with pytest.raises(WebSocketDisconnect) as excinfo:
+        with client.websocket_connect("/api/agent/ws?token=anything") as websocket:
+            websocket.receive_json()
+    assert excinfo.value.code == 4001
+    assert len(adapter._active_connections) == 0
+
+
+def test_websocket_wrong_token_rejected():
+    from fastapi.websockets import WebSocketDisconnect
+
+    app = FastAPI()
+    adapter = AgentWebSocketAdapter(internal_secret="super-secret")
+
+    @app.websocket("/api/agent/ws")
+    async def ws_endpoint(websocket: WebSocket):
+        await adapter.handle_websocket(websocket)
+
+    client = TestClient(app)
+
+    with pytest.raises(WebSocketDisconnect) as excinfo:
+        with client.websocket_connect("/api/agent/ws?token=wrong") as websocket:
+            websocket.receive_json()
+    assert excinfo.value.code == 4001
+
+
 def test_websocket_authenticated_integration():
     app = FastAPI()
     adapter = AgentWebSocketAdapter(internal_secret="super-secret")