diff --git a/HANDOVER.md b/HANDOVER.md
index 59fd524..1aaa855 100644
--- a/HANDOVER.md
+++ b/HANDOVER.md
@@ -4,81 +4,118 @@ _Last updated: 2026-03-09_
 
 ## Current status
 
-The wizard runs end-to-end. The generated project (`thewiztest`) starts the container
-and opens the Electron UI. **The last fix was NOT yet confirmed working by the user.**
-The session ended before the user could test it.
+**Fully working end-to-end.** The wizard generates a project, `./start.sh` starts the container,
+registers the project, launches the Electron UI, agents start cleanly, and kai's terminal opens
+and primes without any trust dialog or startup errors.
 
 ## What was fixed this session (newest first)
 
-### 1. `bin/claude` — workdir fallback (UNVERIFIED — last fix, not yet tested)
+### 1. Trust dialog bypass
+**File:** `lib/workflow.sh` → agent `.claude/settings.json`
+**Symptom:** Claude Code shows "Quick safety check — do you trust this folder?" on every start,
+blocking `/prime` injection.
+**Fix:** Added `"hasTrustDialogAccepted": true` to every generated agent's `.claude/settings.json`.
+
+### 2. CLI ready marker — container/host gap
 **File:** `lib/container.sh` → generated `bin/claude`
-**Symptom:** `[Server:err] claude-code is required but not found` → `[Server] Exited with code 1` → all agents fail to start
-**Root cause:** When Electron spawns `node core/start.js`, its cwd is `~/.context-studio/core`. The `bin/claude` wrapper used `--workdir "$PWD"` in `podman exec`. That directory isn't mounted in the container → podman fails → returns non-zero → claude appears "missing".
-**Fix:** If `$PWD` is not under `$PROJECT_DIR`, fall back to `$PROJECT_DIR` as the container workdir.
-**Also patched:** `thewiztest/bin/claude`
+**Symptom:** `[CLI] Timeout waiting for CLI ready marker for kai (10000ms)` — 10s delay before `/prime`.
+**Root cause:** CS Core polls `/tmp/cs-ready-<agentId>` on the **host**. Claude runs inside the
+container, so it can't create this file on the host.
+**Fix:** `bin/claude` detects when it's being invoked interactively from an agent PTY
+(`$PWD == $PROJECT_DIR/workflow/agents/*`), extracts the agent ID from `basename "$PWD"`,
+and spawns a background job `(sleep 3 && touch /tmp/cs-ready-<agentId>)` before running podman exec.
 
-### 2. Container runs as root → `--dangerously-skip-permissions` rejected
+### 3. Project not registered → `localhost:null` for all agents
 **File:** `lib/container.sh` → generated `start.sh`
-**Symptom:** `--dangerously-skip-permissions cannot be used with root/sudo privileges`
-**Fix:** Added `--user "$(id -u):$(id -g)"` and `-e HOME="$HOME"` to `podman run`
-**Why it works:** Host user `elmar` = uid 1000 = `node` user in `node:22` image → permissions match
-**Also patched:** `thewiztest/start.sh`
+**Symptom:** All agent SSE URLs are `http://localhost:null/...` → DOMException, no agent communication.
+**Root cause (deep):**
+- `waitForServers()` in `server-management.js` polls `runtimeConfig.findRuntimeByWorkflowDir()`
+- That function maps workflowDir → project UUID → lock file at
+  `~/.config/context-studio/projects/locks/<uuid>.json`
+- `acquireProjectLock()` in `launcher.js` writes the lock file — but **silently skips** it if the
+  project is not registered in `~/.config/context-studio/projects/<uuid>.json`
+- Without the lock file, `waitForServers()` always times out → `applyRuntimePorts()` never called
+  → all agent ports remain `null`
+**Fix:** `start.sh` now auto-registers the project before launching Electron. It scans
+`~/.config/context-studio/projects/` for an existing entry matching `$PROJECT_DIR/workflow`,
+and if none is found, writes a new `<uuid>.json` registration file using python3.
 
-### 3. Electron manages server startup — removed redundant headless node
+### 4. `serverStartupTimeout` is in `electron-config.js`, not `system.json`
+**File:** `~/.context-studio/core/config/electron-config.js`
+**Symptom:** 30s startup timeout even when servers start in ~3s.
+**Root cause:** The timeout is read from `ctx.getElectronConfig()?.startup.serverStartupTimeout`.
+This comes from `electron-config.js` in the core config dir, NOT from the workflow's `system.json`.
+**Fix:** Changed `serverStartupTimeout` from `30000` to `90000` in `electron-config.js`.
+Note: `electron-config.js` is a config file (not source), so editing it is appropriate.
+
+### 5. Credential mounts — `~/.claude/.credentials.json` not mounted
 **File:** `lib/container.sh` → generated `start.sh`
-**Symptom:** Would have caused port conflicts
-**Fix:** Removed `node start.js --ui-mode=headless &` from start.sh. The Electron app's `server-management.js` checks the lock file and spawns servers itself.
-**Also patched:** `thewiztest/start.sh`
+**Symptom:** Claude Code inside container can't authenticate to Anthropic.
+**Root cause:** `start.sh` was only mounting `~/.anthropic` (empty on this system).
+Actual credentials are at `~/.claude/.credentials.json` (or `$CLAUDE_CONFIG_DIR/.credentials.json`,
+`~/.claude.json`, `$CLAUDE_CONFIG_DIR/.claude.json`).
+**Fix:** `start.sh` now builds `_CREDS_ARGS` array and conditionally mounts whichever credential
+files exist on the host. All known Claude Code credential locations are checked.
 
-### 4. Electron must be launched separately
-**File:** `lib/container.sh`
-**Symptom:** UI never opened — servers ran but no window
-**Root cause:** `node core/start.js --ui-mode=electron` does NOT launch Electron. It logs "Electron app started separately" and only manages A2A servers.
-**Fix (later superseded):** Direct Electron launch via `$CS_CORE/app/node_modules/.bin/electron $CS_CORE/app`
+### 6. API key check in wizard
+**File:** `lib/prereqs.sh` → `ensure_api_key()`
+**Symptom:** Agents fail if `ANTHROPIC_API_KEY` not set and no credentials file mounted.
+**Fix:** Added `ensure_api_key()` called from `check_prerequisites()`. Checks in order:
+`ANTHROPIC_API_KEY` env var → `$CLAUDE_CONFIG_DIR/.credentials.json` → `~/.claude/.credentials.json`
+→ `~/.claude.json` → `$CLAUDE_CONFIG_DIR/.claude.json` → `~/.anthropic/.credentials.json`.
+If none found, prompts for API key with masked input and offers to save to shell profile.
 
-## What still needs verifying
-
-1. **Does the server now start without the `claude-code missing` error?**
-   - Run `./start.sh` in `thewiztest/`
-   - Watch for `[12:xx:xx] ✅ All agent servers started` (no `Server startup failed`)
-   - The Electron UI should open and kai's terminal should start without root errors
-
-2. **`localhost:null` network error** — this is downstream of (1). If servers start cleanly, the registry port gets written to the lock file and `localhost:null` disappears.
-
-3. **Kai can't connect to the internet** — mentioned by user but not investigated. Could be:
-   - Container network settings (Podman default: slirp4netns, should have internet)
-   - ANTHROPIC_API_KEY not set or not passed into container
-   - Proxy/VPN issue on the host network
+### 7. Ctrl+C graceful abort with cleanup
+**File:** `wizard.sh`
+**Fix:** `trap 'handle_sigint' INT` in `main()`. On Ctrl+C: shows `gum confirm` popup.
+If confirmed: removes `$PROJECT_DIR` (if created) and `$CS_CORE_DIR` (if cloned this session).
+State flags: `_WIZARD_PROJECT_CREATED` and `_WIZARD_CORE_CLONED` (set at moment of action).
+Installed packages (git, podman) are never reverted.
 
 ## Key architecture facts
 
-### How CS Core + Electron work together
-- `electron app/` starts the UI
-- Electron's `server-management.js` checks `workflow/data/` for a lock file
-- If no lock file → it spawns `node core/start.js --ui-mode=headless` as a child process
-- Child process inherits Electron's `process.env` including PATH (with `bin/claude`)
-- When the requirements check runs `claude --version`, it finds `bin/claude` in PATH
-- `bin/claude` proxies to `podman exec cs-<slug> claude --version`
-- Container must be running BEFORE Electron is launched (start.sh handles this)
+### Lock file mechanism (critical for startup)
+- Lock file: `~/.config/context-studio/projects/locks/<uuid>.json`
+- Project registration: `~/.config/context-studio/projects/<uuid>.json`
+- `start.sh` auto-registers the project before launching Electron
+- Without registration, `acquireProjectLock()` silently skips writing the lock file
+- Without the lock file, all agent ports remain `null` → `localhost:null` errors
 
-### Path that must be mounted in container
-Only `$PROJECT_DIR` is mounted (at the same absolute path). NOT:
-- `~/.context-studio/core`
-- `~/.anthropic` (mounted read-only separately)
-- Any other host path
+### How CS Core + Electron work together
+- `start.sh` starts the container, registers the project, then launches Electron
+- Electron's `server-management.js` spawns `node core/start.js --ui-mode=headless`
+- That process starts all A2A agent servers on the **host** (not in container)
+- Servers register with the registry (port 8000), write the lock file
+- `waitForServers()` polls until lock file appears + health check passes
+- `applyRuntimePorts()` is called → agent ports loaded from lock file
+- CLI ready marker (`/tmp/cs-ready-<agentId>`) created by `bin/claude` after 3s delay
+
+### Credential lookup order (container mounts)
+1. `ANTHROPIC_API_KEY` env var (passed via `-e`)
+2. `~/.anthropic/` (mounted read-only, always)
+3. `$CLAUDE_CONFIG_DIR/.credentials.json` or `~/.claude/.credentials.json` (mounted if exists)
+4. `~/.claude.json` (mounted if exists)
+5. `$CLAUDE_CONFIG_DIR/.claude.json` (mounted if exists)
+
+### What runs where
+- **Container:** Claude Code binary only (`sleep infinity` + `podman exec`)
+- **Host:** Electron UI, CS Core, all A2A agent server processes (node)
+- **`bin/claude`:** bridges host agent calls → container claude
 
 ### Generated files per project
-- `bin/claude` — wrapper with hardcoded `PROJECT_DIR` and `CONTAINER_NAME`
-- `start.sh` — starts container as `$(id -u):$(id -g)`, exports PATH, launches Electron
+- `bin/claude` — wrapper: workdir fallback, credential routing, ready marker creation
+- `start.sh` — starts container, mounts credentials, registers project, launches Electron
 - `stop.sh` — force-removes container
 - `update.sh` — git pull core, npm update claude-code in container, apt upgrade
 
 ## File locations
 - Wizard: `/home/elmar/Projects/ContextStudioWizard/`
-- Test project: `/home/elmar/Projects/thewiztest/`
-- Core (read-only!): `/home/elmar/.context-studio/core/`
-- Wizard repo remote: check `git remote -v` in ContextStudioWizard
+- Core config (editable): `/home/elmar/.context-studio/core/config/electron-config.js`
+- Core (read-only source): `/home/elmar/.context-studio/core/` (never modify source, never push)
+- CS settings: `~/.config/context-studio/projects/`
+- Lock files: `~/.config/context-studio/projects/locks/`
 
 ## What NOT to do
-- Never modify `~/.context-studio/core` — it is read-only
+- Never modify `~/.context-studio/core/` source files — read-only
+- `electron-config.js` in `core/config/` is an exception — it is a config file, safe to edit
 - Never commit or push to the core repo
diff --git a/README.md b/README.md
index ba9267f..caa43cf 100644
--- a/README.md
+++ b/README.md
@@ -21,7 +21,7 @@ One command. A complete project drops out:
 └── update.sh                    ← 🔄 update core, claude-code, OS packages
 ```
 
-Run `./start.sh` → Podman container starts → Electron UI opens → talk to your agent team.
+Run `./start.sh` → Podman container starts → Electron UI opens → agents prime automatically → talk to your team.
 
 ---
 
@@ -30,14 +30,14 @@ Run `./start.sh` → Podman container starts → Electron UI opens → talk to y
 ```
 HOST                                    CONTAINER (cs-<project>)
 ─────────────────────────────           ────────────────────────
-Context Studio Core (Electron)   ←───  All Claude Code agents
-  starts A2A servers                    mounted at same abs path
-  opens Electron UI                     bin/claude wrapper routes
-  manages workflow                      every claude call here
+Context Studio Core (Electron)          Claude Code binary
+  A2A agent servers (node)       ←───  bin/claude wrapper routes
+  Electron UI                          every claude call here
+  workflow management                  mounted at same abs path
 ```
 
-- **CS Core** runs on the host (Electron UI, no display issues)
-- **All agents** run inside the Podman container as the current user (not root)
+- **CS Core + all agent servers** run on the host (no display issues, native Electron)
+- **Claude Code binary** runs inside the Podman container as the current user (not root)
 - **`bin/claude`** is prepended to PATH so CS Core's server process finds it automatically
 - **Paths match** host ↔ container (project mounted at the same absolute path)
 
@@ -50,7 +50,7 @@ Context Studio Core (Electron)   ←───  All Claude Code agents
 | 🔀 | `git` | Auto-installed if missing |
 | 🐳 | `podman` _(preferred)_ or `docker` | Auto-installed if missing |
 | 🔑 | SSH key → `github.com` | Must be set up manually |
-| 🗝️ | `ANTHROPIC_API_KEY` | Must be set in your environment |
+| 🗝️ | Anthropic credentials | API key or `claude auth login` — wizard checks automatically |
 
 **Auto-install supported on:** Arch · Debian/Ubuntu · RHEL/Fedora · openSUSE
 
@@ -64,22 +64,41 @@ Context Studio Core (Electron)   ←───  All Claude Code agents
 
 The wizard guides you through:
 
-1. 🔧 **Core setup** — clones `context-studio-core` to `~/.context-studio/core/` (once, shared)
-2. 📝 **Project name & location**
-3. ⚙️ **Workflow** — generate from scratch _or_ clone an existing repo
-4. 🤖 **Agent preset** _(if generating)_
+1. 🔧 **Prerequisites** — checks git, podman/docker, Anthropic credentials
+2. 🔧 **Core setup** — clones `context-studio-core` to `~/.context-studio/core/` (once, shared)
+3. 📝 **Project name & location**
+4. ⚙️ **Workflow** — generate from scratch _or_ clone an existing repo
+5. 🤖 **Agent preset** _(if generating)_
 
    | Preset | Agents |
    |--------|--------|
    | `minimal` | 5 agents: coordinator · 2× coder · researcher · tester |
    | `standard` | 9 agents: 2× coordinator · 3× coder · 2× researcher · tester · reviewer |
 
-5. ✅ **Done** — project created, git repo initialized
+6. ✅ **Done** — project created, git repo initialized
+
+**Ctrl+C at any point** → confirmation prompt → reverts all created files (installed packages kept).
 
 > ⚠️ **First `./start.sh`** builds the container image. Expect 5–15 min. Subsequent starts are instant.
 
 ---
 
+## 🔑 Anthropic credentials
+
+The wizard checks for credentials in this order:
+
+1. `ANTHROPIC_API_KEY` environment variable
+2. `~/.claude/.credentials.json` (default `claude auth login` location)
+3. `~/.claude.json`
+4. `$CLAUDE_CONFIG_DIR/.credentials.json` / `$CLAUDE_CONFIG_DIR/.claude.json`
+5. `~/.anthropic/.credentials.json`
+
+If none found, you're prompted to enter an API key (masked input) with an option to save it to `~/.zshrc` / `~/.bashrc`.
+
+All credential files that exist are **mounted read-only** into the container so Claude Code can authenticate.
+
+---
+
 ## ▶️ Starting your project
 
 ```bash
@@ -89,9 +108,11 @@ cd ~/Projects/my-project
 
 This will:
 1. Build the container image if not yet built (first run: 5–15 min)
-2. Start the Podman container (agents run here as your user, not root)
-3. Launch the Context Studio Electron UI
-4. When you close the UI → container is stopped automatically
+2. Start the Podman container (agents run inside as your user, not root)
+3. Register the project with Context Studio (required for server startup)
+4. Launch the Context Studio Electron UI
+5. Agents start, prime automatically — no trust dialogs, no manual steps
+6. When you close the UI → container is stopped automatically
 
 ```bash
 ./stop.sh      # force-stop container without closing the UI
@@ -106,8 +127,8 @@ CS Core on the host needs to call `claude`. Instead of using the host's `claude`
 
 1. Checks the agents container is running
 2. Relays the call into the container via `podman exec`
-3. Passes the working directory (falling back to project root if cwd isn't mounted)
-4. All agents — including the main coordinator — run **inside the container**
+3. Falls back to project root as workdir if cwd isn't mounted in the container
+4. Creates the CS Core ready marker (`/tmp/cs-ready-<agentId>`) after 3s so `/prime` injection isn't delayed
 
 ---
 
@@ -117,7 +138,7 @@ CS Core on the host needs to call `claude`. Instead of using the host's `claude`
 🧙 wizard.sh               ← entry point
 📁 lib/
    utils.sh                ← prompts, colors, gum helpers
-   prereqs.sh              ← auto-install git + podman/docker
+   prereqs.sh              ← auto-install git + podman/docker + API key check
    core.sh                 ← global core install/update
    project.sh              ← devcontainer + project scaffold
    workflow.sh             ← generate or clone workflow config
@@ -136,8 +157,9 @@ CS Core on the host needs to call `claude`. Instead of using the host's `claude`
 
 ## ⚠️ Hard rules
 
-- **`~/.context-studio/core` is READ-ONLY** — never modify, commit, or push to it
-- It is a shared global dependency; the wizard only clones and `npm install`s it
+- **`~/.context-studio/core/` source is READ-ONLY** — never modify source files, commit, or push to it
+- `~/.context-studio/core/config/electron-config.js` is a config file — safe to edit (e.g. timeouts)
+- The wizard only clones and `npm install`s core; it never touches core source
 
 ---
 
diff --git a/lib/container.sh b/lib/container.sh
index 224b1c5..5ca4ef5 100644
--- a/lib/container.sh
+++ b/lib/container.sh
@@ -41,6 +41,21 @@ else
     WORKDIR="\$PROJECT_DIR"
 fi
 
+# ── CS Core ready marker ────────────────────────────────────────────────
+# CS Core polls /tmp/cs-ready-<agentId> on the host to know when the CLI
+# banner is visible and /prime can be injected. Claude runs inside the
+# container so it cannot create this file on the host itself.
+# We infer the agent ID from the PTY working directory (CS Core sets it to
+# workflow/agents/<agentId>) and create the marker after a short delay.
+_is_interactive=true
+for _arg in "\$@"; do
+    case "\$_arg" in --version|--help|-h|-v) _is_interactive=false; break ;; esac
+done
+if [[ "\$_is_interactive" == "true" && "\$PWD" == "\$PROJECT_DIR/workflow/agents/"* ]]; then
+    _AGENT_ID="\$(basename "\$PWD")"
+    (sleep 3 && touch "/tmp/cs-ready-\$_AGENT_ID") &
+fi
+
 # Pass through TTY if available, relay working directory into container
 if [ -t 0 ]; then
     exec "\$RUNTIME" exec -it --workdir "\$WORKDIR" "\$CONTAINER_NAME" claude "\$@"
@@ -82,6 +97,22 @@ fi
 # ── Ensure ~/.anthropic exists (Claude Code stores auth/config here) ─────
 mkdir -p "\$HOME/.anthropic"
 
+# ── Build credential mounts ───────────────────────────────────────────────
+# Claude Code may store credentials in various locations depending on version
+# and whether CLAUDE_CONFIG_DIR is set. Mount whichever files exist.
+_CREDS_ARGS=()
+_CREDS_ARGS+=("-v" "\$HOME/.anthropic:\$HOME/.anthropic:ro")
+_claude_dir="\${CLAUDE_CONFIG_DIR:-\$HOME/.claude}"
+if [[ -f "\$_claude_dir/.credentials.json" ]]; then
+    _CREDS_ARGS+=("-v" "\$_claude_dir/.credentials.json:\$_claude_dir/.credentials.json:ro")
+fi
+if [[ -f "\$HOME/.claude.json" ]]; then
+    _CREDS_ARGS+=("-v" "\$HOME/.claude.json:\$HOME/.claude.json:ro")
+fi
+if [[ -n "\${CLAUDE_CONFIG_DIR:-}" && -f "\$CLAUDE_CONFIG_DIR/.claude.json" ]]; then
+    _CREDS_ARGS+=("-v" "\$CLAUDE_CONFIG_DIR/.claude.json:\$CLAUDE_CONFIG_DIR/.claude.json:ro")
+fi
+
 # ── Start agents container ───────────────────────────────────────────────
 # Mount project at the same absolute path so host and container paths match.
 # CS Core sets agent working dirs to host paths; the wrapper relays PWD.
@@ -92,7 +123,7 @@ echo "→ Starting agents container '\$CONTAINER_NAME'..."
     --name "\$CONTAINER_NAME" \\
     --user "\$(id -u):\$(id -g)" \\
     -v "\$PROJECT_DIR:\$PROJECT_DIR" \\
-    -v "\$HOME/.anthropic:\$HOME/.anthropic:ro" \\
+    "\${_CREDS_ARGS[@]}" \\
     -e ANTHROPIC_API_KEY="\${ANTHROPIC_API_KEY:-}" \\
     -e CS_WORKFLOW_DIR="\$PROJECT_DIR/workflow" \\
     -e PROJECT_ROOT_DIR="\$PROJECT_DIR" \\
@@ -121,6 +152,45 @@ if [[ ! -d "\$CS_CORE/app/node_modules" ]]; then
     (cd "\$CS_CORE" && npm install) || { echo "npm install failed." >&2; exit 1; }
 fi
 
+# ── Register project with Context Studio (required for lock file to be written) ──
+# CS Core's acquireProjectLock() skips writing the lock file if the project
+# isn't registered in ~/.config/context-studio/projects/<uuid>.json.
+# Without the lock file, waitForServers() can never find the registry port
+# and always times out — causing localhost:null errors in the UI.
+_CS_PROJECTS_DIR="\$HOME/.config/context-studio/projects"
+mkdir -p "\$_CS_PROJECTS_DIR"
+_WORKFLOW_DIR="\$PROJECT_DIR/workflow"
+_already_registered=false
+for _f in "\$_CS_PROJECTS_DIR"/*.json; do
+    if [[ -f "\$_f" ]] && python3 -c "
+import json,sys
+d=json.load(open(sys.argv[1]))
+sys.exit(0 if d.get('workflowDir') == sys.argv[2] else 1)
+" "\$_f" "\$_WORKFLOW_DIR" 2>/dev/null; then
+        _already_registered=true
+        break
+    fi
+done
+if [[ "\$_already_registered" == "false" ]]; then
+    _UUID=\$(python3 -c "import uuid; print(uuid.uuid4())")
+    _NOW=\$(python3 -c "from datetime import datetime,timezone; print(datetime.now(timezone.utc).isoformat())")
+    python3 -c "
+import json, sys
+data = {
+    'id': sys.argv[1],
+    'name': sys.argv[2],
+    'workflowDir': sys.argv[3],
+    'user': 'default',
+    'created': sys.argv[4],
+    'lastOpened': sys.argv[4]
+}
+with open(sys.argv[5], 'w') as f:
+    json.dump(data, f, indent=2)
+    f.write('\n')
+" "\$_UUID" "\$(basename "\$PROJECT_DIR")" "\$_WORKFLOW_DIR" "\$_NOW" "\$_CS_PROJECTS_DIR/\$_UUID.json"
+    echo "→ Registered project with Context Studio"
+fi
+
 # ── Check display for Electron UI ───────────────────────────────────────
 if [[ -z "\${DISPLAY:-}" && -z "\${WAYLAND_DISPLAY:-}" ]]; then
     echo "⚠  No display detected (DISPLAY / WAYLAND_DISPLAY not set)."
diff --git a/lib/core.sh b/lib/core.sh
index ecb1dbb..8bc5f7c 100644
--- a/lib/core.sh
+++ b/lib/core.sh
@@ -21,6 +21,7 @@ setup_core() {
 
     info "Cloning context-studio-core → $CS_CORE_DIR"
     mkdir -p "$CS_HOME"
+    _WIZARD_CORE_CLONED=true
     spin "Cloning context-studio-core..." \
         git clone "$CS_CORE_REPO" "$CS_CORE_DIR" \
         || die "Failed to clone context-studio-core. Check your SSH key and network."
diff --git a/lib/prereqs.sh b/lib/prereqs.sh
index 404451a..0053562 100644
--- a/lib/prereqs.sh
+++ b/lib/prereqs.sh
@@ -98,8 +98,73 @@ _install_docker() {
     warn "Added $USER to docker group — log out and back in for it to take effect."
 }
 
+ensure_api_key() {
+    if [[ -n "${ANTHROPIC_API_KEY:-}" ]]; then
+        success "ANTHROPIC_API_KEY is set"
+        return
+    fi
+
+    # Resolve config dir — CLAUDE_CONFIG_DIR overrides default ~/.claude
+    local claude_dir="${CLAUDE_CONFIG_DIR:-$HOME/.claude}"
+
+    local creds_file=""
+    if   [[ -f "$claude_dir/.credentials.json" ]]; then creds_file="$claude_dir/.credentials.json"
+    elif [[ -f "$HOME/.claude.json"             ]]; then creds_file="$HOME/.claude.json"
+    elif [[ -f "$claude_dir/.claude.json"       ]]; then creds_file="$claude_dir/.claude.json"
+    elif [[ -f "$HOME/.anthropic/.credentials.json" ]]; then creds_file="$HOME/.anthropic/.credentials.json"
+    fi
+
+    if [[ -n "$creds_file" ]]; then
+        success "Anthropic credentials found ($creds_file)"
+        return
+    fi
+
+    warn "ANTHROPIC_API_KEY is not set."
+    echo ""
+    gum style --foreground "$C_SKY" --margin "0 4" \
+        "Claude Code needs an Anthropic API key to run inside the container." \
+        "Get one at: https://console.anthropic.com/settings/api-keys"
+    echo ""
+
+    local key
+    key=$(gum input \
+        --password \
+        --placeholder "sk-ant-..." \
+        --prompt "  › " \
+        --prompt.foreground "$C_MAUVE" \
+        --cursor.foreground "$C_MAUVE" \
+        --header "  Anthropic API key" \
+        --header.foreground "$C_SKY" \
+        --width 70) || true
+
+    if [[ -z "$key" ]]; then
+        warn "No API key entered — set ANTHROPIC_API_KEY before running ./start.sh"
+        return
+    fi
+
+    export ANTHROPIC_API_KEY="$key"
+    success "ANTHROPIC_API_KEY set for this session"
+
+    ask_yn _save "Save to shell profile (~/.zshrc / ~/.bashrc)?" "y"
+    if [[ "$_save" == "y" ]]; then
+        local profile
+        if [[ "$SHELL" == */zsh ]]; then
+            profile="$HOME/.zshrc"
+        else
+            profile="$HOME/.bashrc"
+        fi
+        if grep -q "ANTHROPIC_API_KEY" "$profile" 2>/dev/null; then
+            warn "ANTHROPIC_API_KEY already in $profile — not adding again."
+        else
+            printf '\nexport ANTHROPIC_API_KEY="%s"\n' "$key" >> "$profile"
+            success "Saved to $profile"
+        fi
+    fi
+}
+
 check_prerequisites() {
     header "Prerequisites"
     ensure_git
     ensure_container_runtime
+    ensure_api_key
 }
diff --git a/lib/workflow.sh b/lib/workflow.sh
index 9f41ebe..70415af 100644
--- a/lib/workflow.sh
+++ b/lib/workflow.sh
@@ -127,6 +127,7 @@ _create_agent_dir() {
 {
   "model": "sonnet",
   "spinnerTipsEnabled": false,
+  "hasTrustDialogAccepted": true,
   "permissions": {
     "allow": [
       "Bash(*)",
@@ -162,5 +163,6 @@ clone_workflow() {
     spin "Cloning workflow..." \
         git clone "$repo_url" "$workflow_dir" \
         || die "Failed to clone workflow repo: $repo_url"
+
     success "Workflow cloned"
 }
diff --git a/wizard.sh b/wizard.sh
index 1db7f76..e5a89f9 100755
--- a/wizard.sh
+++ b/wizard.sh
@@ -155,10 +155,56 @@ confirm_summary() {
     if [[ "$_ok" != "y" ]]; then die "Aborted."; fi
 }
 
+# ── Cleanup state tracking ────────────────────────────────────────────────
+_WIZARD_CORE_CLONED=false
+_WIZARD_PROJECT_CREATED=false
+
+wizard_cleanup() {
+    local removed=false
+    if [[ "$_WIZARD_PROJECT_CREATED" == "true" && -n "${PROJECT_DIR:-}" && -d "$PROJECT_DIR" ]]; then
+        rm -rf "$PROJECT_DIR"
+        gum style --foreground "$C_GREEN" "  ✓  Removed $PROJECT_DIR"
+        removed=true
+    fi
+    if [[ "$_WIZARD_CORE_CLONED" == "true" && -d "$CS_CORE_DIR" ]]; then
+        rm -rf "$CS_CORE_DIR"
+        gum style --foreground "$C_GREEN" "  ✓  Removed $CS_CORE_DIR"
+        removed=true
+    fi
+    if [[ "$removed" == "false" ]]; then
+        gum style --foreground "$C_SURFACE" "  Nothing to revert."
+    fi
+}
+
+handle_sigint() {
+    trap '' INT  # block further Ctrl+C while dialog is open
+    echo ""
+    if gum confirm \
+        --affirmative "Yes, quit" \
+        --negative "No, continue" \
+        --default=No \
+        --prompt.foreground "$C_YELLOW" \
+        --selected.background "$C_RED" \
+        --selected.foreground "$C_BASE" \
+        --unselected.foreground "$C_TEXT" \
+        "  Abort setup and revert changes?"; then
+        echo ""
+        gum style --foreground "$C_YELLOW" --bold "  Reverting changes..."
+        wizard_cleanup
+        echo ""
+        gum style --foreground "$C_RED" --bold "  Setup aborted."
+        echo ""
+        exit 130
+    fi
+    trap 'handle_sigint' INT  # re-arm after "No, continue"
+}
+
 # ── Build ─────────────────────────────────────────────────────────────────
 build_project() {
     header "Building Project"
 
+    _WIZARD_PROJECT_CREATED=true
+
     local slug
     slug="$(slugify "$PROJECT_NAME")"
 
@@ -230,6 +276,7 @@ print_next_steps() {
 
 # ── Main ──────────────────────────────────────────────────────────────────
 main() {
+    trap 'handle_sigint' INT
     show_banner
     check_prerequisites
     setup_core
@@ -238,6 +285,7 @@ main() {
     confirm_summary
     build_project
     print_next_steps
+    trap - INT
 }
 
 main "$@"