diff --git a/.planning/phases/01-session-process-foundation/01-RESEARCH.md b/.planning/phases/01-session-process-foundation/01-RESEARCH.md
new file mode 100644
index 0000000..6e7a0cd
--- /dev/null
+++ b/.planning/phases/01-session-process-foundation/01-RESEARCH.md
@@ -0,0 +1,565 @@
+# Phase 1: Session & Process Foundation - Research
+
+**Researched:** 2026-02-04
+**Domain:** Python asyncio subprocess management, Claude Code CLI integration, Telegram bot architecture
+**Confidence:** HIGH
+
+## Summary
+
+Phase 1 requires spawning and managing Claude Code CLI subprocesses from a Telegram bot written in Python using python-telegram-bot 22.5 and asyncio. The core technical challenge is safely managing subprocess I/O without deadlocks while handling concurrent Telegram messages.
+
+Research confirms that asyncio provides robust subprocess management primitives, and Claude Code CLI's `--output-format stream-json` provides structured, parseable output ideal for subprocess consumption. The standard pattern is pipes with concurrent stream readers using `asyncio.gather()`, not PTY, as Claude Code doesn't require interactive terminal features for this use case.
+
+Key findings: (1) Always use `communicate()` or concurrent stream readers to avoid pipe deadlocks, (2) Claude Code sessions are directory-based and persistent via `--resume`, (3) python-telegram-bot 22.5 handles async natively but requires careful handler design to avoid blocking, (4) Process cleanup must use `terminate()` + `wait()` to prevent zombie processes.
+
+**Primary recommendation:** Use `asyncio.create_subprocess_exec()` with `PIPE` for stdout/stderr, concurrent `asyncio.gather()` for stream reading, and Claude Code's `--output-format stream-json --verbose` for structured output. Skip PTY complexity unless future phases need interactive features.
+
+## Standard Stack
+
+The established libraries/tools for this domain:
+
+### Core
+| Library | Version | Purpose | Why Standard |
+|---------|---------|---------|--------------|
+| python-telegram-bot | 22.5 | Telegram bot framework | Industry standard for Python Telegram bots, native async/await, comprehensive API coverage |
+| asyncio | stdlib (3.12+) | Async subprocess management | Python's official async framework, subprocess primitives prevent deadlocks |
+| Claude Code CLI | 2.1.31+ | AI agent subprocess | Official CLI with --resume, session persistence, stream-json output |
+
+### Supporting
+| Library | Version | Purpose | When to Use |
+|---------|---------|---------|-------------|
+| json | stdlib | Parse stream-json output | Every subprocess output line (NDJSON format) |
+| pathlib | stdlib | Session directory management | File/directory operations for `~/telegram/sessions/` |
+| typing | stdlib | Type hints for session metadata | Code clarity and IDE support |
+
+### Alternatives Considered
+| Instead of | Could Use | Tradeoff |
+|------------|-----------|----------|
+| asyncio.create_subprocess_exec | pty.spawn + asyncio | PTY adds complexity (terminal emulation, signal handling) without benefit for non-interactive CLI |
+| python-telegram-bot | aiogram | aiogram is also async but has smaller ecosystem, PTB is more mature |
+| Pipes | PTY (pseudo-terminal) | PTY needed only for programs requiring terminal features (color codes, cursor control) - Claude Code works fine with pipes |
+
+**Installation:**
+```bash
+# Already installed on mgmt container
+source ~/venv/bin/activate
+pip show python-telegram-bot  # Version: 22.5
+which claude  # /home/mikkel/.local/bin/claude
+claude --version  # 2.1.31 (Claude Code)
+```
+
+## Architecture Patterns
+
+### Recommended Project Structure
+```
+telegram/
+├── bot.py                    # Existing bot entry point
+├── sessions/                 # NEW: Session storage
+│   ├── <name>/              # Per-session directory
+│   │   ├── metadata.json    # Session state (PID, timestamps, persona)
+│   │   └── .claude/         # Claude Code session data (auto-created)
+├── personas/                # NEW: Persona library
+│   ├── brainstorm.json      # Shared persona templates
+│   ├── planner.json
+│   └── research.json
+├── session_manager.py       # NEW: Session lifecycle management
+└── claude_subprocess.py     # NEW: Subprocess I/O handling
+```
+
+### Pattern 1: Concurrent Stream Reading (CRITICAL)
+**What:** Read stdout and stderr concurrently using `asyncio.gather()` to prevent pipe buffer overflow
+**When to use:** Every subprocess with `PIPE` for stdout/stderr
+**Example:**
+```python
+# Source: https://docs.python.org/3/library/asyncio-subprocess.html
+import asyncio
+
+async def read_stream(stream, callback):
+    """Read stream line by line, invoke callback for each line."""
+    while True:
+        line = await stream.readline()
+        if not line:
+            break
+        callback(line.decode().rstrip())
+
+async def run_claude(session_dir, message):
+    proc = await asyncio.create_subprocess_exec(
+        'claude', '-p', message,
+        '--output-format', 'stream-json',
+        '--verbose',
+        stdout=asyncio.subprocess.PIPE,
+        stderr=asyncio.subprocess.PIPE,
+        cwd=session_dir
+    )
+
+    # Concurrent reading prevents deadlock
+    await asyncio.gather(
+        read_stream(proc.stdout, handle_stdout),
+        read_stream(proc.stderr, handle_stderr)
+    )
+
+    await proc.wait()
+```
+
+### Pattern 2: Session Directory Isolation
+**What:** Each session gets its own directory; Claude Code automatically manages session state
+**When to use:** Every session creation/switch
+**Example:**
+```python
+# Source: Phase context + Claude Code CLI reference
+from pathlib import Path
+import json
+
+def create_session(name: str, persona: str = None):
+    """Create new session with isolated directory."""
+    session_dir = Path.home() / 'telegram' / 'sessions' / name
+    session_dir.mkdir(parents=True, exist_ok=True)
+
+    metadata = {
+        'name': name,
+        'created': datetime.now().isoformat(),
+        'persona': persona,
+        'pid': None,
+        'status': 'idle'
+    }
+
+    # Write metadata
+    (session_dir / 'metadata.json').write_text(json.dumps(metadata, indent=2))
+
+    # Copy persona if specified
+    if persona:
+        persona_file = Path.home() / 'telegram' / 'personas' / f'{persona}.json'
+        if persona_file.exists():
+            (session_dir / 'persona.json').write_text(persona_file.read_text())
+
+    return session_dir
+```
+
+### Pattern 3: Stream-JSON Event Handling
+**What:** Parse newline-delimited JSON events from Claude Code output
+**When to use:** Processing subprocess output in real-time
+**Example:**
+```python
+# Source: https://code.claude.com/docs/en/headless + stream-json research
+import json
+
+def handle_stdout(line: str):
+    """Parse and route stream-json events."""
+    try:
+        event = json.loads(line)
+        event_type = event.get('type')
+
+        if event_type == 'assistant':
+            # Claude's response
+            content = event['message']['content']
+            for block in content:
+                if block['type'] == 'text':
+                    send_to_telegram(block['text'])
+
+        elif event_type == 'result':
+            # Task complete
+            session_id = event['session_id']
+            update_session_state(session_id, 'idle')
+
+        elif event_type == 'system':
+            # System events (hooks, init)
+            pass
+
+    except json.JSONDecodeError:
+        logger.warning(f"Invalid JSON: {line}")
+```
+
+### Pattern 4: Process Lifecycle Management
+**What:** Spawn on session switch, suspend (don't kill), rely on Phase 3 timeout for cleanup
+**When to use:** Session switching, process termination
+**Example:**
+```python
+# Source: Asyncio subprocess best practices + Phase context decisions
+import asyncio
+import signal
+
+async def switch_session(new_session: str):
+    """Switch to new session, suspend current process."""
+    current = get_active_session()
+
+    # Mark current as suspended (don't kill)
+    if current and current.proc:
+        current.status = 'suspended'
+        save_metadata(current)
+        # Process stays alive, Phase 3 timeout handles cleanup
+
+    # Activate new session
+    new = load_session(new_session)
+    if not new.proc or new.proc.returncode is not None:
+        # No process or dead - spawn new one
+        new.proc = await spawn_claude(new.session_dir)
+
+    set_active_session(new)
+
+async def terminate_gracefully(proc, timeout=10):
+    """Terminate subprocess with timeout, prevent zombies."""
+    # Source: Python asyncio subprocess best practices research
+    try:
+        proc.terminate()  # Send SIGTERM
+        await asyncio.wait_for(proc.wait(), timeout=timeout)
+    except asyncio.TimeoutError:
+        proc.kill()  # Force SIGKILL
+        await proc.wait()  # CRITICAL: Always await to prevent zombies
+```
+
+### Pattern 5: Non-Blocking Telegram Handlers
+**What:** Use `block=False` for handlers that spawn long-running tasks
+**When to use:** Message handlers that interact with Claude Code subprocess
+**Example:**
+```python
+# Source: https://github.com/python-telegram-bot/python-telegram-bot/wiki/Concurrency
+from telegram.ext import Application, MessageHandler, filters
+
+async def handle_message(update, context):
+    """Handle incoming Telegram messages."""
+    session = get_active_session()
+
+    if not session:
+        await update.message.reply_text("No active session. Use /new <name>")
+        return
+
+    # Queue message to subprocess (non-blocking)
+    await session.send_message(update.message.text)
+
+# Register with block=False for concurrency
+app.add_handler(MessageHandler(
+    filters.TEXT & ~filters.COMMAND,
+    handle_message,
+    block=False
+))
+```
+
+### Anti-Patterns to Avoid
+- **Direct stream reading without concurrency:** Calling `await proc.stdout.read()` then `await proc.stderr.read()` sequentially will deadlock if stderr fills up first
+- **Using `wait()` with pipes:** `await proc.wait()` deadlocks if stdout/stderr buffers fill; always use `communicate()` or concurrent stream readers
+- **Killing processes without cleanup:** `proc.kill()` without `await proc.wait()` creates zombie processes
+- **PTY for non-interactive programs:** PTY adds signal handling complexity; Claude Code CLI works fine with pipes
+
+## Don't Hand-Roll
+
+Problems that look simple but have existing solutions:
+
+| Problem | Don't Build | Use Instead | Why |
+|---------|-------------|-------------|-----|
+| Concurrent stream reading | Manual threading or sequential reads | `asyncio.gather()` with StreamReader | Prevents deadlocks, handles backpressure, battle-tested |
+| JSON Lines parsing | Custom line-by-line JSON parser | `json.loads()` per line with try/except | Standard library is fast, handles edge cases |
+| Session ID generation | Custom UUID logic | `uuid.uuid4()` from stdlib | Cryptographically secure, standard format |
+| Process termination | Manual signal handling | `proc.terminate()` + `asyncio.wait_for(proc.wait())` | Handles timeout, cleanup, zombie prevention |
+
+**Key insight:** Asyncio subprocess management has well-documented pitfalls (deadlocks, zombies). Use standard patterns from official docs rather than custom solutions.
+
+## Common Pitfalls
+
+### Pitfall 1: Pipe Deadlock from Sequential Reading
+**What goes wrong:** Reading stdout then stderr sequentially causes deadlock if stderr fills buffer first
+**Why it happens:** OS pipe buffers are finite (~64KB). If stderr fills while code waits on stdout, child process blocks writing, parent blocks reading - deadlock.
+**How to avoid:** Always read stdout and stderr concurrently using `asyncio.gather()`
+**Warning signs:** Subprocess hangs indefinitely, no output, high CPU usage from blocked I/O
+
+```python
+# WRONG - Sequential reading
+stdout_data = await proc.stdout.read()  # Blocks forever if stderr fills first
+stderr_data = await proc.stderr.read()
+
+# RIGHT - Concurrent reading
+async def read_all(stream):
+    return await stream.read()
+
+stdout_data, stderr_data = await asyncio.gather(
+    read_all(proc.stdout),
+    read_all(proc.stderr)
+)
+```
+
+### Pitfall 2: Zombie Processes from Missing wait()
+**What goes wrong:** Process terminates but stays in zombie state (shows as `<defunct>` in ps)
+**Why it happens:** Parent must call `wait()` to let OS reclaim process resources. Forgetting this after `terminate()`/`kill()` leaves zombies.
+**How to avoid:** ALWAYS `await proc.wait()` after termination, even after `kill()`
+**Warning signs:** `ps aux` shows increasing number of `<defunct>` processes, eventual resource exhaustion
+
+```python
+# WRONG - Zombie process
+proc.terminate()
+# Process is now zombie - resources not reclaimed
+
+# RIGHT - Clean termination
+proc.terminate()
+await proc.wait()  # CRITICAL - reaps zombie
+```
+
+### Pitfall 3: Blocking Telegram Bot Event Loop
+**What goes wrong:** Long-running subprocess operations freeze bot, no messages processed
+**Why it happens:** Telegram handlers run on main event loop. Blocking operations (like `communicate()` on long-running process) block all handlers.
+**How to avoid:** Use `block=False` in handler registration, or spawn background tasks with `asyncio.create_task()`
+**Warning signs:** Bot becomes unresponsive during Claude Code processing, commands queue up
+
+```python
+# WRONG - Blocks event loop
+async def handle_message(update, context):
+    stdout, stderr = await proc.communicate()  # Blocks for minutes
+    await update.message.reply_text(stdout)
+
+# RIGHT - Non-blocking handler
+app.add_handler(MessageHandler(
+    filters.TEXT,
+    handle_message,
+    block=False  # Runs as asyncio.Task
+))
+```
+
+### Pitfall 4: Assuming Claude Code Session Isolation
+**What goes wrong:** Spawning multiple Claude Code processes in same directory causes session conflicts
+**Why it happens:** Claude Code manages session state in `.claude/` subdirectory. Multiple processes in same directory share session state, corrupting history.
+**How to avoid:** Each session must have its own directory (`~/telegram/sessions/<name>/`). Change `cwd` parameter when spawning subprocess.
+**Warning signs:** Session history mixed between conversations, `--resume` loads wrong context
+
+```python
+# WRONG - Shared directory
+proc = await asyncio.create_subprocess_exec('claude', '-p', msg)
+
+# RIGHT - Isolated directory per session
+session_dir = Path.home() / 'telegram' / 'sessions' / session_name
+proc = await asyncio.create_subprocess_exec(
+    'claude', '-p', msg,
+    cwd=str(session_dir)
+)
+```
+
+### Pitfall 5: Ignoring stream-json Event Types
+**What goes wrong:** Only handling 'assistant' events misses errors, tool confirmations, completion status
+**Why it happens:** stream-json emits multiple event types (system, assistant, result). Parsing only one type loses critical information.
+**How to avoid:** Handle all event types in stream parser, especially 'result' for completion status and 'system' for errors
+**Warning signs:** Missing error notifications, unclear when Claude finishes processing, tool use not tracked
+
+```python
+# WRONG - Only handles assistant messages
+if event['type'] == 'assistant':
+    send_to_telegram(event['message'])
+
+# RIGHT - Handle all event types
+if event['type'] == 'assistant':
+    send_to_telegram(event['message'])
+elif event['type'] == 'result':
+    mark_session_complete(event)
+elif event['type'] == 'system' and event.get('subtype') == 'error':
+    notify_user_error(event)
+```
+
+## Code Examples
+
+Verified patterns from official sources:
+
+### Creating and Managing Subprocess
+```python
+# Source: https://docs.python.org/3/library/asyncio-subprocess.html
+import asyncio
+from pathlib import Path
+
+async def spawn_claude_subprocess(session_dir: Path, initial_message: str):
+    """Spawn Claude Code subprocess for session."""
+    proc = await asyncio.create_subprocess_exec(
+        'claude',
+        '-p', initial_message,
+        '--output-format', 'stream-json',
+        '--verbose',
+        '--continue',  # Resume session if exists
+        stdout=asyncio.subprocess.PIPE,
+        stderr=asyncio.subprocess.PIPE,
+        cwd=str(session_dir)
+    )
+    return proc
+```
+
+### Concurrent Stream Reading
+```python
+# Source: https://docs.python.org/3/library/asyncio-subprocess.html
+async def read_stream(stream, callback):
+    """Read stream line-by-line, invoke callback for each line."""
+    while True:
+        line = await stream.readline()
+        if not line:
+            break
+        callback(line.decode().rstrip())
+
+async def run_with_stream_handlers(proc, stdout_handler, stderr_handler):
+    """Run subprocess with concurrent stdout/stderr reading."""
+    await asyncio.gather(
+        read_stream(proc.stdout, stdout_handler),
+        read_stream(proc.stderr, stderr_handler),
+        proc.wait()
+    )
+```
+
+### Graceful Process Termination
+```python
+# Source: Python asyncio subprocess research (multiple sources)
+import asyncio
+
+async def terminate_process(proc, timeout: int = 10):
+    """Terminate subprocess gracefully, prevent zombie."""
+    if proc.returncode is not None:
+        return  # Already terminated
+
+    try:
+        proc.terminate()  # Send SIGTERM
+        await asyncio.wait_for(proc.wait(), timeout=timeout)
+    except asyncio.TimeoutError:
+        proc.kill()  # Force SIGKILL
+        await proc.wait()  # CRITICAL: Always reap zombie
+```
+
+### Session Directory Management
+```python
+# Source: Phase context + research
+from pathlib import Path
+import json
+from datetime import datetime
+
+def create_session_directory(name: str, persona: str = None) -> Path:
+    """Create isolated session directory with metadata."""
+    session_dir = Path.home() / 'telegram' / 'sessions' / name
+    session_dir.mkdir(parents=True, exist_ok=True)
+
+    metadata = {
+        'name': name,
+        'created': datetime.now().isoformat(),
+        'persona': persona,
+        'pid': None,
+        'status': 'idle',
+        'last_active': None
+    }
+
+    metadata_file = session_dir / 'metadata.json'
+    metadata_file.write_text(json.dumps(metadata, indent=2))
+
+    return session_dir
+```
+
+### Parsing stream-json Output
+```python
+# Source: https://code.claude.com/docs/en/headless
+import json
+import logging
+
+logger = logging.getLogger(__name__)
+
+def parse_stream_json_line(line: str):
+    """Parse single line of stream-json output."""
+    try:
+        event = json.loads(line)
+        return event
+    except json.JSONDecodeError:
+        logger.warning(f"Invalid JSON line: {line}")
+        return None
+
+async def handle_claude_output(stream, telegram_chat_id, bot):
+    """Handle Claude Code stream-json output."""
+    while True:
+        line = await stream.readline()
+        if not line:
+            break
+
+        event = parse_stream_json_line(line.decode().rstrip())
+        if not event:
+            continue
+
+        event_type = event.get('type')
+
+        if event_type == 'assistant':
+            # Extract text from assistant message
+            content = event.get('message', {}).get('content', [])
+            for block in content:
+                if block.get('type') == 'text':
+                    text = block.get('text', '')
+                    await bot.send_message(chat_id=telegram_chat_id, text=text)
+
+        elif event_type == 'result':
+            # Task completion
+            if event.get('is_error'):
+                await bot.send_message(
+                    chat_id=telegram_chat_id,
+                    text="Claude encountered an error."
+                )
+```
+
+## State of the Art
+
+| Old Approach | Current Approach | When Changed | Impact |
+|--------------|------------------|--------------|--------|
+| PTY for all subprocess interaction | Pipes with concurrent reading for non-interactive programs | Python 3.6+ asyncio maturity | Simpler code, fewer edge cases, better error handling |
+| Sequential stdout/stderr reading | Concurrent `asyncio.gather()` | Python 3.5 async/await | Eliminates deadlocks from buffer overflow |
+| Manual signal handling for termination | `terminate()` + `wait_for()` with timeout | Python 3.7+ | Graceful shutdown with fallback to SIGKILL |
+| Thread-based Telegram bots | Async python-telegram-bot 20.0+ | v20.0 (2023) | Native async/await, better performance |
+| File-based Claude interaction | Stream-json subprocess with live parsing | Claude Code 2.0+ (2024) | Real-time responses, lower latency |
+
+**Deprecated/outdated:**
+- **python-telegram-bot sync mode (< v20):** Deprecated, removed in v20. All new code must use async/await.
+- **subprocess.PIPE without concurrent reading:** Known deadlock risk since Python 3.4, documented as anti-pattern
+- **PTY for Claude Code:** Unnecessary; Claude Code designed for pipe interaction, handles non-TTY gracefully
+
+## Open Questions
+
+Things that couldn't be fully resolved:
+
+1. **Claude Code auto-restart behavior with --resume**
+   - What we know: `--resume` loads session by ID, `--continue` loads most recent in directory
+   - What's unclear: If Claude Code crashes mid-response, can we auto-restart with `--continue` and it resumes cleanly? Or do we need to track message history ourselves?
+   - Recommendation: Test crash recovery behavior. Likely safe to use `--continue` in session directory after crash - Claude Code manages history in `.claude/` subdirectory.
+
+2. **Optimal buffer limit for long-running sessions**
+   - What we know: `limit` parameter on `create_subprocess_exec()` controls StreamReader buffer size (default 64KB)
+   - What's unclear: Should we increase for Claude Code's potentially long responses? What's the memory tradeoff?
+   - Recommendation: Start with default (64KB). Monitor in Phase 4. Claude Code stream-json outputs line-by-line, so readline() should prevent buffer buildup.
+
+3. **Handling concurrent messages during Claude processing**
+   - What we know: User might send multiple messages while Claude is responding
+   - What's unclear: Queue to subprocess stdin (if using `--input-format stream-json`)? Or wait for completion and send as new turn?
+   - Recommendation: Phase context says "queue messages, send after response completes." For Phase 1, buffer messages in Python and send as new `claude -p` invocation after previous completes. Phase 2+ might use `--input-format stream-json` for live piping.
+
+4. **Session metadata beyond basics**
+   - What we know: Need name, PID, timestamps, persona at minimum
+   - What's unclear: Should we track message count, last message timestamp, token usage, Claude Code session ID?
+   - Recommendation: Keep it minimal for Phase 1. Metadata schema:
+     ```json
+     {
+       "name": "session-name",
+       "created": "2026-02-04T14:20:00Z",
+       "last_active": "2026-02-04T15:30:00Z",
+       "persona": "brainstorm",
+       "pid": 12345,
+       "status": "active|suspended|idle"
+     }
+     ```
+     Add fields in later phases as needed (token tracking in Phase 4, etc.)
+
+## Sources
+
+### Primary (HIGH confidence)
+- [Python asyncio subprocess documentation](https://docs.python.org/3/library/asyncio-subprocess.html) - Official Python 3.14 docs
+- [Claude Code CLI reference](https://code.claude.com/docs/en/cli-reference) - Official Anthropic documentation
+- [Claude Code headless mode](https://code.claude.com/docs/en/headless) - Official programmatic usage guide
+- [python-telegram-bot Concurrency wiki](https://github.com/python-telegram-bot/python-telegram-bot/wiki/Concurrency) - Official PTB documentation
+
+### Secondary (MEDIUM confidence)
+- [Super Fast Python - Asyncio Subprocess](https://superfastpython.com/asyncio-subprocess/) - Practical examples verified against official docs
+- [Python asyncio subprocess termination best practices](https://www.slingacademy.com/article/python-asyncio-how-to-stop-kill-a-child-process/) - Community best practices, verified with official docs
+- [Claude Code session management guide](https://stevekinney.com/courses/ai-development/claude-code-session-management) - Educational resource on Claude sessions
+- [Stream-JSON chaining wiki](https://github.com/ruvnet/claude-flow/wiki/Stream-Chaining) - Community documentation on stream-json format
+
+### Tertiary (LOW confidence)
+- WebSearch results on asyncio best practices - Multiple sources, cross-referenced but not deeply verified
+
+## Metadata
+
+**Confidence breakdown:**
+- Standard stack: HIGH - All libraries verified in use on mgmt container, versions confirmed
+- Architecture: HIGH - Patterns sourced from official Python and Claude Code documentation
+- Pitfalls: HIGH - Documented in Python subprocess docs, verified through official warnings
+
+**Research date:** 2026-02-04
+**Valid until:** 2026-03-04 (30 days - Python asyncio and Claude Code are stable, slow-moving APIs)