zwarm 2.0.0

Multi-Agent CLI Orchestration Research Platform

zwarm

Multi-agent CLI for orchestrating coding agents. Spawn, manage, and converse with multiple Codex sessions in parallel.

Installation

# From the workspace
cd /path/to/labs
uv sync

# Or install directly
uv pip install -e ./zwarm

Requirements:

• Python 3.13+
• codex CLI installed and authenticated

Environment:

export OPENAI_API_KEY="sk-..."        # Required for Codex
export WEAVE_PROJECT="entity/zwarm"   # Optional: Weave tracing

Two Modes

zwarm has two ways to orchestrate coding agents:

Mode	Who's in charge	Use case
zwarm interactive	You	Manual control, experimentation
zwarm orchestrate	LLM	Autonomous task execution

Both use the same underlying session manager - the orchestrator LLM has access to the exact same tools you do.

---

Interactive Mode

You are the orchestrator. Spawn sessions, check on them, continue conversations.

zwarm interactive

Commands

Command	Description
spawn "task"	Start a session (waits for completion)
spawn --async "task"	Start async (returns immediately)
spawn -d /path "task"	Start in specific directory
ls	List all sessions
? <id>	Quick peek: status + latest message
show <id>	Full details: all messages, tokens, etc.
c <id> "msg"	Continue conversation (waits)
ca <id> "msg"	Continue async (returns immediately)
kill <id>	Stop a running session
rm <id>	Delete session entirely
killall	Stop all running sessions
clean	Remove sessions older than 7 days
q	Quit

Example Session

$ zwarm interactive

> spawn "Add a login function to auth.py"
✓ Started session a1b2c3d4, waiting...
[a1b2c3d4] codex (completed) - 32s
Response: I've added a login function with JWT support...

> spawn --async "Fix the type errors in utils.py"
✓ Session: b2c3d4e5 (running in background)

> spawn --async "Add unit tests for auth.py"
✓ Session: c3d4e5f6 (running in background)

> ls
1 running | 2 done

ID         │   │ T │ Task                        │ Updated │ Last Message
a1b2c3d4   │ ✓ │ 1 │ Add a login function...     │ 2m      │ I've added a login function...
b2c3d4e5   │ ✓ │ 1 │ Fix the type errors...      │ 30s ★   │ Fixed 3 type errors in...
c3d4e5f6   │ ● │ 1 │ Add unit tests...           │ 5s      │ (working...)

> ? b2c3d4e5
✓ b2c3d4e5 Fixed 3 type errors: Optional[str] -> str | None, added missing...

> c a1b2c3d4 "Now add password hashing with bcrypt"
Continuing session a1b2c3d4...
[a1b2c3d4] codex (completed) - 28s
Response: Done! I've updated the login function to use bcrypt...

> rm b2c3d4e5
✓ Deleted session b2c3d4e5

> q

Session Status Icons

Icon	Status
●	Running
✓	Completed
✗	Failed
○	Killed
★	Recently completed (< 60s)

---

Orchestrate Mode

An LLM is the orchestrator. Give it a task and it delegates to coding agents.

zwarm orchestrate --task "Build a REST API with authentication"

The orchestrator LLM uses the same tools available in interactive mode:

Tool	Description
delegate(task, ...)	Start a new session
converse(id, msg)	Continue a conversation
peek_session(id)	Quick status check
check_session(id)	Full session details
list_sessions()	List all sessions with needs_attention flags
end_session(id, delete=False)	Kill/delete a session

Task Input

# Direct
zwarm orchestrate --task "Build a REST API"

# From file
zwarm orchestrate --task-file task.md

# From stdin
echo "Fix the bug in auth.py" | zwarm orchestrate

---

Configuration

zwarm looks for config in this order:

1. --config flag
2. .zwarm/config.toml
3. config.toml in working directory

Minimal Config

[weave]
enabled = true
project = "your-entity/zwarm"

[executor]
adapter = "codex_mcp"
model = "gpt-5.1-codex-mini"

Full Config Reference

[orchestrator]
lm = "gpt-5-mini"
max_steps = 100

[orchestrator.compaction]
enabled = true
max_tokens = 100000
threshold_pct = 0.85
target_pct = 0.7

[executor]
adapter = "codex_mcp"
model = "gpt-5.1-codex-mini"
sandbox = "workspace-write"
timeout = 300

[weave]
enabled = true
project = "your-entity/zwarm"

[watchers]
enabled = true
watchers = [
    { name = "progress" },
    { name = "budget", config = { max_steps = 50, max_sessions = 10 } },
    { name = "delegation_reminder", config = { threshold = 10 } },
]

---

Session Management

Sessions are the core abstraction. Each session is a conversation with a Codex agent.

Lifecycle

spawn → running → completed/failed
                      ↓
              continue → running → completed
                              ↓
                        continue → ...

Storage

.zwarm/sessions/<uuid>/
├── meta.json           # Status, task, model, messages, tokens
└── turns/
    ├── turn_1.jsonl    # Raw codex output for turn 1
    ├── turn_2.jsonl    # Output after first continue
    └── ...

Sync vs Async

Mode	Spawn	Continue	Use case
Sync	spawn "task"	c id "msg"	Sequential work, immediate feedback
Async	spawn --async "task"	ca id "msg"	Parallel work, batch processing

Async sessions return immediately. Poll with ls or ? to check status.

---

Watchers

Watchers monitor agent behavior and intervene when needed.

Watcher	Purpose
progress	Detects stuck/spinning agents
budget	Enforces step/session limits
scope	Detects scope creep
delegation_reminder	Nudges orchestrator to delegate

Configure in config.toml:

[watchers]
enabled = true
watchers = [
    { name = "progress" },
    { name = "budget", config = { max_steps = 50 } },
]

---

CLI Reference

zwarm init              # Initialize .zwarm/ in current directory
zwarm interactive       # Start interactive REPL
zwarm orchestrate       # Start LLM orchestrator
zwarm exec              # Run single executor directly (testing)
zwarm status            # Show current state
zwarm history           # Show event history
zwarm clean             # Remove old sessions

---

Project Structure

zwarm/
├── src/zwarm/
│   ├── sessions/           # Session management (core)
│   │   ├── manager.py      # CodexSessionManager
│   │   └── __init__.py
│   ├── tools/
│   │   └── delegation.py   # Orchestrator tools (delegate, converse, etc.)
│   ├── cli/
│   │   └── main.py         # CLI commands and interactive REPL
│   ├── core/
│   │   ├── config.py       # Configuration loading
│   │   ├── compact.py      # Context window management
│   │   └── state.py        # State persistence
│   ├── watchers/           # Trajectory alignment
│   └── orchestrator.py     # Orchestrator agent
├── docs/
│   └── INTERNALS.md        # Technical architecture
└── README.md