Warp Oz

Dispatch coding tasks to Warp Oz cloud agents and chain them into multi-agent pipelines — all from chat. Includes a bash wrapper covering every Oz API endpoint (runs, schedules, artifacts, agents) and a Python pipeline orchestrator that chains specialized agents (e.g., architect → developer → security → red-teamer) with automatic severity-based retry loops. Use when you want to kick off cloud coding agents, poll run status, manage cron schedules, run multi-turn conversations in a shared sandbox, or orchestrate multi-agent review pipelines. NOT for local Warp terminal usage.

Oz Cloud Agent Orchestration

Turn your OpenClaw agent into a development team manager. Dispatch coding tasks to Warp Oz cloud agents from chat, chain them into multi-stage pipelines, and get PRs back — no IDE required.

What you get:

• oz-api.sh — bash wrapper covering every Oz API endpoint (runs, polls, schedules, artifacts, agents)
• orchestrator.py — Python pipeline runner that chains agents with automatic severity-based retry loops
• references/agent-roles.md — battle-tested base prompts for 6 specialized agent roles

Example pipeline: "Build a market maker" →

1. Architect designs the system → ARCHITECTURE.md + PR
2. Trading engineer implements it → 10 modules, 36 tests, PR
3. Security engineer reviews → finds 1 critical, 2 high
4. Red teamer attacks → finds 2 more exploits
5. Orchestrator auto-loops developer on critical findings

All from a single command. Each agent runs in an isolated Docker container with your repo cloned.

Setup

1. Get a Warp API key from app.warp.dev → Settings → API Keys
2. export WARP_API_KEY=wk-... (or set OP_WARP_REFERENCE for 1Password)
3. Create an environment at oz.warp.dev and note the ID

Quick Start

# Kick off a single agent
oz-api.sh run "Implement user authentication" --env YOUR_ENV_ID --name developer

# Poll until done
oz-api.sh poll <run_id>

# Run a full pipeline (architect → developer → security → red-teamer)
python3 orchestrator.py \
  --env YOUR_ENV_ID \
  --task "Build a REST API for user management" \
  --stages architect,developer,security-engineer,red-teamer \
  --skill-prefix owner/repo-name

Commands

# Runs
oz-api.sh run "prompt" [--env ID] [--name N] [--base-prompt T] [--model M] [--title T] [--team] [--skill S] [--conversation-id ID] [--interactive]
oz-api.sh status <run_id>
oz-api.sh list [--state S] [--limit N] [--name N] [--source S]
oz-api.sh poll <run_id> [--interval 10] [--timeout 600]
oz-api.sh cancel <run_id>

# Artifacts & sessions
oz-api.sh artifacts <artifact_uid>
oz-api.sh agents
oz-api.sh session-link <session_uuid>

# Schedules
oz-api.sh schedule-create "prompt" --cron "EXPR" --name "N" [--env ID] [--base-prompt T] [--disabled]
oz-api.sh schedule-list
oz-api.sh schedule-get <id>
oz-api.sh schedule-update <id> --cron "EXPR" --name "N" [--prompt T] [--env ID] [--enabled true|false]
oz-api.sh schedule-delete <id>
oz-api.sh schedule-pause <id>
oz-api.sh schedule-resume <id>

Multi-Turn Conversations

Use --conversation-id to continue where a previous run left off (same sandbox, same context):

# First run returns conversation_id in response
oz-api.sh run "Build the auth module" --env ID --name developer
# Continue in same session
oz-api.sh run "Now add rate limiting" --conversation-id <conversation_id>

Pipeline Orchestrator

orchestrator.py chains agent runs in sequence with:

• Automatic context forwarding — each stage gets the previous stage's summary + session link
• Severity detection — parses status messages for CRITICAL/HIGH/MEDIUM/LOW findings
• Retry loops — auto-loops back to the developer when reviewers find critical/high issues (configurable max retries)
• Shared sandbox — optional conversation_id sharing so agents continue in the same environment
• JSON summary — outputs structured pipeline results to stdout

python3 orchestrator.py \
  --env YOUR_ENV_ID \
  --task "Your task description" \
  --stages architect,developer,security-engineer,red-teamer \
  --skill-prefix owner/repo-name \
  --poll-interval 15 \
  --poll-timeout 5400 \
  --max-retries 2 \
  --no-conversation  # optional: isolate each stage's sandbox

Custom Agent Roles

Define specialized agents by pushing .agents/skills/<name>/SKILL.md to your repo. Oz auto-discovers them. See references/agent-roles.md for 6 production-tested role prompts (architect, trading-engineer, quant, security-engineer, red-teamer, risk-manager).

Known Limitations

• Cancel returns 422 on Docker sandbox workers (Warp-side limitation)
• Artifacts endpoint returns 500 (not 404) for nonexistent UIDs
• No API endpoint to list/manage environments — use oz.warp.dev GUI

Reference

• API endpoints & schemas: references/api.md
• Agent role base prompts: references/agent-roles.md
• Creating Warp skills in repos: references/api.md → "Warp Skills via Git" section