v0.3.0 — Research-Backed Multi-Agent Intelligence

Highlights

This release brings research-backed intelligence features that no competitor offers, plus production-grade reliability primitives.

🧠 Multi-Agent Intelligence (Phase 1)

• Max 4 Workers warning — backed by Google/MIT empirical research (arXiv:2512.08296)
• Intent-based prompts — military C2 Auftragstaktik: give agents intent + end_state + constraints, not step-by-step instructions
• Boids emergence rules — Reynolds 1986 flocking rules adapted for LLM agent coordination
• Metacognitive self-assessment — agents tag their confidence levels

💰 Reliability + Cost Dashboard (Phase 2)

• Cost Dashboard MVP — real-time token/cost aggregation by agent, model, and task. clawteam board cost — no competitor has this
• Circuit Breaker — healthy → degraded → open tri-state with half-open probing
• Retry with exponential backoff — resilient agent spawning via spawn_with_retry()
• Idempotency keys — deduplication for create() and send()

🎯 Per-Agent Model Resolution (Phase 3)

• 7-level priority chain: CLI > agent model > agent tier > template strategy > template model > config default > None
• Mix Claude, GPT, Qwen, and other models in the same team
• spawn --model, launch --model-strategy CLI support

🔄 Upstream Cherry-Picks

• Runtime live injection — runtime inject/state/watch for tmux inbox messaging
• Waiter zero-tasks fix
• Windows os.replace() compatibility

🛠 OpenClaw 4.2 Compatibility

• Workspace isolation for workers (prevents SOUL.md/AGENTS.md pollution)
• Allowlist path hints at spawn time
• --agent flag detection and graceful fallback

---

Full changelog: CHANGELOG.md

453 tests | ruff clean | Python 3.10-3.12

pip install git+https://github.com/win4r/ClawTeam-OpenClaw.git