🦞 OpenClaw Soft Engine v2.0.0
This is the first major release of openclaw-soft-engine. The project has evolved from a simple "model matrix template" into a complete Soft Execution Engine solution for AI Agents.
🚀 Core Updates
1. Architectural Refactoring: DAG Scheduling Protocol
- Introduced the
AGENTS.mdcore behavioral protocol to support task scheduling based on DAG (Directed Acyclic Graph). - Strengthened the "planning first" principle, enabling concurrent execution and dependency management of complex engineering tasks via
[PLAN]and[CHECKPOINT]mechanisms.
2. Intelligent Routing Gateway (Matrix Proxy)
- Multi-Expert Dispatch:
proxy.jsautomatically routes payloads to the most suitable model expert based on task modality (Text, Tool, Vision, Reason, Code, Embed, Rerank). - DeepSeek-R1 Deep Reasoning: Supports independent routing for reasoning chains and an ultra-long 600s logical thinking timeout.
- Cross-Modal Data Sanitization: A bottom-layer interception mechanism that automatically strips image payloads from pure-text models to prevent system crashes (400 Bad Request).
3. System-Level Security Hardening
- Resource Fusing Protection: Container-level 512MB RAM hard limit to prevent OOM, and a built-in 45s
AbortControllertimeout lock for external tools. - Path Traversal Defense: Hardened
proxy.jsmedia parsing logic to prevent path traversal vulnerability risks. - API Isolation Pool: Supports 6 independent API Key configurations, completely resolving the 429 rate limit bottleneck under high concurrency.
4. Infrastructure & Documentation
- Provider Agnostic: Completely decoupled from single cloud platform bindings; supports any OpenAI-compatible interface.
- Dual-Engine Fetcher:
web_fetch.cjsnow supports intelligent switching between Jina and Firecrawl. - Documentation System: Rewrote
README.md, including a 10-step practical deployment guide and a 14-item deep troubleshooting archive.
🛠️ Deployment
Please refer to the README.md for local environment initialization and container deployment instructions.