taichi

太极架构多 Agent 协作框架，支持集中式、分布式，元混合三种执行模式。基于 Redis 消息总线，实现 Planner/Drafter/Validator/Dispatcher 四个阶段的工作流。

This package appears to be what it says: a Redis-based multi-agent orchestration framework. Before installing: (1) review configs/configs/skills/manifest.yaml and configs/communication.yaml to ensure allowed_commands and permission rules are strict; (2) inspect install.sh and start.sh and run them in an isolated environment or container (do not run as root); (3) secure Redis (bind to localhost or require AUTH) to avoid exposing the message bus; (4) be aware that tasks can cause the framework to run shell commands — template substitution is simple string replacement and commands are executed via the shell, so untrusted task input or a lax skill manifest can lead to command injection or arbitrary command execution; (5) if you will accept tasks from external/untrusted sources, harden the allowed_commands whitelist and/or use non-shell executors. If you want a more thorough risk assessment, provide the contents of configs/skills/manifest.yaml and configs/communication.yaml and the orchestrator invocation options.

Type: OpenClaw Skill Name: taichi Version: 2.1.0 The 'Taichi Framework' bundle contains a critical Remote Code Execution (RCE) vulnerability due to unsafe command execution in 'taichi-framework/core/skills/skill_executor.py'. The 'bash_executor' skill defined in 'taichi-framework/configs/skills/manifest.yaml' uses a template that performs simple string substitution of user-controlled parameters into a shell command. The associated whitelist check is flawed as it only validates the first word of the resulting command string (using shlex.split), which can be easily bypassed using command chaining characters like ';' or '&&'. This vulnerability is reachable through the framework's standard multi-agent workflow, where input from the PlannerAgent is passed to the DrafterAgent and eventually executed by a worker.