技术方案书全自动写作

技术方案书全自动写作 - 多智能体协作、断点续作、完全自动化

This skill appears to implement exactly what it claims: automated, multi-agent document generation using files and trigger files under ~/.openclaw. Before installing or enabling automatic operation, do the following: - Inspect the bundled scripts (continue-executor.js, trigger-continue.js, scripts/*) to confirm you understand what they read and write. They will scan and modify files in ~/.openclaw/workspace/projects and write logs to ~/.openclaw/logs. - Keep autoContinue disabled (set autoContinue:false in progress.json) until you've tested with a small sandbox project and --dry-run modes (scripts support dry-run). - The guide recommends adding Cron or heartbeat checks; only enable those if you trust the code. Cron/heartbeat will cause periodic autonomous behavior and session spawns on your instance. - Run the scripts in a sandbox account or VM first, and back up any existing projects before migrating. - Note minor code quality issues in some examples (e.g., path expansion usage) — these are bugs, not evidence of exfiltration, but test to ensure scripts behave as expected on your platform. - If you need to be extra cautious, remove or restrict the cron/heartbeat entries and run continuation manually. If you want, I can point out exact lines in the scripts that create trigger files, update progress.json, or recommend a minimal safe test procedure.

Type: OpenClaw Skill Name: tech-proposal-autopilot Version: 1.0.1 The skill bundle implements an autonomous multi-agent workflow for large-scale document generation with built-in persistence and auto-continuation. It utilizes cron jobs and 'heartbeat' mechanisms to trigger new agent sessions without user intervention, which is a high-risk capability. Several Node.js scripts (continue-executor.js, check-and-continue.js) interact with the filesystem and include placeholders for command execution (execSync). While the intent appears aligned with the stated purpose of writing technical proposals, the autonomous lifecycle management and instructions to bypass human oversight in SKILL.md and GUIDE.md create a significant attack surface for unintended or persistent unauthorized actions.