Rocky Know-How

Learning knowledge skill v3.4.8 — vector search with bge-large-zh-v1.5, auto-fallback to keyword search, preprocessing for numbers/English/spell-correction....

What to consider before installing: - Review the code first: handler.js and scripts perform reading of session messages and write them to ~/.openclaw/.learnings; inspect these files for any unexpected network endpoints or obfuscated behavior. - Credential access: the hook attempts to read OpenClaw config and auth stores (openclaw.json, auth-profiles.json) and may use stored API keys / OAuth tokens to call model providers. If you do not want a skill to use agent-managed credentials, do not install it or run it in a sandbox. - Privacy: conversation content and errors may be persisted to disk (experiences.md, drafts, pending). If conversations include sensitive data, this will be captured locally and accessible to other local users/agents sharing the same directory. - Prompt injection: the skill injects text into the agent's systemPrompt (documented). That can bias or alter agent behavior across sessions — evaluate whether you want that. - Inconsistencies: docs and hook metadata differ (events like agent_end vs before_compaction, and code comments show different flows). Treat the published documentation as possibly out-of-date and prefer inspecting handler.js for actual behavior. Recommended actions: - If you still want to use it, run it in an isolated environment (container or dedicated user) so persisted learnings and any credential reads are confined. - Manually inspect handler.js (resolveProviderInfo, callLLMApi) and auto-review.sh/record.sh for any network calls beyond model providers or any code that writes outside ~/.openclaw/.learnings. - Backup or restrict access to openclaw auth files or remove credentials you do not wish the skill to use. - When in doubt, decline installation or request a version from a verified source with clearer manifest declarations of required config and credential access.

Type: OpenClaw Skill Name: rocky-know-how Version: 3.4.8 The skill implements an automated knowledge-retention system with high-risk capabilities, including the modification of the core OpenClaw configuration file (`openclaw.json`) during installation (`install.sh`) and the execution of shell scripts using data processed by LLMs (`handler.js`). While the bundle includes extensive security measures such as shell escaping, path traversal validation (`search.sh`), and atomic file locking (`common.sh`), the complexity of the automated recording loop and the reliance on shell execution for core logic create a significant attack surface. The system also performs network requests to local embedding APIs (`vectors.sh`) and remote LLM providers, which is consistent with its stated purpose but remains a high-privilege behavior.