Wikisage

A Karpathy-style persistent LLM wiki. Use when: (1) user says '加进wiki/ingest/摄入', (2) user says '查wiki/wiki里有没有', (3) user says '整理wiki/lint', (4) answering...

What to check before installing/use: - Confirm WIKI_ROOT and WIKI_SKILL_DIR defaults point to a directory you control; the skill will read and write files there (index.md, pages/, log.md, .ingest-cache.json, .lint-history/). - If you enable embeddings (embed.py), be aware it uses boto3 to call Bedrock and Secrets Manager and expects a Secrets Manager secret (WIKI_EMBED_SECRET) containing OpenSearch credentials — only provide these if you trust the skill and understand the network activity to AWS. - The registry metadata lists no required env vars, but SKILL.md/README rely on env vars (MCPORTER_CONFIG, WIKI_ROOT, etc.). That mismatch should be resolved: confirm required env vars and how they are provided to the agent. - The skill strongly recommends an Obsidian filesystem MCP (mcporter) to sandbox reads/writes. If you do not run MCP, the skill will fall back to direct file reads/writes and exec grep — review that fallback behavior and ensure the agent is allowed only to touch the intended wiki directory. - Review the example cron/webhook lines in README: these are user-level examples that can post lint summaries to webhooks; the skill itself does not send network webhooks unless you wire them in your scheduler or run embed.py. - If you plan to enable embedding/semantic search, audit embed.py and only grant the minimum AWS IAM permissions needed (SecretsManager:GetSecretValue, Bedrock invoke, and OpenSearch access) and ensure the secret contains expected fields. - If you need more assurance, ask the publisher to update the registry metadata to list the environment variables and optional permissions (AWS) explicitly, and to state if any network calls are made by default. If you cannot validate the maintainer, run the skill in an isolated environment or with MCP sandboxing enabled.

Type: OpenClaw Skill Name: wikisage Version: 1.0.0 The wikisage skill is a well-structured implementation of a markdown-based knowledge management system inspired by the 'Karpathy LLM wiki' pattern. It provides clear instructions for an AI agent to manage a local wiki using Obsidian-style links and includes utility scripts for deduplication (dedup.py), maintenance (lint.py), and optional semantic search (embed.py). While the scripts possess capabilities such as reading local files, fetching URLs for hashing, and interacting with AWS services (Bedrock/OpenSearch), these behaviors are explicitly documented and directly support the skill's stated purpose of knowledge ingestion and retrieval. No evidence of malicious intent, unauthorized data exfiltration, or hidden persistence mechanisms was found.