Wiki Knowledge Agent

🌐 한국어 문서 (Korean README)

Turn chat-pasted text/links into a verified, translated, searchable wiki knowledge base. Works on any agent platform (Hermes, Claude Code, Codex, Cursor) with zero external dependencies — just curl + Python stdlib.

• 🔎 Verify links on arrival (HTTP status, redirects) and classify content
• 📢 Ad detection — labels pure ads vs. genuinely useful info (no auto-deletion, provenance kept)
• 🌐 Translate + summarize into your language (default: the user's native language)
• 🗂️ Save to a wiki inbox with structured metadata
• 🔔 Local-first alerts — time-sensitive info (security, expiry, project risk) is queued locally and reported in chat; optional webhook/email/ntfy if you add them later
• 📚 RAG-ready — agents search the wiki as a knowledge channel, with an auto-maintained topic→file map

Why

Many people paste links/text into chat and expect the AI to remember them. Agents forget; wikis don't. This skill gives agents a repeatable procedure for capturing, judging, translating, storing, and surfacing knowledge — on any platform, with no SaaS account required.

Requirements

• Python 3.9+
• curl (for link verification)
• A model that can follow the SKILL.md instructions (any capable LLM agent)

Optional: pyyaml for config read/write (falls back to a minimal parser/writer without it).

Quickstart

# 1. Clone / copy this repo into your agent's skills directory
#    (or load SKILL.md into your agent's skill folder)

# 2. Run onboarding to create the config
python3 scripts/onboarding.py

Install per agent (skill folder locations)

Example install for Claude Code:

git clone https://github.com/atukunare/wiki-knowledge-agent.git
mkdir -p ~/.claude/skills/wiki-knowledge-agent
cp -R wiki-knowledge-agent/SKILL.md wiki-knowledge-agent/scripts \
      wiki-knowledge-agent/references wiki-knowledge-agent/templates \
      ~/.claude/skills/wiki-knowledge-agent/

Troubleshooting — "the agent says it doesn't have this skill"

• Hermes / Claude Code / Cursor: verify the SKILL.md is inside the correct skills folder (paths above). Skills are usually scanned at session start — restart the session after installing.
• Codex CLI: it does not auto-scan skill folders in all versions. Add a line to AGENTS.md:

  ## Wiki saves
  Use the wiki-knowledge-agent skill at ~/.codex/skills/wiki-knowledge-agent/SKILL.md
  when the user asks to save/archive a link or text.
  

• Still not found? Check the folder name matches the skill name (wiki-knowledge-agent), and that SKILL.md sits directly in that folder (not nested). Then re-run onboarding (python3 scripts/onboarding.py) and confirm ~/.config/wiki-knowledge-agent/config.yaml exists.
• The scripts themselves don't need the agent to find them — you can also call python3 scripts/ingest.py --verify <url> directly from a terminal; only the classification/translation step requires an LLM that has read SKILL.md.

Onboarding asks:

1. Wiki root path → default ~/wiki
2. Input channels → any, or a specific list (discord/slack/…); if you pick nothing, the current chat becomes the default — but content pasted into other channels is still judged and saved
3. Target language → default ko (change to your native language)
4. Alert channel → optional webhook/email; without one, alerts go to the local queue + current chat. Add alerts later anytime: just tell the model "알림 채널 추가해줘" — it updates the config.

Usage

Ingest (when content arrives in chat)

The model follows SKILL.md: verify → classify → translate/summarize → save → maybe alert.

Script helpers:

# Verify a URL
python3 scripts/ingest.py --verify "https://example.com/article"
# ✅ 200 → https://example.com/article

# Save a prepared note
python3 scripts/ingest.py --save /tmp/note.md \
  --source-url "https://..." --channel discord --classification useful --language ko
# ✅ 인박스: ~/wiki/knowledge/inbox/2026-08-26-example.md

# Alert queue
python3 scripts/ingest.py --notify "API key expiring 2026-09-01" --category account_expiry
python3 scripts/ingest.py --alerts            # read
python3 scripts/ingest.py --alerts --clear    # archive

Storage layout (under wiki_root)

<wiki_root>/
├── knowledge/
│   ├── inbox/YYYY-MM-DD-<topic>.md   ← collected notes (distilled, not verbatim)
│   └── <topic>.md                    ← integrated knowledge
├── alerts/
│   ├── notifications.md              ← Tier-0 alert queue
│   └── archive/                      ← cleared alerts
└── knowledge-base-map.md             ← RAG inventory (topic → file)

RAG (agents using the wiki as a knowledge channel)

• Search: search_files(pattern=..., path=<wiki_root>/knowledge)
• Fast mapping: read knowledge-base-map.md
• Optional weekly cron: verify inbox → merge into knowledge/<topic>.md → archive processed files

Configuration

~/.config/wiki-knowledge-agent/config.yaml (override with WIKI_AGENT_CONFIG).

Full reference: templates/config.example.yaml

wiki_root: "~/wiki"
target_language: "ko"
translate: true
input:
  sources: ["any"]            # or ["discord", "slack", ...]
  default_channel: "current"
notify:
  tier0: true                 # always on
  webhook: ""                 # optional
  email: ""                   # optional
  ntfy_topic: ""              # optional
alert_on:
  security: true
  account_expiry: true
  project_risk: true
  interest: false

Docs

• SKILL.md — the agent-facing procedure (load this into your agent)
• references/ad-detection.md — ad-vs-useful classification guide with worked examples
• templates/config.example.yaml — full config reference

License

MIT