--- name: wiki-knowledge-agent description: "Turn chat-pasted text/links into a verified, translated, searchable wiki knowledge base. When a user pastes a link or text (in ANY channel — Discord, Slack, CLI, etc.), fetch+verify it, classify it (useful vs ad), translate+summarize it into the target language, save it to the wiki, and optionally alert on time-sensitive information. Also acts as a RAG channel: agents search the wiki for prior knowledge. Use whenever the user pastes content for archiving, asks to save a link/bookmark, asks '알림 있어?', 'any alerts?', or asks the agent to look up knowledge saved in the wiki." category: note-taking triggers: # English - "save this" - "save this link" - "bookmark" - "bookmark this" - "add to wiki" - "remember this" - "archive this" - "store this" - "clip this" - "note this" - "paste this" - "wiki" - "knowledge base" - "any alerts?" - "check alerts" - "notifications" - "save for later" # 한국어 - "북마크" - "링크 저장" - "링크 저장해줘" - "저장해줘" - "위키에 저장" - "위키에 넣어줘" - "붙여넣기 정리" - "인박스" - "알림 있어?" - "알림 확인" ---

# Wiki Knowledge Agent

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 beyond `curl` + Python stdlib.

## When to Use

- User pastes a link or text into **any chat channel** and expects it to be captured. - User says "저장해줘", "링크 저장", "북마크", "save this", "add to wiki". - User asks "알림 있어?" — read the local alert queue and report. - An agent needs prior knowledge from the wiki (RAG usage — search `wiki_root`). - A scheduled job runs the ingest/notify cycle.

## Design Principles

1. **Platform-independent** — works with a single model (no multi-agent setup required) or many. 2. **Zero external dependencies** — `curl` + Python stdlib only. No webhook/app required for core function. 3. **Local-first alerts** — alerts always land in a local file + CLI output. Optional webhook/email/ntfy if configured. 4. **Any-channel input** — a message arriving in any channel is eligible. The onboarding default channel is "the current chat". 5. **LLM does the reasoning** — the scripts fetch/verify/structure; the model classifies ads, translates, and summarizes.

## Configuration

Config file: `~/.config/wiki-knowledge-agent/config.yaml` (default), overridable via `WIKI_AGENT_CONFIG` env var.

Run onboarding to create it interactively:

```bash python3 scripts/onboarding.py ```

Example config (full reference: `templates/config.example.yaml`):

```yaml wiki_root: "~/wiki" target_language: "ko" translate: true

input: sources: ["any"] # "any" = every channel/CLI; or a specific list e.g. ["discord", "slack"] default_channel: "current" # channel used when no preference is set (onboarding default)

notify: tier0: true # always: local file + CLI webhook: "" # optional: Slack/Discord/Telegram webhook URL email: "" # optional: SMTP DSN ntfy_topic: "" # optional: ntfy.sh topic

alert_on: security: true account_expiry: true project_risk: true interest: false ```

### Onboarding flow (interactive questions)

1. **Wiki root path** → default `~/wiki` 2. **Input channels** → choose from known platforms (discord/slack/weixin/telegram/cli…) **or** "any"; if the user doesn't pick, default is **"current chat"** (`default_channel: current`). Content pasted into *other* channels is still judged and saved (never drop input because it arrived in the "wrong" channel). 3. **Target language** → default user's native language (e.g. `ko`). 4. **Alert channel** → if the user has a webhook/email, capture it; otherwise default is the **current chat** (the agent reports alerts in the same conversation). Alerts can be added later anytime: just ask the model "알림 채널 추가해줘" — the LLM updates the config. 5. Write `config.yaml` and print the summary.

## Ingest Workflow (when a message arrives)

``` [input] pasted text or link in any channel │ ▼ 1. VERIFY (script: scripts/ingest.py --verify <url-or-file>) - Link: curl -s -I -L → HTTP status; 2xx/3xx = valid, 404/410/timeout = dead - Text: mark as unverified text │ ▼ 2. CLASSIFY (LLM, guided by references/ad-detection.md) - 📢 Ad: self-promotion is the core intent, "try it" pitch, link-dump without substance - 📚 Useful: real methodology/data/tool with transferable value - Output: label + one-line reason │ ▼ 3. PRE-SAVE FEEDBACK (LLM — before saving, add value from what you already know) a) Project/context match: if the content relates to anything you know about the user (their projects, goals, interests, past notes), say so in one line: "📌 이거 <프로젝트>에 유용하겠네요" / "📌 This looks useful for <project>" b) Date/event cue: if the content has a date-related element (holiday, birthday, anniversary, deadline, event) and you know the relevant date, offer ONE question: "📅 <이벤트>(<날짜>) <N>일 전에 알려줄까요?" / "📅 <event> is in <N> days — remind you?" - If the user says yes → record it (e.g. add to config important_dates, or note it) - If unsure of the exact date → don't guess; skip the offer c) Keep it light: at most ONE feedback line + ONE date offer. Never block saving on it. │ ▼ 4. TRANSLATE + SUMMARIZE (LLM) - Foreign content → target_language - ≤5-line summary + 3–5 key points - If already in target_language, summarize only │ ▼ 5. SAVE (script: scripts/ingest.py --save ...) - <wiki_root>/knowledge/inbox/YYYY-MM-DD-<topic>.md - Frontmatter: date, source_url, source_channel, classification (ad/useful), language - Tag: related project or "일반" │ ▼ 6. ALERT? (LLM decides against alert_on rules + script: scripts/ingest.py --notify) - If alert-worthy → append to <wiki_root>/alerts/notifications.md + CLI output "[알림] …" + optional webhook/email - Else → silent save; reply "✅ 인박스: <file>" ```

> **Note on reminders**: the skill itself stays light — the PRE-SAVE FEEDBACK step is > LLM reasoning only (no code). If the user wants automated day-before reminders, use > your platform's native scheduling (cron / scheduled messages / agent reminders); > this skill just makes the *offer* and records the date.

### Ingest script usage

```bash # Verify a URL (returns HTTP status) python3 scripts/ingest.py --verify "https://example.com/article"

# Verify a pasted text (saved to temp file first, then verified as content) python3 scripts/ingest.py --verify-content /tmp/pasted.txt

# Save a structured note (after LLM produced the markdown body) python3 scripts/ingest.py --save /tmp/note.md --source-url "https://..." --channel "discord" --classification useful --language ko

# Append to the local alert queue python3 scripts/ingest.py --notify "API key expiring 2026-09-01" --category security

# Read the alert queue (for "알림 있어?") python3 scripts/ingest.py --alerts ```

## RAG Usage (agents searching the wiki)

- Search knowledge: `search_files(pattern=..., path=<wiki_root>/knowledge)` - Fast topic mapping: read `<wiki_root>/knowledge-base-map.md` (auto-updated inventory) — topic → file - Recurring integration: a weekly cron can run the curation flow (see `wiki-inbox-curation` pattern): verify → merge into `knowledge/<topic>.md` → move processed inbox files to `inbox/archive/`.

## Alert Notifications

- **Tier 0 (always)**: append `<wiki_root>/alerts/notifications.md` + print `[알림] …` to stdout. - **Tier 1 (optional)**: webhook (Slack/Discord/Telegram), email (SMTP), or ntfy.sh topic — only if configured. - Alerts are *time-sensitive* items: security issues, account/domain expiry, project risk, and (if enabled) user interests. - To add an alert channel later: user says "알림 채널 추가해줘" → model reads config, asks for the webhook/email, updates config.yaml.

## Pitfalls

- **Never drop input because it came from a "non-default" channel** — judge it and save; the channel is metadata, not a gate. - **Don't store full verbatim text** — store the distilled note (summary + points), keeping the wiki lean. - **Verify before trusting** — a 404 after redirects is a dead link; a ~500-byte stub is a bot-block page, not content. - **Ads are not automatically deleted** — label them 📢 and save only the transferable insight (if any); keep provenance. - **Personal info / secrets in pasted content** — strip identifiers before saving. - **Config path** — respect `WIKI_AGENT_CONFIG` env var if set; never hardcode `~/wiki` into scripts. - **Alert queue growth** — mark items read/archived after reporting (`--alerts --clear` or move to `alerts/archive/`).

## Files

- `scripts/onboarding.py` — interactive config wizard - `scripts/ingest.py` — verify / save / notify / alerts CLI - `templates/config.example.yaml` — full config reference - `references/ad-detection.md` — ad-vs-useful classification rules with worked examples - `README.md` — public repo docs (EN/KR)

--- name: wiki-knowledge-agent description: "Turn chat-pasted text/links into a verified, translated, searchable wiki knowledge base. When a user pastes a link or text (in ANY channel — Discord, Slack, CLI, etc.), fetch+verify it, classify it (useful vs ad), translate+summarize it into the target language, save it to the wiki, and optionally alert on time-sensitive information. Also acts as a RAG channel: agents search the wiki for prior knowledge. Use whenever the user pastes content for archiving, asks to save a link/bookmark, asks '알림 있어?', 'any alerts?', or asks the agent to look up knowledge saved in the wiki." category: note-taking triggers: # English - "save this" - "save this link" - "bookmark" - "bookmark this" - "add to wiki" - "remember this" - "archive this" - "store this" - "clip this" - "note this" - "paste this" - "wiki" - "knowledge base" - "any alerts?" - "check alerts" - "notifications" - "save for later" # 한국어 - "북마크" - "링크 저장" - "링크 저장해줘" - "저장해줘" - "위키에 저장" - "위키에 넣어줘" - "붙여넣기 정리" - "인박스" - "알림 있어?" - "알림 확인" ---

# Wiki Knowledge Agent

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 beyond `curl` + Python stdlib.

## When to Use

- User pastes a link or text into **any chat channel** and expects it to be captured. - User says "저장해줘", "링크 저장", "북마크", "save this", "add to wiki". - User asks "알림 있어?" — read the local alert queue and report. - An agent needs prior knowledge from the wiki (RAG usage — search `wiki_root`). - A scheduled job runs the ingest/notify cycle.

## Design Principles

1. **Platform-independent** — works with a single model (no multi-agent setup required) or many. 2. **Zero external dependencies** — `curl` + Python stdlib only. No webhook/app required for core function. 3. **Local-first alerts** — alerts always land in a local file + CLI output. Optional webhook/email/ntfy if configured. 4. **Any-channel input** — a message arriving in any channel is eligible. The onboarding default channel is "the current chat". 5. **LLM does the reasoning** — the scripts fetch/verify/structure; the model classifies ads, translates, and summarizes.

## Configuration

Config file: `~/.config/wiki-knowledge-agent/config.yaml` (default), overridable via `WIKI_AGENT_CONFIG` env var.

Run onboarding to create it interactively:

```bash python3 scripts/onboarding.py ```

Example config (full reference: `templates/config.example.yaml`):

```yaml wiki_root: "~/wiki" target_language: "ko" translate: true

input: sources: ["any"] # "any" = every channel/CLI; or a specific list e.g. ["discord", "slack"] default_channel: "current" # channel used when no preference is set (onboarding default)

notify: tier0: true # always: local file + CLI webhook: "" # optional: Slack/Discord/Telegram webhook URL email: "" # optional: SMTP DSN ntfy_topic: "" # optional: ntfy.sh topic

alert_on: security: true account_expiry: true project_risk: true interest: false ```

### Onboarding flow (interactive questions)

1. **Wiki root path** → default `~/wiki` 2. **Input channels** → choose from known platforms (discord/slack/weixin/telegram/cli…) **or** "any"; if the user doesn't pick, default is **"current chat"** (`default_channel: current`). Content pasted into *other* channels is still judged and saved (never drop input because it arrived in the "wrong" channel). 3. **Target language** → default user's native language (e.g. `ko`). 4. **Alert channel** → if the user has a webhook/email, capture it; otherwise default is the **current chat** (the agent reports alerts in the same conversation). Alerts can be added later anytime: just ask the model "알림 채널 추가해줘" — the LLM updates the config. 5. Write `config.yaml` and print the summary.

## Ingest Workflow (when a message arrives)

``` [input] pasted text or link in any channel │ ▼ 1. VERIFY (script: scripts/ingest.py --verify <url-or-file>) - Link: curl -s -I -L → HTTP status; 2xx/3xx = valid, 404/410/timeout = dead - Text: mark as unverified text │ ▼ 2. CLASSIFY (LLM, guided by references/ad-detection.md) - 📢 Ad: self-promotion is the core intent, "try it" pitch, link-dump without substance - 📚 Useful: real methodology/data/tool with transferable value - Output: label + one-line reason │ ▼ 3. PRE-SAVE FEEDBACK (LLM — before saving, add value from what you already know) a) Project/context match: if the content relates to anything you know about the user (their projects, goals, interests, past notes), say so in one line: "📌 이거 <프로젝트>에 유용하겠네요" / "📌 This looks useful for <project>" b) Date/event cue: if the content has a date-related element (holiday, birthday, anniversary, deadline, event) and you know the relevant date, offer ONE question: "📅 <이벤트>(<날짜>) <N>일 전에 알려줄까요?" / "📅 <event> is in <N> days — remind you?" - If the user says yes → record it (e.g. add to config important_dates, or note it) - If unsure of the exact date → don't guess; skip the offer c) Keep it light: at most ONE feedback line + ONE date offer. Never block saving on it. │ ▼ 4. TRANSLATE + SUMMARIZE (LLM) - Foreign content → target_language - ≤5-line summary + 3–5 key points - If already in target_language, summarize only │ ▼ 5. SAVE (script: scripts/ingest.py --save ...) - <wiki_root>/knowledge/inbox/YYYY-MM-DD-<topic>.md - Frontmatter: date, source_url, source_channel, classification (ad/useful), language - Tag: related project or "일반" │ ▼ 6. ALERT? (LLM decides against alert_on rules + script: scripts/ingest.py --notify) - If alert-worthy → append to <wiki_root>/alerts/notifications.md + CLI output "[알림] …" + optional webhook/email - Else → silent save; reply "✅ 인박스: <file>" ```

> **Note on reminders**: the skill itself stays light — the PRE-SAVE FEEDBACK step is > LLM reasoning only (no code). If the user wants automated day-before reminders, use > your platform's native scheduling (cron / scheduled messages / agent reminders); > this skill just makes the *offer* and records the date.

### Ingest script usage

```bash # Verify a URL (returns HTTP status) python3 scripts/ingest.py --verify "https://example.com/article"

# Verify a pasted text (saved to temp file first, then verified as content) python3 scripts/ingest.py --verify-content /tmp/pasted.txt

# Save a structured note (after LLM produced the markdown body) python3 scripts/ingest.py --save /tmp/note.md --source-url "https://..." --channel "discord" --classification useful --language ko

# Append to the local alert queue python3 scripts/ingest.py --notify "API key expiring 2026-09-01" --category security

# Read the alert queue (for "알림 있어?") python3 scripts/ingest.py --alerts ```

## RAG Usage (agents searching the wiki)

- Search knowledge: `search_files(pattern=..., path=<wiki_root>/knowledge)` - Fast topic mapping: read `<wiki_root>/knowledge-base-map.md` (auto-updated inventory) — topic → file - Recurring integration: a weekly cron can run the curation flow (see `wiki-inbox-curation` pattern): verify → merge into `knowledge/<topic>.md` → move processed inbox files to `inbox/archive/`.

## Alert Notifications

- **Tier 0 (always)**: append `<wiki_root>/alerts/notifications.md` + print `[알림] …` to stdout. - **Tier 1 (optional)**: webhook (Slack/Discord/Telegram), email (SMTP), or ntfy.sh topic — only if configured. - Alerts are *time-sensitive* items: security issues, account/domain expiry, project risk, and (if enabled) user interests. - To add an alert channel later: user says "알림 채널 추가해줘" → model reads config, asks for the webhook/email, updates config.yaml.

## Pitfalls

- **Never drop input because it came from a "non-default" channel** — judge it and save; the channel is metadata, not a gate. - **Don't store full verbatim text** — store the distilled note (summary + points), keeping the wiki lean. - **Verify before trusting** — a 404 after redirects is a dead link; a ~500-byte stub is a bot-block page, not content. - **Ads are not automatically deleted** — label them 📢 and save only the transferable insight (if any); keep provenance. - **Personal info / secrets in pasted content** — strip identifiers before saving. - **Config path** — respect `WIKI_AGENT_CONFIG` env var if set; never hardcode `~/wiki` into scripts. - **Alert queue growth** — mark items read/archived after reporting (`--alerts --clear` or move to `alerts/archive/`).

## Files

- `scripts/onboarding.py` — interactive config wizard - `scripts/ingest.py` — verify / save / notify / alerts CLI - `templates/config.example.yaml` — full config reference - `references/ad-detection.md` — ad-vs-useful classification rules with worked examples - `README.md` — public repo docs (EN/KR)

--- name: wiki-knowledge-agent description: "Turn chat-pasted text/links into a verified, translated, searchable wiki knowledge base. When a user pastes a link or text (in ANY channel — Discord, Slack, CLI, etc.), fetch+verify it, classify it (useful vs ad), translate+summarize it into the target language, save it to the wiki, and optionally alert on time-sensitive information. Also acts as a RAG channel: agents search the wiki for prior knowledge. Use whenever the user pastes content for archiving, asks to save a link/bookmark, asks '알림 있어?', 'any alerts?', or asks the agent to look up knowledge saved in the wiki." category: note-taking triggers: # English - "save this" - "save this link" - "bookmark" - "bookmark this" - "add to wiki" - "remember this" - "archive this" - "store this" - "clip this" - "note this" - "paste this" - "wiki" - "knowledge base" - "any alerts?" - "check alerts" - "notifications" - "save for later" # 한국어 - "북마크" - "링크 저장" - "링크 저장해줘" - "저장해줘" - "위키에 저장" - "위키에 넣어줘" - "붙여넣기 정리" - "인박스" - "알림 있어?" - "알림 확인" ---

# Wiki Knowledge Agent

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 beyond `curl` + Python stdlib.

## When to Use

- User pastes a link or text into **any chat channel** and expects it to be captured. - User says "저장해줘", "링크 저장", "북마크", "save this", "add to wiki". - User asks "알림 있어?" — read the local alert queue and report. - An agent needs prior knowledge from the wiki (RAG usage — search `wiki_root`). - A scheduled job runs the ingest/notify cycle.

## Design Principles

1. **Platform-independent** — works with a single model (no multi-agent setup required) or many. 2. **Zero external dependencies** — `curl` + Python stdlib only. No webhook/app required for core function. 3. **Local-first alerts** — alerts always land in a local file + CLI output. Optional webhook/email/ntfy if configured. 4. **Any-channel input** — a message arriving in any channel is eligible. The onboarding default channel is "the current chat". 5. **LLM does the reasoning** — the scripts fetch/verify/structure; the model classifies ads, translates, and summarizes.

## Configuration

Config file: `~/.config/wiki-knowledge-agent/config.yaml` (default), overridable via `WIKI_AGENT_CONFIG` env var.

Run onboarding to create it interactively:

```bash python3 scripts/onboarding.py ```

Example config (full reference: `templates/config.example.yaml`):

```yaml wiki_root: "~/wiki" target_language: "ko" translate: true

input: sources: ["any"] # "any" = every channel/CLI; or a specific list e.g. ["discord", "slack"] default_channel: "current" # channel used when no preference is set (onboarding default)

notify: tier0: true # always: local file + CLI webhook: "" # optional: Slack/Discord/Telegram webhook URL email: "" # optional: SMTP DSN ntfy_topic: "" # optional: ntfy.sh topic

alert_on: security: true account_expiry: true project_risk: true interest: false ```

### Onboarding flow (interactive questions)

1. **Wiki root path** → default `~/wiki` 2. **Input channels** → choose from known platforms (discord/slack/weixin/telegram/cli…) **or** "any"; if the user doesn't pick, default is **"current chat"** (`default_channel: current`). Content pasted into *other* channels is still judged and saved (never drop input because it arrived in the "wrong" channel). 3. **Target language** → default user's native language (e.g. `ko`). 4. **Alert channel** → if the user has a webhook/email, capture it; otherwise default is the **current chat** (the agent reports alerts in the same conversation). Alerts can be added later anytime: just ask the model "알림 채널 추가해줘" — the LLM updates the config. 5. Write `config.yaml` and print the summary.

## Ingest Workflow (when a message arrives)

``` [input] pasted text or link in any channel │ ▼ 1. VERIFY (script: scripts/ingest.py --verify <url-or-file>) - Link: curl -s -I -L → HTTP status; 2xx/3xx = valid, 404/410/timeout = dead - Text: mark as unverified text │ ▼ 2. CLASSIFY (LLM, guided by references/ad-detection.md) - 📢 Ad: self-promotion is the core intent, "try it" pitch, link-dump without substance - 📚 Useful: real methodology/data/tool with transferable value - Output: label + one-line reason │ ▼ 3. PRE-SAVE FEEDBACK (LLM — before saving, add value from what you already know) a) Project/context match: if the content relates to anything you know about the user (their projects, goals, interests, past notes), say so in one line: "📌 이거 <프로젝트>에 유용하겠네요" / "📌 This looks useful for <project>" b) Date/event cue: if the content has a date-related element (holiday, birthday, anniversary, deadline, event) and you know the relevant date, offer ONE question: "📅 <이벤트>(<날짜>) <N>일 전에 알려줄까요?" / "📅 <event> is in <N> days — remind you?" - If the user says yes → record it (e.g. add to config important_dates, or note it) - If unsure of the exact date → don't guess; skip the offer c) Keep it light: at most ONE feedback line + ONE date offer. Never block saving on it. │ ▼ 4. TRANSLATE + SUMMARIZE (LLM) - Foreign content → target_language - ≤5-line summary + 3–5 key points - If already in target_language, summarize only │ ▼ 5. SAVE (script: scripts/ingest.py --save ...) - <wiki_root>/knowledge/inbox/YYYY-MM-DD-<topic>.md - Frontmatter: date, source_url, source_channel, classification (ad/useful), language - Tag: related project or "일반" │ ▼ 6. ALERT? (LLM decides against alert_on rules + script: scripts/ingest.py --notify) - If alert-worthy → append to <wiki_root>/alerts/notifications.md + CLI output "[알림] …" + optional webhook/email - Else → silent save; reply "✅ 인박스: <file>" ```

> **Note on reminders**: the skill itself stays light — the PRE-SAVE FEEDBACK step is > LLM reasoning only (no code). If the user wants automated day-before reminders, use > your platform's native scheduling (cron / scheduled messages / agent reminders); > this skill just makes the *offer* and records the date.

### Ingest script usage

```bash # Verify a URL (returns HTTP status) python3 scripts/ingest.py --verify "https://example.com/article"

# Verify a pasted text (saved to temp file first, then verified as content) python3 scripts/ingest.py --verify-content /tmp/pasted.txt

# Save a structured note (after LLM produced the markdown body) python3 scripts/ingest.py --save /tmp/note.md --source-url "https://..." --channel "discord" --classification useful --language ko

# Append to the local alert queue python3 scripts/ingest.py --notify "API key expiring 2026-09-01" --category security

# Read the alert queue (for "알림 있어?") python3 scripts/ingest.py --alerts ```

## RAG Usage (agents searching the wiki)

- Search knowledge: `search_files(pattern=..., path=<wiki_root>/knowledge)` - Fast topic mapping: read `<wiki_root>/knowledge-base-map.md` (auto-updated inventory) — topic → file - Recurring integration: a weekly cron can run the curation flow (see `wiki-inbox-curation` pattern): verify → merge into `knowledge/<topic>.md` → move processed inbox files to `inbox/archive/`.

## Alert Notifications

- **Tier 0 (always)**: append `<wiki_root>/alerts/notifications.md` + print `[알림] …` to stdout. - **Tier 1 (optional)**: webhook (Slack/Discord/Telegram), email (SMTP), or ntfy.sh topic — only if configured. - Alerts are *time-sensitive* items: security issues, account/domain expiry, project risk, and (if enabled) user interests. - To add an alert channel later: user says "알림 채널 추가해줘" → model reads config, asks for the webhook/email, updates config.yaml.

## Pitfalls

- **Never drop input because it came from a "non-default" channel** — judge it and save; the channel is metadata, not a gate. - **Don't store full verbatim text** — store the distilled note (summary + points), keeping the wiki lean. - **Verify before trusting** — a 404 after redirects is a dead link; a ~500-byte stub is a bot-block page, not content. - **Ads are not automatically deleted** — label them 📢 and save only the transferable insight (if any); keep provenance. - **Personal info / secrets in pasted content** — strip identifiers before saving. - **Config path** — respect `WIKI_AGENT_CONFIG` env var if set; never hardcode `~/wiki` into scripts. - **Alert queue growth** — mark items read/archived after reporting (`--alerts --clear` or move to `alerts/archive/`).

## Files

- `scripts/onboarding.py` — interactive config wizard - `scripts/ingest.py` — verify / save / notify / alerts CLI - `templates/config.example.yaml` — full config reference - `references/ad-detection.md` — ad-vs-useful classification rules with worked examples - `README.md` — public repo docs (EN/KR)

--- name: wiki-knowledge-agent description: "Turn chat-pasted text/links into a verified, translated, searchable wiki knowledge base. When a user pastes a link or text (in ANY channel — Discord, Slack, CLI, etc.), fetch+verify it, classify it (useful vs ad), translate+summarize it into the target language, save it to the wiki, and optionally alert on time-sensitive information. Also acts as a RAG channel: agents search the wiki for prior knowledge. Use whenever the user pastes content for archiving, asks to save a link/bookmark, asks '알림 있어?', 'any alerts?', or asks the agent to look up knowledge saved in the wiki." category: note-taking triggers: # English - "save this" - "save this link" - "bookmark" - "bookmark this" - "add to wiki" - "remember this" - "archive this" - "store this" - "clip this" - "note this" - "paste this" - "wiki" - "knowledge base" - "any alerts?" - "check alerts" - "notifications" - "save for later" # 한국어 - "북마크" - "링크 저장" - "링크 저장해줘" - "저장해줘" - "위키에 저장" - "위키에 넣어줘" - "붙여넣기 정리" - "인박스" - "알림 있어?" - "알림 확인" ---

# Wiki Knowledge Agent

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 beyond `curl` + Python stdlib.

## When to Use

- User pastes a link or text into **any chat channel** and expects it to be captured. - User says "저장해줘", "링크 저장", "북마크", "save this", "add to wiki". - User asks "알림 있어?" — read the local alert queue and report. - An agent needs prior knowledge from the wiki (RAG usage — search `wiki_root`). - A scheduled job runs the ingest/notify cycle.

## Design Principles

1. **Platform-independent** — works with a single model (no multi-agent setup required) or many. 2. **Zero external dependencies** — `curl` + Python stdlib only. No webhook/app required for core function. 3. **Local-first alerts** — alerts always land in a local file + CLI output. Optional webhook/email/ntfy if configured. 4. **Any-channel input** — a message arriving in any channel is eligible. The onboarding default channel is "the current chat". 5. **LLM does the reasoning** — the scripts fetch/verify/structure; the model classifies ads, translates, and summarizes.

## Configuration

Config file: `~/.config/wiki-knowledge-agent/config.yaml` (default), overridable via `WIKI_AGENT_CONFIG` env var.

Run onboarding to create it interactively:

```bash python3 scripts/onboarding.py ```

Example config (full reference: `templates/config.example.yaml`):

```yaml wiki_root: "~/wiki" target_language: "ko" translate: true

input: sources: ["any"] # "any" = every channel/CLI; or a specific list e.g. ["discord", "slack"] default_channel: "current" # channel used when no preference is set (onboarding default)

notify: tier0: true # always: local file + CLI webhook: "" # optional: Slack/Discord/Telegram webhook URL email: "" # optional: SMTP DSN ntfy_topic: "" # optional: ntfy.sh topic

alert_on: security: true account_expiry: true project_risk: true interest: false ```

### Onboarding flow (interactive questions)

1. **Wiki root path** → default `~/wiki` 2. **Input channels** → choose from known platforms (discord/slack/weixin/telegram/cli…) **or** "any"; if the user doesn't pick, default is **"current chat"** (`default_channel: current`). Content pasted into *other* channels is still judged and saved (never drop input because it arrived in the "wrong" channel). 3. **Target language** → default user's native language (e.g. `ko`). 4. **Alert channel** → if the user has a webhook/email, capture it; otherwise default is the **current chat** (the agent reports alerts in the same conversation). Alerts can be added later anytime: just ask the model "알림 채널 추가해줘" — the LLM updates the config. 5. Write `config.yaml` and print the summary.

## Ingest Workflow (when a message arrives)

``` [input] pasted text or link in any channel │ ▼ 1. VERIFY (script: scripts/ingest.py --verify <url-or-file>) - Link: curl -s -I -L → HTTP status; 2xx/3xx = valid, 404/410/timeout = dead - Text: mark as unverified text │ ▼ 2. CLASSIFY (LLM, guided by references/ad-detection.md) - 📢 Ad: self-promotion is the core intent, "try it" pitch, link-dump without substance - 📚 Useful: real methodology/data/tool with transferable value - Output: label + one-line reason │ ▼ 3. PRE-SAVE FEEDBACK (LLM — before saving, add value from what you already know) a) Project/context match: if the content relates to anything you know about the user (their projects, goals, interests, past notes), say so in one line: "📌 이거 <프로젝트>에 유용하겠네요" / "📌 This looks useful for <project>" b) Date/event cue: if the content has a date-related element (holiday, birthday, anniversary, deadline, event) and you know the relevant date, offer ONE question: "📅 <이벤트>(<날짜>) <N>일 전에 알려줄까요?" / "📅 <event> is in <N> days — remind you?" - If the user says yes → record it (e.g. add to config important_dates, or note it) - If unsure of the exact date → don't guess; skip the offer c) Keep it light: at most ONE feedback line + ONE date offer. Never block saving on it. │ ▼ 4. TRANSLATE + SUMMARIZE (LLM) - Foreign content → target_language - ≤5-line summary + 3–5 key points - If already in target_language, summarize only │ ▼ 5. SAVE (script: scripts/ingest.py --save ...) - <wiki_root>/knowledge/inbox/YYYY-MM-DD-<topic>.md - Frontmatter: date, source_url, source_channel, classification (ad/useful), language - Tag: related project or "일반" │ ▼ 6. ALERT? (LLM decides against alert_on rules + script: scripts/ingest.py --notify) - If alert-worthy → append to <wiki_root>/alerts/notifications.md + CLI output "[알림] …" + optional webhook/email - Else → silent save; reply "✅ 인박스: <file>" ```

> **Note on reminders**: the skill itself stays light — the PRE-SAVE FEEDBACK step is > LLM reasoning only (no code). If the user wants automated day-before reminders, use > your platform's native scheduling (cron / scheduled messages / agent reminders); > this skill just makes the *offer* and records the date.

### Ingest script usage

```bash # Verify a URL (returns HTTP status) python3 scripts/ingest.py --verify "https://example.com/article"

# Verify a pasted text (saved to temp file first, then verified as content) python3 scripts/ingest.py --verify-content /tmp/pasted.txt

# Save a structured note (after LLM produced the markdown body) python3 scripts/ingest.py --save /tmp/note.md --source-url "https://..." --channel "discord" --classification useful --language ko

# Append to the local alert queue python3 scripts/ingest.py --notify "API key expiring 2026-09-01" --category security

# Read the alert queue (for "알림 있어?") python3 scripts/ingest.py --alerts ```

## RAG Usage (agents searching the wiki)

- Search knowledge: `search_files(pattern=..., path=<wiki_root>/knowledge)` - Fast topic mapping: read `<wiki_root>/knowledge-base-map.md` (auto-updated inventory) — topic → file - Recurring integration: a weekly cron can run the curation flow (see `wiki-inbox-curation` pattern): verify → merge into `knowledge/<topic>.md` → move processed inbox files to `inbox/archive/`.

## Alert Notifications

- **Tier 0 (always)**: append `<wiki_root>/alerts/notifications.md` + print `[알림] …` to stdout. - **Tier 1 (optional)**: webhook (Slack/Discord/Telegram), email (SMTP), or ntfy.sh topic — only if configured. - Alerts are *time-sensitive* items: security issues, account/domain expiry, project risk, and (if enabled) user interests. - To add an alert channel later: user says "알림 채널 추가해줘" → model reads config, asks for the webhook/email, updates config.yaml.

## Pitfalls

- **Never drop input because it came from a "non-default" channel** — judge it and save; the channel is metadata, not a gate. - **Don't store full verbatim text** — store the distilled note (summary + points), keeping the wiki lean. - **Verify before trusting** — a 404 after redirects is a dead link; a ~500-byte stub is a bot-block page, not content. - **Ads are not automatically deleted** — label them 📢 and save only the transferable insight (if any); keep provenance. - **Personal info / secrets in pasted content** — strip identifiers before saving. - **Config path** — respect `WIKI_AGENT_CONFIG` env var if set; never hardcode `~/wiki` into scripts. - **Alert queue growth** — mark items read/archived after reporting (`--alerts --clear` or move to `alerts/archive/`).

## Files

- `scripts/onboarding.py` — interactive config wizard - `scripts/ingest.py` — verify / save / notify / alerts CLI - `templates/config.example.yaml` — full config reference - `references/ad-detection.md` — ad-vs-useful classification rules with worked examples - `README.md` — public repo docs (EN/KR)