OpenSec

Single-owner personal AI assistant workspace for news, research, coding, and automation

Deterministic engines where they matter. Bounded LLM help where it adds leverage. One workspace that can grow from a news brief into a real personal control plane.

Korean README • Architecture • News Engine • DB Schema

What OpenSec Is

OpenSec is not just a news bot.

It is a public-safe foundation for a single-owner personal assistant system:

• news-bot/ gives you deterministic news and signal pipelines
• skills/ gives OpenClaw structured entry points for news, repo work, coding, memory, and system tasks
• workspace-template/ gives you a long-lived personal workspace shape
• docs/ keeps the architecture, plans, and product memory durable

The core idea is simple:

treat retrieval, state, and operations as first-class system components, not as side effects of one big prompt

That lets the repo support more than one use case:

• scheduled briefs
• bounded follow-up research
• repo and coding tasks
• memory capture and distillation
• small personal automations
• future private sibling assistants, such as a training bot, without collapsing everything into one hidden prompt

The Stack

Think of the system in three layers:

Layer	Role	Examples
Front doors	Where you talk to the system	Discord, Telegram fallback, shell, Codex
Workspace control plane	Routing, approvals, memory, skills, cron	OpenClaw workspace, skills/, workspace-template/
Deterministic engines and scripts	Actual domain logic	news-bot/, watcher flows, repo scripts, future project engines

Practical interpretation:

• OpenClaw is the always-on gateway and coordinator
• Codex is a direct coding/operator surface against the same repo
• this repository contains the engines, skills, scripts, and docs that make those surfaces useful

What You Can Do Today

Capability lane	What exists now
Daily news brief	Deterministic tech and finance digests with SQLite state, scoring, resend suppression, and Korean rendering
Follow-up Q&A	expand N, show sources for N, today themes, ask <질문>, research <질문>
Repo and coding work	code_ops, repo_ops, and system_ops skill scaffolds for controlled remote execution
Memory loop	Daily note capture and curated long-term memory flow through workspace-template/ and skills/memory_ops/
Specialized automation	Separate bounded runtimes such as the Xiaohongshu housing watcher
Cost-aware LLM use	Task-tiered routing, usage telemetry, and budget controls in the news engine
Global token audit	One ledger that collects OpenClaw-wide session usage plus direct app llm_runs
Safe extension path	New skills, scripts, and deterministic engines can be added without redesigning the whole system

Mental Model

flowchart TB
    U["You"]

    subgraph Surfaces["User surfaces"]
        D["Discord / DM"]
        T["Telegram fallback"]
        C["Codex"]
        S["Shell / cron"]
    end

    subgraph Workspace["Personal workspace / control plane"]
        O["OpenClaw coordinator"]
        K["skills/"]
        M["memory + notes"]
        A["approval + routing policy"]
    end

    subgraph Engines["Project engines and scripts"]
        N["news-bot/"]
        R["repo and coding workflows"]
        W["watchers and automations"]
        P["future private sibling assistants"]
    end

    U --> D
    U --> T
    U --> C
    U --> S

    D --> O
    T --> O
    S --> O
    C --> N
    C --> R

    O --> K
    O --> M
    O --> A

    K --> N
    K --> R
    K --> W
    K --> P

Loading

The important point is that news-bot/ is one engine inside a broader personal assistant shape.

Design Stance

These are the repository-level design choices:

• deterministic retrieval before model enrichment
• local state and evidence preserved in SQLite
• non-LLM fallback must remain usable
• one visible coordinator is better than many visible bots
• skills and scripts should expose bounded actions instead of prompt-only magic
• future private assistants should live beside this repo, not be smuggled into it

This is why the news system is structured the way it is:

• daily candidate discovery is deterministic
• dedupe and resend suppression happen in the data layer
• full-read article context is cached locally
• LLM calls are routed by task tier
• follow-up answers can reuse stored evidence instead of starting from zero

The same philosophy also makes the repo usable for non-news work.

Why This Structure Scales

OpenSec is meant to grow with the owner.

Typical progression:

1. Start with local digest generation.
2. Add OpenClaw and a private Discord front door.
3. Add repo and system skills for remote work.
4. Add memory capture and daily distillation.
5. Add more bounded automations or private sibling assistants.

Because the workspace, skills, engines, and docs are separated, new capabilities do not need to be bolted onto the news prompt itself.

Examples of clean extension paths:

• add a new deterministic engine under projects/ or this repo
• add a new skill under skills/
• add a new bounded automation script under scripts/
• add a private sibling workspace or private repo for a training bot, while keeping this public repo clean

Repository Map

Path	Purpose
news-bot/	Deterministic news engine with scoring, evidence, follow-up, LLM routing, and delivery-friendly rendering
skills/	OpenClaw-facing skills for news, code, repo, memory, and system tasks
workspace-template/	Base personal workspace scaffold, including memory files and operator documents
scripts/	Workspace bootstrap, daily note helpers, VPS bootstrap, and other operational scripts
docs/design-docs/	Long-lived system design notes
docs/product-specs/	Product behavior and operating model specs
docs/exec-plans/	Active and completed implementation plans
docs/generated/	Derived references such as the database schema

Current Engines And Lanes

1. News engine

news-bot/ currently provides:

• source adapters for curated feeds
• normalization, canonicalization, and dedupe
• SQLite-backed raw items, normalized items, digests, follow-up context, and telemetry
• deterministic ranking
• recent 72-hour resend suppression
• full-read article or repo context extraction
• Korean digest rendering
• bounded LLM enrichment and research

2. Workspace skills

The current public workspace skills are:

• skills/ai_news_brief/
• skills/code_ops/
• skills/repo_ops/
• skills/memory_ops/
• skills/system_ops/

These are intentionally simple, explicit entry points. The goal is to make remote execution understandable and auditable.

3. Memory loop

The memory model is deliberately conservative:

• raw daily notes go into memory/YYYY-MM-DD.md
• stable preferences and durable operating facts go into MEMORY.md

OpenSec does not assume every conversation should become long-term memory.

4. Private extensions

This public repo is designed to coexist with private siblings.

For example:

• a private training bot
• private owner memory
• private operational exports
• private secrets and hidden rules

Those should live in a separate private workspace or private repo, not inside OpenSec itself.

OpenClaw And Codex

If you are using both OpenClaw and Codex, the clean boundary is:

Tool	Best use
OpenClaw	Always-on gateway, Discord/DM front door, approvals, cron, memory, tool orchestration
Codex	Direct implementation work inside the repo, code edits, debugging, tests, documentation updates
OpenSec repo	Shared system of record for deterministic logic, skills, scripts, and docs

That means:

• you can run the repo locally with Codex and shell only
• you can also mount the same repo under an OpenClaw workspace for always-on assistant behavior
• you do not need to choose between "news bot" and "coding assistant"; the same workspace can support both

Quick Start

1. Run the deterministic engine locally

cd ./news-bot
pnpm install
pnpm approve-builds
cp .env.example .env
pnpm test
pnpm digest -- --profile tech --mode am
pnpm followup -- --profile tech "expand 1"

If pnpm approve-builds prompts for native packages, approve better-sqlite3 and esbuild.

2. Bootstrap the personal workspace

bash ./scripts/setup-personal-workspace.sh

That sets up the base workspace shape used by OpenClaw:

• operator documents
• memory files
• public skills
• expected project layout

3. Attach it to OpenClaw

Key files:

• openclaw.personal.example.jsonc
• scripts/setup-personal-workspace.sh
• scripts/ensure-daily-memory-note.sh
• workspace-template/

Then:

1. copy the example OpenClaw config
2. fill in your Discord or Telegram credentials
3. start the OpenClaw gateway
4. point it at the personal workspace
5. bind channels and approvals
6. install cron jobs or recurring automations

Suggested Reading Order

1. ARCHITECTURE.md
2. news-bot/README.md
3. docs/generated/db-schema.md
4. docs/design-docs/openclaw-personal-control-plane.md
5. docs/product-specs/discord-personal-control-plane.md
6. docs/product-specs/llm-assisted-digest.md

Who This Repo Is For

OpenSec is a fit if you want:

• a single-owner assistant system, not a multi-tenant SaaS
• deterministic engines plus bounded LLM help
• one workspace that can cover news, research, coding, and automation
• direct evidence and local state instead of opaque browsing-only behavior
• a public-safe repo that can sit next to private sibling assistants

It is not trying to be:

• an unconstrained autonomous browsing agent
• a general SaaS chatbot backend
• a pure prompt-only system with no persistent operational state