724 Office

7/24 Office is a personal assistant agent that runs continuously behind a messaging platform. It is written in roughly 10,000 lines of Python across 20 files with no agent framework underneath — the standard library plus a handful of small packages — on the stated principle that every line stays visible and debuggable.

The core is a tool-use loop over an OpenAI-compatible function-calling API, retrying automatically and running up to 20 iterations per conversation, with 36 built-in tools split across seven domain modules covering messaging, admin and file operations, search, video, page rendering, and behavioural reports. A nudge system detects when the model has tools available but is not calling them and injects a hint, and a circuit breaker disables a tool after three consecutive failures in a session. Dynamic tool filtering selects one of five context profiles — voice, scheduler, group, diagnostic or default — to keep the tool list short, and system-prompt assembly tracks a token budget.

Memory is three layers. The session layer holds the last 40 messages per session as JSON files and compresses on overflow. The compressed layer has the model extract structured facts from evicted messages, deduplicates them by cosine similarity at a 0.92 threshold, and stores them as vectors in LanceDB. The retrieval layer embeds each incoming message, searches those vectors, and injects the top matches into the system prompt within the token budget.

Around that sit the operational parts: an MCP client speaking JSON-RPC over stdio or HTTP with reconnection and hot reload, runtime tool creation where the agent writes and loads new Python tools itself, daily self-checks and error-log analysis with owner notification, cron and one-shot scheduling that survives restarts and skips dormant users after three days, and a Docker-based router that provisions one container per user and rebuilds missing ones from its routing table at startup. Group chats get their own container with @-mention gating and a 20-message context buffer.

A successor rewrite, xiaowang-v2, lives in the same repository: a single Node process over one WAL SQLite database with a hand-written agentic loop, adding durable cross-day execution, layered memory with explicit forgetting, turn assembly for burst messages, photo-then-caption and voice input, and a split where the model does fuzzy routing while the harness handles execution feedback, schema locks, state exposure, approval gates and idempotency. Its documentation is in Chinese.

Features

• Tool-use loop: OpenAI-compatible function calling with automatic retry, capped at 20 iterations per conversation
• Three-layer memory: session history, model-compressed long-term facts deduplicated at 0.92 cosine similarity, and vector retrieval from LanceDB
• MCP and plugins: external MCP servers over JSON-RPC on stdio or HTTP, hot-reloaded without a restart
• Runtime tool creation: the agent writes, saves and loads new Python tools while running
• Self-repair: daily self-check, session health diagnostics, error-log analysis and automatic notification on failure
• Cron scheduling: one-shot and recurring tasks, timezone-aware, persistent across restarts, skipped for users dormant three days
• Multi-tenant router: one Docker container per user, auto-provisioned, health-checked and reconciled from the routing table
• Multimodal: images, video, files and voice, with speech-to-text and vision through base64 input
• Web search: multiple engines with automatic routing and a dual-engine default
• Visualization: line, bar, pie, radar, table and report pages rendered as interactive HTML
• Behavioural reports: a profile report and a future-self dialogue mode built from session history
• Edge deployment: designed to run on a Jetson Orin Nano with 8 GB of RAM and a budget under 2 GB