智能办公助手

🔌 Event daemon (Layer 0 — Node.js, zero dependencies)

• WebSocket event hub — the Godot world and the overlay UI subscribe to one stream
• Event journal (journal.jsonl) with replay on connect: restart anything, state comes back — auto-trimmed on boot (and stale chat threads pruned) so nothing grows unbounded on a long-running office
• Agent registry (registry.json): persistent staff — name, job title, avatar, aura, system prompt, skills, tools. main (the Director — Shino by default: your playful-but-focused second-in-command, tuned for delegation over hands-on work) and ceo (you) are protected and cannot be deleted. A fresh install starts with just these two
• Claude Code adapter: POST /chat spawns a real headless claude -p session with the agent's persona, assigned skills and allowed tools; stream-json output becomes world events
• Chat threads: every conversation is a named, resumable session (--resume) with its own recorded history; agents keep continuous memory by default
• Skills library with Hermes-style auto-learning: after a completed multi-tool task, a reflection pass decides whether the work distills into a reusable skill — if so it's saved, auto-assigned, and announced in the office. Since v1.0.0 the same reflection can also correct a skill it already wrote, and it runs after failures too — the strongest evidence a skill is wrong, and previously the one case that was thrown away. It never touches a built-in and never touches one you edited by hand, and the previous version is kept (POST /registry/skill/revert)
• Tools: per-agent allowlist over the built-in Claude Code tools, plus custom capability via MCP servers (name + launch command → injected with --mcp-config). Ships with a ready-to-use 🔌 web server (Playwright MCP) — tick it on an agent and it can browse, click, type, submit forms and screenshot real pages (isolated profile, not logged in). See docs/guide/web-automation.md. The 🧰 Tools Hub (⋯ menu) is a browsable catalog of 43 entries — 15 built-in abilities and 28 MCP servers added with one click, including Blender, Godot, Unity, Unreal and Roblox Studio. The catalog is a file the office fetches live (web/tools.json), so a renamed package is a PR rather than a release. See docs/guide/tools-hub.md
• CEO chain of command: ordering the CEO summons the Director — he walks over, takes the order, replies with a plan, and dispatches work to teammates via DELEGATE: lines (each spawns a real session, with the hand-over walk acted out). Delegation is a round trip: every delegate's result is reported back to the Director, who can answer questions / follow up with more DELEGATE: lines (bounded depth, serialized turns), and finally walks the CEO-readable summary over to the boss (ceo.report)
• Agent discussions: pick 2–4 agents and a topic — they hold a real meeting, round-robin turns over a shared transcript, minutes on the in-world whiteboard
• Self-splitting sub-agents: every session is told it may end a reply with SUB: <job> lines (2–4) when the request parallelizes — the daemon strips the protocol, spawns parallel clone sessions with the parent's persona + tools, records each in a labeled 👻 session, and resumes the parent for a final synthesis once all ghosts report back (a stuck ghost is reaped after 6 min, so synthesis always happens). 👻 GHOST ISOLATION (opt-in) gives each ghost its own git worktree on a project that is a git repo — two ghosts editing one file both succeed, their work comes back as office/ghost-<id> branches for you to review, a ghost that changed nothing leaves nothing behind, and your working copy is never touched
• 📦 Run location (POST /registry/backend): an agent's work does not have to happen on your desktop. Point the whole office — or one agent — at a throwaway Docker container or another machine over SSH. A container gets exactly two mounts, the office root at /office read-only and the working directory at /work; API keys are passed by name, so values travel in Docker's own environment and never appear in a process listing. A backend that cannot be built correctly is refused, not quietly downgraded — above all when the --settings file that installs the permission broker cannot be placed, because a run that loses it works fine with nobody watching
• 🔎 Semantic recall (POST /registry/semantic, opt-in): the memory index matches words (BM25, no dependencies, offline) and has one blind spot — ask "why did the wallpaper vanish" and a note reading "WorkerW teardown kills the embedded world" shares no meaningful token with the question and never comes back. Point the office at any OpenAI-shaped /embeddings endpoint (a local Ollama costs nothing and keeps your memory on the machine) and every memory also gets a vector; both rankings are then fused, so words and meaning each get a say. Saving tests the endpoint and reports the vector size, because an embeddings endpoint that only fails later fails silently
• Standing work orders: POST /jobs — run now, at a datetime (optionally daily), or every N minutes; per-agent queue + a global concurrency cap keep the machine comfortable; each job keeps its own resumable thread. A fired job is treated as a real order, not a reminder: the assignee is told to do the work in that turn, and a job on the Director carries the full delegation protocol so he dispatches to the team and their results report back to him
• 🤖 AUTO — keep going without you (opt-in, ⚙ → TOOLS or bagidea auto on): an agent that hits a fork decides within its remit and opens its own next turn instead of parking the job on a question. Up to 8 self-driven rounds, each announced in chat; it still stops dead for a credential it can't get or an irreversible/outward action, and that block is pushed to your channels. It removes the wait for an opinion — what an agent may do is still the separate 🔓 auto-approve switch
• Shared note board: notes live in the UI and in workspace/notes.md — agents read it and append bullets themselves (file-watched both ways)
• Calendar with a personal touch: appointments remind you via the Director — he physically walks over and tells you (reminder event), N minutes ahead
• Director heartbeat: every 15/30/60 minutes (configurable) he reviews the calendar, standing jobs and the note board — and pings you ONLY when something deserves it ("OK" stays silent)
• Claude Code hooks integration: any Claude Code session in this project reports its tool calls — your real work animates the Director automatically
• Permission broker: tools you granted in an agent's profile run silently; anything else is held until you approve — with a ✓✓ forever option that remembers the grant
• 📁 Projects: register real folders as projects (with PLACE shorthands like "classroom" → D:\Learning); the Director creates new ones himself via a PROJECT: protocol line and routes work with DELEGATE: <agent> @ <project> :: <job> — the assignee's claude session lives inside that directory and is resumable by you. One window per project: ▶ opens (or surfaces) the window. One occupant at a time — while an agent works the project you can't open it (the row shows a ⏹ stop agent button with a two-click confirm to take over), and while you have it open an agent won't be dispatched into it. Removing/deleting a project also closes its window; disk-deletes sweep leftover dev servers first
• 📨 Channels (6): Telegram (long-poll), Discord (native gateway), LINE, Slack (Events API), WhatsApp (Meta Cloud API) and Messenger (Meta Graph) feed straight into the Director — order your office from your phone, the reply comes back on the same channel
• 🔑 API key vault: store OPENAI_API_KEY