bound-reference

name	bound-reference
description	Self-reference for the bound agent — what you are, the tools you have, how your memory and architecture work.
allowed_tools	skill-read query memory schedule

Bound Self-Reference

You are an agent running inside bound: a persistent, model-agnostic personal agent daemon that runs on the operator's own infrastructure. This skill is your own reference manual. Read it when you are unsure what you are, what tools you have, or how your own machinery works — instead of guessing.

This SKILL.md is the index. The detail lives in references/:

• references/tools.md — the tools you can call: the native agent tools, the built-in file tools, and how MCP-dispatched commands work. Start here when a tool call fails or you are unsure a tool exists — a hallucinated tool call fails hard, so confirm against this catalog rather than inventing a name.
• references/memory.md — semantic memory: tiers, key conventions, edges, and how Working Knowledge / Discoverable Archive / Live State are assembled into your context.
• references/architecture.md — the shape of the system you live in: the agent loop, context assembly, the scheduler, sync, and the hub/spoke cluster.
• references/connectors.md — platform connectors (Discord and the like): the connector tool, event subscriptions, connector handles, leader election, connector-authored instructions, and webhooks. Read this when an event woke you and you are unsure where the message came from or which tools you have.
• references/sandbox.md — the virtual filesystem your read/write/edit/bash tools act on: the VFS, the 256 KiB tool-result cap, the deferred files-table snapshot, and the boundless real-disk distinction.
• references/config.md — operator-owned host configuration: where it lives (config/ on host disk, not the sandbox), strict schemas, and why your file tools cannot reach it.
• references/trace-queries.md — the canonical join keys and common-question query recipes for reconstructing "who ran what, where, and why" across the cluster through the query tool. Read this before writing an ad-hoc multi-table join — the join paths are already mapped.

What bound gives you

• Cross-session memory that persists across conversations, hosts, and surfaces (web UI, Discord, and the boundless terminal client). The SQLite database is the single source of truth.
• Autonomous task execution — you can schedule deferred, recurring (cron), and event-driven work that runs in its own thread with its own context window.
• Tools — native agent tools with structured JSON-schema parameters, built-in file tools, and any MCP servers the operator has connected.
• Multi-host operation — state replicates between hosts over Ed25519-signed, encrypted WebSocket sync. One host is the hub; others are spokes.

How to orient when you wake up

1. The stable system prompt already describes your environment, persona, and the live database schema. Read it first.
2. Your volatile context carries Working Knowledge (pinned rules + summaries), a Discoverable Archive (titles you can search), and Live State (threads, files, tasks). Treat summaries as pointers, not ground truth — verify against the database with the query tool before acting on a claim.
3. If you need a capability you are not sure you have, check references/tools.md. Do not assume a tool exists because it would be convenient.

Maintaining this reference

The source of truth for this skill is the markdown under packages/agent/src/bundled-skills/bound-reference/ in the bound repository. It is seeded to the files table on startup and embedded into the compiled binary at build time. To change what this skill says, edit the markdown in the repo — not the seeded copy. The tool catalog is guarded by a test that fails if a native agent tool is added or renamed without updating references/tools.md.