knowledge

name	knowledge
description	ALWAYS load this skill when the user mentions alpi (the Python agent project, NOT a generic assistant). Read the bundled docs FIRST — never answer from training; alpi shipped after your cutoff.
category	meta
version	0.1.0
origin	bundled
tools	["skill"]

alpi:knowledge

You are answering a question about alpi itself — how to use it, how it's configured, how the protocol works, how to deploy it, where things live, what the security model is, what changed in a specific version. This skill bundles alpi's user-facing documentation so you can answer without a web_search round-trip and without guessing from your training data (which predates alpi).

How to use this skill

1. Decide which reference matches the user's question (the routing table below maps topics to files).
2. Read the file with:

   skill(action="view", name="@alpi/knowledge",
         file="references/<filename>.md")
   

3. Read more than one file when the question crosses topics (e.g. "how do I install alpi on a Linux server with the schedule daemon?" spans install.md + deployments.md).
4. Synthesize a concise answer. Quote the relevant fragment when it's short; summarize and link when it's long. Do not invent details that aren't in the bundled docs.

If the question genuinely isn't covered (e.g. troubleshooting a specific stack trace, asking about a third-party library), say so plainly and only then fall back to web_search or general reasoning.

Routing table — topic → file

If the user asks about…	Read
What alpi is, the project pitch, common commands	readme.md
First-time setup walkthrough, install + first message	quickstart.md
Install methods (uv tool, pipx, dev install), update path, uninstall, troubleshooting, supported platforms	install.md
Profiles (creating, switching, isolation, identity, keys, memory layout)	profiles.md
Skills system: bundled vs user, frontmatter, security scanner, where credentials live, the skill tool actions	skills.md
Models — picking a provider, tier guidance for tool-heavy use, local Ollama setup	models.md
ALP protocol — pinned identity, signed envelopes, peer capabilities, workgroups, group keys, transcript shape, error codes	alp.md
Internals — code structure, turn loop, gateway, scheduler, MCP, logging, env vars (ALPI_HOME, ALPI_PROFILE, ALPI_SKIP_UPDATE_CHECK, ALPI_UPDATE_INDEX)	architecture.md
Config knobs — every YAML field, its default, what it controls (TUI theme, sandbox, budget, gateway, schedule…)	config.md
Security model — approval system, SSRF, prompt-injection, sensitive-path denylist, sandbox	security.md
Deployment — launchd on macOS, systemd on Linux, gateway/schedule daemon shape, keep-alive, log paths	deployments.md
Day-2 ops — doctor, diagnostics, log rotation, backup, recovery, upgrade workflow	operations.md

When NOT to use this skill

• Questions about generic Python, the user's own codebase, or third-party services (OpenAI, Anthropic, Telegram API, Ollama internals). Those go to web_search or normal reasoning.
• Questions about the roadmap or unreleased features. The bundled references only cover what shipped — the roadmap file is intentionally NOT bundled because it's planning, not knowledge. Say so plainly if asked.
• Questions about what changed in a specific version or "what's the latest version?". The CHANGELOG is intentionally NOT bundled because it stales every release. Point the user at:
  • alpi --version for the version they have installed.
  • alpi update --check for whether a newer one exists.
  • The GitHub releases page (https://github.com/satoshi-ltd/alpi/releases) for full release notes.
• Operational status of the user's own setup (what their config currently is, whether their service is running). That's alpi doctor and read_file against their config.

Tone

Answer in the user's language. Be concrete: cite the doc you read, quote the relevant snippet, give the exact command if one applies. Never make up flag names, paths, or behaviours that aren't in the references.