heliox

Operate Helio artifacts with heliox. Use when the user explicitly asks to create or publish an artifact, or to list, republish, restore, delete, or schedule a Helio artifact by name, ID, or app.helio.im/a link.

Use `heliox assistant ...` for AI teammate lifecycle: show an AI's profile/channels, create or delete an AI teammate (optionally from an agent template via `template list --query`), pick a model provider, or inspect runtime nodes. Trigger whenever the task involves spawning, retiring, or inspecting an AI teammate, hiring from a template, choosing a model source (helio / host / BYO), or finding node ids. **To list AI teammates** use `heliox workspace members list --json` (`assistant list` is retired); **for AI DMs** use `heliox message list/send @<handle>` (assistant message verbs are retired).

Create a Helio automation for AI work that should run later. Use when the user asks for a reminder, a recurring report or digest, wants you to watch or monitor something, notify people when an event happens, or follow up later, even if they never say automation. Infer Scheduled, Pushed, or Observed privately. Do not use for a one-off task now, a human calendar event, or an SOP with no future AI run. Send any automation from a prior conversation to automation-refiner, whether or not it was enabled. If the request includes an automation-rec-*.zip screen-recording package, use automation-recorder instead.

Use when a message arrives whose attachment is a single zip named `automation-rec-*.zip` produced by Helio's Record flow (a screen-recording package: manifest.json + events.jsonl + img/*.jpg inside). Turn the demonstration into a working automation. Do not use for any other purpose.

Refine a Helio automation from a prior conversation or one already delivered: repair it when it stopped working, change what a recurring job produces or who receives it, or review how it has been running and improve it. Covers automations from earlier conversations regardless of their enabled state. Creator's context is gone once that conversation ends. Use whenever someone is unhappy with a recurring report, a scheduled job keeps breaking or returning errors, a run exposed a problem in its own procedure, or they return to one from an earlier conversation, even if they never say 'automation'. Do not use to build a new automation in this conversation (that is automation-creator), or to just read back one run's error or the schedule.

Use Heliox channel commands for channel CRUD (create, list, show, update, delete), membership management, and attachments. Trigger whenever the assistant needs to inspect or modify a channel object, change members, or recover attachments. **Message verbs (send / cede / list / get) are under `heliox message *`**: run `heliox message --help` for that surface; do not look for them here.

Use FIRST in an organic group channel where more than one teammate (AI or human) shares a goal and the runtime prompt does not already contain a managed team operating context. It establishes how the team coordinates: if the channel already has a team charter, it points you to your role and how to operate under it; if there is none, it runs a short formation ritual (learn the roster, agree roles, a measurable north-star, and independent verification) and records a charter the whole team operates under. A managed hired-team context is already the channel's hidden operating agreement: follow it silently and do not run this ritual. Do not use in 1:1 DMs or single-agent channels.

Use `heliox document ...` for Helio collaborative documents: create a document with an initial markdown body, read a live document as markdown, make exact in-place text edits, or handle `helio://document/<id>` references. Trigger when the user asks to create, inspect, or edit a Helio document, task-description document, or document URI.

Use `heliox email ...` for everything involving the AI user's own email: reading or filtering the inbox (`--from`/`--subject`/`--since`), loading a full thread for context, reading a message body, waiting for an inbound mail (signup verification, OTP, vendor reply), sending new mail, or replying within a thread. Trigger whenever a task touches the assistant's inbox or outbound mail, even when the user never says the word 'email' (e.g. 'did the vendor get back to us', 'grab the code they sent me', 'let them know we're ready'). This is the only mail path; the runtime has no separate mail client.

Use after substantive work or an automation result when new evidence may warrant a non-urgent update a person should see later, or one distinct task you can perform after they accept it. Trigger on meaningful state changes, requested or subscribed digests, unassigned AI-executable follow-ups, and repeatable work; skip current-turn work, urgent risks, routine success without a delivery obligation, human-only decisions, and work already owned or underway.

Use `heliox tool image ...` to generate images from text prompts or edit existing images with Helio-managed models (GPT-Image, Gemini). Trigger whenever the task needs a picture created OR modified: an illustration, logo, icon, poster, banner, avatar, wallpaper, diagram concept, any 画图 / 生成图片 / 文生图 request, or an edit like 换背景 / 改颜色 / restyle / remove-object on an image the user provided, even when the user never says 'image' ('make me a logo', '把背景换成白色'). The result is a durable helio://attachment/<id> URI; this is the only image surface available to AI runtimes.

Install or uninstall a capability in the engine you are running inside. Trigger when the user asks to install/add or uninstall/remove a skill, plugin, or MCP server, e.g. an install prompt like `Use heliox:installer to install X (skill:foo) from source github/owner/repo@rev/path` or an uninstall prompt `Use heliox:installer to uninstall X (skill:foo)` from the read view's +/trash button. You act ENGINE-NATIVELY via the hosting engine's own CLI. INSTALL needs a SOURCE carried in the prompt: no source, FAIL FAST, never guess, never web-search. UNINSTALL needs only the machine key in the prompt: ambiguous or absent, FAIL FAST. Per-engine commands live in claude-code.md and codex.md.

Use `heliox me` for the caller AI user's own account: `me show` to view your profile, `me set name|handle` to change editable outward-facing fields, `me activity` for your places (DMs / groups with last activity), `me turns` for your recent turns. Trigger when the assistant needs to see or change its own display name or @handle, inspect its avatar, email, model, runtime id, or status, or review its own places/turns. For another AI use `heliox assistant`; for workspace metadata or members use `heliox:workspace`.

Use `heliox memory ...` when prior memory may change the answer: user preferences, earlier decisions, recent agent activity, facts from another channel/entity, or a thin injected memory block. Trigger before re-asking, repeating work, or answering from uncertain recollection.

Use `heliox message ...` for the Helio message plane: sending channel/DM messages, reading history (list / --around / --grep / --since), native threads, declining a turn (cede), and deep-reading one message with its turn pivots. Trigger whenever the job involves posting to a channel or DM, catching up on conversation, replying in a thread, or checking whether a message was processed.

Use `heliox task ...` for Helio task lifecycle: creating, listing, showing, updating status, assigning, commenting on, or deleting tasks. Trigger whenever a task number, task card, task assignment, status transition, or task comment is part of the job.

Connected tools portal: use `heliox tool ...` to see which external tools (Gmail, Slack, Notion, GitHub, Discord, LinkedIn, X, ...) are connected to you, ask the user to connect one (authorize link), and call a connected tool with credentials injected automatically. Covers the tool approval gate (`APPROVAL_REQUIRED` → `heliox approval request` → replay with `--approval`) and `heliox tool browser`: drive the user's paired local Chrome to browse the web. Trigger whenever a task needs an external service account (reading or sending the user's email, posting to Slack, editing Notion pages, working with GitHub repos), when a tool command exits with APPROVAL_REQUIRED, or when the user asks to connect/disconnect a tool.

Use when a user asks about Helio ITSELF: 'how do I do X in Helio', where a button or setting lives, whether Helio supports a feature, what a channel / task / Inbox / Vault / skill is, why an AI teammate didn't do something (product behavior, not work content), what's new, or any onboarding 'what is this app / what can you do here' moment. Answers MUST be grounded in the live Help Center, cited, and given as exact UI steps, never from prior knowledge of Helio or of similar apps. Not for questions about the user's own content or ongoing work.

Own the user wiki, the living document about your owner: who they are, how they write, who they work with, what they have in flight. Use this skill every time the wiki is touched. An onboarding kickoff directive names it (first build); the owner connects a new tool (refresh); the owner says anything like 'my wiki', 'update my profile', 'redo what you know about me'; or you learned something about your owner that belongs on the page. The wiki is the owner's heliox document, visible and editable by them and updated in place. It is never your private brain wiki.

Use whenever credentials, API tokens, passwords, private keys, vault delegation, requestable credential search, approval polling, or owner/grantee credential flows require `heliox vault ...` or `heliox approval ...`.

Use `heliox workspace ...` for the active org's workspace: showing current workspace metadata, renaming it, editing description/slug, listing workspace members, and inviting teammates. Trigger whenever the assistant needs to introspect or mutate workspace-level org state (name, description, slug, members) or invite a human teammate. For the AI's own profile (display name, avatar, status) use `heliox:me`; for runtime hosts use `heliox assistant node`.