explain-ace

Explain how to work with ACE and what keeps it safe, answered from the live system. Use when someone is new to ACE, or asks how to reach it, who reviews its replies, who sees what, or what data it can

SKILL.md

GitHub

Explain ACE — answered from the live system

People only use ACE comfortably once they understand how they interact with it and what keeps it safe (Jon, 2026-10-07). Those answers change week to week — routing, interfaces, turn modes, Slack behaviour are all live configuration and fast-moving code — so never answer from a stored write-up, this file, or memory. Read the current sources below, then answer.

1. Always share Canopy's own explainer

Canopy is the system ACE runs on, and it maintains its own explainer pages. Put both in the reply, first time and whenever the question is about how the system works:

Canopy's address is whatever canopy-web says it is today: CANOPY_PUBLIC_BASE_URL in config/settings/connectlabs.py (and the Deployment line of canopy-web's CLAUDE.md). It moved from labs.connect.dimagi.com/canopy/ to canopy.dimagi.com on 2026-10-05; the old address still 302s browser pages, so a curl -L check "passes" on a stale link. Check a link with plain curl -s -o /dev/null -w '%{http_code} %{redirect_url}' — a 302 to another host means the link is stale. Never take a base URL from a memo, a note or an older doc.

Do not write or link an ACE-specific explainer doc as the answer. A dated snapshot is fine as reference for one analysis (filed under an ACE project on canopy-web), never as the source.

2. Read the current state before answering

Read only what the question needs, and say in the answer what you read and when.

Question Read
Who may make ACE do what; what outsiders can do get_interface (slug ace) and canopy-web docs/architecture/access.md (the decision table, agent roles)
Does a person have to approve replies; who can get_agent (turn_mode, owner, workspace) + access.md rows on turn mode
Whose runner / which mode a person's or channel's work goes to list_agent_actor_routes, list_agent_runner_rules (slug ace)
How Slack works (threads, DMs, --history, Stop, status lines) canopy-web apps/slack/services.py, views.py, window.py, status.py
What Labs data a viewer or ACE sees connect-labs CLAUDE.md § Permission Model, connect_labs/labs/models.py (UserConnectToken)
ACE's own rails (send path, deny rails, tenancy guard and its current warn/enforce mode) this repo's CLAUDE.md § Agent operating model, config/gating.json, config/tenancy-targets.json

Read canopy-web and connect-labs at current origin/main — a local checkout may be days stale (git fetch first), or use gh api repos/dimagi-internal/<repo>/contents/<path>.

3. What a good answer covers

Interaction model: the channels (email to ace@dimagi-ai.com, Slack @canopy ace in a thread or a DM, canopy chat, the Connect Labs agent panel); that Slack and canopy chat show the whole session live while email returns only the output (Jon, 2026-10-07 — some people prefer one, some the other); one thread per task, and separate threads run in parallel; approvals and who gives them; that routing a person or channel to another runner moves who oversees it.

Acting as you. To have ACE change Connect Labs (workflows, dashboards, the semantic layer) as yourself rather than as ACE, say "as me" in the request. Before any labs write ACE names the account it will write as, and "as me" only goes through a connection signed in as you; with none it stops and says how to add one. Only labs can be "as me" today. Connect, HQ, OCS and Drive changes are made as ACE's own accounts. Rule and setup: skills/_acting-as-you.md.

ACE runs in your own Claude Code. Anyone with ACE access who uses Claude Code works with ACE in their own session: questions, reviews, Labs reads, even an opp run. They set it up once with canopy agent use ace (read its --help for the current steps; it ends with the line that starts ACE locally). It brings ACE's skills and env to their computer, and follow-ups stay in the same session. canopy agent dispatch --slug ace is not how to reach ACE. It hands work to a runner to do UNATTENDED, for when you want to walk away from it. Each dispatch starts a fresh cloud session that remembers nothing earlier, and the reply read back through canopy agent turns is cut at 2,000 characters. Measured 2026-10-23: two editors with no local ACE sent 17 questions this way in a day. 12 replies were cut at exactly 2,000 characters, and about half the sessions only re-asked for the cut text, each one redoing the work.

Security model: who counts as whom and that a From: line proves nothing; that outsiders get a confined session; manual mode; the rails enforced in code; and data.

On data, the rule is the owner's (2026-10-07): share with ACE only what you would share with your own AI assistant. ACE runs in an ordinary Claude session that Jon or another operator controls. There is no extra data control or security layer on top of that, so ACE is not a safer or more restricted place than a person's own Claude. Say this plainly when someone asks what they may share, and do not suggest ACE adds protections. Zero-data-retention (ZDR) mode exists, but only Jon may use it. Nobody else can ask for it, so for everyone else the normal AI rules apply. Labs reports still read Connect as each viewer, and the Labs agent panel acts as the person who sent, whoever they are (owner included): on a labs page ACE reads labs only through their act-as-you permission, never its own account (skills/_labs-on-a-page.md). Do not present Slack as riskier than ACE itself; the real risk worth naming is Canopy itself making a mistake.

Honest limits: say what is early, unverified, or not live, as the sources show it today.

4. Who changes things

Turn mode, admins, routing and the interface are set by ACE's owner or an admin on canopy-web, never by ACE. Say so, and route the request to the owner rather than promising it.