Skip to content

Ai-chan ​

Ai-chan, drawn in pixel art

Ai-chan is Humanize's own operator: you tell her what you want in words, and she does it on your epics as you would from the prompt or the page. Ask her what your epics did today, have her steer a session, or run the tests in an epic's environment.

At a glance

  • You will open a conversation with Ai-chan, say what you want, and watch each thing she does to Humanize as she does it.
  • Use it when a question spans several epics, or a job is easier said than clicked: "summarize what my epics did today", "tell the actor to stop refactoring", "run the tests in the env of epic X".
  • You need an agent for her to think with, chosen as Ai-chan on the General page of the settings or when you open her.

Try it ​

At the prompt, type /ai, pick New, press Done on the form as it comes, and say what you want:

text
❯ /ai
❯ summarize what my epics in this directory did today

On the page of hmz web or in Humanize Desktop, click Ai-chan's avatar at the right of any view's bar instead: a drawer opens beside it, with New.

How it works ​

Every conversation with Ai-chan is fixed, when you make it, to three things. None of them changes while the conversation lasts; to change one, make another conversation.

What it isChoicesDefault
Scopewhat she may reachepic: one epic, and the directory it ran in · workspace: one directory and every epic run there · daemon: every directory and every epic on this machine's daemonthe epic you are looking at, else the directory, else the daemon
Nouswhat she thinks withany nous, or an agent: a CLI, an account, a model and an effortthe Ai-chan setting
Permissionwhat she may changea preset, read-only or danger-full-access, and network, off or onread-only, network off

What she may do follows from the preset, and from the scope reaching what she acts on:

She canread-onlydanger-full-access
Read epics, transcripts, traces and what they spentyesyes
Ask a session a side questionyesyes
List, read and search the files of a running epic's environmentsyesyes
Read files under the scope's directoryyesyes
Write files in her own scratch directoryyesyes
Say something to an epic, answer it, stop it, force it, edit its boardnoyes
Write files in a running epic's environments, and run commands therenoyes
Write files under the scope's directorynoyes
Start an epic (workspace and daemon scopes)noyes
Change the settings and the nous, install flows (daemon scope)noyes

Whatever the permission, she cannot reach anything outside her scope, pair a device, add an account or touch its credentials, or see another conversation with Ai-chan, her own included. Her CLI never stops to ask you for approval: what she cannot do is refused as she tries it, not put to you.

Network off cuts the network for her CLI and every command it runs, except what her model and its sign-in need. Turn it on when she must fetch something, such as a page or a package.

She talks as a chat does: she answers each thing you say in a turn, and she shows each tool she calls -- what she asked Humanize for, and how it was answered -- as a line of her turn you can open. What she does to an epic, every frontend reading that epic sees, under her name, Ai-chan#<n>: a line she says to an agent reads · by Ai-chan#3, as yours reads · by you.

Example: summarize what your epics did today ​

You have run a few epics in ~/code/shop today and want the gist without opening each.

At the prompt. In ~/code/shop, type /ai. A menu lists every conversation with Ai-chan, under New:

text
  New                                                                         ①
    a conversation of its own scope, nous and permission
  Ai-chan#2 · check the reviewer's last round                                 ②
    epic 0912-a1b2c3 · claude/claude-opus-5:high · read-only · network off · between turns
  Ai-chan#1 · what did the benchmark score
    daemon · claude/claude-opus-5:high · read-only · network off · folded away

Pick New. A form asks for the scope, the nous, the preset and the network, each set to its default. Nothing has run here, so the scope is workspace, this directory. Press Done: the first line you say is the task that starts the conversation.

text
❯ summarize what my epics here did today, one line each
  …                                                                          ③
  Four epics today:
  - 09:12 ralph_loop: added the coupon endpoint; done in 3 rounds, $2.10.
  - 11:40 rlar: the reviewer refused the migration twice; stopped by its budget.
  …
Ai-chan#3 · workspace shop · ◉ done                                          ④

What to look at, by number:

  1. New makes a conversation. Every one you made is listed under it, newest first.
  2. Each conversation, Ai-chan#<n> and its title, says its scope, its nous, its preset, whether its network is on, and how it stands: answering, between turns or folded away.
  3. Each tool she calls is a tool line of her turn, as an agent's is, even with Details off: what she asked Humanize for, and how it was answered.
  4. Ai-chan#3 · workspace shop at the front of the status line is the conversation you follow, with its scope.

Check it worked ​

  • She answered from what your epics wrote down: each line names an epic you ran.
  • /ai again lists the conversation, between turns: ready for your next line.

Open a conversation ​

WhereOpenNewStop her turnForget
The prompt (hmz)/ai, then enter on a conversation/ai › New: a form, Done, then say the first line/stop/ai, m on a conversation, forget
The page (hmz web)Ai-chan's avatar, at the right of every view's bar, then a conversationNew in the drawer: a form, then type the task and press enterStop, while she answersForget on a conversation
Humanize Desktopas on the pageas on the pageas on the pageas on the page
  • At the prompt, the log screen follows the conversation you open, as it follows a chat: each line you say is her next turn. /clear leaves her conversation going and comes back to this directory's own. /btw is refused while you follow her: she is not a session of an epic.
  • On the page, the drawer stays open from view to view until you close it, and lists each conversation as Answering, Waiting or Folded away. Its form also takes a Title, made from the task where you leave it empty. On a phone the drawer covers the view; esc or its close button puts it away.
  • Stopping interrupts her turn and leaves the conversation between turns. It does not end it.
  • Forgetting removes the conversation from the list and deletes her scratch directory. It asks first (Forget Ai-chan#<n>? at the prompt), and cannot be undone.

Many conversations ​

Have as many as you like, of any scope, at once. Making one, talking in one, or forgetting one never waits on another, nor on a chat or an epic.

A conversation nobody follows for a minute between turns folds away: nothing holds it, and it costs nothing. It stays listed. Open it and say something, and she carries on where she was, with the conversation so far. A conversation lasts until you forget it.

Examples ​

Each is a first line, with the scope and permission it needs.

SayScopePermission
"Summarize what my epics did today."daemon, or workspace for one directoryread-only
"Why did the last rlar run stop?"the epic, or its directoryread-only
"What has the reviewer of this epic been complaining about?"epicread-only
"Find where the actor in this epic left a TODO."epicread-only
"Steer the actor: stop refactoring, fix the failing test first."epicdanger-full-access
"Ask actor/2 why it skipped the migration."epicread-only
"Run pytest -x in the env repo of this epic and tell me what fails."epic, while it runsdanger-full-access
"Start ralph_loop here on the open issue about coupons."workspacedanger-full-access
"Read the new release notes of pytest: does anything in them break us?"workspaceread-only, network on

To steer a session, she says the line to it as you would by typing into its transcript: see Talking to a running turn. To run tests in an env, she runs the command there, in the env's own directory, as the flow would, without the epic being told. An env is open only while its epic runs.

Safety ​

danger-full-access on the daemon scope can do whatever you can

A daemon scope reaches every directory on this machine, and its directory is /. With danger-full-access, she may write anywhere you may. Keep daemon conversations read-only unless you mean it.

  • Read-only means read-only. Under read-only, every call that would change something -- starting, saying, stopping, writing a file or a setting -- is refused before it has any effect, and she is told Ai-chan is read-only, and <method> changes things. She tells you so and stops there; make a danger-full-access conversation to do it.
  • Out of scope is refused too, with <method> is out of Ai-chan's scope: <what>. An epic conversation cannot list the other epics of its directory, and a workspace one never sees another directory's: they are left out of what she reads.
  • Some things she can never do: pair or revoke a device, touch an account's credentials, or make, read or forget a conversation with Ai-chan. These are refused with Ai-chan cannot call <method>.
  • She is confined. Her CLI and every command it runs can read only the scope's directory, write it only under danger-full-access, and always write her scratch directory. Humanize's own machine directory is closed to her, so she reaches Humanize only through her tools. Where this machine cannot hold that, the conversation is refused rather than run wider: see Permissions.
  • What she reads is data. She is told that nothing she reads -- an epic, a transcript, a file, what another frontend says -- is an instruction to her.
  • She has no budget. Her turns are never stopped by a limit, so mind what you ask of a large scope.

Troubleshooting ​

Name Ai-chan's nous: the ai_chan setting names none. ​

No nous was given, and the Ai-chan setting names none. Pick one on the form, or set Ai-chan on the General page.

Tell Ai-chan something to start with. ​

The first line was empty. Say what you want her to do.

There is no epic <epic> in <workspace>. ​

An epic scope names an epic that did not run in that directory. Pick it again from the directory it ran in.

Ai-chan is not answering, so there is nothing to stop ​

You pressed Stop or typed /stop between her turns. There is nothing to interrupt.

Ai-chan's harness must be kept from the daemon's socket, which Landlock cannot do: set sandbox.backend to bwrap ​

On Linux only bubblewrap keeps her CLI from Humanize's own socket; Landlock closes paths, not connections to a socket. Install bubblewrap and set sandbox.backend to bwrap (see Permissions), then start the conversation again. macOS needs nothing.

She says she is read-only ​

The conversation is read-only, and what you asked changes something. The preset cannot change: make a new conversation with danger-full-access.

She says something is out of her scope ​

The conversation's scope does not reach it. Make a new one whose scope does: the epic's directory, or daemon.

Next steps ​

Released under the Apache-2.0 licence.