Glossary

The words the rest of these docs use, one entry each, with where to read more.

epic ─── one run of a flow, written down
 │
flow ─── the loop: who is asked what, and until when
 │
 ├── role "builder" ─── agent: backend, model, effort
 │                       └── session ─── turn, turn, turn …
 │
 └── role "reviewer" ── agent: …
                         └── session ─── turn …

One run, in these words

Most of the words below meet in a single command. This one runs the rlar flow, where an actor works and a reviewer checks each of its turns:

hmz exec -f rlar \
    -a actor=claude/claude-opus-5-5:high \
    -a reviewer=codex/gpt-5.6-sol:high \
    -p budget.duration=2h,budget.cost=10 \
    "fix the flaky payment test"

1. -f rlar names the flow: the loop that says who is asked what, and when to stop.
2. actor= and reviewer= are its two roles. Each is filled with an agent: a backend (claude, codex), a model, and an effort (high). With no @, each runs as the provider you signed the CLI in as.
3. -p budget.duration=2h,budget.cost=10 is the budget: the run stops at two hours or ten dollars, whichever comes first.
4. The quoted line is the task.

While it runs, the actor keeps one session and the reviewer opens a fresh one every round; each exchange inside a session is a turn. All of them work in the workspace, the directory you ran the command in, which is the run's one environment. The whole run is written down as an epic, and afterwards its trace shows every turn on one timeline.

simulated

simulated

The same run at the prompt is /flow, the same choices made in menus, and the task typed at ❯. See Your first run.

Agent

What fills one role of a flow: a backend, a model and an effort, and optionally the provider it runs as. On a command line it is written after the role it fills:

reviewer=claude@work/claude-opus-4-8:high
   │       │     │         │           └── effort
   │       │     │         └── model
   │       │     └── provider; left out, as you signed in
   │       └── backend
   └── role

An agent holds no conversation of its own: it is the settings every session it opens runs at. Two agents set up alike are still two agents. Set one up in /flow: Your first run.

Backend

A coding agent CLI that Humanize drives: agy, claude, codex, cursor-agent, dsh, grok, kimi, mcode, mimo, omp, opencode, pi or qwen. It is the cli row of an agent. See Installation and Many backends, one agent.

Budget

What a run may spend before it is stopped: a duration, a cost in US dollars, or a number of output_tokens. Every flow except chat needs one. See Every run has a budget. A single turn can also have a budget of its own, set by the flow: A turn can be cut off.

Effort

How hard the model thinks, in the backend's own words: high, max, off and so on. Each model has its own list. See Efforts.

Environment

Where a turn's work lands: this directory, another directory, or a directory on another machine. A flow declares the ones it needs, and most need only this directory. The flow names one each time it runs a turn, so one session may work in more than one; a turn that names none works in the workspace. See Remote execution and Containers.

Epic

One run of one flow, written down as it happens: the flow, its agents, and every session they opened. /epics lists the runs of this directory. See Tracing and Picking a run up.

Flow

The loop: Python that says which agent is asked what, in what order, and when to stop. You choose one at /flow or name it with $. See Your first run and Flows.

Flowverse

An index of flows: a git repository listing each release of a flow, and the repository and commit it is. A flow it lists runs once you install it from /flow. official is Humanize's own; add others on /flow's Flowverses page. A flow built in or installed from official is named as official lists it -- aot for humanfia's own, <user>/<flow> for anybody else's -- and one installed from another flowverse @<flowverse>/<flow>. See Flowverses.

Outworlder

You, as a flow sees you: whoever is at the prompt, when the flow asks something. /afk says you are away. See Being away.

Permission

What an agent may touch, declared by the flow for each role: its working directory, the rest of the home directory, the machine, and the web. See Permissions and Security.

Provider

A named account for one backend, kept apart from the CLI's own login: a login, an API key, or a gateway. It is the account row of an agent, where as local means the CLI as you signed it in. See Accounts.

Role

A named slot a flow declares, such as agent, builder or reviewer. You give each agent role an agent; the flow says what the role is for and what it may touch. See Your first run.

Session

One conversation with one agent, kept across its turns. When to start a fresh one is the biggest choice a flow makes, because it decides what the agent remembers: a Ralph loop starts a new session every round, so the agent starts from the task each time. See Many conversations at once.

Trace

A whole run as one timeline, which you open in Perfetto. See Tracing.

Turn

One exchange with the model: it is told something, works with its tools, and answers. A line you type while a turn is running goes into that turn. See Talking to a running turn.

Weaver

Whoever writes a flow. Writing Flows is for them; the User Guide never asks for Python.

Workspace

The directory hmz runs in. What is remembered, the run you can leave and come back to, and the runs /epics lists all belong to one workspace. See Settings.

---

Writing a flow of your own? The flow API has the Python.