Hearting
Fleet · part of Hearting

Every agent, on one screen.

Background agents leave you guessing. Is it still running? Is it stuck? Is it waiting for you? Did it finish? Fleet answers in one terminal view, for every Claude Code, Codex, and OpenCode session on your machine and every worker Hearting dispatched from them.

Fleet walk-through in five steps: the OVERVIEW of all sessions, the ROUTES list with one route opened in place, one session's page, that session's own window with its Hearting panel, and one request's stage graph with the owner's raw output.
OVERVIEW → ROUTES with a route opened → a session page → that session’s own panel → a request's raw output. Every picture on this page is drawn by Hearting’s own renderers from synthetic demo data, never from a real machine; a conversation beside a panel is illustrative.
The problem

A background agent is a black box until you open its terminal.

With several sessions across three runtimes, and workers dispatched under them, you end up opening windows one by one to learn the same three things.

ALIVE?

What is running

Working, idle, and waiting-for-you have different marks. Each session shows what it is doing now and for how long.

STUCK?

What went quiet

A stalled worker shows up as a stalled worker, not as a parent session that went quiet. Rows whose parent is gone are shown, not dropped.

WAITING?

What needs you

QUESTIONS lists every agent waiting for your answer or approval, since when, and for which request.

The screens

One screen per question.

Press Tab to move between views. Select a session or a request to open its page.

OVERVIEW

Every session, grouped by project.

Each project's line says where it stands, and each session shows what it is doing right now.

  • Per project: open requests, questions waiting for you, what is left to decide, and when it last changed.
  • Per session: runtime, model and effort, a short title, context left, what it is doing now, and for how long.
  • Marks: working, idle, and waiting-for-you each look different.
  • Top row: each runtime's usage-window gauges, and how many sessions are working, idle, or waiting.
Fleet's OVERVIEW at the same moment: eight sessions in four projects across Claude Code, Codex, and OpenCode. Session [4f] runs Rate limiting at 46%, the Codex session [b2] runs the Test run for [4f] with 31 of 40 suites passing, and [9c] waits for an answer; below, the two open routes with their stage lines, staged 2/4 execute and staged 2/4 eval-asr.
PEERS

Who works with whom.

One card per session, and the newest link of each kind between them: a request, a message, a watch, and when.

  • The same exchanges each session’s panel lists, at the same minutes.
Fleet's PEERS: one card per session, grouped by project, with each owner's route and stage marks; below, the five links with their kind and time: a reply from [e8] to [d1], a steer from [4f] to [9c], a steer from [d1] to [e8], the PEER REQUEST from [4f] to [b2], and a reply from [b2] to [4f].
ROUTES

Each request, stage by stage.

A stage line such as plan › execute › test › report, with the workers running under it. Press a route to open its stages in place.

  • See which stage a request is in and which worker, on which model, is running it.
Fleet's ROUTES with the web-api route opened in place: Rate limiting, staged 2/4 with execute running, the owner Opus 5.5 xhigh and the execute worker Sonnet 5.5; its stage boxes plan done, execute running, test and report pending, each with who runs it, when, its gate, and its output; under it the ml-train route at eval-asr.
QUESTIONS

Everything waiting for your answer.

Every agent waiting for an answer or an approval: which session asks, what kind of question, for which request, and since when.

  • Older waits stay listed until you answer or the work ends. Nothing drops off for its age.
  • Answer from Fleet when the runtime offers a reply path. Otherwise Fleet points you to the session.
QUESTIONS: one decision a Claude Code session has been waiting on for seven minutes, with since when and how to answer it; one older approval wait from a Codex session listed under it; and three questions answered today.
SESSION PAGE

One session, in full.

State, context left, what it is doing now, its routes and requests, the messages it sent and received, who supervises whom (its PEERS row) and the sessions it is linked with, and the GPU it holds, in the session panel's words.

  • Select any session in the tree to open its page.
A session page: one Claude Code session's state, context gauge, and current activity, its route, the message it sent the OpenCode session [e8], and the GPU it holds.
RESOURCES

What is using the machine.

CPU, memory, and per-GPU use on this machine and on any hosts you configured, with the session that owns each GPU process.

  • Even without configured hosts, Fleet shows this machine's CPU and GPU use.
RESOURCES: two hosts with per-GPU utilisation and memory, the session and command behind each GPU process, and CPU and memory use.
REQUEST PAGE

A request's stage graph and raw output.

Select a request to see its stages, and under them the owner's raw output: tool calls and results in order, without opening the runtime itself.

  • Setup done, one stage running, the rest pending, and what the owner just did.
A request's page: the stage graph with setup done, eval-asr running, eval-sep and report pending, and below it the owner's raw output listing an edit, a shell command, and their results.

HISTORY

Earlier sessions of each terminal seat, so work that continued in a new session stays traceable.

MEMORY

Per project: the newest decisions, open work by stage, and the latest records saved to shared memory.

MODELS · SYSTEM

Which real model each budget maps to, and the installed Hearting release.

It fits a narrow terminal

The same OVERVIEW, laid out for a 100-column window. Click any picture on this page for the full-size image.

Tabview/filterEnteropen uusageqquit
Fleet's OVERVIEW in a 100-column terminal: the same sessions, projects, and request lines in a narrower layout.
Beside each session

The same values, in the panel beside each session.

Every new Claude Code, Codex, and OpenCode terminal opens with a Hearting panel. It reads the same records Fleet reads, so the panel’s head and the session’s Fleet row agree. After Ctrl+], f opens Fleet in that window; on a session page in Fleet, Ctrl+] opens that exact session’s panel.

Session [4f]'s window in project web-api with the STATUS button on its REPO row pressed: the project status document in a popup in the middle of the window, with the one line, the decisions needed, what is next, and what closed recently; behind it the conversation and the panel in gray, and on the popup's bottom line STATUS, ROUTE, and MESSAGES. On the window's last line the working folder and hearting v4.46.0.

What it shows

Project and Git (with a PEERS row when the session supervises or is supervised), model and context, the title and what the session is doing now; then a few rows each of the questions waiting for you, this session’s routes and requests in one box of three zones (direct, dispatch, request), messages, and saved memory with its scope and kind; then the usage windows of the three agents. The STATUS › button opens the project’s status document. Press the questions, the route box, or MESSAGES and its whole table opens in a popup in the middle of the window, the rest gray behind it; MEM and USAGE open nothing.

A question you can answer here stands in a rose-bordered box under QUESTIONS. Press Ctrl+] and it opens in the popup: answer with Tab, 1–9, and Enter, and the popup closes. In a narrow window the panel folds into a strip under the agent’s screen; opened there by hand, it sits below. The conversation in these pictures is illustrative.

hearting session shell disable turns it off for new terminals. Keys, settings, and per-runtime details →

The bottom layout in a 100-column window: the panel under the native area, with the session's head, the 2 QUESTIONS box with a rose border, the first question, its three choices as text, the first recommended, and answer, Ctrl+] at the right; then messages and the usage of the three agents.

Sessions hand each other work

A session can hand another session work as a request: it gets a number, the other session opens the body itself, partial reports keep it open, and a final report closes it and wakes the session that asked. Both panels list it in the route box’s request zone until it closes — → handed out, ◂ handed to you, ✓ finished.

Messages and watching work the same way, and none of it needs herdr or tmux. Fleet lists a session’s requests on its session page and shows the newest as a PEERS link; it has no list of every request across sessions, and you cannot send one from Fleet yet.

Two panels side by side. Left, the session [4f] that asked: in its route box, under the request rule, an arrow to [b2] with its harness, model, and effort, Test run, partial answer, and a finished review; under it MESSAGES with a reply from [b2] and a note to [9c]. Right, the session [b2] that was asked: in its route box, under the request rule, an arrow from [4f] with its harness, model, and effort, Test run, partial answer; under it its reply to [4f].
Why you can trust it

It shows what the work recorded, and nothing it made up.

01

Built from the work’s own records

Fleet reads the attempt registry the dispatcher writes and each runtime’s own session state. It does not guess from what a terminal printed.

02

Unknown stays unknown

A value Fleet could not read is shown as ?. Never as zero, never as success.

03

Watching changes nothing

Observation never completes, retries, or cleans up work. Fleet acts only on what you select, such as answering a waiting question.

Run it

Installed with Hearting. One word to start.

Press Ctrl+] then f in any session’s window, or run fleet in any terminal.

fleet              # the Explorer, full screen (needs curses)
fleet --once       # one plain snapshot
fleet --json       # the same state as JSON, for scripts
fleet-monitor      # the earlier group/process view

Titles and Fleet’s own words follow your language. Set title_language in ${XDG_CONFIG_HOME:-~/.config}/hearting/fleet.json to override it.

--once and --json need no curses, so they also work on Windows under Git Bash (experimental).

Known limits

Fleet is not light yet.

On a busy machine with a long work history and many live sessions, one Explorer window:

≈ 1/3of one CPU core

Measured use. Lowering its CPU use continues in v4.x.

≈ 0.8GB of memory

Measured use of one Explorer window, on that machine.

earlyfirst screen

A new window draws its first screen as soon as the session list is read, and marks the cells still loading details wait.

Answering a question from Fleet works only where the runtime offers a reply path. Where it does not, Fleet points you to the session instead.

See every agent without opening a single terminal.

Fleet comes with Hearting. Install once, then run fleet.