beady-eye

bdi is one unblinking eye over every beads tracker you point it at. bd is all it needs. It will also look at whatever else you leave lying about, and each thing you let it see gives it more to do.

• It shows you the work. Each tracker's work is drawn as a tree, and a bead waiting on another project's work has that work drawn beneath it, whichever tracker holds it. Add herdr, and beside every bead an agent has claimed sits the live pane that agent is in.
• It wakes whoever is waiting. bdi watch keeps watch over every tracker at once, and tells anything connected to it the moment a bead it watches changes. Add Claude Code with the beady-eye plugin, and a sleeping session is woken by the bead it waits on.
• It settles waits on pull requests. Add gh, and bdi gates asks GitHub where each awaited pull request stands. A merge ends the bead's wait. A failing check, a conflict, a review or a comment is passed on to the bead that waits.

This is arkham with herdr running. Three of its beads have a live agent beside them, and the band at the foot is what the selected one's pane is saying. ark-5 is what else falls out of watching beads and herdr at once: beads says an agent claimed it, herdr has no pane for that agent, and the eye says so. It has seen a pane die on a Tuesday before.

The other two jobs happen while you are looking elsewhere. When a pull request moves, the news finds whoever waits on it:

    a pull request merges, fails a check, or is reviewed
                              │
                              ▼
    bdi gates closes the bead's wait, or comments on the bead
                              │
                              ▼
    bdi watch sees the bead change in its tracker
                              │
               ┌──────────────┴──────────────┐
               ▼                             ▼
    the plugin wakes the            the view redraws
    Claude Code session             in front of you
    waiting on that bead

Nobody polls GitHub by hand, and no agent sleeps through the answer.

The eye only looks. Waking an agent is still looking, only louder. When there is something it must do, it grows a pseudopod for the job, the way a shoggoth grows a limb from its own body, and so far it has grown two. bdi gates is one. The other carries a person's answer to the bead that asked for it. Changing the work itself is still bd's job, and the eye finds this arrangement acceptable.

Summoning

Homebrew fetches a built binary, for macOS and Linux on Intel and arm64:

$ brew install codeforbreakfast/tap/bdi

crates.io, if you would rather compile it yourself:

$ cargo install beady-eye

herdr, as a plugin that fetches the released binary for your platform:

$ herdr plugin install CodeForBreakfast/beady-eye

That puts nothing on screen by itself, because herdr has no command palette. Bind a key to the plugin's action in ~/.config/herdr/config.toml, and that key opens the eye in a split beside the focused pane, on that pane's project:

[[keys.command]]
key = "prefix+i"
type = "plugin_action"
command = "codeforbreakfast.beady-eye.open"

The plugin keeps its bdi to itself. To run bdi anywhere else, install it another way as well.

Nix, with flakes on, builds the tip of main. The flake serves Linux on Intel and arm64, and macOS on Apple silicon. An Intel Mac takes the Homebrew or release binary instead:

$ nix run github:CodeForBreakfast/beady-eye

From your own flake, pin a release tag and take the package or the overlay:

inputs.beady-eye.url = "github:CodeForBreakfast/beady-eye/v0.25.0";

beady-eye.packages.${system}.default                # the package
nixpkgs.overlays = [ beady-eye.overlays.default ];  # pkgs.beady-eye

CI pushes every main build for x86_64 Linux to a Cachix cache, so with the cache in your nix.settings neither of those compiles anything:

substituters = [ "https://codeforbreakfast.cachix.org" ];
trusted-public-keys = [ "codeforbreakfast.cachix.org-1:W96fHCqTzLZ7Vj2TKMMRLOkWzUn7N3LxKyBuF255XKg=" ];

Both are built against this flake's own nixpkgs, which is what the cache holds. pkgs.beady-eye-rebuilt through the overlay builds bdi against your nixpkgs instead, and compiles.

Or take a binary from the latest release. There is one per platform with a .sha256 beside it, and the Linux ones are static. Rename it bdi and put it on your PATH. Apple has not been asked to sign it, so a copy a browser downloaded needs xattr -d com.apple.quarantine bdi before macOS will let it open its eye.

The crate is beady-eye. The command is bdi.

Opening the eye

Stand in a repository beads tracks and run it:

$ bdi

No config. It reads that project, draws its trees, and keeps them fresh. Three bands: the forest at the top, the tail of the selected bead's pane under it, and a foot row with notices on the left and keys on the right.

Keys to start with. ? lists the lot.

key	does
↑ ↓ j k	move
← → h l	collapse or expand; again to move to the parent or first child
Enter	open the bead, and from there, focus its pane
f	focus the selected bead's pane
a	every tree, not only the ones with a live agent
F	only the selected bead and the work beneath it; again for the whole forest
/ n N	search ids and titles
y	copy the bead id (OSC 52, so it survives ssh and a multiplexer)
^R	read the trackers again now
q	avert the eye

bdi --json writes the same snapshot to stdout instead of drawing it. That is also what to reach for when stdout is not a terminal; bdi | cat says so and exits.

bdi --beads writes each unfinished bead once instead of the forest: whether it is ready, what blocks it, the agent on it, and its labels and description. Unlike bd ready, it counts a blocker in another project, so bdi --beads | jq '.beads[] | select(.ready)' is the work that waits on nothing. With --all-projects, select(.labels | index("human")) picks out every project's beads labelled human in one query.

Several trackers

A config file opens the eye on all of them:

[[projects]]
name = "arkham"
path = "/home/you/arkham"

[[projects]]
name = "dunwich"
path = "/srv/work/dunwich"
environment_command = "nix develop -c"

That lives at ~/.config/beady-eye/config.toml, or wherever --config says, and an edit takes effect while bdi runs. Start it under one of the projects and it reads that one alone. Start it anywhere else, or pass --all-projects, and it reads them all. --project NAME picks.

Each project is read with its own bd, entered the way you would enter it yourself. An .envrc and direnv need nothing said. Anything else, say it with environment_command.

A bead can wait on a bead in another project. Write the edge in the waiting bead's own tracker, with the other bead's bare id:

$ bd dep add ark-5 dun-7

bdi places dun-7 in the project whose beads carry the dun prefix, or in the project whose entry states prefix = "dun", and draws it beneath ark-5. A qualified form such as dunwich:dun-7 is not a bead id, so bdi cannot place it and draws it as not in any configured project.

bd dep list cannot see ark-5 waiting on dun-7, because the edge is stored in arkham's tracker. The forest can. Every unfinished bead is in some tree, with each bead it depends on drawn directly beneath it. So this lists every unfinished bead in any project that depends on dun-7, with its status, and goes on listing them after dun-7 closes:

$ bdi --json --all --all-projects | jq --arg project dunwich --arg id dun-7 '
    [.trees[].nodes
     | foreach .[] as $node ([]; .[:$node.depth] + [$node])
     | select(.[-1] | .project == $project and .id == $id and .edge == "blocks")
     | .[-2] | select(.status != "closed" and .status != "pinned")
     | {project, id, status}]
    | unique'

bdi watch reads every configured project and holds what it read, polling as the eye does and taking the same reports on a socket of its own. Run one per machine under whatever supervises your processes, such as a systemd user unit or a launchd agent. The view, bdi --json and bdi --beads then answer from what it holds and run no bd of their own, and read for themselves whenever it is not there. "Running the watcher" in docs/configuration.md has an example of each.

Anything can watch beads through the watcher. A consumer connects to its socket and says what it watches. It is sent each of those beads as one line of JSON, first as they stand and then each time one changes. This one watches a project and prints every bead it hears of:

import json, os, socket, sys

project = sys.argv[1]
at = os.path.join(os.environ["XDG_RUNTIME_DIR"], "beady-eye", "watcher.sock")

watcher = socket.socket(socket.AF_UNIX)
watcher.connect(at)        # refused: no watcher, so nothing is known
watcher.settimeout(60)     # an alive line is due every 20 seconds
watcher.sendall(f"watch {project}\n".encode())

try:
    for line in watcher.makefile():
        said = json.loads(line)
        if said["line"] == "bead":
            print(said["row"]["id"], said["row"]["status"], flush=True)
        elif said["line"] == "gone":
            print(said["id"], "gone", flush=True)
        elif said["line"] == "freshness" and said["tracker"] != "ok":
            print(project, "unreachable, last read", said["as_of"], flush=True)
except TimeoutError:
    sys.exit("the watcher has wedged")
sys.exit("the watcher has gone")

It starts from the beads that are not closed. A bead filed after it connected arrives as a line of its own, and so does a bead that closes:

$ python3 watch.py summit-works
smt-4kd3p open
smt-4kd3p.20 blocked
smt-4kd3p.21 open
smt-4kd3p.20 closed

The script tells a quiet watcher from one that has gone. A watcher that is up sends an alive line every 20 seconds, however quiet its trackers are. So a refused connection, a closed one, or a minute of silence means nothing is watching, and the script stops rather than act on what it has not been told. A tracker the watcher cannot reach is reported in that project's freshness line, and the beads already sent stand as the last known. Where a project's entry sets events_journal = true, the watcher also sends each of bd's event records as an event line, which says who made a change and carries the text of a comment. docs/design.md has every line the watcher sends, under "Watching".

A Claude Code session can be one of those consumers. The beady-eye plugin gives it watch, unwatch and watching tools, and wakes it when a bead it watches changes status, becomes ready, is commented on or goes. Its using-beady-eye skill tells the session how to read a change and when to unwatch. It needs Node and a watcher on the same machine. Add this repository as a marketplace pinned to a plugin release, and install the plugin from it:

$ claude plugin marketplace add CodeForBreakfast/beady-eye#plugin-v0.2.2
$ claude plugin install beady-eye@beady-eye

Where npm's min-release-age is set, npm refuses to start the plugin's server until each release has aged past it. The plugin's NPM_MIN_RELEASE_AGE setting replaces that age for the plugin's server alone, and 0 starts a new release at once.

Claude Code delivers a plugin's messages only to a session started with its channel allowed. A session started any other way still has the tools, and is never woken:

$ claude --dangerously-load-development-channels plugin:beady-eye@beady-eye

docs/configuration.md has the rest: badges drawn from what a bead carries, credentials, extra roots, intervals, the light theme, and the socket you can poke to say a tracker changed so the eye stops polling it.

Pseudopods

Each pseudopod does one job in a tracker the config names, with that project's own bd, and nothing else.

bdi bd <project> human respond <bead> <response> carries a person's answer to a bead that asked for one. It runs that project's bd the way the eye reads it, so whatever asks needs no map of trackers or credentials. No other bd command is passed, and no flag but -r/--response. A response that starts with a dash goes after --. Nothing after the project can pick another tracker, so a permission rule on bdi bd <project> holds a caller to that one project.

bdi gates settles the beads waiting on pull requests. Whoever opens the pull request records the wait as a beads gate on its number, and bdi gates never creates one:

$ bd gate create --type=gh:pr --blocks dun-7 --await-id=12

The gate names its repository in repo metadata, as OWNER/REPO. bd gate create copies it from the bead the gate blocks, and where that bead has none, bd update <gate> --set-metadata repo=dunwich/arkham sets it. The eye draws the gate under the bead it blocks, and links it to its pull request only through a badge in your config:

[[badges]]
key      = "metadata.repo"
when     = { await_type = "gh:pr", await_id = "[0-9]+" }
match    = "(?<owner>[A-Za-z0-9_.-]+)/(?<name>[A-Za-z0-9_.-]+)"
render   = "⇢ {name} #{await_id}"
short    = "⇢ #{await_id}"
link     = "https://github.com/{owner}/{name}/pull/{await_id}"
drawn_on = "blocked"

bdi gates looks at every configured project's open gh:pr gates and asks GitHub, through gh, where each pull request stands. A merge closes the gate, so the bead it held back becomes ready. A gate with awaits=ready_for_review in its metadata closes sooner, once the pull request leaves draft, and so does one with awaits=approved, once GitHub's review decision is approved. A close without a merge leaves the gate open and comments once on that bead, so whoever waits on it hears. So do failing checks on the pull request's head commit, a conflict between that commit and its base, each review submitted, and each comment on the pull request's conversation, once for each commit, review or comment. Then it looks again a minute later, until it is stopped. Run it under whatever supervises your processes, as you would the watcher. Given --listen and the webhook's secret, it also takes GitHub's deliveries about pull requests, their reviews and comments, check suites and statuses, and settles a pull request when GitHub says it moved rather than at the next look. "Settling pull-request gates" in docs/configuration.md has a unit to run it under, the webhook, and every line it prints.

Being seen

The eye draws only what it can find. An agent that wants to be found tells the bead which pane it is in. One that does not still toils, but toils among the unaccounted-for below the trees, and no amount of staring will move it up.

docs/agents.md is addressed to the agent rather than to you. Hand it over, and let it fold the lines into whatever it already obeys.

What it needs

• Linux or macOS. Unix sockets and unix signals. Windows would need a different ritual entirely.

• bd 1.1.0 or newer. An older one is reported on the project's line rather than obeyed. Packaged builds lag, so check what yours says.

• A tracker bd can open. Server or embedded. A Dolt server wants a password, and it reaches bd in BEADS_DOLT_PASSWORD, from your shell or from a project's credential_command.

• herdr, for the agents. Without it you get the trees and the claims, and every tree is drawn. With it you get the point: which claim has a live pane behind it, which pane toils on nothing any bead has heard of, and the tail. With bdi on your PATH, a binding in herdr's config opens it as a popup on the focused pane's project, and q puts it away again:

  [[keys.command]]
  key = "prefix+shift+i"
  type = "popup"
  command = "bdi"

• git, for worktrees. Without it the project is named after its directory and a pane cannot be placed by worktree.

• gh, for bdi gates. It settles what the account gh is signed in to can see, and its token needs the read access the configuration reference lists. Nothing else in bdi asks GitHub anything.

• A terminal that honours OSC 52, for y. Terminal.app does not, and says nothing about it.

Status

Released, and gazed into daily by the people who wrote it. It has not yet gazed back. The design is in docs/design.md. It is made at Code For Breakfast, where it has a page of its own.

Versions are 0.x, and a breaking change bumps the minor. Pin a tag. 1.0.0 arrives when the shape has settled, not when something breaks.