crowsnest

One session that watches your other Claude Code sessions.

A machine running many Claude Code sessions has a question nobody answers: what are they all doing, and which of them needs me? Each session knows only itself, the terminal tabs are silent until you click them, and the answer lives in forty scrollbacks.

crowsnest reads what Claude Code already writes, the registry it keeps for every running session and the transcript each one appends to, and answers in three tiers, cheapest first. It can start a new named session (crowsnest spawn), but it never sends into, kills, or otherwise writes into a session that already exists.

Start here

pip install crowsnest          # Python 3.10+. Puts a `crowsnest` command on your PATH.
crowsnest install-skills       # link the skills and the scout subagent into ~/.claude
crowsnest init --hooks         # in the directory you will run the lookout from

init writes that directory's CLAUDE.md (the rules that keep a watching session small enough to be cleared at any moment), creates ~/.local/share/crowsnest/, and — with --hooks — adds three lines to your settings.json, after backing it up and removing nothing. Without --hooks it prints them for you to paste. The two that push events (Stop, Notification) are registered async, so watching costs the watched sessions no wall-clock; the SessionStart one blocks, because its roster is meant to be read.

Then open a session, name it, and ask in whatever words you would have used anyway:

claude -n lookout

what are my sessions doing, and is anything waiting on me?

## Waiting on you
- xa_needed_or_not (xa, 25m) — input needed

## Just finished
- monitor (proj, 10m) — "The sweep landed. It's the repo-side view, and it's complementary…"

## Working
- openloops: one session, mid-turn on the one-command test.

## Headline
1 waiting on you, 1 just finished, 1 working, 29 idle.

That answer is a synthesis. Every row behind it is one the crowsnest command printed, and the skill tells the session how to go deeper when a row is not enough: read a session's earlier turns, ask the session itself, or arm the event stream and be told.

The three tiers

The skill's rule is to never spend a costlier tier when a cheaper one answers, and never to ask a session that is busy or waiting.

The command

crowsnest                          who is alive: waiting on you first, then busy, then idle
crowsnest show <session>           one session: last asked, last said, running now, pending question
crowsnest turns <session> -n 5     the last five turns, oldest first; --before N pages back
crowsnest brief <session>          openloops' dated digest for one session; reads no transcript
crowsnest lineage                  who started whom, as a tree; --backfill recovers it from transcripts once
crowsnest report [--out FILE]      the roster as one phone-readable HTML page, no stylesheet or script
                                   (--fragment: without the document wrapper, for publishing as an artifact;
                                    --interactive: buttons per row and a Refresh, live when published with the db capability)
crowsnest seen|done <session>      you read it / you handled it: until what it asks for changes
crowsnest later <session> 1h       put it off: 1h, evening, tomorrow, or change (--plan "next step")
crowsnest note|undo <session>      a note to yourself; one step back
crowsnest attention export|import  your attention records as JSON, last write wins
crowsnest watch                    one line per change, forever (started, exited, idle, busy, waiting, error)
crowsnest ledger [<session>]       one session's durable page, or all of them with ages
crowsnest hook <event>             called by your Stop and Notification hooks; reads their JSON on stdin
crowsnest spawn <name> --cwd <dir> start a named session in <dir>, and wait for it to show up
                                   (--profile NAME: on another account; --binary PATH: another claude)
crowsnest open <session>           raise its terminal on the desktop, or say where it runs
crowsnest init                     this session's CLAUDE.md, the data directory, the hook lines
crowsnest install-skills           link the skills and the scout subagent into ~/.claude

open is a desktop command -- it looks for an iTerm tab or a tmux session and has nothing to raise from a web page. So wherever crowsnest report cannot link a session (no Remote Control, so no claude.ai URL), the page shows the crowsnest open command that reaches it. You paste that into a terminal; the page never brings the terminal to the front for you.

<session> is the name you gave the session with claude -n <name>, a unique prefix of one, a session-id prefix, or a pid.

waiting  25m  xa_needed_or_not      xa           input needed · Let me split that into smaller steps.
busy     14m  session_monitor       openloops    → Bash: Run the one-command test
idle     10m  monitor               proj         "The sweep landed. It's the repo-side view…"
-- 32 live: 1 waiting, 1 busy, 30 idle

What needs you, what is safe to close

The roster is not the useful answer. Thirty sessions sorted by status is still thirty things to read; the question underneath it is always the same three.

crowsnest triage --quiet

NEEDS YOU (5)
  cn-mergeset          mergeset    [decision] The history rewrite and the PyPI deletions — awaiting go-ahead…
  cosmo_3d_02          cosmograph  [action] attach the two GIFs to #604/#616 bodies (no API for that)…

SAFE TO CLOSE (4)
  cn-cosm-synth        cosm        Nothing outstanding

WORKING (1)
UNCLASSIFIED — has not said where it stands: 48

-- 5 need you (1 to decide, 3 to do), 4 safe to close, 1 working, 48 unclassified

[decision] is a minute of thought; [action] is a trip to another window. The reason is the session's own words, so you can check the verdict without opening the session.

unclassified is the honest group, and it is the point. A wrong "safe to close" is the expensive error — somebody closes a terminal on unfinished work and nothing ever tells them — so a verdict is only reached on positive evidence. A session that has not said where it stands is reported as not having said. Nothing infers "finished" from silence, because silence is also what an interrupted session leaves behind.

To be reported well, a session writes the ledger's open questions field, or says "nothing outstanding" in its notes when it is done. The shipped crowsnest-worker skill teaches both, so the signal improves as sessions run.

crowsnest report organises the page the same way: Needs you first, then Safe to close, then what is working and what is quiet.

What you have already dealt with

Triage says what each session needs; it cannot know that you read three of them, put two off until this evening, and handled a fourth. That is a second record, kept apart from the first: yours, not the sessions', one small JSON document per item under ~/.local/share/crowsnest/attention/.

crowsnest seen shipper                              read: dimmed until what it asks for changes
crowsnest later shipper evening --plan "after the deploy"
crowsnest done shipper                              handled: hidden until what it asks for changes
crowsnest note shipper "check the benchmark first"  a note to yourself, never an instruction
crowsnest undo shipper                              one step back

Seen and done are pinned to a revision of the item, not a flag. The revision is a hash of what you would have to decide again — the group, why, and the ask in its own words — and never a timestamp, a tool name, or the session's latest chatter and the links in it. A session that asks a different question comes back as changed; one that merely runs another command does not. Later wakes when its time passes or as soon as the item changes, whichever comes first (--ignore-changes keeps it asleep through a change), and change as the preset means no time at all. Nothing is scheduled: what you see is computed from the record whenever it is shown.

Items are keyed by session id, not name, so a resumed session keeps its record and a new session given an old name starts fresh. The hours behind evening and tomorrow are an [attention] table in the config file (evening_hour = 18, morning_hour = 9). crowsnest attention export and import move the records as JSON, the newer updated_at winning, which is how a published page's copy and this machine's are kept in step. The report page showing this record is the next step (#55).

Every reference, as a link

A session working in i2mint/mergeset that writes #17 means that repository's issue 17. Nothing in the text says so — but crowsnest knows the session's working directory, and that is enough to resolve the commonest, least resolvable reference a session writes. So the report renders references as links rather than as text you have to reconstruct a URL from: markdown links the session already wrote out, bare URLs typed by where they point (issue, pull request, discussion, commit, CI run, artifact, Slack), owner/repo#N, bare #N, and bare commit shas.

#45 and a link to pull/45 are one reference, not two, and collapse to the label its author chose. A discussion of the same number does not — it is a different object that happens to share a number. Nothing is fetched: a link is constructed from what the text says and what the directory implies, so a page of two hundred references still costs no network calls.

crowsnest show <session>       # ends with a References section
crowsnest report               # every row's references, linked

Who started whom

A fleet of forty reads as a list of forty until you can see that six of them are one dispatcher's children. The registry does not record that, so crowsnest spawn writes it down itself, at the one moment it is free and certain — the session that ran the command is the parent, and one spawn line goes into lineage.jsonl naming both ends. That file is deliberately not the hook's events.jsonl, which rotates at 4 MiB: a machine busy enough to have an interesting graph is exactly the machine whose graph would vanish.

crowsnest lineage

cn                                . idle    cn
|-- cn-mergeset                   . idle    mergeset
|-- cn-priv-manifest              . idle    priv
`-- cn-cosm-align                 . idle    cosm
    |-- cn-cosm-supply            . idle    cosm
    `-- cn-cosm-synth             . idle    cosm
crowsnest                         x gone
`-- cn-c                          . idle    c ~
    |-- c-surfaces                x gone     ~
    `-- c-traversal               x gone     ~

A parent that has exited stays in the picture (x gone) so its children stay a fleet rather than becoming unrelated roots.

crowsnest report draws the same forest as inline SVG, laid out in Python — no graph library, no script, nothing fetched, which is what lets the page keep its offline promise. Sessions that started nobody and were started by nobody are left out (the roster above is already that list), a parent's childless children collapse once there are more than a handful — the ones that need you kept, the rest summarised as "7 more · 7 idle" — and the figure is drawn at a fixed size so its labels are 12px whatever the screen, scrolling sideways inside its own box rather than shrinking.

Recording only works forwards. For a machine that has been running sessions since before this shipped, crowsnest lineage --backfill recovers what it can, once, by scanning every transcript for the crowsnest spawn commands that created today's sessions, and writes what it finds into lineage.jsonl so the cheap reader has it from then on. That is inference — the command may have failed — so those edges are marked ~ wherever they are shown, and a recorded edge is never overwritten by a guessed one. A mention is not a spawn: the command has to be the head of a shell segment, so grepping for the phrase or writing a commit message about it does not invent a parent. --dry-run says what it would add and writes nothing.

Several accounts and machines in one roster

Reading crosses accounts and machines; messaging does not, so a crowsnest session operates the fleet of its own account and machine and can read all the others. What it spawns lands on its own account too: crowsnest spawn starts the new session the way this one runs — same account (its CLAUDE_CONFIG_DIR, stated absolutely on the command line, because a tmux server hands a new session its environment rather than the client's) and same claude binary ($CLAUDE_CODE_EXECPATH, not whatever a login shell's PATH resolves) — so a second-account crowsnest never opens sessions under the default account by mistake, and ANTHROPIC_API_KEY is dropped so the child signs in as its account, not as an API bill. --profile NAME starts one on another account, --home DIR spells that home out, and $CROWSNEST_PROFILE is that choice made once. List the homes once:

# ~/.config/crowsnest/config.toml
[[homes]]
name = "main"
path = "~/.claude"

[[homes]]
name = "work"
path = "~/.claude-work"          # a second account on this machine

[[homes]]
name = "server"
path = "~/.cache/xa/remotes/server"   # a copy synced down with `xa sync`
remote = true                         # its pids are not ours: alive while fresh

Every report row links to the session on claude.ai when it runs with Remote Control, and otherwise shows the crowsnest open command that reaches it. The spawn tree's list does the same for every session it names. Rows also link to the session's repository, and to the issues and pull requests it mentioned. Then crowsnest --all-homes prints every home with a column saying which — which is also how a session spawned with --profile shows up, on the account it actually runs on — crowsnest show name@home --all-homes picks one when a name exists in two, and crowsnest watch --all-homes streams events from all of them, each tagged name@home.

A profile name is resolved against the [[homes]] names above first, then against a claude-profile dir <name> command on your PATH if you keep one; a name neither knows is an error, never a quiet fall back to the default account, and a home marked remote is refused — those are another machine's, read-only.

Which claude gets started

By default, the one this session runs ($CLAUDE_CODE_EXECPATH) — so a spawned session is the same build, signed in the same way, even mid-upgrade. Say otherwise when your Claude Code is not the claude a login shell finds first:

export CROWSNEST_CLAUDE_BIN=/opt/claude-next     # this shell

# ~/.config/crowsnest/config.toml — once and for all.
# Above the first [[homes]]: TOML gives every key after a table header to that
# table, so a claude_bin at the bottom of the file belongs to the last home and
# does nothing. crowsnest refuses that arrangement rather than ignoring it.
claude_bin = "claude-next"

and --binary PATH names one for a single spawn. --binary is handed to the spawner verbatim and unchecked, because a spawner may run the line on another machine where this one's PATH means nothing — so an absolute path or a name only the target resolves both survive it. The corollary is that --binary claude means the bare name, which a local spawner then resolves through the settings above; to force the plain one, say --binary "$(which claude)".

The two persistent settings are checked, because they are claims about this machine: $CROWSNEST_CLAUDE_BIN first, then claude_bin, then the inherited binary.

Either has to be something that can actually be executed — a script on your PATH or an absolute path (a relative path is refused in the config file, which is read from every directory). A shell alias is not. Aliases exist only inside an interactive shell, and a session is started by tmux or a bare subprocess, neither of which reads your startup files; naming one is refused with an error that says so, rather than opening a terminal that prints command not found and closes. If you have an alias you want to reuse, make it a two-line script instead.

Before you do: a wrapper whose whole job is --dangerously-skip-permissions and env -u ANTHROPIC_API_KEY — the usual shape — has nothing to add, because crowsnest spawn already does both.

The ledger: what a session leaves behind

The roster and the transcript both describe now. Neither survives a /clear in the watching session, and neither says what a session decided or what it is still waiting on you for. So each session gets a small markdown file — ~/.local/share/crowsnest/ledger/<name>.md — that it writes and a watcher reads:

# lookout

state: working
last asked: 2026-09-06T18:12:00+00:00 · fix the widget
last said: 2026-09-06T18:14:22+00:00 · Fixed and merged; PR 12 is green.
open questions:
- squash or rebase for the release?
decisions:
- the ledger lives under ~/.local/share/crowsnest

## Notes

Anything at all. Nothing in crowsnest ever rewrites this part.

last asked and last said are mechanical — the Stop hook below writes them from the transcript tail, so they are true without anyone deciding anything. state, open questions and decisions are judgements, and only the session whose ledger it is writes those. A write rewrites the named fields and leaves every other byte alone, so a hook and a human can edit the same file minutes apart.

Being told instead of polling

crowsnest watch notices a change within a poll and has to read a transcript to guess why. Claude Code knows both exactly and immediately, and will say so — if you give it a line to say it on:

// ~/.claude/settings.json
"Stop":         [{"hooks": [{"type": "command", "command": "crowsnest hook stop", "async": true}]}],
"Notification": [{"hooks": [{"type": "command", "command": "crowsnest hook notification", "async": true}]}]

With those two lines, a turn ending becomes a stopped line in crowsnest watch carrying the session's last words, and a permission prompt or a question becomes a needs-you line carrying the message — both a poll sooner than the registry could notice, and both with the reason rather than a guess at it. crowsnest hook prints nothing and exits 0 whatever happens; async means Claude Code does not wait for it either, so watching costs the watched sessions no wall-clock. Without the hooks installed, crowsnest watch is exactly the registry diff it always was. crowsnest init --hooks writes all of this for you.

One walk, end to end

A morning with a fleet, from the lookout session. Everything in italics is something you say; everything else it does.

1. Open it. claude -n lookout in the directory you ran crowsnest init in. Its CLAUDE.md and the SessionStart hook put the roster in front of it before you type anything.
2. what needs me? — it sends the crowsnest-scout subagent, which reads the ledgers, then the roster, and hands back a page: who is waiting, who just finished, what is working. The rosters never enter the lookout's own context.
3. start something on the parser tests — the crowsnest-dispatch skill: it names the session, writes a brief that is pointers rather than prose (the issue URL, the acceptance line, the reply contract), and runs crowsnest spawn parser-tests --cwd ~/proj/parser --prompt "…". A new terminal session appears, named, in that directory. It subscribes once with notify_when_idle and stops looking.
4. It is told when that finishes — from the crowsnest watch stream under the Monitor tool, or from the one idle notice. Either way you hear about it without asking.
5. what did it do? — crowsnest ledger parser-tests. The worker wrote that file itself, under the crowsnest-worker skill, which every session on the machine has; the lookout only reads it. crowsnest brief parser-tests adds openloops' dated digest of the same session, still without opening a transcript.
6. give me a page — the crowsnest-report skill: crowsnest report --fragment --out fleet.html, published with the Artifact tool, one link, stable across re-publishes.
7. From your phone, later: you highlight the row for parser-tests, comment ask it what is left, and send it to Claude. The lookout wakes on the comment, answers from the ledger, replies in the thread, resolves it, and re-publishes the page.

At no point does the lookout edit a file in any of those repositories, or read a transcript itself. That is the whole design: the corpus sessions hold the context, and the lookout stays small enough to /clear at any moment.

What it ships for agents

crowsnest install-skills links all of them into ~/.claude (or --target) and never overwrites anything that is not already ours. The worker skill installs everywhere by default, because any session on the machine may be asked.

What the hooks add

crowsnest init --hooks registers two user-wide hooks, both asynchronous so they never delay a turn: Stop (a turn ended; the session's last words go to its ledger and the event log) and Notification (a session asked for something). It puts the roster-on-start hook only in the crowsnest directory's own .claude/settings.json, so no other session is handed a roster. An idle_prompt notification (a session merely sitting idle) is recorded but never streamed as needs-you.

What it reads

• ~/.claude/sessions/<pid>.json: written while a session runs. Name, session id, working directory, busy / idle / waiting, and when waiting, what for. Checked against a live process before it is reported, because a crash leaves the file behind.
• ~/.claude/projects/<slug>/<session-id>.jsonl: the transcript. Read from the end, widening until one human prompt is in view; turns reads the whole file on request.

What a transcript's content means is openloops' business, and crowsnest calls it rather than re-implementing it. openloops deliberately never looks at whether a process is running; crowsnest is that other half.

What it writes

Nothing into another session, and nothing into a repository. Everything crowsnest writes is its own and lives under ~/.local/share/crowsnest ($CROWSNEST_DATA_DIR or $XDG_DATA_HOME if you set either):

• ledger/<name>.md: one per session, as above.
• events.jsonl: one line per hook event, append-only, rotated by size. crowsnest watch tails it.
• hook.log: one line for anything crowsnest hook swallowed, so "the hook did nothing" is a question with an answer.
• attention/<item-id>.json: one per item you have marked seen, put off, handled or noted. Shared by every account on this machine, because the data directory is per user.

The only other writes in the package are crowsnest spawn, which starts a session, and install-skills, which writes symlinks.

From Python

from crowsnest import (
    roster,
    show,
    turns,
    brief,
    events,
    live_sessions,
    spawn,
    read_ledger,
    update_ledger,
)

roster()["counts"]  # {'waiting': 1, 'busy': 1, 'idle': 30, 'other': 0}
show("monitor")["activity"]["last_assistant_text"]
brief("monitor")["digest"]  # openloops' dated digest, or None if there is none yet
update_ledger("monitor", state="working", open_questions=["squash or rebase?"])
read_ledger("monitor")["fields"]["last_said"]
for event in events(interval=5):  # forever
    ...
spawn("demo", cwd="/path/to/repo", prompt="run the tests")["pid"]

Every function takes home= (the Claude Code config directory; a synced copy of another machine's works the same way), the readers take is_alive= (how a registry pid is confirmed running), and everything that writes takes ledger_dir= or events_path=.

Not in crowsnest

Killing or resuming a session is xa's job (xa spawn is also the pointed replacement for crowsnest spawn's spawner seam, adding hosts and a phone web UI). Asking a session a question is Claude Code's own SendMessage; the skill says when.