dotagents

herdr — agent skill

before using this skill, check that HERDR_ENV=1. if it is not set to 1, say you are not running inside a herdr-managed pane and stop. do not inspect or control the focused herdr pane from outside herdr.

you are running inside herdr, a terminal-native agent multiplexer. herdr gives you workspaces, tabs, and panes — each pane is a real terminal with its own shell, agent, server, or log stream — and you can control all of it from the cli.

this means you can:

the herdr binary is available in your PATH. its workspace, tab, pane, and wait commands talk to the running herdr instance over a local unix socket.

if you need the raw protocol or full api reference, read the socket api docs.

survey before you dispatch — mandatory

before spawning an agent, creating a workspace/tab for a topic, or writing a mission brief for another agent, look. spawning is the last step, not the first.

run all four. do not skip one because the topic “feels new”:

herdr workspace list          # is a workspace already labeled for this topic?
herdr pane list               # which agents are live, and what are they doing?
tmux ls; for u in <other users>; do sudo -n -u "$u" tmux ls; done   # herdr is not the only multiplexer on the box
ls /tmp/*<topic>* 2>/dev/null # are the deliverables already written?

then, on every match:

herdr agent read <pane> --source recent --lines 120

read what it concluded. an agent that is idle, done, or sitting at a prompt is not spare capacity — it may be a finished piece of work nobody has collected, or an agent stopped mid-task waiting on a human’s answer. dispatching a second agent at that topic destroys the first one’s result and asks the human the same question twice.

never assert unverified facts in a brief

every commit hash, file path, branch name, and divergence count you put in a mission brief must come from command output you read in the same turn you write it. do not carry facts forward from an earlier turn, another agent’s summary, or memory.

a brief that opens with “context already verified” and contains a hash that git cat-file -t rejects is a fabrication. the receiving agent cannot tell your invented hash from a real one, so it acts on it and writes the fiction into its own output as provenance. this is unrecoverable downstream: bad provenance is indistinguishable from good provenance once it is written.

if you have not run the command, write “unverified” next to the claim, or leave it out.

prefer messaging a live lane over spawning its twin

if a lane already exists for the topic, talk to it:

herdr agent send <pane> "<your message>"
herdr pane send-keys <pane> Enter

duplicate lanes are a coordination defect, not extra coverage. two agents on one topic produce two partial truths and no owner.

concepts

workspaces are project contexts. each workspace has one or more tabs. unless manually renamed, a workspace’s label follows the first tab’s root pane — usually the repo name, otherwise the root pane’s current folder name.

tabs are subcontexts inside a workspace. each tab has one or more panes.

panes are terminal splits inside a tab. each pane runs its own process — a shell, an agent, a server, anything.

agent status is detected automatically by herdr. the api exposes one public field for it:

done means the agent finished, but you have not looked at that finished pane yet.

plain shells still exist as panes, but herdr’s sidebar agent section intentionally focuses on detected agents rather than listing every shell.

ids — as of herdr 0.7.4 (live-verified 2026-07-16, ep137 smoke): workspace ids look like w1, w2; tab ids like w1:t1; pane ids like w1:p1, w1:p2. (Earlier versions of this doc said 1, 1:1, 1-1 — the dashed pane form does not match live herdr 0.7.) Pane rows carry terminal_title/terminal_title_stripped (no title field); a label field appears only after herdr pane rename. These are compact public ids for the current live session.

important: ids can compact when tabs, panes, or workspaces are closed. do not treat them as durable ids. re-read ids from workspace list, tab list, pane list, or create/split responses when you need a current id. do not guess that an older 1-3 is still the same pane later.

discover yourself

see what panes exist and which one is focused:

herdr pane list

the focused pane is yours. other panes are your neighbors.

list workspaces:

herdr workspace list

tab management

list tabs in the current workspace:

herdr tab list --workspace 1

create a new tab:

herdr tab create --workspace 1

without --label, the new tab keeps the default numbered tab name.

create and name it in one step:

herdr tab create --workspace 1 --label "logs"

rename it:

herdr tab rename 1:2 "logs"

focus it:

herdr tab focus 1:2

close it:

herdr tab close 1:2

read another pane

see what is on another pane’s screen:

herdr pane read 1-1 --source recent --lines 50

split a pane and run a command

split your pane to the right and keep focus on your current pane:

herdr pane split 1-2 --direction right --no-focus

that prints json with the new pane nested at result.pane.pane_id. parse that value, then run a command in that pane:

NEW_PANE=$(herdr pane split 1-2 --direction right --no-focus | python3 -c 'import sys,json; print(json.load(sys.stdin)["result"]["pane"]["pane_id"])')
herdr pane run "$NEW_PANE" "npm run dev"

split downward instead:

herdr pane split 1-2 --direction down --no-focus

wait for output

block until specific text appears in a pane. useful for waiting on servers, builds, and tests.

for --source recent, matching uses unwrapped recent terminal text, so pane width and soft wrapping do not break matches. pane read --source recent still shows the pane as rendered. if you want to inspect the same transcript that the waiter matches, use pane read --source recent-unwrapped.

herdr wait output 1-3 --match "ready on port 3000" --timeout 30000

with regex:

herdr wait output 1-3 --match "server.*ready" --regex --timeout 30000

if it times out, exit code is 1.

wait for an agent status

block until another agent reaches a specific status:

herdr wait agent-status 1-1 --status done --timeout 60000

use this when you want the same done / idle distinction the UI shows.

send text or keys to a pane

send text without pressing Enter:

herdr pane send-text 1-1 "hello from claude"

press Enter or other keys:

herdr pane send-keys 1-1 Enter

pane run sends the text and then a real Enter key in one request:

herdr pane run 1-1 "echo hello"

workspace management

create a new workspace:

herdr workspace create --cwd /path/to/project

without --label, the new workspace keeps the default cwd-based name.

create and name one in one step:

herdr workspace create --cwd /path/to/project --label "api server"

create one without focusing it:

herdr workspace create --no-focus

focus a workspace:

herdr workspace focus 2

rename:

herdr workspace rename 1 "api server"

close:

herdr workspace close 2

close a pane

herdr pane close 1-3

recipes

run a server and wait until it is ready

NEW_PANE=$(herdr pane split 1-2 --direction right --no-focus | python3 -c 'import sys,json; print(json.load(sys.stdin)["result"]["pane"]["pane_id"])')
herdr pane run "$NEW_PANE" "npm run dev"
herdr wait output "$NEW_PANE" --match "ready" --timeout 30000
herdr pane read "$NEW_PANE" --source recent --lines 20

run tests in a separate pane and inspect the result

herdr pane split 1-2 --direction down --no-focus
herdr pane run 1-3 "cargo test"
herdr wait output 1-3 --match "test result" --timeout 60000
herdr pane read 1-3 --source recent --lines 30

check what another agent is working on

herdr pane list
herdr pane read 1-1 --source recent --lines 80

watch another pane robustly

use this pattern when you need to coordinate with a sibling pane:

# inspect what is already there
herdr pane read 1-3 --source recent --lines 40

# wait only for the next output you expect
herdr wait output 1-3 --match "ready" --timeout 30000

# if you need to inspect the same transcript the waiter matched,
# read the unwrapped recent text directly
herdr pane read 1-3 --source recent-unwrapped --lines 40

spawn a new agent and give it a task

herdr pane split 1-2 --direction right --no-focus
herdr pane run 1-3 "claude"
herdr wait output 1-3 --match "shortcuts|Welcome" --regex --timeout 30000
herdr pane run 1-3 "review the test coverage in src/api/"
herdr pane send-keys 1-3 Enter    # agent TUIs often swallow the Enter that pane run appends — see "driving agent panes reliably"
herdr pane list                    # confirm the pane's agent_status is now "working"

coordinate with another agent

herdr wait agent-status 1-1 --status done --timeout 120000
herdr pane read 1-1 --source recent --lines 100

driving agent panes reliably

sends are fire-and-forget: send-text, send-keys, and run print nothing on success and give no delivery guarantee. when a send matters, verify it landed — read the pane or check agent_status. three failure modes recur in real orchestration:

1. agent TUIs swallow the Enter that pane run appends. interactive agent REPLs (claude code and similar) treat rapidly-injected text as a paste; the trailing Enter gets absorbed into the composer instead of submitting, and the prompt sits there while the agent stays idle. after sending a prompt to an agent pane, always follow with an explicit Enter, then confirm the agent actually started:

herdr pane run 1-3 "review the test coverage in src/api/"
herdr pane send-keys 1-3 Enter
herdr pane list   # this pane's agent_status must now be "working", not "idle"

if it is still idle, read the pane: the prompt is probably sitting un-submitted in the composer — send Enter again rather than re-sending the text (re-sending duplicates the prompt).

2. long compound strings can arrive clipped. a long cd /path && long-command sent with pane run can arrive as just the trailing command, silently running in the wrong cwd. send cd as its own run, verify the prompt line shows the new cwd, then send the command. for genuinely long invocations (agent CLIs with many flags), write a small launcher script and send bash path/to/launch.sh instead of the flag soup.

3. freshly created panes lose early keystrokes. a new pane’s shell takes a moment to initialize; text sent immediately after split / tab create can be partially eaten or echoed without executing. wait for the shell prompt to render (wait output, or read the pane) before the first run.

4. approval dialogs stall agents silently. even a generous pre-approved tool list will not cover everything: commands with shell expansions ($(…), variables), compound chains, or novel subcommands still raise an approval dialog, and the agent sits blocked until someone answers — a mission can stall for an hour on one unanswered “Do you want to proceed?”. treat blocked as a first-class state you own as orchestrator:

# auto-approve permission dialogs on agent panes (poll-loop core)
txt=$(herdr pane read "$pane" --source visible --lines 25)
if echo "$txt" | grep -q "Do you want to proceed?"; then
  if echo "$txt" | grep -q "ask again"; then herdr pane send-keys "$pane" 2
  else herdr pane send-keys "$pane" 1; fi
fi

orchestrating several agents (the loop that works)

  1. create panes; per pane: cd (verify cwd in the prompt line) → launch the agent via a launcher script → wait for the REPL to render.
  2. send the task prompt → send-keys Enter → confirm agent_status is working.
  3. watch with wait agent-status <pane> --status done (long timeout). done means finished and not yet looked at; reading the pane acknowledges it. watch blocked in parallel (or run an auto-approver poll loop — see failure mode 4) so a permission dialog never stalls the mission.
  4. on wake: read the pane tail or the agent’s report file, verify its claims yourself (run the tests), then continue.

when several agents work one repo, give them strict file ownership and a shared handoff file on disk — pane text is for driving, files are for coordination.

notes