| name | prowl-cli |
| description | Use the Prowl CLI (`prowl`) to inspect or control a running Prowl GUI app and the agent sessions it hosts. Prowl runs several coding agents in parallel, each in its own pane/tab/worktree, so reach for this whenever the user wants to act on a pane other than the current one — check on, coordinate, read from, focus, send text or keys to, open, or close another pane, tab, worktree, split, window, or sibling/neighboring agent. Covers colloquial framings that never say "prowl": "check what the agent in my other window is doing", "are any of my agents running side by side still working or idle?", "tell the agent in my left split to rerun the tests", "send npm run build to the build tab and grab the output", "open ~/proj in a fresh tab", "close that scratch tab I left open". Not for ordinary editing or building inside the Prowl source repo, and not for how-to questions about Prowl's settings, preferences, or keybindings — only when the task is to actually drive panes in the live Prowl app. |
| metadata | {"prowl-summary":"Lets an agent drive the running Prowl app through the prowl command — list panes, read or message sibling agents, send text and keys, create or close tabs and panes, launch profiles, and wait for a task to finish. Link it into a runtime's skill folder so the agent knows when and how to use prowl."} |
Prowl CLI
Use prowl only when the task is to inspect or control the running Prowl GUI app: read panes, check sibling agents, focus a pane, open a repo/path in Prowl, send text, send keys, or create/close panes and tabs. Do not use it merely because the current shell is inside the Prowl repo.
The authoritative per-command reference is Prowl's manual, components/cli.md under the docs folder. That folder is docs/ in a Prowl source checkout; otherwise it ships inside the app bundle, which you can locate from the installed CLI (normally /Applications/Prowl.app/Contents/Resources/docs):
prowl_docs="$(dirname "$(dirname "$(readlink -f "$(command -v prowl)")")")/docs"
ls "$prowl_docs/components/"
Other docs/components/*.md references below live in that same folder.
This skill ships inside the app: prowl skills install prowl-cli links it into every detected agent skill folder (~/.claude/skills, ~/.codex/skills, ~/.agents/skills — whichever parent directories exist; add --target <claude|codex|agents> to pick or create specific ones) so it stays current across app updates. prowl skills list shows per-target status; the command is local and needs no running app.
Who You Are
Every Prowl pane exports its own identity to the processes inside it:
PROWL_PANE_ID — this pane's UUID, identical to pane.id in prowl list --json.
PROWL_WORKTREE_PATH, PROWL_ROOT_PATH — this pane's worktree directory and repository root.
Use $PROWL_PANE_ID as your own selector and as the guard against operating on yourself; resolve your tab and worktree from it when you need them:
me="$(prowl list --json | jq -c --arg p "$PROWL_PANE_ID" '.data.items[] | select(.pane.id == $p)')"
if [ -z "$me" ]; then
echo "no pane matches PROWL_PANE_ID=[$PROWL_PANE_ID] — unset, or prowl reached another Prowl instance; stop, do not guess" >&2
else
printf '%s\n' "$me" | jq -r '.tab.id, .worktree.id, .worktree.name, .worktree.path'
fi
Gate every action on a target behind that lookup result ($me), never behind the bare variable — a stale id that points at another instance would otherwise pass — and keep the dependent commands inside the branch, because a bare predicate line does not stop an interactive shell:
if [ -z "$me" ] || [ "$pane" = "$PROWL_PANE_ID" ]; then
echo "refusing: self identity is unverified, or \$pane is me" >&2
else
prowl send --pane "$pane" 'git status --short' --capture --timeout 30 --json
fi
The variable is inherited, not verified: it is missing after sudo/ssh/containers and can name the wrong pane inside a tmux/screen session attached from elsewhere. A set value that matches no pane.id usually means prowl is talking to a different Prowl instance than the one hosting your pane (two apps running; see PROWL_CLI_SOCKET under Pitfalls). A match only proves that pane exists, not that you are running in it: trust the value only when your process ancestry reaches the pane's shell (no tmux/screen server or detached wrapper in between); under tmux/screen or a detached wrapper, identify your pane by other means — prowl agents --json for the pane hosting your own agent session, or a unique pane.cwd — and pass it explicitly. If it is unset or matches nothing, stop rather than guess: pane.cwd only narrows the candidates — several panes usually share one cwd — and may stand in for you only when the match is unique. Never assume the focused pane is you — open and focus move focus, and the user may be looking anywhere.
Safe Default Workflow
Resolve a concrete pane before read, send, key, focus, or close, and pass it explicitly:
prowl list --json
prowl agents --json
prowl read --pane "$pane" --last 80 --wait-stable --json
prowl send --pane "$pane" 'printf "PWD:%s\n" "$PWD"' --capture --timeout 30 --json
prowl key --pane "$pane" enter --json
prowl focus --pane "$pane" --json
Pick targets by pane.id, tab.id, worktree.id/name/path, and pane.cwd. Never trust tab titles: they are free-form and can lag or lie. For a Git worktree, worktree.name is the checked-out branch and stops matching after a checkout — automation should pass worktree.id (the path). Text prowl list / prowl agents also print short handles (p7, t6) that work in any target position for the life of the app process (read p7, close t6). UUIDs are the canonical identity of a live pane or tab, not a durable one: after an app restart restored tabs keep their tab UUID but panes are new surfaces with new UUIDs — never cache handles or pane UUIDs across a restart; re-run prowl list.
For a currently active Codex or Claude Code agent, prowl agents read p7 --json returns an immediate semantic snapshot: .data.agent.status, .data.blocker.text when blocked, and .data.result — a result is trustworthy only when .data.result.state == "complete".
Report an event only for the pane running the current process; agents signal has no target or focus fallback:
prowl agents signal turn-ended --detail "Review complete" --json
prowl agents signal needs-input --session session-1 --json
The app attributes the socket peer PID through process ancestry. turn-ended means a runtime turn edge, not task completion; only an agents dispatch-complete receipt proves an assigned task finished. Public --origin is claimed metadata and never upgrades trust. Check .data.signal.binding on the receipt: current means the signal is wait/dispatch evidence; unbound (with a signal_unbound entry in .data.warnings[]) means it was only recorded as diagnostics — the pane has no detected agent yet, the caller is outside that agent's process tree, or --session names a different session than the one Prowl knows. Report distinct events from distinct turns: waits see the latest terminal signal, so needs-input followed within 200 ms by turn-ended leaves only the turn-ended.
Common Recipes
Open a split beside yourself (or any positively identified anchor) and capture the new pane:
pane="$(prowl create pane "$PROWL_PANE_ID" --direction right --json | jq -r '.data.target.pane.id')"
Directions are right, left, up, down; the anchor must be a pane UUID or current pN, and .data.anchor.pane.id echoes it. The new pane inherits the anchor's working directory, becomes focused, and Prowl selects its worktree and tab (as create tab does). Without --profile there is no --background, so the split always takes focus and keystrokes a person is typing at that moment land in it — while someone is working in the app, prefer a Profile launch with --background. Run input afterwards with an explicit prowl send --pane "$pane" ….
Launch a reviewer beside yourself after the identity guard in Who You Are has verified $me:
launch="$(prowl create pane "$PROWL_PANE_ID" --direction right --profile Reviewer --prompt - --json <<'EOF'
Review the current branch against its base. Report only actionable findings with file and line references.
EOF
)"
pane="$(printf '%s\n' "$launch" | jq -r '.data.target.pane.id')"
dispatch="$(printf '%s\n' "$launch" | jq -r '.data.dispatch.id')"
if result="$(prowl agents wait --dispatch "$dispatch" --include-screen 40 --json)"; then
printf '%s\n' "$result" | jq -r '.data.receipt.summary, .data.target.pane.id'
else
printf '%s\n' "$result" | jq '.error.code, .error.details'
fi
The returned pane is the launched agent; .data.launch records the resolved Profile and
.data.dispatch is the exact assignment receipt. --prompt - requires a pipe or heredoc
(never interactive stdin); Prowl carries up to 256 KiB outside initial PTY input through a
command portable across zsh, bash, and fish. Put larger requirement sets in a repository file
and prompt the Profile to read it. Add --background when the split must not change focus or
select a hidden anchor's tab/worktree.
Only a succeeded dispatch receipt proves that prompted assignment completed. The receipt may
arrive before the TUI paints its final response — and, for Codex, a second or two before its
own turn-ended, so prowl agents / agents read right after a receipt can still report
working / pending; if the next action sends another prompt to the same pane, wait for an
idle condition or read a stable screen first.
Reuse one reviewer across rounds instead of launching a fresh Profile per round: launch once,
then hand each later round to the same pane with agents dispatch, which keeps the reviewer's
context and still returns an exact receipt per round.
launch="$(prowl create pane "$PROWL_PANE_ID" --direction right --profile Reviewer --prompt - --json <<'EOF'
Round 1: review the current branch against main. Write findings to /tmp/review-1.md.
EOF
)"
pane="$(printf '%s\n' "$launch" | jq -r '.data.target.pane.id')"
dispatch="$(printf '%s\n' "$launch" | jq -r '.data.dispatch.id')"
prowl agents wait --dispatch "$dispatch" --json | jq -r '.data.receipt.summary'
prowl agents wait "$pane" --until idle --timeout 30 --json >/dev/null
round="$(prowl agents dispatch "$pane" --prompt - --json <<'EOF'
Round 2: re-review the diff against main. Report only findings not already fixed; write them to /tmp/review-2.md.
EOF
)"
dispatch="$(printf '%s\n' "$round" | jq -r '.data.dispatch.id')"
prowl agents wait --dispatch "$dispatch" --json | jq -r '.data.receipt.summary'
agents dispatch needs a pane whose detected agent is idle: it applies the same evidence
rules as --until idle (a corroborated turn-ended resolves at once; a detector-only idle
view must hold for two seconds) and refuses a working or blocked agent with
DISPATCH_TARGET_BUSY instead of merging text into the running turn — hence the --until idle
between rounds, which also covers Codex's late turn-ended. When that wait resolved on a fresh
runtime turn-ended a moment before the screen caught up, the dispatch waits up to five
seconds for the screen to agree before refusing, so the two commands work back to back. A pane
holds at most one pending dispatch: a second dispatch fails with DISPATCH_PENDING (the
record is in .error.details.record) until you wait for or dispatch-abandon the previous
one. The prompt is piped stdin (multi-line is fine; it arrives as one message), and the
reviewer completes with the usual agents dispatch-complete — from its own pane, no id needed.
Run a workflow bundle instead of scripting the rounds yourself (workflow list shows what is
visible to your worktree; workflow validate <bundle.pwlworkflow> checks a new one locally):
run="$(prowl workflow run review --role reviewer=Codex --input max_rounds=3 --json)"
printf '%s\n' "$run" | jq -r '.data.id, .data.status.state, .data.self_initiated.line'
A workflow with a current role runs from the pane you call it in (that pane becomes the
role; SOURCE_REQUIRED outside a pane, PANE_BUSY when it already belongs to a run). Launch
roles freeze a Profile before anything happens: --role <role>=<profile name|UUID|auto>, else
the remembered, suggested, or Recommended one (PROFILE_NOT_FOUND when nothing qualifies);
pick roles need --role <role>=pN. When the first step messages your own pane, nothing is
typed: .data.self_initiated.line is the instruction to follow now, and its completion command
delivers your output. Every step Prowl types into a pane ends with the exact command that
completes it — run it with the output on stdin when your work for that step is fully done:
PROWL_WORKFLOW_TOKEN=… prowl workflow deliver - <<'EOF'
…
EOF
prowl workflow status --json | jq '.data.activation'
deliver answers after the output is saved: .data.delivery.state is delivered, or
provisional when the body had issues a non-strict step tolerates (.data.delivery.warnings[])
— then the run waits for the user, not for another deliver. A pane whose step moved on gets
STEP_NOT_EXPECTING; a wrong token TOKEN_INVALID; agents dispatch-complete in a workflow
pane is refused with WORKFLOW_DELIVERY_REQUIRED and the command to run instead. prowl workflow cancel <run-id> ends a run and keeps every pane and output; status <run-id> reads a
run even after an app restart (.data.source: record).
Create a fresh tab in a listed worktree:
pane="$(prowl create tab "$worktree" --json | jq -r '.data.target.pane.id')"
Pass the worktree.id (its path) from prowl list --json: worktree.name is the checked-out branch for Git worktrees and stops matching after a checkout, and a hand-typed path is easy to get wrong. --path only sets the new tab's working directory inside that worktree. prowl open /path is navigation — it may reuse an existing pane — so use create tab/create pane when you need a guaranteed new shell.
Run a command and capture its output and exit code:
out="$(prowl send --pane "$pane" 'git status --short' --capture --timeout 30 --json)"
printf '%s\n' "$out" | jq -r '.data.capture.text, .data.wait.exit_code'
Deliver input without waiting, or pre-fill and submit later:
prowl send --pane "$pane" 'long-running command' --no-wait --json
prowl send --pane "$pane" 'echo ready' --no-enter --no-wait --json && prowl key --pane "$pane" enter --json
Send multiline input from stdin:
printf '%s\n' 'echo first' 'echo second' | prowl send --pane "$pane" --capture --timeout 30 --json
Close what you created:
prowl close "$pane" --json
prowl close --tab "$tab" --force --json
close requires an explicit pane or tab and has no focus or worktree fallback.
Parsing JSON Output
Every --json response is { "ok", "command", "schema_version", "data": {...} }; failures are { "ok": false, "command", "schema_version", "error": { "code", "message", "details"? } }. Wait failures use governed .error.details for the retained dispatch record or last condition evidence. Parser errors (bad flags) print plain text even with --json, so check the exit code before piping into jq. When JSON sits in a shell variable, use printf '%s\n' "$json" | jq … — zsh echo can turn \u001B escapes back into control characters. Pass shell values into jq with --arg.
Key fields by command:
list → .data.items[] with .worktree.{id,name,path,root_path,kind}, .tab.{id,title,selected}, .pane.{id,title,cwd,focused,agent}, .task.status (running|idle|null).
agents → .data.agents[] with .status, .raw_state, .detection_reason, .type, .name, .pane.{id,focused,cwd}, .tab, .worktree, .project.{name,branch,path}.
agents read → .data.agent, .data.blocker.text, .data.result.{state,text} — pending, unavailable, missing, incomplete, too_large carry no partial text; pending is returned whenever the agent is working or blocked, even if an earlier turn completed.
agents signal → .data.pane.{id,worktree_id}, .data.signal.{event,source,confidence,binding,at,session_id,detail,claimed_origin}, optional .data.warnings[] (code=signal_unbound); optional fields are omitted.
agents dispatch → .data.target and the new pending .data.dispatch.{id,state,created_at}; refusals carry .error.details.{target,record,observation,signals}.
agents wait --dispatch → .data.receipt, immutable .data.target, .data.signals, optional .data.screen; nonzero results retain the record and evidence under .error.details.
agents wait <pane> --until … → .data.observation.{status,raw_state,source,confidence,at,revision}, .data.signals, and optional .data.screen. source/confidence are the evidence that satisfied the condition; status/raw_state are what the screen detector saw at that moment, so satisfied by a signal may still show .
Terminal text is .data.text (read) and .data.capture.text (send) — never .content, .output, or .stdout.
Reading Agent Output
- For Codex/Claude Code,
prowl agents read beats scraping: check .data.agent.status, inspect .data.blocker.text before answering a prompt with send/key (read and write are not atomic), and only trust .data.result.text when state == "complete". The result belongs to the latest completed turn and is pending while the agent works or is blocked, so after re-prompting a pane wait for --until changed or idle before reading it. --result-only prints the raw trusted result and fails otherwise; it cannot combine with --json.
- Prowl-launched Claude Code, Codex, Copilot, Droid, Qoder, Pi, Oh My Pi, and OpenCode
Profiles may expose
verified_live channels with source=hook_<runtime>; manually typing
those runtimes does not. A managed hook
turn-ended proves only a runtime turn edge, never assigned-task completion. If Profile
creation returns managed_hook_degraded, keep the successful pane but expect honest
heuristic/cooperative fallback for that session.
- After launching an agent by hand (
prowl send --pane "$pane" 'claude' --no-wait), wait until prowl agents --json lists that pane before arming waits or sending the first prompt: a condition wait tolerates ten seconds of detector latency and then fails with AGENT_NOT_FOUND, and text typed into a runtime that is still starting can merge with the next message.
- For an unpaired or manually launched agent, use one condition wait instead of a polling loop:
result="$(prowl agents wait "$pane" --until idle --include-screen 40 --timeout 600 --json)"
printf '%s\n' "$result" | jq '.data.observation, .data.screen'
--until idle|blocked observe the current state: a signal that already existed when the wait was armed counts only if the screen detector agrees, a signal arriving afterwards counts on its own, and an already-idle agent with such a signal returns immediately. Detection-only evidence (no hook or cooperative signal, the usual case for a manually launched agent) resolves only after the state has stayed unchanged for two seconds. The same stabilized fallback applies to a Profile agent whose verified_live channel holds no terminal signal yet — a freshly launched, unprompted Profile has only reported session-start — so --until idle before the first prompt resolves with confidence: heuristic; once the channel holds a turn-ended or needs-input, that runtime evidence decides and the screen never overrides it. Give those waits a --timeout of at least a few seconds. To wait for the next turn edge (for example after sending a new prompt), use --until changed; with a verified_live hook channel it returns at the next runtime signal, not at a screen change.
Exact/high evidence can establish the requested observable condition. If
jq -e '.data.observation.confidence == "heuristic"' matches, inspect the included stable
screen and, when needed, prowl agents read "$pane" --json. A finished answer with an empty
prompt is positive evidence; a spinner/tool footer means working; a permission dialog or
explicit question means blocked. Always use task context, and never treat heuristic evidence as task completion or perform destructive follow-up from it alone. A timeout leaves the task unresolved: inspect the pane, then re-arm the wait rather than assuming completion.
DISPATCH_NEEDS_INPUT means the exact worker needs intervention. DISPATCH_INCOMPLETE means
its turn ended without the required receipt. Both retain a pending receipt, so inspect
.error.details and do not immediately re-arm the dispatch wait. When possible, arm
prowl agents wait "$pane" --until changed --timeout 120 --json before the send / key
intervention; otherwise use read --wait-stable afterward. Re-arm the strict dispatch wait
only after newer activity or screen evidence shows the intervention took effect.
- Rendered screens can truncate or fold content. When you need an agent's complete output, have the command write a file (
… > /tmp/out.txt) and read that; shell redirection avoids the agent's own sandbox prompts.
read returning fewer lines than --last with truncated: false means the pane simply has less history — do not retry. --source detection returns the exact detector input instead of the viewport; it exists for diagnosing agent-state detection (see components/agent-detection.md in the docs folder), not for everyday reading.
Targeting & Arguments
- Selectors are mutually exclusive:
--pane <uuid|pN>, --tab <uuid|tN>, --worktree <id|name|path>, or -t/--target (auto: pN, tN, then UUID, then worktree). A stale handle fails rather than falling back to a same-named worktree.
send and key positionals are count-sensitive: send 'text' and key enter go to the focused pane, send p7 'text' / key p7 enter to p7, and stdin replaces the text argument. Avoid positional targeting in automation.
send --capture waits for completion and sends Enter; it cannot combine with --no-wait or --no-enter. --capture needs shell integration (OSC 133) on the target pane.
key --repeat <1-100> repeats a token, e.g. prowl key --pane "$pane" down --repeat 10.
- Quote payloads with outer single quotes when variables should expand in the target pane:
prowl send --pane "$pane" 'printf "PWD:%s\n" "$PWD"'.
- In zsh, never name a variable
status — it is readonly.
Pitfalls
open /path may refocus an existing pane; it is not a create command.
- Focus is not stable and is not you:
open and focus change it, and the user clicks around.
send --capture captures a screen diff; multiline input may include command echo.
- The CLI talks to one socket owner. With two Prowl instances running, the default
prowl reaches whichever owns the standard socket; a manually launched dev instance and every CLI call must share the same PROWL_CLI_SOCKET=/tmp/name.sock and the CLI built with that app (./.build/debug/prowl from the same checkout, or Prowl Debug.app/Contents/Resources/prowl-cli/prowl) — the version string does not reveal a mismatch, a missing command does. Sandboxed agents must be allowed to connect to that Unix socket.
- A newer CLI talking to an older app can fail at transport level (
TRANSPORT_FAILED) — confirm the running app was built with the command.
Error Handling
APP_NOT_RUNNING: Prowl is not reachable or the socket is stale — ask before restarting the app.
SOCKET_PERMISSION_DENIED: the sandbox or filesystem blocked connect(); report a permission problem, not an app-liveness problem.
TRANSPORT_FAILED: the connection broke or the socket path is invalid (ENOTSOCK, too-long PROWL_CLI_SOCKET).
TARGET_NOT_FOUND / TARGET_NOT_UNIQUE: re-run prowl list --json and pass an explicit UUID or a current pN.
PROFILE_NOT_FOUND / PROFILE_NOT_UNIQUE: re-run prowl profiles list --json; choose an enabled Profile UUID.
NO_ACTIVE_PANE: focused-pane targeting found nothing — pass --pane. SOURCE_REQUIRED: a caller-owned command (agents signal, selector-free handoff) could not map process ancestry to a Prowl pane.
EMPTY_INPUT, INVALID_ARGUMENT, UNSUPPORTED_KEY, INVALID_REPEAT: fix the arguments (prowl <cmd> --help).
CAPTURE_UNSUPPORTED: drop --capture and use read --wait-stable or file redirection.
WAIT_TIMEOUT: inspect .error.details, then re-arm the wait if the task remains active.
AGENT_NOT_FOUND (agents wait, agents dispatch): no detected agent in the pane (a wait tolerates ten seconds; a dispatch does not) — confirm the pane with prowl agents --json. DISPATCH_NOT_FOUND: no such dispatch record, or dispatch-complete ran in a pane that never held one (records reset on app restart). DISPATCH_CONTEXT_REQUIRED: dispatch-complete ran from a process outside any Prowl pane. DISPATCH_ALREADY_TERMINAL: dispatch-abandon hit a record that already completed, was abandoned, or is gone.
DISPATCH_PENDING / DISPATCH_TARGET_BUSY (agents dispatch): the pane still holds a pending record (wait for it or abandon it), or its agent is working/blocked (wait , then retry). Nothing was typed into the pane.
Handing Off Your Task
The built-in prowl.handoff workflow is also available: it requests a briefing and saves
context before optionally launching a receiver. See the prowl-workflow skill for that
flow. The direct commands below remain available.
prowl handoff to <agent> --brief - hands your task to another agent. Run it from your own pane (the calling pane is the source — no selector needed) and pipe your briefing on stdin. Prowl finds the calling pane through process ancestry, so any descendant of the pane's shell (an agent, its tool shell) works; under tmux/screen or a detached wrapper that resolution fails with SOURCE_REQUIRED, and in exactly those setups $PROWL_PANE_ID is not trustworthy either (it names the pane the tmux server started in, which may still exist) — identify your pane by other means (prowl agents --json, a unique pane.cwd) and pass it with --pane explicitly.
prowl handoff to codex --brief - <<'EOF'
…
…
…
EOF
Required sections are ## Objective, ## Current State, and ## Next Steps; optional ones are ## What Has Been Done, ## Open Questions, ## Risks / Watch Out, and ## Suggested Prompt For Next Agent. The receiver launches in a background tab of the same worktree; your session stays open. prowl handoff save --brief - checkpoints the same briefing without launching anyone; --no-brief is for an intentional context-only handoff; --pane hands off a pane other than your own. Details: components/handoff.md in the docs folder.
Command Set
list, agents, agents read, agents signal, agents dispatch, agents dispatch-complete, agents dispatch-abandon, agents wait, profiles list, skills list|install|uninstall|path (local-only), workflow list|run|status|deliver|cancel (workflow validate|schema local-only), read, send, key, focus, create tab, create pane, close, handoff to, handoff save, and open (default). There is no CLI quit; close temporary tabs or panes with an explicit close. tab create, tab close, and pane close remain deprecated aliases for one release.