PwrAgent Dev Profile

Use this skill to manage the local PwrAgent Electron app for the dev profile from the checkout Codex is currently working in.

Commands

Run commands from the repository root unless the user asks for a different checkout.

Start or restart the dev-profile app:

.agents/skills/pwragent-dev-profile/scripts/pwragent-dev-profile.zsh restart --root "$PWD"

Close the dev-profile app if it is running:

.agents/skills/pwragent-dev-profile/scripts/pwragent-dev-profile.zsh close --root "$PWD"

Check whether the dev-profile app is running:

.agents/skills/pwragent-dev-profile/scripts/pwragent-dev-profile.zsh status --root "$PWD"

List the profile's recorded app instances and messaging lease owner:

.agents/skills/pwragent-dev-profile/scripts/pwragent-dev-profile.zsh leases

Verify that a previously started instance is still up:

.agents/skills/pwragent-dev-profile/scripts/pwragent-dev-profile.zsh verify --root "$PWD"

Workflow

1. Run status first when the user asks what is running or when closing processes may be surprising.
2. Run restart when the user asks to start the dev profile; it closes the prior managed dev-profile instance, starts PWRAGENT_PROFILE=dev PWRAGENT_INSTANCE_ROOT="$PWD" pnpm dev detached from the checkout, and waits until the app writes a matching profile runtime record.
3. For visual QA, use the emitted Computer Use target line. Target its exact checkout-local appPath, then verify the native window title is PwrAgent and the AX URL matches the emitted rendererUrl. Never target the generic Electron display name, shared com.github.Electron bundle id, or installed com.pwrdrvr.pwragent app: those can select a sibling project, another checkout, or a packaged build that does not contain the code under test.
4. Run close when the user only wants the dev-profile app stopped. Never kill the Electron PID by hand — see "Closing the app" below.
5. Relay the script output and the log path to the user. The default log is .local/pwragent-dev-profile.log.

Closing the app

close signals the app's own process, waits for it to actually exit (polling until gone or 30 seconds pass), and only then cleans up the pnpm dev supervisor chain. Do not replace the wait with a fixed sleep, do not kill the Electron PID yourself, and do not collapse the two phases back into one signal to every matched pid — SIGTERMing the supervisor or a renderer alongside the main process cuts short the very drain the wait exists to protect.

A SIGTERM'd app does not exit promptly. Note what SIGTERM does not do: installProcessShutdownHandlers in index.ts calls allowImmediateQuit(), so the signal path deliberately skips the quit-confirmation dialog and its 10-second countdown. That countdown belongs to Cmd-Q and window close, not to this script.

What the wait is for is the phased drain that follows. disposeMainProcessResources runs the before-quit phases under a 12-second global budget (MAIN_PROCESS_SHUTDOWN_TIMEOUT_MS), and the app-server phase alone may take 7.5s, with messaging and federation at 4s each and renderer teardown adding 2s. The messaging and federation runtimes are stopped in those phases, so a SIGKILL landing mid-drain strands the profile's leases: the next instance boots with federation off, names the dead PID as the lease holder in Settings → Federation, and shows zero peers until the dead-owner grace expires. Any federation-dependent UI then renders its empty state and looks broken when it is not — observed on 2026-08-12 after a hand-rolled kill that did not wait.

If the 30-second budget is exhausted, close escalates to SIGKILL but says loudly that it did, and warns that the next launch may report federation off.

Script Notes

• The script defaults to --profile dev, --root "$PWD", .local/pwragent-dev-profile.pid, and .local/pwragent-dev-profile.log.
• Prefer restart over hand-running PWRAGENT_PROFILE=dev pnpm dev; the script passes PWRAGENT_INSTANCE_ROOT, starts a detached daemon helper, then uses the app's lease-backed runtime metadata in ~/.pwragent/profiles/dev/state/state.db to find the Electron owner process.
• The detached daemon helper stops the pnpm dev supervisor when the first lease-backed Electron instance it started exits, so closing the spawned app does not leave a dev supervisor relaunching it.
• The daemon clears inherited ELECTRON_EXEC_PATH, ELECTRON_CLI_ARGS, and ELECTRON_MAJOR_VER before launching. An agent hosted by another Pwr-family Electron app can otherwise silently start this checkout through the host app's Electron installation.
• The script only targets the app instance whose recorded root hash matches the requested checkout, plus that instance's bounded pnpm dev / electron-vite parent chain.
• Use leases when debugging which process owns the profile messaging lease.
• If verification fails, inspect the last log lines printed by the script before retrying.