@dumbmice/pi-mailbridge 0.0.0-stage → 0.1.1

package/src/extension.ts

@@ -0,0 +1,1381 @@
+// pi-mailbridge — pi extension entry.
+//
+// Everything here runs INSIDE the agent process. The mailbox is not touched: the
+// daemon owns IMAP, and this file talks to it over loopback.
+//
+// Dependencies are injectable so the registration surface (which events, which
+// commands, which tools) and the settle behaviour can be asserted with a fake
+// `pi`, no daemon and no mailbox.
+
+import { statSync } from "node:fs"
+import { Type } from "@sinclair/typebox"
+
+import { loadConfig, hasUsableMailConfig, type ResolvedConfig } from "./core/config.ts"
+import { shouldReloadConfig } from "./core/config-freshness.ts"
+import { configPath } from "./store/paths.ts"
+import { ensureSecret, readSecret } from "./store/secret.ts"
+import { formatScan, type ScanSummary } from "./core/status.ts"
+import { createSettleNotifier, type SettleContext, type SettleNotifier } from "./core/settle.ts"
+import { createDeliveryRunner, type DeliveryRunner } from "./core/deliveries.ts"
+import { toUserContent, type UserContent } from "./core/attachments.ts"
+import { parseMailCommand } from "./core/mail-command.ts"
+import { API_VERSION, apiCompatible } from "./daemon/api.ts"
+import { probeHealth, probeHealthDetail } from "./daemon/spawn.ts"
+import { resolveDaemonScript } from "./daemon/script.ts"
+import { daemonBaseUrl, ensureDaemon, createDaemonClient, type DaemonClient } from "./daemon/spawn.ts"
+import { parseMailbridgeCommand, USAGE } from "./core/commands.ts"
+import { runSetup } from "./core/setup.ts"
+import { runSmoke } from "./core/smoke.ts"
+import { buildTokenProvider } from "./core/token-source.ts"
+import { createOpenidClient } from "./core/oauth-openid.ts"
+import { createDeviceCodeProvider } from "./core/oauth-device.ts"
+import { fileTokenStore } from "./store/oauth-store.ts"
+import { ensureRouteForSession, registerRoute } from "./store/routes.ts"
+import { existsSync } from "node:fs"
+import { fileURLToPath } from "node:url"
+
+const POLL_MS = 10_000
+/** How often the poll may try to start a dead daemon. */
+const RESPAWN_BACKOFF_MS = 60_000
+/** How often to ask a healthy daemon which build it is. */
+const VERSION_CHECK_MS = 60_000
+
+/** The slice of the pi runtime this extension uses. */
+export interface ExtensionLike {
+	on(event: string, handler: (event: Record<string, unknown>, ctx: ExtensionCtxLike) => unknown): void
+	registerCommand(name: string, spec: { description: string; handler: (args: string, ctx: ExtensionCtxLike) => Promise<void> }): void
+	registerTool(spec: {
+		name: string
+		label: string
+		description: string
+		parameters: unknown
+		execute: (
+			toolCallId: string,
+			params: Record<string, unknown>,
+			signal: unknown,
+			onUpdate: unknown,
+			ctx: ExtensionCtxLike,
+		) => Promise<{ content: { type: "text"; text: string }[]; details: Record<string, unknown> }>
+	}): void
+	/** Used to answer a reply inside the session. */
+	sendUserMessage(
+		content: string | UserContent[],
+		options?: { deliverAs?: "steer" | "followUp"; expandPromptTemplates?: boolean },
+	): void
+}
+
+export interface ExtensionCtxLike {
+	cwd: string
+	sessionManager: {
+		getSessionId(): string
+		/** The real API. `getName` was a mistake here and always returned undefined. */
+		getSessionName?(): string | undefined
+		getName?(): string | undefined
+		/** Current branch entries; assistant text and turn count come from here. */
+		getBranch?(): unknown[]
+	}
+	ui: { notify(message: string, level?: "info" | "warning" | "error"): void }
+}
+
+export interface SettleTimers {
+	setTimeout(handler: () => void, ms: number): unknown
+	clearTimeout(handle: unknown): void
+}
+
+export interface MailbridgeDeps {
+	env?: NodeJS.ProcessEnv
+	loadConfig?: () => ResolvedConfig
+	secret?: () => Buffer | undefined
+	/** Build a daemon client for a base URL and bearer token. */
+	makeClient?: (baseUrl: string, token: string) => DaemonClient
+	/** Make sure the daemon is up. */
+	ensureDaemon?: (
+		baseUrl: string,
+		scriptPath: string,
+	) => Promise<
+		boolean | { healthy: boolean; started?: boolean; incompatible?: boolean; error?: string; apiVersion?: number }
+	>
+	/** Injected by tests: is something answering on this base URL? */
+	probeHealth?: (options: { baseUrl: string }) => Promise<boolean>
+	/** As above, but reporting the protocol that answered. */
+	probeHealthDetail?: (options: { baseUrl: string }) => Promise<{ ok: boolean; apiVersion?: number }>
+	/** How often to re-check the daemon's protocol. Defaults to a minute. */
+	versionCheckMs?: number
+	/** How often the poll may respawn a dead daemon. Defaults to a minute. */
+	respawnBackoffMs?: number
+	/** How long to wait for a stopped daemon's port to close. Defaults to five seconds. */
+	stopWaitMs?: number
+	/** How long a command waits for a tick it interrupted. Defaults to eight seconds. */
+	joinTickMs?: number
+	timers?: SettleTimers
+	/** Poll interval for pending replies. */
+	pollMs?: number
+}
+
+const defaultTimers: SettleTimers = {
+	setTimeout: (handler, ms) => setTimeout(handler, ms),
+	clearTimeout: (handle) => clearTimeout(handle as ReturnType<typeof setTimeout>),
+}
+
+export default function mailbridge(pi: ExtensionLike, deps: MailbridgeDeps = {}): void {
+	const env = deps.env ?? process.env
+	const load = deps.loadConfig ?? (() => loadConfig(env))
+	let config: ResolvedConfig = load()
+	/** The address the current client was built with; may lag `config` after a port edit. */
+	let clientBase: string | undefined
+	const secretFn = deps.secret ?? (() => readSecret(env))
+	const timers = deps.timers ?? defaultTimers
+	const pollMs = deps.pollMs ?? POLL_MS
+
+	let ctx: ExtensionCtxLike | undefined
+	/** Set by stopPolling(): a tick already in flight must not touch the daemon. */
+	let pollPaused = false
+	/** The tick that may be inside an await right now; see trackTick/settleTick. */
+	let tickInFlight: Promise<void> | undefined
+	let client: DaemonClient | undefined
+	let runner: DeliveryRunner | undefined
+	let pollHandle: unknown
+	let notifier: SettleNotifier | undefined
+	/** Mirror of the daemon's global mode. Default off: never mail unprompted. */
+	let emailMode: "on" | "off" = "off"
+	/** Stable id for this pi process, used for claims and for liveness pings. */
+	const instanceId = `pi-${process.pid}`
+	/** Throttle for respawning a daemon that exited for idleness. */
+	let lastRespawnAttempt = 0
+	/** Throttles the post-update daemon replacement; a failed respawn must not loop. */
+	let lastVersionCheck = 0
+	/** Set when the daemon rejects our bearer: said once, not every poll tick. */
+	let tokenRejected = false
+	/** Set when a respawn still spoke another protocol: only a new pi session can fix that. */
+	let protocolStuck = false
+
+	// Computed per use, not once: `config` is replaced whenever a /mailbridge command
+	// reloads it, and a frozen URL means editing daemon.port leaves every probe (including
+	// /mailbridge restart, the documented remedy) dialling the old port forever.
+	const currentBaseUrl = (): string => daemonBaseUrl(config.daemon.port, config.daemon.host)
+
+	/**
+	 * Last assistant text and turn count from the session branch.
+	 *
+	 * Without this the notification body is just a placeholder — and the point of the
+	 * email is to read the answer from whatever device you happen to be on. Branch
+	 * entries are `{ type: "message", message: { role, content } }`; assistant
+	 * content is `(TextContent | ThinkingContent | ToolCall)[]`, so only `text`
+	 * parts are joined (thinking and tool calls are not the answer).
+	 */
+	function readSessionSnapshot(manager: ExtensionCtxLike["sessionManager"]): {
+		lastAssistantText?: string
+		turns?: number
+	} {
+		const branch = manager.getBranch?.() ?? []
+		let lastAssistantText: string | undefined
+		let turns = 0
+
+		for (let i = branch.length - 1; i >= 0; i--) {
+			const entry = branch[i] as { type?: string; message?: { role?: string; content?: unknown } } | undefined
+			if (entry?.type !== "message" || entry.message?.role !== "assistant") continue
+			turns++
+			if (lastAssistantText) continue
+
+			const content = entry.message.content
+			if (typeof content === "string" && content.trim()) {
+				lastAssistantText = content.trim()
+				continue
+			}
+			if (!Array.isArray(content)) continue
+			const text = content
+				.filter((part): part is { type: "text"; text: string } => {
+					const p = part as { type?: unknown; text?: unknown }
+					return p?.type === "text" && typeof p.text === "string"
+				})
+				.map((part) => part.text)
+				.join("\n")
+				.trim()
+			if (text) lastAssistantText = text
+		}
+
+		return { ...(lastAssistantText ? { lastAssistantText } : {}), ...(turns > 0 ? { turns } : {}) }
+	}
+
+	function sessionContext(): SettleContext | undefined {
+		if (!ctx) return undefined
+		const sessionId = ctx.sessionManager.getSessionId()
+		// `getSessionName()` is the documented method; `getName` never existed.
+		const name = ctx.sessionManager.getSessionName?.() ?? ctx.sessionManager.getName?.()
+		return {
+			sessionId,
+			...(name ? { sessionName: name } : {}),
+			device: config.device.name,
+			cwd: ctx.cwd,
+			...readSessionSnapshot(ctx.sessionManager),
+		}
+	}
+
+	/**
+	 * Park the loop.
+	 *
+	 * Clearing the timer is not enough on its own: a tick that entered just before this call is
+	 * already inside `tick()`, and it would find the daemon this command is stopping, respawn it,
+	 * and manufacture the "still answering" evidence the command is measuring. The flag makes a
+	 * tick in flight a no-op.
+	 */
+	function stopPolling(): void {
+		pollPaused = true
+		if (pollHandle !== undefined) {
+			timers.clearTimeout(pollHandle)
+			pollHandle = undefined
+		}
+	}
+
+	function startPolling(): void {
+		stopPolling()
+		pollPaused = false
+		schedulePoll()
+	}
+
+	/**
+	 * Start a tick, and remember it while it runs.
+	 *
+	 * A pause cannot stop a tick that is already inside an await - and one of those awaits spawns
+	 * a daemon. The command that parks the loop waits for the tick it interrupted instead, so the
+	 * daemon that tick started is on the port before the command looks at it.
+	 */
+	function trackTick(running: Promise<void>): void {
+		tickInFlight = running
+		// `.then(clear, clear)`: a rejecting tick must not surface as an unhandled rejection.
+		void running.then(
+			() => {
+				if (tickInFlight === running) tickInFlight = undefined
+			},
+			() => {
+				if (tickInFlight === running) tickInFlight = undefined
+			},
+		)
+	}
+
+	/**
+	 * Wait for an interrupted tick, bounded. Returns true when the budget ran out first.
+	 *
+	 * The budget has to exceed what a tick can spend inside `ensureDaemon` (a spawn, then probes
+	 * for `waitMs`), or the join expires exactly in the window it exists to cover. The timer is
+	 * cleared and uses the injected clock: a leaked five-second timer per command would keep the
+	 * process alive for nothing.
+	 */
+	async function settleTick(): Promise<boolean> {
+		const running = tickInFlight
+		if (running === undefined) return false
+		// A real timer, not the injected (fake) clock: it bounds a wait on another promise, so
+		// nothing else can resolve it. Cleared below, so it never outlives the command.
+		let handle: ReturnType<typeof setTimeout> | undefined
+		const expired = new Promise<boolean>((resolve) => {
+			handle = setTimeout(() => resolve(true), deps.joinTickMs ?? 8_000)
+		})
+		try {
+			return await Promise.race([running.then(() => false), expired])
+		} finally {
+			if (handle !== undefined) clearTimeout(handle)
+		}
+	}
+
+	/**
+	 * Arm the timer without touching `pollPaused`.
+	 *
+	 * For callers inside an await - `replaceOutdatedDaemon` runs within a command that parked the
+	 * loop, and clearing the pause there would un-park it from underneath.
+	 */
+	function schedulePoll(): void {
+		const tick = async () => {
+			if (pollPaused) return
+			try {
+				// Heartbeat: the daemon exits when every session stops pinging, so a quiet
+				// daemon is a stopping daemon.
+				const ping = client ? await client.ping(instanceId) : undefined
+				// Checked again after the await: a tick that entered before `stopPolling()` was
+				// suspended right here, and with the client nulled by the command it would read as
+				// "dead daemon" and respawn inside the wait that is measuring the port.
+				if (pollPaused) return
+				if (!client || (ping && !ping.ok)) {
+					// No client is not only "dead daemon": it can also be "a daemon on another
+					// protocol that this session never managed to replace". Only the no-client case
+					// is checked here - with a live client, a failed ping means it is gone, and the
+					// protocol is the else-branch's business.
+					if (!client && (await replaceOutdatedDaemon())) return
+					if (pollPaused) return
+					// The daemon exited for idleness (or crashed). Nothing else would bring it
+					// back for this session, so the poll becomes the supervisor.
+					const now = Date.now()
+					if (now - lastRespawnAttempt > (deps.respawnBackoffMs ?? RESPAWN_BACKOFF_MS)) {
+						lastRespawnAttempt = now
+						const healthy = await ensureDaemonSafe()
+						// The spawn itself can take seconds. A paused tick that returns here has
+						// already created the daemon the restart is about to measure.
+						if (pollPaused) return
+						if (healthy) {
+							// A settle waiting on its debounce is the one mail this product exists
+							// to send. startRuntime() rebuilds the notifier, so flush the old one
+							// after the client is live again - never cancel it.
+							await replaceRuntime()
+						}
+					}
+					// Fall through to the reschedule below: returning here is what stopped the
+					// poll loop permanently after a single failed ping.
+				} else {
+					// After `pi update` the daemon can still be running the previous build: it
+					// mints nothing and resolves nothing, so every reply after the upgrade would
+					// be rejected with the cursor already past it. Replace it once the build
+					// differs, throttled so a failing respawn cannot loop.
+					if (await replaceOutdatedDaemon()) return
+					if (pollPaused) return
+					const mode = await client?.mode()
+					if (pollPaused) return
+					if (mode?.ok) emailMode = mode.data.mode === "on" ? "on" : "off"
+					// `/mailbridge off` closes the channel: no claim, no injection while it is off.
+					// Anything already queued stays pending and is delivered on the next
+					// `on`, which is why the docs say so explicitly.
+					if (emailMode === "on") await runner?.drain()
+				}
+			} catch {
+				/* drain already swallows; a poll must never throw into the runtime */
+			} finally {
+				// Re-arm whenever this session is alive - not only when a runner exists: a
+				// daemon replacement clears the runner for a moment, and tying the loop to it
+				// stopped polling exactly when the daemon was missing.
+				if (ctx && pollHandle === undefined && !pollPaused) {
+					pollHandle = timers.setTimeout(() => {
+						pollHandle = undefined
+						trackTick(tick())
+					}, pollMs)
+				}
+			}
+		}
+		// Clearing first: arming twice leaves the previous handle live and untracked, so both
+		// fire and the next re-arm adds a third. (Unreachable from the two current callers, which
+		// always run with no handle - this keeps it that way if a third appears.)
+		if (pollHandle !== undefined) {
+			timers.clearTimeout(pollHandle)
+			pollHandle = undefined
+		}
+		pollHandle = timers.setTimeout(() => {
+			pollHandle = undefined
+			trackTick(tick())
+		}, pollMs)
+	}
+
+	// ---- lifecycle ---------------------------------------------------------
+
+	/**
+	 * Is the running daemon this build? Returns true when it replaced one.
+	 *
+	 * The extension and the daemon are separate processes with separate lifetimes: `pi update`
+	 * reloads this module while the daemon keeps executing what it started with. A version it
+	 * cannot resolve routing ids against is worse than no daemon, because replies are consumed.
+	 */
+	async function replaceOutdatedDaemon(known?: { apiVersion?: number; incompatible?: boolean }): Promise<boolean> {
+		// Once a respawn has proved the mismatch is not ours to fix, stop trying.
+		if (protocolStuck) return false
+		if (Date.now() - lastVersionCheck < (deps.versionCheckMs ?? VERSION_CHECK_MS)) return false
+		lastVersionCheck = Date.now()
+		// No client yet (a fresh session that found a daemon from another install) is exactly the
+		// case this must handle: use what the spawn probe already saw, or ask the daemon directly.
+		// `undefined` means *nothing answered* - the dead-daemon path owns that, and treating it as
+		// a protocol mismatch would "replace" a daemon that is not there instead of respawning it.
+		const health = client ? await client.health() : undefined
+		let theirs = known?.incompatible ? (known.apiVersion ?? 0) : known?.apiVersion
+		if (theirs === undefined && health?.ok) theirs = health.data.apiVersion ?? 0
+		if (theirs === undefined && !client) {
+			// No client and no client to ask: ask the port. `ok:false` (nothing answered) stays
+			// undefined so the dead-daemon path keeps ownership of a daemon that is not there.
+			const found = await probeDetail({ baseUrl: clientBase ?? currentBaseUrl() })
+			if (found.ok) theirs = found.apiVersion ?? 0
+		}
+		if (theirs === undefined || apiCompatible(API_VERSION, theirs)) return false
+		// The awaits above can take seconds; a command may have parked the loop in the meantime.
+		if (pollPaused) return false
+
+		ctx?.ui.notify(
+			`mailbridge: replacing the running daemon (protocol ${theirs ?? "0"} → ${API_VERSION})`,
+			"info",
+		)
+		// To stop it we need a client, even if this session has no runtime yet.
+		const stopping = client ?? buildClient()
+		await stopping?.shutdown().catch(() => undefined)
+		client = undefined
+		runner = undefined
+
+		// `/shutdown` answers before it closes the listener, so an immediate probe finds the
+		// dying daemon and "reuses" it - the replacement then reports success while the old
+		// protocol is still being served. Wait for the port to stop answering.
+		await waitForPortToClose(clientBase ?? currentBaseUrl())
+		// Stopped, and parked: the command that parked the loop owns the respawn, and spawning
+		// here would hand it the survivor it is measuring.
+		if (pollPaused) return false
+
+		// A failed respawn must not end the poll loop: with `runner` gone, a re-arm tied to it
+		// would stop the only thing that notices the daemon is missing.
+		const ensured = await ensureDaemonDetailed()
+		if (pollPaused) return false
+		if (!ensured.healthy || !(await replaceRuntime())) {
+			ctx?.ui.notify(
+				ensured.incompatible
+					? `mailbridge: the daemon is still on protocol ${ensured.apiVersion ?? "0"} (this session: ${API_VERSION}) — /mailbridge restart forces it`
+					: `mailbridge: the daemon did not come back after the update${ensured.error ? ` (${ensured.error})` : ""} — retrying, and /mailbridge restart forces it`,
+				"warning",
+			)
+			return true
+		}
+
+		// If the respawn brought back another protocol again, this session is older than the
+		// package on disk. Replacing forever would kill the bridge once a minute, silently.
+		const after = await (client as DaemonClient | undefined)?.health()
+		if (after?.ok && !apiCompatible(API_VERSION, after.data.apiVersion)) {
+			protocolStuck = true
+			ctx?.ui.notify(
+				`mailbridge: this pi session speaks protocol ${API_VERSION} and the installed daemon speaks ` +
+					`${after.data.apiVersion ?? "0"} — restart pi to load the matching extension`,
+				"error",
+			)
+			return true
+		}
+		schedulePoll()
+		return true
+	}
+
+	const probe = deps.probeHealth ?? probeHealth
+	const probeDetail =
+		deps.probeHealthDetail ??
+		(deps.probeHealth
+			? async (options: { baseUrl: string }): Promise<{ ok: boolean; apiVersion?: number }> =>
+					// A seam that only says "something answers" cannot also say "and it is protocol 0",
+					// so it means the protocol this build speaks - otherwise every such stub looks
+					// like an outdated daemon and gets shut down.
+					({ ok: await deps.probeHealth!(options), apiVersion: API_VERSION })
+			: probeHealthDetail)
+
+	// The pi manifest points at this TypeScript file inside the installed package, so the daemon
+	// beside it must be the compiled build wherever one exists (Node refuses type stripping in
+	// node_modules). See daemon/script.ts.
+	const daemonScript = (): string => resolveDaemonScript(import.meta.url)
+
+	/**
+	 * Wait for the listener to close, bounded, and report whether something still answers.
+	 *
+	 * `/shutdown` answers before it closes: the handler releases the lock, stops the watcher and
+	 * logs out of IMAP first, so a wedged connection keeps `/health` alive well past the reply.
+	 * The return value is the *last* probe, so a caller does not ask a second time.
+	 */
+	async function waitForPortToClose(base: string): Promise<boolean> {
+		const deadline = Date.now() + (deps.stopWaitMs ?? 5_000)
+		while (Date.now() < deadline && (await probe({ baseUrl: base }))) {
+			await new Promise((resolve) => setTimeout(resolve, 100))
+		}
+		return probe({ baseUrl: base })
+	}
+
+	/**
+	 * Build the client, notifier and delivery runner. Called on session_start and again
+	 * if the poll tick finds the daemon gone, so a session recovers without a reload.
+	 */
+	/**
+	 * Rebuild the runtime, carrying a debounced settle across the gap.
+	 *
+	 * Every replacement path (dead daemon, version mismatch, /mailbridge restart) needs the same
+	 * three steps, and each used to do them itself - which is why the same dropped-notification
+	 * bug had to be fixed three times. It lives here now.
+	 */
+	async function replaceRuntime(): Promise<boolean> {
+		const stale = notifier
+		notifier = undefined
+		if (!(await startRuntime())) return false
+		await stale?.flush().catch(() => undefined)
+		return true
+	}
+
+	/** A client for the current base URL, or undefined when there is no secret yet. */
+	function buildClient(secret: Buffer | undefined = secretFn()): DaemonClient | undefined {
+		if (!secret) return undefined
+		const token = secret.toString("base64url")
+		const makeClient =
+			deps.makeClient ??
+			((base: string, tok: string) => createDaemonClient({ baseUrl: base, token: tok, device: config.device.name }))
+		return makeClient(clientBase ?? currentBaseUrl(), token)
+	}
+
+	async function startRuntime(): Promise<boolean> {
+		const current = ctx
+		if (!current) return false
+		const secret = secretFn()
+		if (!secret) return false
+
+		clientBase = currentBaseUrl()
+		client = buildClient(secret)
+		if (!client) return false
+		const active = client
+
+		const initialMode = await active.mode()
+		// A session_shutdown can land during that await. Without this the runtime is rebuilt
+		// around a dead ctx: the poll keeps pinging and the daemon's idle exit can never fire.
+		if (!ctx) {
+			client = undefined
+			return false
+		}
+		// A 401 is not "the mode is off": the daemon rejected this session's bearer, which means a
+		// rotated/removed `secret.key`. Flipping the mirror to "off" there silently stopped every
+		// notification and drain, and the poll then respawned a daemon that kept rejecting us.
+		if (!initialMode.ok && initialMode.status === 401 && !tokenRejected) {
+			tokenRejected = true
+			ctx.ui.notify(
+				"mailbridge: the daemon rejected this session's bearer token — restart pi (or /mailbridge restart) to load the current secret",
+				"error",
+			)
+		}
+		if (initialMode.ok) emailMode = initialMode.data.mode === "on" ? "on" : "off"
+
+		notifier = createSettleNotifier({
+			mode: () => emailMode,
+			context: sessionContext,
+			send: async (message) => {
+				const sent = await active.send({
+					subject: message.subject,
+					text: message.text,
+					// Exempt from the notification budget: a burst of `notify_user` mail must not
+					// be able to silence "your agent is waiting".
+					kind: "settle",
+					...(ctx ? { cwd: ctx.cwd, sessionId: ctx.sessionManager.getSessionId() } : {}),
+				})
+				// A non-ok result used to be discarded here, so a refused or timed-out send
+				// vanished: no toast, no log, no retry. Throwing routes it to onError.
+				// `sent === undefined` means there is no client at all; reporting that as a
+				// success would make "nothing was sent" indistinguishable from "sent".
+				if (!sent) throw new Error("send failed (daemon unavailable)")
+				if (!sent.ok) {
+					throw new Error(
+						sent.code === "send_budget"
+							? "notification budget exhausted (10 a minute)"
+							: `send failed (${sent.code ?? sent.status})`,
+					)
+				}
+			},
+			secret: () => secret,
+			config: { ...config.notify, tokenTtlHours: config.security.tokenTtlHours },
+			tagFor: (context) =>
+				ensureRouteForSession({
+					sessionId: context.sessionId,
+					...(context.device ? { device: context.device } : {}),
+					subject: `pi: ${context.sessionName ?? context.cwd}`,
+					ttlHours: config.security.tokenTtlHours,
+					env,
+				}),
+			timers,
+			onError: (error) => ctx?.ui.notify(`mailbridge: could not send (${error.message})`, "warning"),
+		})
+
+		runner = createDeliveryRunner({
+			client: {
+				// Pass every argument through. This adapter has silently dropped one before
+				// (uidValidity), which disabled the stale-space guard on the only production
+				// client while the store's own test stayed green.
+				claim: (uid, id, uidValidity) => active.claim(uid, id, uidValidity),
+				ack: (uid, id, sessionId, uidValidity) => active.ack(uid, id, sessionId, uidValidity),
+				pending: () => active.pending(),
+			},
+			instanceId,
+			sessionId: () => ctx?.sessionManager.getSessionId(),
+			stillDelivering: () => emailMode === "on",
+			inject: async (body, record) => {
+				// A command by mail runs the same handler the prompt would, and then there is
+				// nothing to inject: the human asked for an action, not for the agent to read a
+				// sentence. Anything we do not run here stays a message, so a quoted "/mailbridge off" or a
+				// refused "/mailbridge setup" is never executed by accident.
+				const command = parseMailCommand(body)
+				if (command) {
+					const handler = commandHandlers.get(command.name)
+					if (handler && !ctx) {
+						// The session is tearing down. Throwing keeps the record pending so the
+						// command runs in a live session; falling through would inject the
+						// command text into a dead session and then ack it, consuming the mail.
+						throw new Error("no active session for a command reply")
+					}
+					if (handler && ctx) {
+						// A command that throws must NOT propagate: deliveries treats an inject
+						// failure as "retry later" and never acks, so the same command would run
+						// again every poll tick - and `/mailbridge smoke` would send a real mail
+						// each time. The command was handled either way; say it failed and let
+						// the delivery be acked.
+						try {
+							await handler(command.args, ctx)
+						} catch (error) {
+							ctx.ui.notify(
+								`mailbridge: /${command.name} from mail failed: ${(error as Error).message}`,
+								"error",
+							)
+						}
+						return
+					}
+				}
+
+				const content = toUserContent(body, record.attachments ?? [])
+				// A reply is text from a human, never a command. With template expansion on, pi
+				// dispatches a message starting with "/" as an extension command and never gives
+				// it to the model - so a reply of "/mailbridge off" would run a command, return without a
+				// turn, and the runner would ack and \Seen it. The human would believe the agent
+				// had read it.
+				pi.sendUserMessage(content, { deliverAs: "followUp", expandPromptTemplates: false })
+			},
+			onError: (error) => ctx?.ui.notify(`mailbridge: reply not delivered (${error.message})`, "warning"),
+		})
+
+		// Ping before the first drain: a drain of several replies can outlast the idle window,
+		// and the daemon reaps itself when no ping arrives in that time. A client without ping
+		// (tests, older shape) is not a reason to skip the drain.
+		try {
+			await active.ping(instanceId)
+		} catch {
+			/* the drain below is what decides */
+		}
+		// Catch-up on connect, but only when the channel is actually open: `off` must not
+		// be bypassed by a session starting up.
+		if (emailMode === "on") await runner.drain()
+		// Same reason as above, after the drain's awaits.
+		if (!ctx) {
+			client = undefined
+			return false
+		}
+		return true
+	}
+
+	/**
+	 * Tell a pre-existing daemon to re-read its config when the file is newer than the
+	 * config it loaded. Best-effort: a failure here must not block the session.
+	 */
+	async function reloadIfStale(): Promise<void> {
+		if (!client) return
+		try {
+			const health = await client.health()
+			if (!health.ok) return
+			const loadedAt = health.data.configLoadedAt
+			const mtime = statSync(configPath(env), { throwIfNoEntry: false })?.mtimeMs ?? null
+			if (!shouldReloadConfig({ configLoadedAt: loadedAt, configMtimeMs: mtime })) return
+			const reloaded = await client.reload()
+			// Never silent: this is the path that decides which address the agent's mail goes
+			// to, so a refused reload (daemon kept the old config) must be visible.
+			if (!reloaded.ok) {
+				ctx?.ui.notify(
+					"mailbridge: the daemon could not reload its config — run /mailbridge restart",
+					"warning",
+				)
+			} else if (reloaded.data.warning) {
+				ctx?.ui.notify(`mailbridge: ${reloaded.data.warning}`, "warning")
+			}
+		} catch {
+			/* best effort: a stale config is not a reason to fail the session */
+		}
+	}
+
+	async function ensureDaemonSafe(): Promise<boolean> {
+		return (await ensureDaemonDetailed()).healthy
+	}
+
+	/**
+	 * Spawn-or-reuse, keeping the reason. `incompatible` means something answered but spoke
+	 * another protocol - the replacement path, not the dead-daemon path.
+	 */
+	async function ensureDaemonDetailed(): Promise<{
+		healthy: boolean
+		/** True when this call spawned the process, rather than finding one. */
+		started?: boolean
+		incompatible?: boolean
+		error?: string
+		apiVersion?: number
+	}> {
+		if (deps.ensureDaemon) {
+			const result = await deps.ensureDaemon(currentBaseUrl(), daemonScript())
+			return typeof result === "boolean" ? { healthy: result } : result
+		}
+		return ensureDaemon({ baseUrl: currentBaseUrl(), scriptPath: daemonScript(), env })
+	}
+
+	pi.on("session_start", async (_event, c) => {
+		ctx = c
+
+		// Spawning a daemon that is guaranteed to exit (no mailbox configured) would just
+		// cost five seconds on every session start, so check the mailbox first.
+		if (!hasUsableMailConfig(config)) {
+			c.ui.notify(
+				"mailbridge: no mailbox configured — run /mailbridge setup <pi-address> to <your-address>",
+				"warning",
+			)
+			return
+		}
+
+		// The daemon owns secret.key, but waiting for a daemon that only this code can start
+		// is a deadlock: on a machine that has never run it, the key does not exist yet and
+		// every session would bail out here. Mint it instead (first-writer-wins, 0600).
+		if (!secretFn()) {
+			try {
+				ensureSecret(env)
+			} catch (error) {
+				c.ui.notify(`mailbridge: could not create the shared secret (${(error as Error).message})`, "error")
+				return
+			}
+		}
+
+		protocolStuck = false
+		tokenRejected = false
+		const ensured = await ensureDaemonDetailed()
+		if (!ensured.healthy) {
+			c.ui.notify(
+				ensured.incompatible
+					? "mailbridge: the running daemon speaks another protocol — restarting it"
+					: `mailbridge: daemon not reachable${ensured.error ? ` (${ensured.error})` : ""}; replies will not arrive`,
+				"warning",
+			)
+			// An incompatible daemon is the replacement path, not a dead one. It succeeds or it
+			// does not; either way this session must not then build a runtime on top of it.
+			if (ensured.incompatible) {
+				await replaceOutdatedDaemon(ensured)
+				// Success re-arms the poll itself; a failure must too, or the session sits with no
+				// runtime and no loop while the toast says it is retrying.
+				if (ctx) startPolling()
+				return
+			}
+			// Keep polling even with no runtime: the tick retries the spawn, which is the only
+			// automatic recovery a session has.
+			if (ctx) startPolling()
+			return
+		}
+
+		if (!(await startRuntime())) {
+			if (ctx) startPolling()
+			return
+		}
+		startPolling()
+
+		// A daemon that was already running may have read the config before the user edited
+		// it (hand-edited config.json, headless --setup). Re-read it rather than mailing the
+		// address they just replaced - but do NOT block the session on an IMAP redial, which
+		// can take imapflow's full connect timeout when the new config is wrong.
+		void reloadIfStale()
+	})
+
+	// The ONLY settle signal. `agent_end` is deliberately not registered: pi may
+	// still retry, compact or run queued follow-ups after it.
+	pi.on("agent_settled", (event) => {
+		notifier?.settled({ aborted: event["aborted"] === true })
+	})
+
+	pi.on("session_shutdown", async () => {
+		// A settle inside the debounce window is a notification the user asked for; dropping it
+		// because pi is closing makes a one-shot run (`pi -p`) mail nothing at all. The daemon
+		// outlives us for the idle timeout, so the send can still land.
+		await notifier?.flush().catch(() => undefined)
+		notifier?.cancel()
+		// Release our liveness slot so an idle daemon can exit promptly rather than
+		// waiting out the grace period.
+		void client?.bye(instanceId)
+		stopPolling()
+		client = undefined
+		runner = undefined
+		ctx = undefined
+	})
+
+	// ---- commands ----------------------------------------------------------
+
+	async function setMode(mode: "on" | "off", c: ExtensionCtxLike): Promise<void> {
+		// No client means this session never built a runtime (it started before a config
+		// existed). Writing mode.json alone would report ON with no notifier and no runner - the
+		// documented "copy config.example.json yourself" path lands here.
+		if (!client && (await startRuntime()) && !pollPaused) startPolling()
+		const viaDaemon = client ? await client.setMode(mode) : undefined
+		if (viaDaemon?.ok) {
+			emailMode = mode
+			c.ui.notify(mode === "on" ? "mailbridge: email mode ON" : "mailbridge: email mode OFF", "info")
+			return
+		}
+		// No daemon: fall back to the local file so the mode change is never a silent no-op.
+		const { writeMode } = await import("./store/mode-store.ts")
+		writeMode(env, mode)
+		// The settle notifier reads this mirror, not the file. Without it the command reports
+		// ON while every notification is dropped as "mode off".
+		emailMode = mode
+		c.ui.notify(`mailbridge: email mode ${mode.toUpperCase()} (daemon unreachable, stored locally)`, "warning")
+	}
+
+	// The same handlers back both the in-pi command and a command arriving as a mail reply, so
+	// the mode subcommands cannot drift between mail and prompt.
+	const commandHandlers = new Map<string, (args: string, c: ExtensionCtxLike) => Promise<void>>()
+	const registerCommand = (
+		name: string,
+		description: string,
+		handler: (args: string, c: ExtensionCtxLike) => Promise<void>,
+	): void => {
+		commandHandlers.set(name, handler)
+		pi.registerCommand(name, { description, handler })
+	}
+
+	registerCommand("mailbridge",
+		"Status, on/off, setup, check, auth, smoke, restart",
+		async (args, c) => {
+			// Re-read config on every command. It can change underneath us: `--setup`
+			// run from a shell, or an edit to config.json, would otherwise leave the
+			// command surface reporting stale settings until pi restarts.
+			config = load()
+			const parsed = parseMailbridgeCommand(args)
+
+			/**
+			 * Build client/notifier/runner and start polling if this session has no runtime.
+			 *
+			 * `session_start` returns early when there is no usable config - which is exactly
+			 * the state on a fresh install. Every command that can make the config usable
+			 * (`setup`, `restart`) must therefore be able to wire the runtime afterwards, or
+			 * the documented first run ends in a session that reports "on" and never mails.
+			 */
+			type UiLevel = "error" | "info" | "warning"
+			const ensureRuntime = async (ui: { notify: (m: string, level?: UiLevel) => void }): Promise<boolean> => {
+				if (client) return true
+				if (!secretFn()) {
+					try {
+						ensureSecret(env)
+					} catch (error) {
+						ui.notify(`mailbridge: could not create the shared secret (${(error as Error).message})`, "error")
+						return false
+					}
+				}
+				if (!(await ensureDaemonSafe())) {
+					ui.notify("mailbridge: daemon not reachable; replies will not arrive", "warning")
+					return false
+				}
+				if (!(await startRuntime())) return false
+				// The loop may be parked by a command that owns it; its own finally re-arms.
+				if (!pollPaused) startPolling()
+				return true
+			}
+
+			const statusReport = async () => {
+				const secret = secretFn()
+				const health = client ? await client.health() : undefined
+				// No client in *this* session is not the same as no daemon on the port.
+				const alive = health?.ok ? true : await probe({ baseUrl: clientBase ?? currentBaseUrl() })
+				const pending = client ? await client.pending() : undefined
+				// The daemon is a separate process with its own lifetime; after `pi update` it can
+				// be an older build, so name it here rather than making the user read /health.
+				const daemonVersion = health?.ok ? (health.data.version ?? "unknown") : "unknown"
+				// The protocol is what matters; the package version is diagnostic.
+				const build = health?.ok
+					? apiCompatible(API_VERSION, health.data.apiVersion)
+						? ` — daemon version ${daemonVersion}, protocol ${health.data.apiVersion}`
+						: ` — daemon protocol ${health.data.apiVersion ?? "0"}, this session is ${API_VERSION} (replacing it)`
+					: ""
+				c.ui.notify(
+					[
+						`daemon: ${health?.ok ? "healthy" : alive ? "running, but not attached to this session" : "unreachable"} (${clientBase ?? currentBaseUrl()})${build}`,
+						`device: ${config.device.name}`,
+						`mode: ${health?.ok ? health.data.mode : "(unknown)"}`,
+						`imap: ${health?.ok ? (health.data as { imap?: string }).imap ?? "unknown" : "(unknown)"}`,
+						`session: ${ctx?.sessionManager.getSessionId() ?? "(none)"}`,
+						`secret: ${secret ? "present" : "missing"}`,
+						`mailbox: ${config.mail.to || "(unset)"}`,
+						`allowlist: ${config.security.allowFrom.join(", ") || "(empty — denies all)"}`,
+						`pending replies: ${pending?.ok ? pending.data.pending.length : "?"}`,
+						// From the same response: /pending reports what it did not offer.
+						...(pending?.ok && (pending.data.withheld ?? 0) > 0
+							? [
+									`held for another device: ${pending.data.withheld} (a renamed device.name, or a copied state directory; see docs/ARCHITECTURE.md)`,
+								]
+							: []),
+						...formatScan(health?.ok ? (health.data.lastScan ?? null) : null),
+					].join("\n"),
+					"info",
+				)
+			}
+
+			switch (parsed.sub) {
+				case "on":
+					return void (await setMode("on", c))
+
+				case "off":
+					return void (await setMode("off", c))
+
+				case "help":
+					return void c.ui.notify(USAGE, "info")
+
+				case "status":
+					return void (await statusReport())
+
+				case "setup": {
+					if (!parsed.address) return void c.ui.notify(`setup needs an address\n\n${USAGE}`, "warning")
+					const result = await runSetup({
+						address: parsed.address,
+						...(parsed.to ? { to: parsed.to } : {}),
+						force: parsed.force,
+						env,
+					})
+					config = load() // pick up what was just written
+					const reloaded = client ? await client.reload() : undefined
+					c.ui.notify(
+						[
+							...result.steps,
+							result.next ? `next: ${result.next}` : "",
+							client ? (reloaded?.ok ? "daemon reloaded" : "daemon did not reload — run /mailbridge restart") : "",
+						]
+							.filter(Boolean)
+							.join("\n"),
+						result.ok ? "info" : "error",
+					)
+					// A session that started before the config existed has no runtime yet.
+					if (result.ok && !client) {
+						const started = await ensureRuntime(c.ui)
+						c.ui.notify(
+							started
+								? "daemon started — /mailbridge smoke, then /mailbridge on"
+								: "config written, but the daemon did not start — /mailbridge restart",
+							started ? "info" : "warning",
+						)
+					}
+					return
+				}
+
+				case "check": {
+					// Mint rather than advise: "start the daemon once" is unfollowable in the only
+					// state where this is missing.
+					let secret = secretFn()
+					let mintedSecret = false
+					if (!secret) {
+						try {
+							secret = ensureSecret(env)
+							mintedSecret = Boolean(secret)
+						} catch {
+							/* reported below */
+						}
+					}
+					const problems = [...config.warnings]
+					if (config.security.allowSelfReplies && config.security.requireDmarcPass) {
+						problems.push(
+							"allowSelfReplies is on but requireDmarcPass is too: providers that do not stamp " +
+								"Authentication-Results on self-addressed mail will reject every reply (dmarc-missing)",
+						)
+					}
+					if (!secret) problems.push("shared secret missing and could not be created")
+					// A running daemon caches the bearer token from startup, so a freshly minted key
+					// is rejected until it restarts. Say so; "config ok" would be a lie.
+					else if (mintedSecret) problems.push("created a new shared secret — restart the daemon so it uses it")
+					const usesOAuth = config.mail.imap.authType === "XOAUTH2" || config.mail.smtp.authType === "XOAUTH2"
+					if (config.oauth && !usesOAuth) {
+						problems.push("oauth is configured but neither account uses XOAUTH2 — it would be ignored")
+					}
+					if (usesOAuth) {
+						const oauthConfigured = Boolean(config.oauth?.issuer && config.oauth.clientId)
+						const hasToken = Boolean(fileTokenStore(env).read()?.refreshToken)
+						const hasCommand = Boolean((config.mail.smtp.tokenCommand ?? config.mail.imap.tokenCommand ?? []).length)
+						if (oauthConfigured && !hasToken) problems.push("oauth configured but not authorised — /mailbridge auth")
+						if (!oauthConfigured && !hasCommand) problems.push("XOAUTH2 needs oauth.* or a tokenCommand")
+					}
+					c.ui.notify(
+						problems.length ? problems.map((p) => `! ${p}`).join("\n") : "config ok — run /mailbridge smoke to prove it",
+						problems.length ? "warning" : "info",
+					)
+					return
+				}
+
+				case "auth": {
+					if (!config.oauth?.issuer || !config.oauth.clientId) {
+						return void c.ui.notify("oauth.issuer and oauth.clientId are required — run /mailbridge setup first", "error")
+					}
+					if (config.oauth.clientId.startsWith("<")) {
+						return void c.ui.notify("oauth.clientId is still the placeholder — register an app (docs/OAUTH.md)", "error")
+					}
+					try {
+						const oauthClient = await createOpenidClient(config.oauth)
+						const authorization = await oauthClient.deviceAuthorization()
+						c.ui.notify(
+							[
+								"Approve this login on any device, then wait here:",
+								authorization.verificationUri,
+								authorization.userCode ? `code: ${authorization.userCode}` : "",
+							]
+								.filter(Boolean)
+								.join("\n"),
+							"info",
+						)
+						// Poll in the background so the TUI stays usable.
+						void oauthClient
+							.pollDeviceAuthorization()
+							.then(async (tokens) => {
+								if (!tokens.refreshToken) {
+									c.ui.notify("no refresh_token returned — add offline_access to oauth.scope", "error")
+									return
+								}
+								fileTokenStore(env).write({
+									refreshToken: tokens.refreshToken,
+									accessToken: tokens.accessToken,
+									expiresAt: tokens.expiresAt,
+								})
+								const reloaded = client ? await client.reload() : undefined
+								c.ui.notify(
+									`authorised; refresh token stored${reloaded?.ok ? ", daemon reloaded" : ""}`,
+									"info",
+								)
+							})
+							.catch((error: Error) => c.ui.notify(`login failed: ${error.message}`, "error"))
+					} catch (error) {
+						c.ui.notify(`login failed: ${(error as Error).message}`, "error")
+					}
+					return
+				}
+
+				case "smoke": {
+					try {
+						// Reload first: smoke is the documented "does it work?" step after editing
+						// the password, and the daemon holds the config it started with. Without
+						// this, smoke passes (it reads the file) while /send keeps using the old
+						// empty password and the first notification fails.
+						const reloaded = client ? await client.reload() : undefined
+						if (client && reloaded?.ok !== true) {
+							c.ui.notify("daemon did not reload; run /mailbridge restart", "warning")
+						} else if (reloaded?.ok && reloaded.data.warning) {
+							c.ui.notify(`mailbridge: ${reloaded.data.warning}`, "warning")
+						}
+						config = load()
+						const provider = await buildTokenProvider(config, env)
+						const result = await runSmoke(config, {
+							...(provider ? { getAccessToken: () => provider.getToken() } : {}),
+						})
+						c.ui.notify(
+							[
+								...result.steps.map((s) => `${s.step} ${s.ok ? "ok" : "FAILED"}: ${s.detail}`),
+								...result.steps.filter((s) => s.hint).map((s) => `hint: ${s.hint}`),
+							].join("\n"),
+							result.ok ? "info" : "error",
+						)
+					} catch (error) {
+						c.ui.notify(`smoke failed: ${(error as Error).message}`, "error")
+					}
+					return
+				}
+
+				case "prime": {
+					// First run only: the cursor is what makes mail visible, so moving it after
+					// the bridge has been watching would skip replies that arrived unscanned.
+					let imap: { logout: () => Promise<void> } | undefined
+					try {
+						const provider = await buildTokenProvider(config, env)
+						const { createImapflowFactory } = await import("./daemon/imapflow.ts")
+						const { advanceCursor, isPrimed, markPrimed, readCursor } = await import("./store/uid-cursor.ts")
+						const { alreadyPrimedMessage, runPrime } = await import("./core/prime.ts")
+						const imapClient = await createImapflowFactory({
+							...(provider ? { getAccessToken: () => provider.getToken() } : {}),
+						})(config, { idleRenewMs: config.tuning.idleRenewMs })
+						imap = imapClient
+						const plan = await runPrime(config, {
+							openMailbox: async (cfg) => {
+								await imapClient.connect()
+								await imapClient.mailboxOpen(cfg.mail.imap.folder || "INBOX")
+								return {
+									search: () => imapClient.search({ uid: "1:*" }, { uid: true }),
+									close: async () => imapClient.logout(),
+								}
+							},
+							readCursor: () => readCursor(env),
+							isPrimed: () => isPrimed(env),
+							markPrimed: () => markPrimed(env),
+							advanceCursor: (uid) => {
+								advanceCursor(env, uid)
+							},
+						})
+						if (!plan.ok) {
+							c.ui.notify(alreadyPrimedMessage(plan.currentUid), "warning")
+							return
+						}
+						const reloaded = client ? await client.reload() : undefined
+						c.ui.notify(
+							`primed: cursor at uid ${plan.highest}; ${plan.count} existing message(s) will not be read${reloaded?.ok ? ", daemon reloaded" : ""}`,
+							"info",
+						)
+					} catch (error) {
+						c.ui.notify(`prime failed: ${(error as Error).message}`, "error")
+					} finally {
+						// A connect that succeeded before mailboxOpen threw would otherwise leak the
+						// connection. logout is idempotent; the CLI does the same.
+						await imap?.logout().catch(() => undefined)
+					}
+					return
+				}
+
+				case "restart": {
+					// A config reload cannot replace running code, so after an update the
+					// daemon would keep executing the old version. Stop the process and let
+					// the supervisor start a fresh one.
+					// Park the poll first. It runs every ten seconds and the wait below lasts up to
+					// five, so a tick landing inside it would hear the dying daemon's ping fail,
+					// spawn or attach a replacement, and make `survivor` true - reporting a failure
+					// for the restart this command just performed. Every exit below re-arms it.
+					stopPolling()
+					let joinExpired = false
+					try {
+					// A tick already past every pause check may be inside a spawn. Wait for it, so
+					// the daemon it started is on the port before this command looks.
+					joinExpired = await settleTick()
+					// To stop it we need a client even when this session has no runtime yet - the
+					// same rule the protocol replacement follows, and the remedy its own warning
+					// names.
+					const stopping = client ?? buildClient()
+					// `ok` only means it acknowledged, and a 401 or a timeout means nothing was
+					// confirmed at all - so none of them decides the outcome. `status: 0` with a
+					// network code is a fetch that never got a reply (refused, or the process died
+					// first): not a failure to stop, and the commonest way to run this command.
+					let nothingToStop = false
+					let survivor = false
+					let rejected = false
+					let refusedBase: string | undefined
+					if (stopping) {
+						const result = await stopping.shutdown().catch(() => undefined)
+						nothingToStop = Boolean(result && !result.ok && result.status === 0 && result.code === "network")
+						// A 401 is not a wedged daemon: retrying cannot help until it exits, and the
+						// advice below has to say so. Only 401: the daemon has no 403 (service.ts
+						// answers 401 for a bad bearer), so a 403 would be something else on the port.
+						rejected = Boolean(result && !result.ok && result.status === 401)
+						if (rejected) {
+							// It answered on the client's base, which may not be the configured one
+							// after a `daemon.port` edit: remember that separately, then check
+							// whether it was still there once the stop had its chance.
+							refusedBase = clientBase ?? currentBaseUrl()
+						}
+						client = undefined
+						runner = undefined
+						// Whether something still answers here is the only evidence that a restart did
+						// or did not happen, and `ensureDaemon` would happily reuse a survivor.
+						// `clientBase` follows the client, which may predate a `daemon.port` edit;
+						// the daemon this command starts will be on the configured one.
+						survivor = await waitForPortToClose(currentBaseUrl())
+					}
+					// A 401 proves only that something answered *then*. If it exited inside the
+					// stop window, the restart worked and must not be reported as a failure.
+					// One probe per base: when the refusing base *is* the configured port, `survivor`
+					// already answers this, and asking twice can disagree.
+					const refusalBase = refusedBase ?? currentBaseUrl()
+					const rejectionSurvived = rejected
+						? refusalBase === currentBaseUrl()
+							? survivor
+							: await probe({ baseUrl: refusalBase })
+						: false
+					const ensured = await ensureDaemonDetailed()
+					if (ensured.healthy) {
+						// Cleared before the rebuild: the rebuild is what reads the key again, so a
+						// rejection after it must be able to say so. A daemon that still cannot be
+						// reached never gets here, and keeps both latches.
+						tokenRejected = false
+						protocolStuck = false
+						lastRespawnAttempt = 0
+						// One rebuild, through the helper: it captures the notifier that holds a
+						// debounced settle, rebuilds, and flushes. Calling ensureRuntime first would
+						// overwrite that notifier and lose the settle.
+						if (!secretFn()) {
+							try {
+								ensureSecret(env)
+							} catch {
+								/* reported by the toast below */
+							}
+						}
+						const attached = await replaceRuntime()
+						if (!attached) {
+							c.ui.notify(
+								stopping
+									? "daemon restarted, but this session could not attach — run /mailbridge restart again"
+									: "the daemon is up, but this session could not attach — run /mailbridge restart again",
+								"warning",
+							)
+							return
+						}
+					}
+					// Tell the user what happened, not what was attempted: an answering daemon that
+					// was never stopped is not a restart.
+					let message: string
+					let level: "info" | "warning" | "error" = "info"
+					if (rejected && rejectionSurvived) {
+						// First: it explains a failed respawn (the rejecting daemon holds the lock,
+						// so the new one exits 0) and it is the only advice that works.
+						level = "error"
+						message =
+							"the daemon rejected this session's key, so it could not be stopped — stop it (or let it idle out) and start pi again"
+					} else if (!ensured.healthy) {
+						level = "error"
+						message = ensured.incompatible
+							? `daemon is still on another protocol (${ensured.apiVersion ?? "0"} → ${API_VERSION}) — close pi and start it again`
+							: `daemon did not come back${ensured.error ? ` (${ensured.error})` : ""} — run /mailbridge restart again`
+					} else if (survivor && joinExpired) {
+						// The poll was still starting a daemon when the budget ran out, so the thing
+						// on the port may be its spawn rather than something that survived the stop.
+						level = "warning"
+						message = "the poll was still starting a daemon when this ran — run /mailbridge restart again"
+					} else if (survivor && nothingToStop) {
+						// The stop was refused, yet something answers: either the connection failed
+						// between the two processes or a supervisor restarted it. Not a failure, but
+						// "with the current code" is not something this session can claim.
+						level = "warning"
+						message =
+							"the old daemon was already gone, but something is still serving its port — try again, or close pi if it keeps answering"
+					} else if (survivor) {
+						// `ok:true` is not proof of exit: a shutdown that is still logging out of IMAP
+						// keeps answering, and so does a timeout. Only the port can say.
+						level = "error"
+						message = "a daemon is still answering after the stop request — try again, or close pi if it keeps answering"
+					} else if (!stopping && ensured.started) {
+						// First run: no key existed, so there was nothing to stop, and this call is
+						// what created both. Reporting that as a warning would greet every headless
+						// setup with a caveat about a daemon that never existed.
+						message = "daemon started with the current code"
+					} else if (!stopping) {
+						// No key and no client when the stop was tried, and a daemon was already
+						// there - so this session cannot say what it was serving. State the
+						// observation, not a cause: it may have minted its key a moment ago.
+						level = "warning"
+						message =
+							"a daemon was already up, but this session had no key when it tried to stop anything — stop it and run this again"
+					} else {
+						message = nothingToStop
+							? "the old daemon was already gone; a new one is up with the current code"
+							: "daemon restarted with the current code"
+					}
+					c.ui.notify(message, level)
+				} finally {
+					// One arm for every exit, including a thrown one: parking the loop must never
+					// outlive the command that parked it.
+					startPolling()
+				}
+				return
+				}
+			}
+		},
+	)
+
+	// ---- agent-facing tools ------------------------------------------------
+
+	pi.registerTool({
+		name: "request_decision",
+		label: "Request decision",
+		description:
+			"Email the user a decision request and end the turn. Use only when email mode is on and you cannot proceed without the user's choice.",
+		parameters: Type.Object({
+			subject: Type.String({ description: "Short subject line" }),
+			question: Type.String({ description: "The decision, with enough context to answer from a phone" }),
+			options: Type.Optional(Type.Array(Type.String())),
+			recommendation: Type.Optional(Type.String()),
+		}),
+		async execute(_id, params, _signal, _update, c) {
+			if (!client || !runner) {
+				return { content: [{ type: "text", text: "Mailbridge daemon is not reachable; ask in the terminal instead." }], details: {} }
+			}
+			const mode = await client.mode()
+			if (!mode.ok || mode.data.mode !== "on") {
+				return { content: [{ type: "text", text: "Email mode is off — ask in the terminal." }], details: {} }
+			}
+			const sessionId = c.sessionManager.getSessionId()
+			const reserved = await client.register(sessionId)
+			if (!reserved.ok) {
+				return { content: [{ type: "text", text: "Another decision is already outstanding for this session." }], details: {} }
+			}
+			if (reserved.data.alreadyPending) {
+				return { content: [{ type: "text", text: "A decision is already outstanding for this session; not mailing twice." }], details: {} }
+			}
+
+			// The slot is reserved above. Everything from here must either send the mail or free
+			// the reservation: a leaked slot answers "already outstanding" for the rest of the
+			// session, because re-registering refreshes its TTL.
+			let sent = false
+			try {
+				const secret = secretFn()
+				if (!secret) return { content: [{ type: "text", text: "Mailbridge secret missing." }], details: {} }
+
+				const options = (params["options"] as string[] | undefined) ?? []
+				const recommendation = params["recommendation"] as string | undefined
+				const body = [
+					String(params["question"] ?? ""),
+					options.length ? `\noptions:\n${options.map((o) => `- ${o}`).join("\n")}` : "",
+					recommendation ? `\nrecommendation: ${recommendation}` : "",
+					"\n",
+					"--",
+					"Reply above this line with your answer.",
+				].join("\n")
+
+				const tag = ensureRouteForSession({
+					sessionId,
+					device: config.device.name,
+					subject: String(params["subject"] ?? "pi: decision"),
+					ttlHours: config.security.tokenTtlHours,
+					env,
+				})
+				const outcome = await client.send({
+					kind: "decision",
+					subject: `[pi:${tag}] ${String(params["subject"] ?? "pi: decision")}`,
+					text: body,
+					cwd: c.cwd,
+					sessionId,
+				})
+				if (!outcome.ok) {
+					const why =
+						outcome.code === "send_budget"
+							? "Could not send the email: the notification budget is exhausted (10 a minute); ask in the terminal instead"
+							: "Could not send the email; ask in the terminal instead"
+					return { content: [{ type: "text", text: why }], details: {} }
+				}
+				sent = true
+			} catch (error) {
+				return {
+					content: [{ type: "text", text: `Could not email the decision (${(error as Error).message}); ask in the terminal.` }],
+					details: {},
+				}
+			} finally {
+				// Free only a reservation that never produced a mail.
+				if (!sent) await client.release(sessionId).catch(() => undefined)
+			}
+			return { content: [{ type: "text", text: "Emailed the decision. End your turn and wait for the reply." }], details: {} }
+		},
+	})
+
+	pi.registerTool({
+		name: "notify_user",
+		label: "Notify user",
+		description: "Send the user a one-way email update. Does not end the turn and does not wait for a reply.",
+		parameters: Type.Object({ subject: Type.String(), message: Type.String() }),
+		async execute(_id, params, _signal, _update) {
+			if (!client) return { content: [{ type: "text", text: "Mailbridge daemon is not reachable." }], details: {} }
+			const mode = await client.mode()
+			if (!mode.ok || mode.data.mode !== "on") {
+				return { content: [{ type: "text", text: "Email mode is off." }], details: {} }
+			}
+			const secret = secretFn()
+			if (!secret) return { content: [{ type: "text", text: "Mailbridge secret missing." }], details: {} }
+			// No session means no session to route back to: minting a route for "" would be
+			// dead state that the scanner treats as no-route anyway.
+			const context = sessionContext()
+			if (!context?.sessionId) {
+				return { content: [{ type: "text", text: "No active session to attach this notification to." }], details: {} }
+			}
+			const tag = ensureRouteForSession({
+				sessionId: context.sessionId,
+				device: config.device.name,
+				subject: String(params["subject"] ?? "pi: update"),
+				ttlHours: config.security.tokenTtlHours,
+				env,
+			})
+			const sent = await client.send({
+				subject: `[pi:${tag}] ${String(params["subject"] ?? "pi: update")}`,
+				text: String(params["message"] ?? ""),
+				...((sessionContext()?.cwd ?? undefined) ? { cwd: sessionContext()!.cwd } : {}),
+				...(sessionContext()?.sessionId ? { sessionId: sessionContext()!.sessionId } : {}),
+			})
+			return {
+				content: [
+					{
+						type: "text",
+						text: sent.ok
+							? "Sent."
+							: sent.code === "send_budget"
+								? "Not sent: the notification budget is exhausted (10 a minute)."
+								: "Could not send.",
+					},
+				],
+				details: {},
+			}
+		},
+	})
+}