@alphacoder-v0/pi-loops 0.22.8 → 0.24.0

package/src/host-entry.mjs

@@ -8,7 +8,9 @@ import { importTs } from "./ts-entry.mjs";
 try {
 	await importTs(import.meta.url, "./host.ts");
 } catch (err) {
-	// stdout and stderr are the host log; a line there is what `pi-loops host start` points a person at.
-	process.stderr.write(`pi-loops host: ${err?.message ?? err}\n`);
-	process.exit(1);
+	// stdout and stderr are the host log; a line there is what `pi-loops start` points a person at.
+	process.stderr.write(`pi-loops: ${err?.message ?? err}\n`);
+	// 78 (EX_CONFIG): it cannot run here — the package or pi is not where it was. A systemd unit with
+	// RestartPreventExitStatus=78 does not restart it, where a crash (1) would be restarted.
+	process.exit(78);
 }

package/src/host-runtime.ts

@@ -1,22 +1,32 @@
 /**
- * What the headless host runs (see host.ts for the process shell): the same LoopScheduler and
- * TriggerRuntime as the extension, wired for a process that has no chat — promotions go to the
- * inbox, control-plane operations by sub-agents are audited into `triggers-audit.jsonl`, and the
- * process leaves as soon as an interactive pi owns the clock. Injected so it can be unit-tested.
+ * What the always-on host runs (see host.ts for the process shell): the same LoopScheduler and
+ * TriggerRuntime as the extension, wired for a process that has no chat. It runs every loop and
+ * checks every rule; a promotion is a request addressed to the conversation that created the rule
+ * (src/requests.ts): its holder delivers it, and this process does when it is closed. Control-plane
+ * operations by sub-agents are audited into `triggers-audit.jsonl`. Injected so it can be unit-tested.
  */
-import type { ToolDefinition } from "@earendil-works/pi-coding-agent";
+import * as fs from "node:fs";
+import { SessionManager, type ToolDefinition } from "@earendil-works/pi-coding-agent";
 import * as path from "node:path";
 import type { LoopsConfig } from "./config.ts";
 import { HookRunner, runEndEvent, runStartEvent } from "./hooks.ts";
 import { PI_BUILTIN_TOOL_NAMES } from "./mcp.ts";
 import { redact } from "./redact.ts";
-import type { SubagentRequest, SubagentRunner } from "./runner.ts";
-import { LoopScheduler, type SessionSnapshot } from "./scheduler.ts";
+import { keepNotice, parseRingNote, type Notice } from "./host-control-channel.ts";
+import type { RunnerResult, SubagentRequest, SubagentRunner } from "./runner.ts";
+import { DELIVERY_JOB_ID, type LeaseHolder, LeaseStore, closedConversation, liveHolderPid, procStart, sameHolder } from "./lease.ts";
+import { CEILING, type EngineSinks, type Executor, type ExecutorView, type FileMark, type QuietMark, bury, changesSomething, conversationHas, conversationView, decideDrain, drain, promotionTarget, promotionView, quietMark, ruleResultRow, sessionFileEntries, stillQuiet } from "./engine.ts";
+import { RequestStore } from "./requests.ts";
+import { PI_LOOPS_VERSION } from "./version.ts";
+import type { PresenceEntry } from "./presence.ts";
+import { DEFAULT_RUN_TIMEOUT_MS, type HeldSession, LoopScheduler, type PlainJobLeases, STOPPED_BY_PERSON, type SessionSnapshot } from "./scheduler.ts";
+import { stamp } from "./schedule.ts";
 import { SubagentSlots } from "./slots.ts";
 import { type LoopJob, newId } from "./store.ts";
 import { type ToolHost, automationTools, createLoopJob } from "./tools.ts";
-import { TriggerRuntime } from "./trigger-runtime.ts";
+import { TriggerRuntime, promotionToInbox } from "./trigger-runtime.ts";
 import { auditCronFinish, auditCronStart, TriggerStore, controlPlanePreflight } from "./triggers.ts";
+import { shortSessionId } from "./session-head.ts";
 
 export interface HostRuntimeDeps {
 	dir: string;
@@ -27,12 +37,18 @@ export interface HostRuntimeDeps {
 	/** MCP tool definitions of the host's own (user-level) servers, for every sub-session. */
 	mcpTools: () => ToolDefinition<any, any>[];
 	/** That project's own MCP servers, connected on demand: the host has no project of its own. */
-	projectMcpTools?: (cwd: string, taken: Set<string>) => Promise<ToolDefinition<any, any>[]>;
+	projectMcpTools?: (cwd: string, taken: Set<string>, trusted?: boolean) => Promise<ToolDefinition<any, any>[]>;
 	/** Whether pi has a saved trust decision for exactly this directory (src/trust.ts). Default: no. */
 	isProjectTrusted?: (cwd: string) => boolean;
+	/** pi's own saved decision for a conversation's directory or a parent: what a held turn there may load. */
+	ownerTrusts?: (cwd: string) => boolean;
 	log: (message: string) => void;
-	/** Called when an interactive pi owns the clock: the host has nothing left to do. */
-	exit: (code: number) => void | Promise<void>;
+	/** Does pi still have this session? Plain jobs of deleted sessions are parked by the host's sweep. */
+	sessionExists?: (sessionId: string) => boolean;
+	/** Every tick, before the rules are checked (the legacy fence lives here). */
+	onTick?: (now: number) => void;
+	/** Is this conversation open in the browser page? Then its pi writes the file, and this process never holds it. */
+	pageHolds?: (sessionId: string) => boolean;
 }
 
 export interface HostRuntime {
@@ -40,14 +56,176 @@ export interface HostRuntime {
 	triggers: TriggerRuntime;
 	/** The tools a sub-session started by this host gets (its cwd and model, not the host's). */
 	customTools: (req: SubagentRequest) => Promise<ToolDefinition<any, any>[]>;
+	/** Say a warning to the windows of `cwd`'s project (or every window, without one); see `notices`. */
+	notice(text: string, cwd?: string): void;
+	/** The latest warnings, for the snapshot a window fetches on each tick. */
+	notices(): Notice[];
+	/** The doorbell: deliver what is queued for a closed conversation now, not on the next tick. Returns at once. */
+	ring(sessionId: string): void;
+	/** The turn this process is running in the conversation, if any: the browser page waits for it. */
+	heldTurn(sessionId: string): { runId: string; label: string } | undefined;
 	start(): void;
 	stop(): Promise<void>;
 }
 
+/** Rings one session may answer between two ticks; a burst of rows is delivered by one ring, so this is far above any real one. */
+const RINGS_PER_TICK = 60;
+
 export function createHostRuntime(deps: HostRuntimeDeps): HostRuntime {
 	const { dir, log } = deps;
+	const kept: Notice[] = [];
+	let noticeSeq = 0;
+	let budgetSaid: string | undefined;
+	const notice = (text: string, cwd?: string) => {
+		noticeSeq = keepNotice(kept, noticeSeq, text, cwd, Date.now());
+	};
 	let triggers: TriggerRuntime;
-	let exiting = false;
+	const leases = new LeaseStore(dir);
+	const requests = new RequestStore(dir);
+	/** The host as a lease holder: it holds a closed conversation for one piece of work, then lets go. */
+	const ownStart = procStart(process.pid);
+	const me = (busy?: LeaseHolder["busy"]): LeaseHolder => ({ kind: "host", pid: process.pid, instance: "host", procStart: ownStart, acquiredAt: stamp(), pushServers: [], ...(busy ? { busy } : {}) });
+	const fileMtime = (file: string | undefined): number | undefined => {
+		if (!file) return undefined;
+		try {
+			return fs.statSync(file).mtimeMs;
+		} catch {
+			return undefined; // gone: nothing for the quiet guard to remember
+		}
+	};
+	const fileMark = (file: string): FileMark => {
+		try {
+			const { size, mtimeMs } = fs.statSync(file);
+			return { size, mtimeMs };
+		} catch {
+			return undefined; // not there: a mark of its own, so its appearance is a change
+		}
+	};
+	/** Per closed conversation, the quiet mark of its last "nothing to do" (src/engine.ts quietMark). */
+	const quiet = new Map<string, QuietMark>();
+	/** Open in a pi-loops 0.22 window, which fires its plain jobs without a lease: never the host's to hold. */
+	let legacyOpen = (_sessionId: string): boolean => false;
+	const heldLeases: PlainJobLeases = {
+		heldEpoch: (sessionId) => (leases.holds(sessionId, me()) ? leases.read(sessionId)?.epoch : undefined),
+		holderOf: (sessionId) => liveHolderPid(leases, sessionId),
+		acquireForWork: (sessionId, runId, jobId) => {
+			if (legacyOpen(sessionId) || deps.pageHolds?.(sessionId)) return undefined;
+			const got = leases.acquireForHost(sessionId, me({ runId, jobId, since: stamp() }));
+			if (!("lease" in got) || !got.lease.sessionFile) return undefined;
+			return { sessionId, epoch: got.lease.epoch, sessionFile: got.lease.sessionFile, cwd: got.lease.cwd };
+		},
+		// With the epoch it took, and the mtime it left: see LeaseStore.release.
+		releaseWork: (sessionId, epoch) => void leases.release(sessionId, me(), { epoch, fileMtimeMs: fileMtime(leases.read(sessionId)?.sessionFile) }),
+	};
+	/** A plain job the host may fire: a conversation pi-loops knows the file of, that nobody holds and the page does not have open. */
+	const canHold = (sessionId: string | undefined): boolean => !(sessionId !== undefined && deps.pageHolds?.(sessionId)) && closedConversation(leases, sessionId, legacyOpen);
+
+	/** The engine's sinks here: the jobs the rows name, the inbox for dead letters, the audit. */
+	const sinks = (): EngineSinks => ({ jobs: { exists: (id) => scheduler.store.load().some((j) => j.id === id), enabled: (id) => scheduler.store.load().find((j) => j.id === id)?.enabled ?? false, goalActive: () => false }, inbox: scheduler.inbox, audit: triggers.store });
+
+	/**
+	 * The held turns this process is in, by session: the run (the lease's `busy.runId`), when it
+	 * started, and whether the runner has the file (`writing`). A stop may name the run from the turn's
+	 * own first drain on; nothing is appended beside the runner once it writes.
+	 */
+	const turnOpen = new Map<string, { requestId: string; startedAt: string; writing: boolean; label: string }>();
+	/**
+	 * The engine's executor while this process holds a closed conversation: an `append` row is written
+	 * into the file at once, each one once (the `[Trigger <id>]` already in the file is the check).
+	 * No attended row runs here (`CEILING.held`); a turn row is not admitted yet either (held turns move
+	 * onto rows in a later stage), so it waits for a window instead of failing here. While a turn of
+	 * this process runs in the conversation its runner is writing the file, so nothing is appended
+	 * beside it: only a stop or a cancel acts, and the inputs wait.
+	 */
+	const heldView = (sessionId: string): ExecutorView => {
+		const open = turnOpen.get(sessionId);
+		const writing = open?.writing ?? false;
+		return { kind: "held", ceiling: CEILING.held, admits: (row) => !writing && row.wake !== "turn", idle: !writing, running: open && { requestId: open.requestId, startedAt: open.startedAt }, inFlight: new Set(), connected: () => true, producerAlive: () => true, mine: () => false };
+	};
+	function heldExecutor(held: HeldSession): Executor {
+		const sm = SessionManager.open(held.sessionFile);
+		return {
+			holds: (sessionId) => leases.holds(sessionId, me(), held.epoch),
+			entries: () => conversationView(() => sm.getEntries()),
+			view: () => heldView(held.sessionId),
+			append: (row) => {
+				sm.appendCustomMessageEntry("pi-loops:trigger", row.payload?.text ?? "", true);
+				// Written into a closed conversation, so no screen showed it: the audit row says so. The
+				// "since you last looked" line counts the rule's finding in the inbox (src/seen.ts).
+				triggers.store.appendAudit({ cwd: row.payload?.cwd ?? "", type: "trigger_promotion", traceId: row.id, state: "appended", details: { session_id: held.sessionId } });
+				// And said in the project's open windows: the person is likely in another conversation of
+				// it now, where a rule's result used to land (0.22), and would not know to look there.
+				// Once an hour per conversation (keepNotice keeps one of each text an hour).
+				notice(`a rule's result was added to conversation ${shortSessionId(held.sessionId)}, which is closed: pi --session ${held.sessionFile} shows it`, row.payload?.cwd);
+			},
+			turn: () => {
+				throw new Error("the background process does not run turn requests yet");
+			},
+			// As `/triggers abort` does: the run's AbortController, through the scheduler.
+			abort: () => {
+				const run = turnOpen.get(held.sessionId);
+				if (run) scheduler.abortRun(run.requestId);
+			},
+			// Never asked: nothing is in flight here (the view's set is empty), a row is in the file or not.
+			abortWhenStarted: () => undefined,
+			// Never asked either: a person row is attended, above this executor's ceiling.
+			refused: () => undefined,
+		};
+	}
+
+	/** Requests queued for a conversation this process holds: delivered now, while it holds it. */
+	async function drainInto(held: HeldSession): Promise<void> {
+		await drain(held.sessionId, requests, heldExecutor(held), Date.now(), sinks());
+	}
+
+	/** One plain job's turn in a closed conversation (src/scheduler.ts launchHeld). */
+	async function heldTurn(job: LoopJob, prompt: string, runId: string, held: HeldSession, signal: AbortSignal): Promise<{ preempted: boolean; result: RunnerResult; landed: boolean }> {
+		const stillHolds = () => leases.holds(held.sessionId, me(), held.epoch);
+		// The run exists from here: a stop that arrives during this first drain names it.
+		// What the page's waiting line calls it: a job's id is 32 characters no person typed.
+		const open = { requestId: runId, startedAt: scheduler.runningRuns().find((r) => r.runId === runId)?.startedAt ?? stamp(), label: job.name ?? runId.slice(0, 12) };
+		turnOpen.set(held.sessionId, { ...open, writing: false });
+		log(`cron ${job.name ?? job.id}: turn ${runId.slice(0, 12)} in session ${held.sessionId.slice(0, 8)}`);
+		let result: RunnerResult;
+		try {
+			try {
+				await drainInto(held);
+			} catch (err: any) {
+				log(`requests for session ${held.sessionId.slice(0, 8)}: ${err?.message ?? err}`);
+			}
+			if (signal.aborted) {
+				// Stopped before it began: nothing to run, and no model call to start for a turn already over.
+				log(`cron ${job.name ?? job.id}: turn ${runId.slice(0, 12)} stopped before it began`);
+				return { preempted: false, result: { ok: false, exitCode: 1, timedOut: false, text: "", errorMessage: "aborted", stopReason: "aborted", usage: { input: 0, output: 0, cost: 0, turns: 0 } }, landed: false };
+			}
+			// As the conversation's owner: pi's own trust for its directory, read now, and the job's declared tools.
+			// Without the conversation's own directory (an older lease), the job's cwd could be one a model
+			// chose: no owner trust then.
+			const cwd = held.cwd ?? job.cwd;
+			turnOpen.set(held.sessionId, { ...open, writing: true });
+			// No model or thinking level: the turn is the conversation's own and runs on what the
+			// conversation is on, as its window turn does (a 0.22 job's captured `model` was never used).
+			// The conversation is the turn's parent session, so the tools act for it: `cron_create
+			// stateful:false` binds to it, `all_projects` lists as a window's does.
+			result = await deps.runner({ cwd, prompt, tools: job.tools, builtinTools: job.builtinTools, trusted: held.cwd ? !!deps.ownerTrusts?.(held.cwd) : false, timeoutMs: job.timeoutMs ?? DEFAULT_RUN_TIMEOUT_MS, signal, hop: 1, kind: "held", jobId: job.id, runId, sessionFile: held.sessionFile, stillHolds, parentSessionId: held.sessionId, parentCwd: cwd });
+		} finally {
+			turnOpen.delete(held.sessionId);
+		}
+		// Taken back only if the turn did not complete: a window that arrives as the answer is saved has
+		// nothing to run again, and rolling the slot back would answer the same prompt twice.
+		const preempted = !result.ok && !stillHolds();
+		const byPerson = !result.ok && signal.aborted && signal.reason === STOPPED_BY_PERSON;
+		log(`cron ${job.name ?? job.id}: ${preempted ? "a window took the session back; its turn stopped" : result.ok ? "turn done" : byPerson ? `turn stopped by you (from ${scheduler.stopFrom(runId) === "page" ? "the page" : "a terminal"})` : `turn FAILED (${redact(result.errorMessage ?? "")})`}`);
+		// Read from the file, not from the result: a turn cut off after pi wrote the prompt has landed
+		// it, and the row it came from is done either way.
+		let landed = result.ok;
+		try {
+			landed = conversationHas(SessionManager.open(held.sessionFile).getEntries(), runId);
+		} catch (err: any) {
+			log(`cron ${job.name ?? job.id}: could not read session ${held.sessionId.slice(0, 8)} after its turn: ${err?.message ?? err}`);
+		}
+		return { preempted, result, landed };
+	}
 
 	/* ------------------------------------------------------ lifecycle hooks */
 
@@ -106,10 +284,13 @@ export function createHostRuntime(deps: HostRuntimeDeps): HostRuntime {
 		getSession: deps.session,
 		getSettings: () => ({ maxConcurrentRuns: deps.config().maxConcurrentRuns, catchUp: deps.config().cronCatchUp, dailyBudgetUsd: deps.config().dailyBudgetUsd }),
 		runner: deps.runner,
-		// No dead-session parking here: that is a decision the interactive leader can show the user.
+		// Every loop, and the plain jobs of closed conversations: a plain job's result is a message in
+		// its conversation, so the host fires one only where no window holds it.
+		owns: (job) => job.stateful || canHold(job.sessionId),
+		leases: heldLeases,
+		sessionExists: deps.sessionExists,
 		hooks: {
-			// Plain (inject) jobs belong to a chat; the host has none, so they stay dormant here.
-			onInject: () => undefined,
+			onHeldTurn: heldTurn,
 			onRunStart: (job, runId) => {
 				// The same rows the interactive extension writes: without them `/triggers audit` is
 				// blank for every hour the host had the clock.
@@ -124,7 +305,7 @@ export function createHostRuntime(deps: HostRuntimeDeps): HostRuntime {
 			onCatchUp: (job) => log(`loop ${job.name ?? job.id}: catching up a missed tick`),
 			onRunFinished: ({ job, record, result }) => {
 				auditCronFinish(triggers.store, job, record, result.stopReason === "aborted");
-				log(`loop ${job.name ?? job.id}: ${record.ok ? "ok" : `FAILED (${redact(record.error ?? "")})`} · ${record.findings} finding(s)${record.usage?.cost ? ` · $${record.usage.cost.toFixed(3)}` : ""}`);
+				log(`loop ${job.name ?? job.id}: ${record.ok ? "ok" : record.stoppedBy === "person" ? `stopped by you (from ${record.stoppedFrom === "page" ? "the page" : "a terminal"})` : `FAILED (${redact(record.error ?? "")})`} · ${record.findings} finding(s)${record.usage?.cost ? ` · $${record.usage.cost.toFixed(3)}` : ""}`);
 				const hooks = runHooks.get(record.runId);
 				if (!hooks) return;
 				// `$PI_RUN_OK` alone answers "did last night's loop fail" — the thing the host exists to
@@ -134,25 +315,242 @@ export function createHostRuntime(deps: HostRuntimeDeps): HostRuntime {
 				const forget = () => void runHooks.delete(record.runId);
 				void hooks.fire(runEndEvent(job, record)).then(forget, forget);
 			},
-			onTick: async (now, leader) => {
-				if (!leader) {
-					if (exiting) return;
-					exiting = true;
-					log("an interactive pi owns the clock; exiting");
-					// Not awaited: exit() stops the scheduler, and scheduler.stop() waits for this tick.
-					Promise.resolve(deps.exit(0)).catch((err: any) => log(`exit failed: ${err?.message ?? err}`));
-					return;
-				}
-				await triggers.tick(now, leader);
+			onTick: async (now) => {
+				deps.onTick?.(now);
+				releaseStrayLeases(now);
+				await deliverToClosed();
+				await triggers.tick(now);
+			},
+			log: (msg, level, job) => {
+				log(redact(msg));
+				if (level !== "info") notice(msg, job?.cwd);
+			},
+			// A loop failing night after night: said in its project's windows, as it was when loops ran in one.
+			onSchedulerError: (msg, job) => {
+				log(redact(msg));
+				notice(msg, job?.cwd);
+			},
+			onBudgetExceeded: (spent, cap) => {
+				// Each due job reports it; a person needs it once a day.
+				const day = stamp().slice(0, 10);
+				if (budgetSaid === day) return;
+				budgetSaid = day;
+				const message = `today's automation has cost $${spent.toFixed(2)} of the $${cap.toFixed(2)} budget; dispatching is paused until tomorrow or a higher [limits] daily_budget_usd`;
+				log(message);
+				notice(message);
 			},
-			log: (msg) => log(redact(msg)),
 		},
 	});
 
-	const toInbox = async (content: string, trigger: { sourceLabel: string; traceId: string; cwd?: string }) => {
-		await scheduler.inbox.append({ source: `trigger:${trigger.sourceLabel}`, text: content.replace(/^\[Trigger [^\]]+\]\s*/, ""), run_id: trigger.traceId, job_id: trigger.sourceLabel, cwd: trigger.cwd ?? "" });
+	legacyOpen = (sessionId) => scheduler.presenceList().some((e: PresenceEntry) => e.sessionId === sessionId && !e.leases && e.kind !== "host");
+
+	const toInbox = async (content: string, trigger: { sourceLabel: string; traceId: string; cwd?: string }, forSession?: string) => {
+		// The check decided this result's finding already (TriggerRuntime's inbox); a push with no rule did not.
+		await promotionToInbox(scheduler.inbox, triggers.store, content, trigger, { cwd: trigger.cwd ?? "", sessionId: forSession });
 		return "inbox" as const;
 	};
+	/**
+	 * A rule's result is a request addressed to a conversation (src/engine.ts promotionTarget): the one
+	 * that created the rule while a window holds it, else one a window holds in the rule's project,
+	 * else the closed creating conversation, which this process appends to while it holds it for that
+	 * one write. A closed conversation this process cannot take this instant (written moments ago,
+	 * busy) keeps the row for a later tick. The inbox is for a result no conversation can take.
+	 */
+	const promote = async (content: string, trigger: { sourceLabel: string; traceId: string; cwd?: string }, forSession?: string): Promise<"queued" | "chat" | "inbox"> => {
+		const route = promotionTarget(forSession, promotionView(leases, trigger.cwd, legacyOpen));
+		if (route.to === "inbox") return toInbox(content, trigger, route.sessionId);
+		requests.append(ruleResultRow(route.sessionId, content, trigger, { pid: process.pid, instance: "host", procStart: ownStart, version: PI_LOOPS_VERSION }));
+		if (!canHold(route.sessionId)) return "queued";
+		return (await deliverQueued(route.sessionId, trigger.traceId)) ? "chat" : "queued";
+	};
+
+	/**
+	 * Take a closed conversation for one write and deliver what is queued for it. True when `traceId`
+	 * (if given) is now in the conversation. The tick calls it for every closed conversation with a
+	 * queue, so a result queued while the file was busy, or for a window that quit before reading it,
+	 * still arrives.
+	 */
+	async function deliverQueued(sessionId: string, traceId?: string): Promise<boolean> {
+		const held = heldLeases.acquireForWork?.(sessionId, traceId ?? "requests", DELIVERY_JOB_ID);
+		if (!held) return false;
+		try {
+			await drainInto(held);
+			if (traceId === undefined) return true;
+			const delivered = conversationHas(SessionManager.open(held.sessionFile).getEntries(), traceId);
+			if (delivered) log(`trigger ${traceId.slice(0, 8)}: appended to session ${sessionId.slice(0, 8)}`);
+			return delivered;
+		} catch (err: any) {
+			log(`requests for session ${sessionId.slice(0, 8)}: ${err?.message ?? err}`);
+			// Giving the lease back below rings this process again; with the same failure waiting
+			// (an inbox it cannot write), that ring would fail the same way, for ever.
+			quietUntilTick.add(sessionId);
+			return false;
+		} finally {
+			heldLeases.releaseWork?.(sessionId, held.epoch);
+		}
+	}
+
+	/**
+	 * A conversation this process still holds for a turn that is no longer running: the release at
+	 * the turn's end failed (a lock timeout, a full disk), and nothing ran or was delivered there again
+	 * until a restart. Given back here, a minute after the turn began, which no turn's setup takes.
+	 */
+	function releaseStrayLeases(now: number): void {
+		const running = new Set(scheduler.runningRuns().map((r) => r.runId));
+		for (const lease of leases.list()) {
+			const busy = lease.holder?.busy;
+			if (!lease.holder || !busy || !sameHolder(lease.holder, me()) || running.has(busy.runId) || now - Date.parse(busy.since) < 60_000) continue;
+			try {
+				leases.release(lease.sessionId, me(), { epoch: lease.epoch, fileMtimeMs: fileMtime(lease.sessionFile) });
+				log(`gave back session ${lease.sessionId.slice(0, 8)}, held for run ${busy.runId} that is no longer running`);
+			} catch (err: any) {
+				log(`lease ${lease.sessionId.slice(0, 8)}: ${err?.message ?? err}`);
+			}
+		}
+	}
+
+	/**
+	 * What is queued for one conversation nobody has open: delivered by this process, not left waiting
+	 * for a resume. A conversation whose file is gone can never take it, so its rows become dead
+	 * letters in the inbox, marked for it, instead of waiting there for ever.
+	 *
+	 * Writes nothing when the drain would change nothing. Taking the lease is itself a write, and this
+	 * process hears its own writes as rings (`watch` below): a delivery that took the lease only to
+	 * find every row waiting would ring itself into the next one, for ever.
+	 */
+	async function deliverSession(sessionId: string): Promise<void> {
+		const rows = requests.list(sessionId);
+		if (!rows.length) {
+			quiet.delete(sessionId);
+			return;
+		}
+		// A turn of this process's own is running there: the conversation is held already, for that
+		// turn, so a stop or cancel is applied through that lease, with the epoch it took, and nothing
+		// is taken a second time. The executor admits no input while the runner has the file.
+		const lease = leases.read(sessionId);
+		if (turnOpen.has(sessionId) && lease?.holder && sameHolder(lease.holder, me()) && lease.sessionFile) {
+			await drainInto({ sessionId, epoch: lease.epoch, sessionFile: lease.sessionFile, cwd: lease.cwd });
+			return;
+		}
+		if (!canHold(sessionId)) {
+			if (liveHolderPid(leases, sessionId) !== undefined || legacyOpen(sessionId) || deps.pageHolds?.(sessionId)) return; // its holder drains it
+			const file = leases.read(sessionId)?.sessionFile;
+			if (file && fs.existsSync(file)) return; // there, only busy or recently written: a later tick
+			const buried = await bury(sessionId, requests, sinks(), "its conversation is gone");
+			if (buried.length) log(`${buried.length} queued request(s) for session ${sessionId.slice(0, 8)} went to the inbox: its conversation is gone`);
+			return;
+		}
+		// The same decision the drain makes, read without the lease (a read needs none). Marked before
+		// the read, so a write that lands during it shows as a change on the next tick, not a miss.
+		const file = leases.read(sessionId)?.sessionFile;
+		if (!file) return;
+		const files = [fileMark(file), fileMark(path.join(dir, "jobs.json"))];
+		const now = Date.now();
+		if (stillQuiet(quiet.get(sessionId), files, rows, now)) return;
+		quiet.delete(sessionId);
+		const entries = sessionFileEntries(file);
+		if (!changesSomething(decideDrain(rows, conversationView(() => entries), sinks().jobs, heldView(sessionId), now))) {
+			quiet.set(sessionId, quietMark(files, rows, now));
+			return;
+		}
+		await deliverQueued(sessionId);
+	}
+
+	/**
+	 * Every closed conversation with something queued, on the tick: the floor under the doorbell, so a
+	 * ring that was lost costs at most one tick. One conversation's failure (a lock that timed out)
+	 * does not keep the others, or the rule checks after this, from their tick.
+	 */
+	async function deliverToClosed(): Promise<void> {
+		quietUntilTick.clear();
+		rungSinceTick.clear();
+		// Image files whose row is gone (a crash between the two writes): one process sweeps them, here.
+		try {
+			const swept = requests.sweepImages();
+			if (swept.length) log(`removed ${swept.length} image file(s) no request names any more`);
+		} catch (err: any) {
+			log(`requests: ${err?.message ?? err}`);
+		}
+		for (const sessionId of requests.sessions()) {
+			try {
+				await deliverSession(sessionId);
+			} catch (err: any) {
+				log(`requests for session ${sessionId.slice(0, 8)}: ${err?.message ?? err}`);
+			}
+		}
+	}
+
+	/* ------------------------------------------------------------ the doorbell */
+
+	/**
+	 * Rings arrive in bursts (every append, every lease write is one), so a session has at most one
+	 * delivery running and one more to run after it: a ring during a delivery is kept as that one
+	 * rerun, which sees whatever the running one read too early to see. Nothing waits on a timer.
+	 */
+	const delivering = new Set<string>();
+	const ringAgain = new Set<string>();
+	/**
+	 * The bell is never the only way a row arrives, so it may stop ringing: after a delivery that
+	 * failed, and past a number of rings no real burst reaches between two ticks. Either way the
+	 * session is the tick's until then. This bounds a loop of self-rings that a check above missed;
+	 * the timers it would starve include the tick's own.
+	 */
+	const quietUntilTick = new Set<string>();
+	const rungSinceTick = new Map<string, number>();
+	function ring(sessionId: string): void {
+		if (quietUntilTick.has(sessionId)) return;
+		const rung = (rungSinceTick.get(sessionId) ?? 0) + 1;
+		rungSinceTick.set(sessionId, rung);
+		if (rung > RINGS_PER_TICK) {
+			quietUntilTick.add(sessionId);
+			log(`doorbell for session ${sessionId.slice(0, 8)}: rang ${RINGS_PER_TICK} times since the last tick; left to the tick`);
+			return;
+		}
+		if (delivering.has(sessionId)) {
+			ringAgain.add(sessionId);
+			return;
+		}
+		delivering.add(sessionId);
+		deliverSession(sessionId)
+			.catch((err: any) => log(`requests for session ${sessionId.slice(0, 8)}: ${err?.message ?? err}`))
+			.then(() => {
+				delivering.delete(sessionId);
+				if (ringAgain.delete(sessionId)) ring(sessionId);
+			})
+			.catch((err: any) => log(`doorbell: ${err?.message ?? err}`));
+	}
+
+	/**
+	 * The file name is the whole message: `<session>.jsonl` in `requests/` is a row written for it,
+	 * `<session>.json` in `leases/` a conversation given back, or taken. Either rings the session.
+	 * fs.watch can miss events and is best-effort by design; the tick still delivers, so a watcher
+	 * that fails to start, or fails later, is logged and not restarted.
+	 */
+	const watchers: fs.FSWatcher[] = [];
+	function watch(): void {
+		if (watchers.length) return;
+		const stores: Array<{ sub: "requests" | "leases"; ext: string; kind: "request" | "lease" }> = [
+			{ sub: "requests", ext: ".jsonl", kind: "request" },
+			{ sub: "leases", ext: ".json", kind: "lease" },
+		];
+		for (const { sub, ext, kind } of stores) {
+			const where = path.join(dir, sub);
+			try {
+				fs.mkdirSync(where, { recursive: true });
+				const watcher = fs.watch(where, (_event, name) => {
+					const file = String(name ?? "");
+					if (!file.endsWith(ext)) return;
+					const note = parseRingNote(`${kind}:${file.slice(0, -ext.length)}`);
+					if (note) ring(note.sessionId);
+				});
+				watcher.on("error", (err) => log(`${sub} watch: ${err?.message ?? err}`));
+				watchers.push(watcher);
+			} catch (err: any) {
+				log(`${sub} watch: ${err?.message ?? err}`);
+			}
+		}
+	}
+
 	triggers = new TriggerRuntime({
 		store: new TriggerStore(dir),
 		jobStore: scheduler.store,
@@ -163,24 +561,25 @@ export function createHostRuntime(deps: HostRuntimeDeps): HostRuntime {
 		dedupFile: path.join(dir, "dedup.json"),
 		slots,
 		budget: () => scheduler.budgetState(),
-		self: scheduler.self,
-		presence: () => scheduler.presenceList(),
-		isLeader: () => scheduler.isLeader,
+		inbox: scheduler.inbox,
 		hooks: {
-			// No chat here: everything a rule would have promoted lands in the inbox.
-			onPromote: toInbox,
-			onInjectAndRun: toInbox,
+			onPromote: promote,
+			// A push that injects and runs reaches the host only when no window has its server (host.ts): the inbox.
+			onInjectAndRun: (prompt, trigger) => toInbox(prompt, trigger),
 			onFinished: (o) => log(`trigger ${o.trigger.traceId.slice(0, 8)} ${o.delivery} ${o.ok ? "ok" : `FAILED (${redact(o.error ?? "")})`}${o.matchedRules.length ? ` matched ${o.matchedRules.length}` : ""}`),
 			log: (msg) => log(redact(msg)),
 		},
 	});
 
 	/** The tools' host for one sub-session: that run's cwd and model, not the process defaults. */
-	const toolHostFor = (req: Pick<SubagentRequest, "cwd" | "model" | "thinking">): ToolHost => {
+	const toolHostFor = (req: Pick<SubagentRequest, "cwd" | "model" | "thinking" | "builtinTools" | "tools" | "kind" | "parentSessionId">): ToolHost => {
+		const held = req.kind === "held";
 		const host: ToolHost = {
 			scheduler,
 			triggers,
-			session: () => ({ ...deps.session(), cwd: req.cwd, model: req.model ?? deps.session().model, thinking: req.thinking ?? deps.session().thinking }),
+			// A job a run makes gets no more built-in tools than that run (createLoopJob trims to them), and no trust of its own.
+			// A held turn's session is the conversation it runs in; a loop's run has none.
+			session: () => ({ ...deps.session(), sessionId: req.parentSessionId, cwd: req.cwd, model: req.model ?? deps.session().model, thinking: req.thinking ?? deps.session().thinking, builtinTools: req.builtinTools ?? req.tools?.filter((t) => PI_BUILTIN_TOOL_NAMES.includes(t)) ?? deps.session().builtinTools, trusted: false }),
 			createJob: (input, scope) => createLoopJob(host, input, scope),
 			// The cron_control_plane audit has no session to live in here; /triggers audit shows it instead.
 			cronControlAudit: (op, actor, before, after) => {
@@ -190,8 +589,9 @@ export function createHostRuntime(deps: HostRuntimeDeps): HostRuntime {
 				log(`cron control plane: ${op} by ${actor} ${job?.id ?? ""}`);
 				return id;
 			},
-			// Nobody can approve here: Prompt-class operations are denied fail-closed.
-			confirmTool: async (_ctx, request, atHop) => controlPlanePreflight({ hop: Math.max(1, atHop), hasUI: false }, request.label),
+			// Nobody can approve here: Prompt-class operations are denied fail-closed. A held turn is the
+			// conversation's own, so its refusal says where the person was not, not "sub-agent".
+			confirmTool: async (_ctx, request, atHop) => controlPlanePreflight({ hop: held ? 0 : Math.max(1, atHop), hasUI: false, held }, request.label),
 			refreshBadge: () => undefined,
 		};
 		return host;
@@ -200,15 +600,28 @@ export function createHostRuntime(deps: HostRuntimeDeps): HostRuntime {
 	return {
 		scheduler,
 		triggers,
+		notice,
+		notices: () => [...kept],
+		ring,
+		heldTurn: (sessionId) => {
+			const open = turnOpen.get(sessionId);
+			return open && { runId: open.requestId, label: open.label };
+		},
 		customTools: async (req) => {
 			const shared = deps.mcpTools();
-			const automation = automationTools({ hop: req.hop, actor: "sub-agent", parentSessionId: req.parentSessionId, parentCwd: req.parentCwd }, toolHostFor(req));
+			// A held turn is the conversation's own turn, as in a window (hop 0, actor "tool"): the
+			// tools list and bind as they would there. A loop's run is a sub-agent's.
+			const automation = automationTools(req.kind === "held" ? { hop: 0, actor: "tool", parentSessionId: req.parentSessionId, parentCwd: req.parentCwd } : { hop: req.hop, actor: "sub-agent", parentSessionId: req.parentSessionId, parentCwd: req.parentCwd }, toolHostFor(req));
 			const taken = new Set([...PI_BUILTIN_TOOL_NAMES, ...shared.map((t) => t.name), ...automation.map((t) => t.name)]);
-			const project = req.cwd ? ((await deps.projectMcpTools?.(req.cwd, taken)) ?? []) : [];
+			const project = req.cwd ? ((await deps.projectMcpTools?.(req.cwd, taken, req.trusted)) ?? []) : [];
 			return [...shared, ...project, ...automation];
 		},
-		start: () => scheduler.start(),
+		start: () => {
+			scheduler.start();
+			watch();
+		},
 		stop: async () => {
+			for (const w of watchers.splice(0)) w.close();
 			await triggers.stop();
 			await scheduler.stop();
 			// The runs are over; their hooks may still be in flight. Same bounded wait as the