@dynamicagents/plugins 0.11.1 → 0.12.0

package/dist/computer/host/workspace.js

@@ -0,0 +1,1258 @@
+import { DurableObject, tracing } from "cloudflare:workers";
+// One Durable Object has one alarm, and this object wakes for more reasons than
+// that. A `Scheduler` is the multiplexer; what it does *not* own is what this
+// object owes on waking, which is the callbacks registered on it below.
+import { installScheduler, namedDeadline } from "@dynamicagents/core/alarm";
+import { Workspace } from "@cloudflare/computer";
+import { CloudflareContainerBackend, withWorkspaceContainer } from "@cloudflare/computer/backends/container";
+import { createGitClient } from "@cloudflare/computer/git";
+import { createCloudflareObserver } from "@cloudflare/computer/observe/cloudflare";
+import { MAX_TOOL_CALL_MS } from "@dynamicagents/core";
+import { ContainerTrust } from "./ca-trust.js";
+import { ContainerDeps } from "./container-deps.js";
+import { ContainerGitIdentity } from "./git-identity.js";
+import { InstallJob } from "./install-job.js";
+import { WorkspaceGitHost } from "./git-host.js";
+import { SYNC_DRAIN_BUDGET_MS, SYNC_DRAIN_RESUME_MS, WorkspaceSync, syncRetryDelayMs } from "./sync.js";
+// Leaf modules rather than `../index.js` — see the same import in
+// {@link file://./install-job.ts} for why the barrel is not reachable from here.
+import { deriveAdvisories } from "../advisory.js";
+import { pathExists } from "../read.js";
+/**
+ * A workspace: one Durable Object, one container, one repository.
+ *
+ * **Shared by every agent in a Worker that has a container.** Each subclasses
+ * {@link WorkspaceObjectBase}, supplies a {@link WorkspaceObjectConfig}, and
+ * inherits everything else. Keep that seam narrow: it is the complete answer to
+ * "what is different about this agent's container", and anything that does not
+ * fit through it becomes a second copy of this file.
+ *
+ * A directory inside the plugin rather than a subpath beside it, so that one
+ * capability stays one import path and a consumer cannot take the tools without
+ * the object they address. The bundle cost of that is nothing: `sideEffects` is
+ * false, so an agent that only calls `sb_exec` carries no container backend and
+ * no isomorphic-git.
+ *
+ * What this file must not do is import the plugin barrel. The barrel re-exports
+ * this module, and the class below is built at import time — a cycle would
+ * evaluate its base as `undefined`. The imports above reach the leaf modules
+ * beside it instead.
+ *
+ * `@cloudflare/computer` pairs a SQLite-backed virtual filesystem in *this*
+ * object's storage with a container running `computerd`, which mounts it over
+ * FUSE at `/workspace`. Commands run against the same tree the Worker reads over
+ * RPC, and the tree outlives the container.
+ *
+ * **One repository per object**, because `computer` is strictly 1 DO ↔ 1
+ * container: the id derives from caller *and* repository (see `workspaceName`),
+ * so two repositories for one caller are two objects and two containers.
+ *
+ * **It constructs `Workspace` rather than using `withWorkspace`.** The mixin
+ * stores the `Workspace` under a module-private symbol the package does not
+ * export, so a method on the object cannot reach it — and what the host is
+ * required to do needs exactly that: driving an outstanding pull (the library
+ * owns no alarm and schedules nothing of its own), and the direct `runtime`
+ * access the detached install needs. So this object owns the `Workspace` and
+ * implements the one method the mixin otherwise provides,
+ * `__getWorkspaceStub`. Callers outside see no difference.
+ *
+ * ## What is here, and what is a module next door
+ *
+ * This file is the Durable Object: the wake map, the deadlines, the RPC surface
+ * and the wiring between them. The work each of those drives has its own home,
+ * and each carries the reasoning for its own decisions:
+ *
+ * - `./install-job.ts` — when a dependency install runs, and every guard on it.
+ * - `./sync.ts` — the host's half of syncing, which nothing else will do.
+ * - `./ca-trust.ts` — making a container able to speak TLS.
+ * - `./git-identity.ts` — who a container's commits belong to.
+ * - `./git-host.ts` — the operations that hold the forge credential.
+ *
+ * Each takes a small explicit seam, and that is the test: a module that needed
+ * the whole object back would be a module that did not want extracting.
+ */
+/** Where every checkout lives, inside the container and in the VFS. */
+export const WORKSPACE_DIR = "/workspace";
+/**
+ * The Durable Object name for one caller's checkout of one repository.
+ *
+ * Exported because several places must agree on it and none can see the others:
+ * the plugin that resolves the stub, the agent that hands it down to subagents,
+ * and the cancellation path. A pipe rather than a slash, so the caller half
+ * cannot forge a repository boundary by containing one.
+ *
+ * `repo` is undefined only before the first `repo_clone` or `scratch_open` of a
+ * session — what a task works in is model-chosen, so there is genuinely nothing
+ * to key on until it has chosen. That window resolves to a caller-level
+ * workspace, which is never worked in: `beforeCheckout` sets the repository
+ * before any git runs, and `scratch_open` sets its sentinel before anything
+ * resolves a name.
+ *
+ * `repo` is not always a repository. A host may pass a sentinel through here as
+ * one — a scratchpad is modelled that way, which is what keys it to its own
+ * object and its own container. Anything that is not a forge name works, as long
+ * as no caller could clone something that collides with it.
+ */
+export function workspaceName(callerKey, repo) {
+    return repo ? `${callerKey}|${repo}` : `${callerKey}|<unassigned>`;
+}
+// --- where the work is ------------------------------------------------------
+/**
+ * The key the checkout record lives under.
+ *
+ * Deliberately **not** the install record's: what is on disk and what was
+ * installed into it have different lifetimes, and one record cannot answer both.
+ * See {@link WorkspaceObjectBase.noteCheckout}.
+ */
+const CHECKOUT_KEY = "checkout";
+// --- what this object wakes for, and when ----------------------------------
+/**
+ * Where the idle-reclaim deadline keeps its current schedule id.
+ *
+ * A schedule is a row the scheduler mints an id for, not a keyed upsert — so
+ * "push this deadline back", which `#touch()` does on every single call into
+ * this object, is cancel-then-set and needs the id of what stands now. See
+ * `namedDeadline`.
+ */
+const IDLE_RECLAIM_ID = "idle-reclaim-id";
+/**
+ * How long a workspace survives without being used.
+ *
+ * A Durable Object is **never** reclaimed by the platform, and a namespace
+ * cannot be enumerated from a Worker, so nothing else is coming to clean up.
+ * Source-only workspaces are small (6.3 MB for slack-gatekeeper), which makes
+ * this hygiene rather than cost control — but unbounded hygiene is still
+ * unbounded.
+ */
+const IDLE_RECLAIM_MS = 7 * 24 * 60 * 60 * 1000;
+/** Where the container-idle deadline keeps its current schedule id. */
+const CONTAINER_IDLE_ID = "container-idle-id";
+/** Where the container-warm wake keeps its current schedule id. */
+const CONTAINER_WARM_ID = "container-warm-id";
+/**
+ * How long a container stays up after the last command **started**.
+ *
+ * Ours to schedule: `withWorkspaceContainer` wraps the runtime's raw
+ * `ctx.container`, not `@cloudflare/containers`' `Container`, so there is no
+ * `sleepAfter` to lean on.
+ *
+ * **This must exceed the longest command the agent allows**, and breaking that
+ * kills work in flight. Measured from when a command *starts*: `#touch()` arms
+ * the clock on the way into this object, and a running command touches nothing
+ * again until it finishes — `handle.result()` is one long await and the FUSE
+ * traffic under it never surfaces as an RPC. Set equal to the computer plugin's
+ * `DEFAULT_TIMEOUT_MS`, the two timers race and whichever fires first destroys
+ * the container the other depends on. Twenty minutes is double that ceiling;
+ * raise it, never lower it, if `sb_exec` is given a longer timeout.
+ *
+ * **The default, not the policy.** "Longest command" is a fact about the agent,
+ * so one whose commands run longer must say so via
+ * {@link WorkspaceObjectConfig.containerIdleMs} — `claude-coder` holds a
+ * `claude -p` session open for its whole 40-minute timeout.
+ */
+const CONTAINER_IDLE_MS = 20 * 60_000;
+/**
+ * Refuse to grow past this, of the 10 GB a Durable Object may hold.
+ *
+ * Source-only workspaces run at ~6 MB, so this should never fire — which is why
+ * it is worth having: if something does start pulling a large tree in, a
+ * sentence naming the number beats a write failing somewhere unrelated.
+ */
+const STORAGE_CAP_BYTES = 8 * 1024 * 1024 * 1024;
+/** Where the pending-pull drain keeps its current schedule id. */
+const SYNC_DRAIN_ID = "sync-drain-id";
+/**
+ * How long a container may be kept alive past its idle deadline so an
+ * outstanding pull can finish.
+ *
+ * The deadline defers while a drain is still moving blocks, because destroying
+ * the container is what makes unpulled work unrecoverable. It cannot defer
+ * forever — a pull that never completes would keep a container billing
+ * indefinitely — so this is the wall on that, measured from the last use.
+ */
+const SYNC_DRAIN_GRACE_MS = 30 * 60_000;
+/**
+ * How long a container start may take before the calls waiting on it are refused.
+ *
+ * A minute under core's per-call limit, and tied to it: past that limit core
+ * abandons the call anyway, the model reads no reason, and the next call joins
+ * the same stuck start. As late as that allows, because refusing a start that
+ * was only slow clears `#readying` while it runs, and the next caller starts a
+ * second one beside it — and a start pushing a large tree has taken nine minutes.
+ */
+const READY_DEADLINE_MS = MAX_TOOL_CALL_MS - 60_000;
+/**
+ * `work`, refused once `ms` pass without it settling. The work itself goes on;
+ * only the wait for it ends — see {@link READY_DEADLINE_MS}.
+ */
+export function readyWithin(work, ms) {
+    let timer;
+    const expired = new Promise((_resolve, reject) => {
+        timer = setTimeout(() => reject(new Error(`the workspace container did not become ready within ${ms / 1000}s; the next call starts it again`)), ms);
+    });
+    // The loser still settles, and a rejection nobody holds fails the request.
+    work.catch(() => { });
+    return Promise.race([work, expired]).finally(() => clearTimeout(timer));
+}
+/**
+ * Set once the synced dependency trees are gone from this object's storage —
+ * see {@link WorkspaceObjectBase.#purgeSyncedTrees}.
+ */
+const SYNCED_TREES_PURGED_KEY = "deps:purged";
+// --- the object -------------------------------------------------------------
+/**
+ * Named rather than the anonymous class expression an app would write here.
+ *
+ * This package emits declarations, and `DurableObject` brings `ctx` and `env` in
+ * protected — which makes the mixin's return type unnameable in a `.d.ts` and
+ * fails the build with TS4094. A named base gives the emitted type something to
+ * refer to. It costs one identifier and is invisible to a consumer.
+ */
+class WorkspaceContainerHost extends DurableObject {
+}
+/**
+ * `withWorkspaceContainer` adds one method, `getWorkspaceContainer()`, over
+ * `this.ctx.container` — the runtime's own container handle. There is no
+ * `@cloudflare/containers` `Container` subclass here and so no `sleepAfter`:
+ * idle shutdown is this object's job, and it lands on the wake map with
+ * everything else.
+ */
+const WorkspaceContainerBase = withWorkspaceContainer(WorkspaceContainerHost);
+/**
+ * One caller's checkout of one repository, and the container that mounts it.
+ *
+ * Abstract because a Durable Object class takes no constructor arguments, so
+ * per-agent configuration cannot arrive that way. {@link workspaceConfig} is the
+ * seam, and it is the shape core's own `RecipeSubagentBase.subagentRuntime`
+ * uses for exactly the same reason.
+ *
+ * ## Why `backend` and `#workspace` are lazy
+ *
+ * **Base class fields run before subclass fields**, so as plain fields they
+ * would read `undefined` from any `workspaceConfig()` that touches a subclass
+ * field — which `ClaudeCoderWorkspaceDO`'s does, for its credential store.
+ * Memoised getters remove the hazard rather than documenting it, which is what
+ * lets a subclass implement the seam however it likes.
+ */
+export class WorkspaceObjectBase extends WorkspaceContainerBase {
+    #configMemo;
+    get #cfg() {
+        return (this.#configMemo ??= this.workspaceConfig());
+    }
+    /** This object's log prefix, so one workspace stays tellable from another. */
+    get #tag() {
+        return this.#cfg.label;
+    }
+    /** This agent's container-idle window — see {@link WorkspaceObjectConfig}. */
+    get #containerIdleMs() {
+        return this.#cfg.containerIdleMs ?? CONTAINER_IDLE_MS;
+    }
+    /**
+     * How long an install may run before it is killed.
+     *
+     * A getter because the fallback has to be the same number at every site an
+     * install is bounded, and a `??` repeated at each is a chance to write a
+     * different one.
+     */
+    get #installTimeoutMs() {
+        return this.#cfg.installPlan.timeoutMs ?? 20 * 60_000;
+    }
+    /**
+     * The one alarm, multiplexed across every reason this object wakes.
+     *
+     * A callback is registered under a **name**, and a schedule row persists that
+     * name rather than a closure: the object is re-created on every wake, so
+     * anything captured here would not survive one. Registration therefore happens
+     * unconditionally, in a field initializer, every time.
+     *
+     * `hostOwns` is the acknowledgement that this class defines `alarm()` and
+     * `fetch()` itself. A lifecycle installs its handlers only where the host has
+     * none, so without saying so here the scheduler would be installed and never
+     * fire, with no error anywhere. `alarm()` below calls through, which is what
+     * makes the declaration true. **`fetch()` deliberately does not** — it serves
+     * `computerd`'s capnweb WebSocket upgrade, and a lifecycle's `fetch` declines
+     * any upgrade no installed capability claims.
+     */
+    #wake = installScheduler(this, {
+        hostOwns: ["alarm", "fetch"],
+        callbacks: {
+            installRun: () => this.#install.onRun(),
+            installWatch: () => this.#install.onWatch(),
+            idleReclaim: () => this.#onIdleReclaim(),
+            containerIdle: (payload) => this.#onContainerIdle(payload),
+            containerWarm: () => this.#onContainerWarm(),
+            syncRetry: (payload) => this.#onSyncDrain(payload)
+        },
+        onError: (err) => {
+            console.error(`[${this.#tag}] a scheduled callback failed for good`, {
+                id: this.ctx.id.toString(),
+                err: String(err)
+            });
+        }
+    });
+    /** Reclaim a workspace nobody has touched. Moved forward by every `#touch`. */
+    #idleReclaim = namedDeadline({
+        storage: this.ctx.storage,
+        scheduler: this.#wake.scheduler,
+        key: IDLE_RECLAIM_ID,
+        callback: "idleReclaim"
+    });
+    /** Stop a container nobody is using. Moved forward by every `#touch`. */
+    #containerIdle = namedDeadline({
+        storage: this.ctx.storage,
+        scheduler: this.#wake.scheduler,
+        key: CONTAINER_IDLE_ID,
+        callback: "containerIdle"
+    });
+    /**
+     * Start a cold container in the background — see {@link #warmIfCold}.
+     *
+     * A deadline rather than a bare `scheduler.set`, for the reason {@link #touch}
+     * gives: armed per tool call, so anything that did not cancel the row standing
+     * would leave one per call.
+     */
+    #containerWarm = namedDeadline({
+        storage: this.ctx.storage,
+        scheduler: this.#wake.scheduler,
+        key: CONTAINER_WARM_ID,
+        callback: "containerWarm"
+    });
+    /**
+     * Come back and move the next block of an outstanding pull.
+     *
+     * **The callback name is on disk in deployed objects, so it may not change.**
+     * A schedule row persists its callback by name, and the scheduler drops a row
+     * whose callback it cannot find with one line in the log — a rename would
+     * strand every pull already scheduled.
+     */
+    #syncDrain = namedDeadline({
+        storage: this.ctx.storage,
+        scheduler: this.#wake.scheduler,
+        key: SYNC_DRAIN_ID,
+        callback: "syncRetry"
+    });
+    /**
+     * The dependency install — see `./install-job.ts`, which owns every guard on it.
+     *
+     * The seam is what only this object can answer: where its storage and its
+     * scheduler are, what the workspace is, and the entry-point bookkeeping an
+     * install still owes on the way in.
+     */
+    #install = new InstallJob({
+        storage: this.ctx.storage,
+        scheduler: this.#wake.scheduler,
+        workspace: () => this.#workspace,
+        plan: () => this.#cfg.installPlan,
+        timeoutMs: () => this.#installTimeoutMs,
+        headroom: () => this.#storageHeadroom(),
+        touch: () => this.#touch(),
+        ready: () => this.#ready(),
+        waitUntil: (promise) => this.ctx.waitUntil(promise),
+        containerGone: () => this.#containerGone(),
+        tag: () => this.#tag,
+        id: () => this.ctx.id.toString()
+    });
+    /**
+     * The forge credential, read at the moment it would be handed over.
+     *
+     * `#`-private, which is the only spelling actually unreachable over RPC — see
+     * {@link WorkspaceGitConfig.tokenBinding} for why that matters here and not
+     * for the rest of the config.
+     *
+     * Indexed rather than named: the subclass says which binding holds it, because
+     * this package cannot name a consumer's `Env`.
+     */
+    #token() {
+        const env = this.env;
+        return env[this.#cfg.git.tokenBinding];
+    }
+    /** Clone, fetch and push, with the credential — see `./git-host.ts`. */
+    #gitHost = new WorkspaceGitHost({
+        git: () => this.#workspace.git,
+        token: () => this.#token(),
+        tag: () => this.#tag
+    });
+    /** The host's half of syncing — see `./sync.ts`. */
+    #sync = new WorkspaceSync({
+        workspace: () => this.#workspace,
+        containerRunning: () => this.ctx.container?.running === true,
+        deadline: () => this.#syncDrain,
+        tag: () => this.#tag,
+        id: () => this.ctx.id.toString()
+    });
+    /**
+     * The container backend.
+     *
+     * `container: () => this` hands the backend this object's own container.
+     * `workspace` is how `computerd` dials *back* in: the runtime builds a loopback
+     * binding from the exported `WorkspaceProxy` class and the values below, which
+     * is why the consuming Worker's entry point must re-export `WorkspaceProxy`
+     * and why dropping that export breaks the container with no compile error.
+     *
+     * Nothing sets `egressHost`; the default `computer.internal` is the host the
+     * container's outbound HTTP is intercepted on, internal to that loopback.
+     *
+     * The binding name and the egress policy are the subclass's — see
+     * {@link WorkspaceObjectConfig}, which carries the warnings on both.
+     */
+    #backendMemo;
+    get backend() {
+        return (this.#backendMemo ??= new CloudflareContainerBackend({
+            container: () => this,
+            workspace: {
+                binding: this.#cfg.binding,
+                id: this.ctx.id.toString()
+            },
+            egress: this.#cfg.egress
+        }));
+    }
+    #workspaceMemo;
+    get #workspace() {
+        return (this.#workspaceMemo ??= new Workspace(this.#workspaceOptions()));
+    }
+    /**
+     * The container's TLS trust, which `#ready` below keeps current.
+     *
+     * See `./ca-trust.ts` for why the answer lives in memory and what clears it.
+     */
+    #trust = new ContainerTrust({
+        workspace: () => this.#workspace,
+        container: () => this.ctx.container,
+        tag: () => this.#tag,
+        id: () => this.ctx.id.toString(),
+        onExit: () => void this.#containerGone()
+    });
+    /**
+     * The identity every repository in the container starts out attributed to.
+     *
+     * See `./git-identity.ts` for why the container needs one of its own when
+     * `/repo` already configures the checkouts it clones.
+     */
+    #gitIdentity = new ContainerGitIdentity({
+        workspace: () => this.#workspace,
+        author: () => this.#cfg.git.author,
+        tag: () => this.#tag,
+        id: () => this.ctx.id.toString()
+    });
+    /** Where installs write — see `./container-deps.ts`. */
+    #deps = new ContainerDeps({
+        workspace: () => this.#workspace,
+        workspaceDir: WORKSPACE_DIR,
+        tag: () => this.#tag,
+        id: () => this.ctx.id.toString()
+    });
+    /** Everything believed about a container, dropped when it goes. */
+    async #containerGone() {
+        this.#trust.forget();
+        this.#gitIdentity.forget();
+        this.#deps.forget();
+        await this.#install.containerGone();
+    }
+    /**
+     * Set once {@link reclaimIfIdle} has emptied storage, until the alarm it armed
+     * resets the isolate. Anything reaching this instance in between is addressing
+     * storage that is gone.
+     */
+    #reclaimed = false;
+    /**
+     * The start already in flight, shared by everyone who asks for one.
+     *
+     * In memory, like {@link file://./ca-trust.ts}'s flag and for the same reason:
+     * it describes a container, and an isolate that lost it asks again.
+     */
+    #readying;
+    /**
+     * Open the workspace, and make sure the container behind it can speak TLS and
+     * installs onto its own disk.
+     *
+     * **Every path that might start a container goes through here**, and the CA
+     * install hangs off container *liveness* rather than off the install job.
+     * Whether a container is new and whether it is due dependencies are different
+     * questions: an install declines to arm for a `skipped` or `idle` state,
+     * returns early when one is already in flight, and again when the workspace is
+     * full — so hanging trust off it leaves a replaced container talking to an
+     * untrusted CA with nothing pending that would fix it, and leaves it worst
+     * exactly when egress is what the agent needs to dig itself out.
+     *
+     * `ctx.container?.running` is read **before** `ready()`, because `ready()` is
+     * what starts a stopped container — afterwards every container looks running
+     * and the distinction is gone.
+     *
+     * Cheap enough to sit on the busiest entry point in the object: at most one
+     * extra exec per isolate, plus one per container it sees start. An unchanged
+     * container costs a boolean.
+     *
+     * Private, like everything else in here that is not an RPC: every non-`#`
+     * method is reachable over RPC — see {@link WorkspaceGitConfig.tokenBinding}.
+     *
+     * **One start at a time**, because {@link #warmIfCold} makes overlap ordinary:
+     * a warm and the first `sb_exec` want the container at the same moment, as do
+     * a warm and the install alarm. Two starts means two trust commands through a
+     * half-established connection, the second pushing the tree again behind the
+     * first. Cleared on settle, so the next caller asks about the container that
+     * is there now — and on {@link READY_DEADLINE_MS}, so a start that hangs holds
+     * its callers for that long and no longer.
+     */
+    async #ready() {
+        this.#readying ??= readyWithin(this.#readyNow(), READY_DEADLINE_MS).finally(() => {
+            this.#readying = undefined;
+        });
+        return this.#readying;
+    }
+    /** One start, run for whoever got there first — see {@link #ready}. */
+    async #readyNow() {
+        // Read **before** `ready()`, which starts a stopped container.
+        if (!this.ctx.container?.running)
+            await this.#containerGone();
+        await this.#workspace.ready();
+        // Before the first exec, which pushes the whole tree into a new container.
+        await this.#purgeSyncedTrees();
+        await this.#trust.ensure();
+        await this.#gitIdentity.ensure();
+        const found = await this.#deps.ensure(await this.#install.dir());
+        if (found)
+            await this.#install.reconcile(found.marker);
+    }
+    /**
+     * Delete every `node_modules` that was synced into this object, once.
+     *
+     * Left in place, each is pushed into every new container — where the bind
+     * mount hides it but computerd still holds it in memory.
+     */
+    async #purgeSyncedTrees() {
+        if (await this.ctx.storage.get(SYNCED_TREES_PURGED_KEY))
+            return;
+        const startedAt = Date.now();
+        const fs = this.#workspace.fs;
+        try {
+            const trees = await fs
+                .find(WORKSPACE_DIR, "**/node_modules", {
+                exclude: ["**/.git", "**/node_modules/*"]
+            })
+                // No workspace directory yet, so nothing was ever synced.
+                .catch((err) => {
+                if (err?.code === "ENOENT")
+                    return [];
+                throw err;
+            });
+            for (const tree of trees) {
+                if (tree.type === "dir")
+                    await fs.rm(tree.path, { recursive: true });
+            }
+            await this.ctx.storage.put(SYNCED_TREES_PURGED_KEY, Date.now());
+            if (trees.length > 0)
+                console.info(`[${this.#tag}] purged synced dependency trees`, {
+                    id: this.ctx.id.toString(),
+                    trees: trees.map((t) => t.path),
+                    ms: Date.now() - startedAt
+                });
+        }
+        catch (err) {
+            // Left unmarked, so the next start tries again; a command must not fail
+            // for it.
+            console.warn(`[${this.#tag}] could not purge synced dependency trees`, {
+                id: this.ctx.id.toString(),
+                err: String(err)
+            });
+        }
+    }
+    #workspaceOptions() {
+        return {
+            // `ctx.storage.sql.exec` returns a narrower row type than
+            // `DurableObjectStorageLike` declares and the two are invariant, so the
+            // cast goes through `unknown`. The runtime shapes match; this is the
+            // pattern the package's own example uses.
+            storage: this.ctx.storage,
+            backends: [this.backend],
+            // One line per committed sync block and one per finished operation, into
+            // the same Workers Observability view as the spans below — blocks,
+            // entries, bytes, and the CPU headroom each block left.
+            //
+            // **It covers the syncs this object drives, not the ones a command drives
+            // for itself.** `pull()` and `push()` log; the push a `runtime.exec` does
+            // before it spawns — the expensive one, since a fresh container takes the
+            // whole tree — logs nothing. Its cost shows up as a `workspace.sync.push`
+            // span, and on `ca-trust.ts`'s line.
+            syncTelemetryEnabled: true,
+            // One span per sync push, sync pull, exec spawn and filesystem op, into
+            // the same Workers Observability view the rest of this Worker traces to.
+            // Needs `observability.traces.enabled` in wrangler.jsonc; without the
+            // feature flag `tracing` is undefined and this degrades to a no-op.
+            observer: createCloudflareObserver({ tracing }),
+            // Git, running **here** rather than in the container.
+            //
+            // This is what lets the forge token stay on this side of the boundary.
+            // `createGitClient` binds isomorphic-git to `provider()` — the local
+            // SQLite store, not the wire — so a clone, fetch or push executes next to
+            // the data it writes, and the container never holds a credential at all.
+            // Credentialed `git` in the container cannot: it executes whatever
+            // `.git/config` and `.git/hooks` in the checkout name.
+            //
+            // Needs `@platformatic/vfs`, an optional peer of `@cloudflare/computer`:
+            // the adapter that wraps `provider()` into an isomorphic-git FsClient
+            // imports it lazily and throws a named error when it is absent.
+            git: createGitClient(),
+            // Only the commit-producing subcommands read this, and the three
+            // operations driven from here — clone, fetch, push — are not among them.
+            // Set anyway so that a `pull` or `merge` added later fails on the merge
+            // itself rather than on `MissingIdentityError`, and set to the same pair
+            // `/repo` writes into the checkout's own config at clone time (see
+            // `author` in each agent's `plugins.ts`), so a commit cannot be
+            // attributed differently depending on which side made it.
+            defaultGitIdentity: this.#cfg.git.author
+        };
+    }
+    /**
+     * Repair a lost alarm on the way in.
+     *
+     * The one failure a scheduler cannot defend against from the inside: the
+     * runtime retries a throwing `alarm()` a bounded number of times and then
+     * stops for good, and a deleted-class migration takes the alarm with the
+     * storage. Both leave schedule rows due with nothing coming for them, and the
+     * symptom is silence.
+     *
+     * So every RPC into this object checks. That fully covers the sync drain,
+     * `install-watch` and `container-idle`, which only matter while somebody is
+     * using the workspace. It does **not** cover `idle-reclaim`, which by
+     * definition fires when nobody is — that one has the agent's weekly cron
+     * poking `reclaimIfIdle` as its backstop.
+     */
+    async #repairAlarm() {
+        try {
+            await this.#wake.start();
+            await this.#wake.rearm();
+        }
+        catch (err) {
+            console.error(`[${this.#tag}] could not repair the alarm`, {
+                id: this.ctx.id.toString(),
+                err: String(err)
+            });
+        }
+    }
+    /**
+     * What `getWorkspace(stub)` calls from outside this object.
+     *
+     * The one piece of `withWorkspace` reimplemented here — see the file comment
+     * for why the mixin is not used. `ready()` first, because the stub is only
+     * meaningful once the workspace has opened its store.
+     */
+    async __getWorkspaceStub() {
+        // The busiest entry point by far, and therefore the one that keeps both the
+        // idle clock and the alarm honest.
+        await this.#touch();
+        await this.#repairAlarm();
+        await this.#ready();
+        await this.#install.armIfTreeMissing();
+        return this.#workspace.stub();
+    }
+    /**
+     * The same workspace, for a caller that will only touch the filesystem.
+     *
+     * **The container is not started, and that is the whole difference.** The
+     * filesystem is this object's own SQLite, so a read or write of it needs no
+     * container and an absent one cannot block it. {@link __getWorkspaceStub} adds
+     * `#ready()`, and the first command in a fresh container re-pushes the whole
+     * tree before it runs, with `timeoutMs` reaching the spawned process and
+     * nothing earlier.
+     *
+     * A stub cannot enforce what the caller does with it: `runtime.exec` on this
+     * one would run against a container with no CA. The exec paths take the other
+     * method — see `openWorkspaceFs` in `@dynamicagents/plugins/computer`.
+     */
+    async __getWorkspaceFsStub() {
+        await this.#touch();
+        await this.#repairAlarm();
+        await this.#warmIfCold();
+        // Indexes mounts and nothing else — no backend connects here, which is what
+        // makes this cheap enough to sit in front of every file tool.
+        await this.#workspace.ready();
+        return this.#workspace.stub();
+    }
+    /**
+     * Ask the alarm to start the container, so the first `sb_exec` finds one that
+     * is up and already holding the tree.
+     *
+     * What this avoids is the push, not the boot — see {@link __getWorkspaceFsStub}.
+     * Moving it to the alarm overlaps it with the model's reading.
+     *
+     * **Only for a workspace with a checkout in it**, since an empty one has
+     * nothing to push and nothing to run, and a container it never uses still
+     * bills.
+     *
+     * Only the arming is awaited. A warm that fails costs the next `sb_exec` the
+     * start it would have paid for anyway.
+     */
+    async #warmIfCold() {
+        // The case on every call but the first, which is what keeps the reads below
+        // off the per-tool-call cost.
+        if (this.ctx.container?.running)
+            return;
+        // Already starting — `#ready()` dedupes that, but arming again would queue a
+        // wake-up for work in progress.
+        if (this.#readying)
+            return;
+        const record = await this.ctx.storage.get(CHECKOUT_KEY);
+        if (!record?.dir)
+            return;
+        await this.#containerWarm.set(new Date());
+    }
+    /**
+     * Clone, fetch and push — the operations that need the forge token.
+     *
+     * The work is `./git-host.ts`, which owns why the credential stays on this
+     * side of the boundary. What stays here is what only this object can do: the
+     * entry-point bookkeeping, and making the workspace current before git reads
+     * it.
+     */
+    async gitClone(req) {
+        const behind = await this.#beforeGit();
+        return behind ?? (await this.#gitHost.clone(req));
+    }
+    async gitFetch(req) {
+        const behind = await this.#beforeGit();
+        return behind ?? (await this.#gitHost.fetch(req));
+    }
+    async gitPush(req) {
+        const behind = await this.#beforeGit();
+        return behind ?? (await this.#gitHost.push(req));
+    }
+    /**
+     * Bring the workspace up to date, or say why git must not run.
+     *
+     * **Git here reads this object's storage**, so it has to be current first. The
+     * commits these operations push were made by `git` *in the container*, and
+     * they reach this side on that command's own post-exec pull. When that pull
+     * did not finish, the workspace still holds the tree as it was before the
+     * commit — and isomorphic-git, running here, would push a branch that does not
+     * include it. That is the one failure worth stopping for: the push reports
+     * success, the forge shows an older tree, and nothing in either account says
+     * why.
+     *
+     * Refused rather than attempted, because every alternative is worse. Pushing
+     * anyway publishes a tree the agent did not produce; pushing "what is here"
+     * silently redefines what the caller asked for. A refusal is recoverable — the
+     * pull resumes on its own schedule and the operation works when retried.
+     *
+     * Returns the refusal to hand back, or `undefined` to proceed.
+     */
+    async #beforeGit() {
+        await this.#touch();
+        await this.#repairAlarm();
+        /**
+         * Drained **before** `#ready()`, which is what starts a stopped container.
+         *
+         * An outstanding pull belongs to the runtime that ran the command. Start a
+         * replacement first and the drain can no longer tell: `pull()` reconnects to
+         * whatever is running now, finds a filesystem that matches this object
+         * because the workspace just pushed it there, and reports a clean
+         * completion — for writes that are gone.
+         *
+         * Asked while the old container is still the only one there, the same call
+         * either moves the writes or says plainly there is nothing to move them
+         * from.
+         */
+        const synced = await this.#sync.drain(SYNC_DRAIN_BUDGET_MS);
+        // `failed` refuses alongside `incomplete`, and the reason is that the two
+        // differ only in what the *drain* should do next. To git they are one thing:
+        // writes that may exist and are not here. Letting a failure through would
+        // push the tree this check exists to refuse.
+        if (synced === "incomplete" || synced === "failed") {
+            // Plain, rather than carrying the failure ladder: a git operation is
+            // somebody waiting, which is fresh evidence that the next attempt is worth
+            // making soon. The ladder rebuilds from there if the fault persists, and
+            // the rate is bounded by how often a model can call a tool.
+            await this.#sync.arm();
+            console.error(`[${this.#tag}] refusing git: the workspace is behind`, {
+                id: this.ctx.id.toString(),
+                outcome: synced
+            });
+            return {
+                ok: false,
+                message: "the container's most recent changes have not reached the workspace " +
+                    "yet, and git runs against the workspace — so this would act on a " +
+                    "tree that is missing them. The transfer resumes on its own; try " +
+                    "again in a moment."
+            };
+        }
+        await this.#ready();
+        return undefined;
+    }
+    /**
+     * Start this checkout's dependency install, and return without waiting.
+     *
+     * Called from `repo_clone` through the repo plugin's `afterCheckout` hook, so
+     * it runs inside a model turn — see `./install-job.ts` for why the drain's owner
+     * is the thing that matters here.
+     */
+    async startInstall(req) {
+        return await this.#install.start(req);
+    }
+    // --- lifecycle ---------------------------------------------------------------
+    /**
+     * Mark this workspace as in use, and push its reclamation back.
+     *
+     * Every entry point calls this, which is what makes the idle clock measure
+     * *use* rather than "when the agent last said this name". The agent hands a
+     * workspace name to a subagent once and then never sees the traffic; the
+     * workspace sees all of it.
+     */
+    async #touch() {
+        // A reclaim emptied this instance and its reset has not landed yet.
+        // Scheduling below would write to a table that is gone, so fail this call
+        // and let the caller come back on a fresh object.
+        if (this.#reclaimed) {
+            this.ctx.abort("workspace reclaimed", { retryAlarm: false });
+        }
+        const now = Date.now();
+        // Everything reaches this object by RPC, which bypasses `fetch` — so this
+        // is where the lifecycle gets started, and without it the scheduler's schema
+        // is never migrated and the first `set` below runs against nothing. Guarded
+        // internally, so calling it on every touch costs one resolved promise.
+        await this.#wake.start();
+        await this.ctx.storage.put("lastUsedAt", now);
+        // Two deadlines *moved*, not two schedules added. This is the hottest path
+        // in the object — every entry point calls it — so a bare `scheduler.set`
+        // here would leave one row per request, every one of them due.
+        await this.#idleReclaim.set(new Date(now + IDLE_RECLAIM_MS));
+        await this.#containerIdle.set(new Date(now + this.#containerIdleMs));
+    }
+    /**
+     * Throw this workspace away if nothing has touched it for `maxIdleMs`.
+     *
+     * Re-checks the clock rather than trusting the caller: the alarm may have been
+     * armed a week ago, and a use since then must win. Safe to call from anywhere
+     * for the same reason, which is what lets the agent's cron poke it as a
+     * backstop without needing to know anything.
+     */
+    async reclaimIfIdle(maxIdleMs = IDLE_RECLAIM_MS) {
+        const lastUsedAt = await this.ctx.storage.get("lastUsedAt");
+        const bytes = this.ctx.storage.sql.databaseSize;
+        // Nothing has ever used this object, so there is nothing to reclaim.
+        //
+        // Load-bearing, not defensive. `lastUsedAt` is written by `#touch()` and
+        // removed by the `deleteAll()` below, so an *already reclaimed* workspace
+        // reads exactly like a brand new one. Defaulting it to 0 makes both "idle
+        // since the epoch", and the weekly sweep then re-reclaims every workspace it
+        // has ever reclaimed, recreating storage just to empty it again.
+        if (lastUsedAt === undefined)
+            return { reclaimed: false, idleMs: 0, bytes };
+        const idleMs = Date.now() - lastUsedAt;
+        if (idleMs < maxIdleMs)
+            return { reclaimed: false, idleMs, bytes };
+        console.info(`[${this.#tag}] reclaiming an idle workspace`, {
+            id: this.ctx.id.toString(),
+            idleDays: Math.round(idleMs / 86_400_000),
+            bytes
+        });
+        await this.#stopContainer("reclaim");
+        // `deleteAll` does not take the alarm with it, so the alarm goes first —
+        // otherwise a reclaimed object wakes once more into empty storage.
+        await this.ctx.storage.deleteAlarm();
+        await this.ctx.storage.deleteAll();
+        // It *also* takes the lifecycle's job queue table, which this instance has
+        // already created and will not create again — so the next schedule set from
+        // this isolate would fail against a table that is gone, and the workspace
+        // would stop arming its timers until something evicted it. The answer is to
+        // stop using this instance: a later call constructs the object fresh on empty
+        // storage, and every table comes back with it.
+        //
+        // **The reset runs from its own alarm invocation, never from here.** An
+        // `abort()` on the way out of an RPC races that RPC's response, and the
+        // weekly sweep reads this result to decide whether to forget the workspace —
+        // so a reset that beat the response would report a failed reclaim for one
+        // that had already emptied the object. A timer would race it the same way;
+        // only a separate invocation is ordered after the response.
+        //
+        // Nothing durable records this. The flag exists to reset *this* isolate, and
+        // an isolate that went away on its own has already achieved that: a fresh
+        // instance creates the tables it needs on the way in.
+        this.#reclaimed = true;
+        await this.ctx.storage.setAlarm(Date.now());
+        return { reclaimed: true, idleMs, bytes };
+    }
+    /**
+     * Stop the container, keeping nothing. The workspace is what persists.
+     *
+     * Logged first because `@cloudflare/computer` reports this exit as
+     * `expected: false`: its own stop is not public, so a raw `destroy()` reads
+     * as a crash in its log.
+     */
+    async #stopContainer(reason) {
+        console.info(`[${this.#tag}] stopping the container`, {
+            id: this.ctx.id.toString(),
+            reason,
+            running: this.ctx.container?.running ?? false
+        });
+        try {
+            await this.ctx.container?.destroy();
+        }
+        catch (err) {
+            // Already gone, most likely, and a container that cannot be stopped must
+            // not turn a clean reclaim into a failed alarm.
+            console.warn(`[${this.#tag}] could not stop the container`, {
+                id: this.ctx.id.toString(),
+                err: String(err)
+            });
+        }
+        await this.#containerGone();
+    }
+    /**
+     * Whether this object is out of room, and by how much.
+     *
+     * Read on **two** paths, and the second is load-bearing. The install job
+     * consults it because an install is the operation that can move the number
+     * meaningfully. {@link advisories} consults it on every call, which is what
+     * lets a full workspace reach commands that have nothing to do with
+     * dependencies — the write being lost is rarely a dependency's, so a capacity
+     * fact delivered only alongside install state reaches everything except what
+     * it is about.
+     *
+     * Cheap enough for that: `databaseSize` is a local property read, not a query.
+     */
+    #storageHeadroom() {
+        const bytes = this.ctx.storage.sql.databaseSize;
+        if (bytes < STORAGE_CAP_BYTES)
+            return undefined;
+        return { bytes, capBytes: STORAGE_CAP_BYTES };
+    }
+    // --- the dependency install ------------------------------------------------
+    /**
+     * Record what was checked out here, and say whether it is actually visible.
+     *
+     * **The single writer of {@link CHECKOUT_KEY}**, called by the two paths that
+     * know a checkout landed: the repo plugin's `afterCheckout` hook, which fires
+     * only once a tree is established, and `scratch_open`, which fires once
+     * `git init` has exited 0.
+     *
+     * **Why the checkout is recorded apart from the install**, which call sites
+     * point here for rather than restating: an install is conditional and a
+     * checkout is not. `resolveInstallCommand` skips a checkout it finds nothing
+     * to install in — any without a `package.json` — so a path written only as
+     * part of an install is missing exactly there. What that costs is not an
+     * optimisation: {@link checkoutDir} is how a delegated session is told where
+     * to work, so a repository can clone perfectly, report itself correctly
+     * through the repo tools, and never be worked in. The install keeps its own
+     * context for the mirror-image reason — what to re-run on a cold container has
+     * a different lifetime from what is on disk.
+     *
+     * Returns the probe as well as the path, so a caller learns in one round trip
+     * whether the tree is visible rather than discovering it a delegation later.
+     */
+    async noteCheckout(req) {
+        await this.#touch();
+        const record = {
+            dir: req.dir,
+            ...(req.repo ? { repo: req.repo } : {}),
+            kind: req.kind,
+            at: Date.now()
+        };
+        await this.ctx.storage.put(CHECKOUT_KEY, record);
+        const present = await this.#isCheckout(req.dir);
+        console.info(`[${this.#tag}] checkout recorded`, {
+            id: this.ctx.id.toString(),
+            dir: req.dir,
+            kind: req.kind,
+            ...(req.repo ? { repo: req.repo } : {}),
+            present
+        });
+        return { dir: req.dir, present };
+    }
+    /**
+     * Where the work is — a directory holding a git repository, or nothing.
+     *
+     * Two callers depend on that being the meaning rather than "a path somebody
+     * wrote down once": a Claude Code session takes it as its cwd, and
+     * `discardWorkingTree` runs `git reset --hard` in it. Both want the same
+     * question answered, and neither can recover from a confident wrong answer.
+     *
+     * So it is **probed, not remembered**. A record is where to look; `.git` being
+     * there is what makes the answer true. That costs one local read of this
+     * object's own SQLite — everything under the checkout is durable, `.git`
+     * included — and it is what turns a stale record into `undefined` instead of
+     * into a session started in a directory that is no longer a checkout.
+     *
+     * The record is the only source. A workspace that has never recorded one
+     * answers `undefined` until its next checkout does, which `repo_clone` does on
+     * any tree it can fetch and reset.
+     */
+    async checkoutDir() {
+        const record = await this.ctx.storage.get(CHECKOUT_KEY);
+        const dir = record?.dir;
+        if (!dir)
+            return undefined;
+        if (await this.#isCheckout(dir))
+            return dir;
+        // Said out loud because the return value cannot say it: a caller reads
+        // `undefined` as "nothing has been cloned", which is also the right answer
+        // for a workspace nobody ever cloned into. Only this line separates them.
+        console.warn(`[${this.#tag}] a recorded checkout is no longer there`, {
+            id: this.ctx.id.toString(),
+            dir,
+            recorded: record?.kind,
+            // The rest of the record, because this line is read during an incident and
+            // "the checkout for acme/spike went missing forty minutes ago" is a
+            // different investigation from a bare path. It is also the only reader
+            // these two fields have.
+            ...(record?.repo ? { repo: record.repo } : {}),
+            ...(record ? { ageMs: Date.now() - record.at } : {})
+        });
+        return undefined;
+    }
+    /**
+     * Whether `dir` holds a git repository, as this object's own storage sees it.
+     *
+     * `.git` rather than the directory: an empty directory is not a checkout, and
+     * every caller of {@link checkoutDir} needs git to be there — to reset the
+     * tree, or to run a session that will commit in it. It is
+     * also the one probe a scratchpad and a clone answer identically, which is
+     * what lets them share every path below this line.
+     */
+    async #isCheckout(dir) {
+        try {
+            return await pathExists(this.#workspace.fs, `${dir}/.git`);
+        }
+        catch (err) {
+            // A read of local SQLite that threw says nothing about the tree. Treat it
+            // as absent: refusing a delegation costs a round, starting a session in a
+            // directory that may not exist costs the run.
+            console.warn(`[${this.#tag}] could not probe a checkout`, {
+                id: this.ctx.id.toString(),
+                dir,
+                err: String(err)
+            });
+            return false;
+        }
+    }
+    /**
+     * Everything currently true about this workspace that a caller must not assume
+     * away — the array `sb_exec`, `sb_write` and `sb_edit` all consult.
+     *
+     * The policy is not here. `deriveAdvisories` decides which facts matter and
+     * how they are worded; this method gathers what only the object can see and
+     * hands it over. That split is why a host cannot get the severity of its own
+     * workspace wrong.
+     *
+     * **Nothing that starts a long job belongs on this path**, and the rule is
+     * sharper here than anywhere else in the object because every tool call reads
+     * it. `startInstall` in particular must never be reached from here: it hands
+     * its drain to `ctx.waitUntil`, whose lifetime is the invocation's, and an
+     * invocation on this path is a tool call that returns in milliseconds. The
+     * drain outlives its owner, dies mid-`npm ci` with "WritableStream RPC stub
+     * was disposed without calling close()", and leaves a half-written tree.
+     *
+     * A missing dependency tree arms an install here, which runs in the alarm,
+     * which owns no request and outlives every RPC. Here because a command reads
+     * this before its `#ready()` starts a container, so the gate holds it for the
+     * install rather than letting it outrun one.
+     */
+    async advisories() {
+        if (!this.ctx.container?.running)
+            await this.#containerGone();
+        await this.#install.armIfTreeMissing();
+        const install = await this.#install.state();
+        const storage = this.#storageHeadroom();
+        // Only when the record says failed: it is the one state whose reading a
+        // queued repair changes, and every other one would pay a storage read for an
+        // answer nothing looks at.
+        const reinstallArmedAt = install.state === "failed"
+            ? await this.#install.reinstallArmedAt()
+            : undefined;
+        return deriveAdvisories({
+            install,
+            ...(storage ? { storage } : {}),
+            ...(reinstallArmedAt === undefined ? {} : { reinstallArmedAt }),
+            // The probe qualifies `deps-broken` and nothing else, and a queued
+            // reinstall turns a failed record into `deps-building`, which does not
+            // read it. `failed` alone is no longer the condition under which the
+            // question is worth asking, so the caller narrows it — see
+            // `InstallJob.treePresentIfItMatters`, whose contract this keeps true.
+            dependencyTreePresent: reinstallArmedAt === undefined
+                ? await this.#install.treePresentIfItMatters(install)
+                : false
+        });
+    }
+    /**
+     * `computerd`'s outbound WebSocket upgrade, on its way back in.
+     *
+     * The container dials the loopback rather than the other way round, so this
+     * object is the server for its own container's capnweb session. Everything
+     * else on this object is RPC; this is the only HTTP it speaks.
+     */
+    fetch(request) {
+        return this.backend.handleFetch(request);
+    }
+    /**
+     * Every durable wake-up this object has, dispatched from the one alarm.
+     *
+     * **This must not throw.** The runtime retries a failing alarm handler a
+     * bounded number of times and then stops for good — so a throw here would
+     * eventually take every *future* wake-up down with it, permanently, and the
+     * only symptom is that nothing ever happens again.
+     *
+     * Nothing else belongs here. Each reason to wake is a registered callback the
+     * scheduler dispatches by name, so this method neither matches on keys nor
+     * catches per-callback failures — the scheduler retries a failing callback and
+     * reports one that fails for good through `onError`. Nor does it sweep for a
+     * schedule that neither rescheduled nor cleared itself: a one-shot row is
+     * dropped when it runs, so it cannot stay due forever.
+     */
+    async alarm() {
+        // The reset a reclaim asked for, in the invocation it asked for it in.
+        // Nothing else may run on this instance: its storage is gone, and the job
+        // queue's table with it.
+        if (this.#reclaimed) {
+            this.ctx.abort("workspace reclaimed", { retryAlarm: false });
+            return;
+        }
+        try {
+            await this.#wake.alarm();
+        }
+        catch (err) {
+            // A reclaim that ran *from* this alarm emptied storage, the job queue with
+            // it, so settling the job that ran it fails. That is the reclaim working,
+            // and the reset it armed is the next thing to happen here.
+            if (this.#reclaimed)
+                return;
+            console.error(`[${this.#tag}] the alarm failed`, {
+                id: this.ctx.id.toString(),
+                err: String(err)
+            });
+            // One missed wake rather than every future one: re-arm from whatever
+            // survived, since the throw above is what the runtime counts against the
+            // bounded retry that ends in the alarm being abandoned.
+            await this.#wake.rearm().catch(() => { });
+        }
+    }
+    /**
+     * An outstanding pull came due.
+     *
+     * The three answers want three different next moves, which is why the drain
+     * distinguishes them. Progress comes straight back for the next block. A
+     * failure backs off, carrying its own count in the schedule's payload — the
+     * object does not survive between wake-ups, so there is nowhere else to keep
+     * it. Nothing reachable ends the attempt: a container that is gone took the
+     * unpulled writes with it.
+     */
+    async #onSyncDrain(payload) {
+        const outcome = await this.#sync.drain(SYNC_DRAIN_BUDGET_MS);
+        if (outcome === "incomplete") {
+            await this.#sync.arm();
+            return;
+        }
+        if (outcome === "failed") {
+            await this.#sync.arm((payload?.attempt ?? 0) + 1);
+        }
+    }
+    /** The idle-reclaim deadline came due. */
+    async #onIdleReclaim() {
+        const { reclaimed, idleMs } = await this.reclaimIfIdle();
+        // Not idle after all — something used it since this was armed, and that
+        // `#touch` already moved the deadline to a row of its own. So there is
+        // nothing to re-arm here, and re-arming would create a second one.
+        if (reclaimed)
+            return;
+        console.info(`[${this.#tag}] idle reclaim deferred`, {
+            id: this.ctx.id.toString(),
+            idleMinutes: Math.round(idleMs / 60_000)
+        });
+    }
+    /**
+     * The container-warm wake came due — see {@link #warmIfCold}.
+     *
+     * **Swallows its own failure.** Every other wake here is work somebody waits
+     * for; this one is early, so a container that will not start must cost the
+     * next `sb_exec` a start rather than put the scheduler's retry ladder behind a
+     * deployment fault that will refuse for as long as it stands.
+     */
+    async #onContainerWarm() {
+        // Started between the arming and now, so there is nothing to be early for.
+        // Trust it may still be missing is `#ready()`'s, on the next call that wants
+        // the container.
+        if (this.ctx.container?.running)
+            return;
+        try {
+            await this.#ready();
+            await this.#install.armIfTreeMissing();
+        }
+        catch (err) {
+            console.warn(`[${this.#tag}] could not warm the container`, {
+                id: this.ctx.id.toString(),
+                err: String(err)
+            });
+        }
+    }
+    /** The container-idle deadline came due. */
+    async #onContainerIdle(payload) {
+        // An install still running is "in use" even though nothing has called in —
+        // stopping the container under it would throw away the work and leave the
+        // gate closed until something noticed.
+        const state = await this.#install.read();
+        if (state.state === "running") {
+            await this.#containerIdle.set(new Date(Date.now() + this.#containerIdleMs));
+            return;
+        }
+        // Re-read the clock rather than trusting the alarm, the same way
+        // `reclaimIfIdle` does. `#touch` moves the deadline forward, but an alarm
+        // already in flight cannot be recalled — so without this a workspace that
+        // was used a second ago can still have its container stopped by a wake-up
+        // that was scheduled before that use.
+        const lastUsedAt = (await this.ctx.storage.get("lastUsedAt")) ?? 0;
+        const idleMs = Date.now() - lastUsedAt;
+        if (idleMs < this.#containerIdleMs) {
+            await this.#containerIdle.set(new Date(lastUsedAt + this.#containerIdleMs));
+            return;
+        }
+        /**
+         * The last chance to move anything the container still holds.
+         *
+         * Destroying it is what turns an outstanding pull from late into lost, and
+         * this is the only moment that is knowable in advance — every other way a
+         * container goes away is something happening *to* it. So the drain runs
+         * here, and the deadline defers while it is still making progress.
+         *
+         * The deferral has a wall, because a pull that cannot finish must not keep
+         * a container billing forever. Past it the container goes and the unpulled
+         * writes go with it, which is the honest outcome: they were unreachable
+         * either way, and the alternative is an idle container nobody stops.
+         */
+        if (await this.#drainBeforeStop(lastUsedAt, payload?.attempt ?? 0))
+            return;
+        await this.#stopContainer("idle");
+    }
+    /**
+     * Whether the idle deadline deferred to let an outstanding pull finish.
+     *
+     * `attempt` is how many times the drain has failed in a row, carried on the
+     * schedule because nothing in memory survives a wake-up.
+     */
+    async #drainBeforeStop(lastUsedAt, attempt) {
+        const outcome = await this.#sync.drain(SYNC_DRAIN_BUDGET_MS);
+        if (outcome === "complete" || outcome === "unavailable")
+            return false;
+        if (Date.now() - lastUsedAt > this.#containerIdleMs + SYNC_DRAIN_GRACE_MS) {
+            console.error(`[${this.#tag}] stopping a container with an unfinished pull`, {
+                id: this.ctx.id.toString(),
+                outcome,
+                graceMinutes: Math.round(SYNC_DRAIN_GRACE_MS / 60_000)
+            });
+            return false;
+        }
+        /**
+         * How soon to come back, and the two outcomes want different answers for the
+         * same reason they do on the drain's own schedule — except that here the
+         * deferral holds a **container** open, so getting it wrong bills as well as
+         * spins. `incomplete` is progress and clears the count; `failed` is a pull
+         * that threw, and returning in a second would put an alarm and an error line
+         * against a broken transport every second until the grace wall, half an hour
+         * later.
+         */
+        const failed = outcome === "failed";
+        const next = failed ? attempt + 1 : 0;
+        console.info(`[${this.#tag}] holding a container for an unfinished pull`, {
+            id: this.ctx.id.toString(),
+            outcome
+        });
+        await this.#containerIdle.set(new Date(Date.now() + (failed ? syncRetryDelayMs(next) : SYNC_DRAIN_RESUME_MS)), { attempt: next });
+        return true;
+    }
+}