@bevel-software/platform-core-backend 0.22.0 → 0.24.0

package/src/modules/workspace/agent-upload.store.ts

@@ -0,0 +1,668 @@
+import { createHash, randomBytes } from 'node:crypto';
+import { promises as fs } from 'node:fs';
+import path from 'node:path';
+import AdmZip from 'adm-zip';
+import { logger } from '../../shared/logging.js';
+import { MAX_UPLOAD_BYTES } from './upload-limits.js';
+
+const log = logger('agent-uploads');
+
+/** How long an issued token, and the bytes sent against it, stay usable. */
+export const UPLOAD_TOKEN_TTL_MS = 15 * 60 * 1000; // 15 minutes
+
+/** How often the store looks for tokens that have expired. */
+const SWEEP_INTERVAL_MS = 60 * 1000;
+
+/**
+ * How long past its expiry a CLAIMED record is pinned against the sweep.
+ *
+ * An apply holds its claim while it resolves the branch (which may clone),
+ * reads the bytes, judges every path and commits — work that can outlast a
+ * token whose TTL was nearly up when the apply started. Deleting the source
+ * mid-apply would make the commit land short with no refusal naming the
+ * reason, so a claim keeps its bytes alive. The grace is what stops that from
+ * being forever: an apply that neither consumed nor released within it has
+ * died with its process (the record is in memory, so there is no third
+ * possibility), and the bytes are reclaimed.
+ */
+const CLAIM_GRACE_MS = 10 * 60 * 1000; // 10 minutes
+
+/**
+ * THE one refusal for every way a token can fail to be usable: it was never
+ * issued, it has already been applied, it has expired, or it belongs to
+ * somebody else. One sentence for all four, deliberately — an answer that
+ * distinguished "no such token" from "not yours" would let anyone holding a
+ * guess learn which guesses exist, and the token is the whole credential this
+ * route has.
+ */
+export const UPLOAD_TOKEN_REFUSAL =
+  'That upload token cannot be used: it is unknown, already applied, expired, or was issued to someone else. ' +
+  'Call `request_file_upload` for a new one.';
+
+/**
+ * How many tokens one user may hold open at once.
+ *
+ * A token is permission to put the deployment's whole upload limit on its
+ * disk for a TTL, and to hold a connection open while it arrives. Without a
+ * bound, one caller asks for a hundred and sends against all of them
+ * together. Ten is far more than the route's own use needs — one token carries
+ * a zip of any number of files — and a token is given back the moment it is
+ * applied or expires.
+ */
+export const MAX_OPEN_UPLOADS_PER_USER = 10;
+
+/** The refusal an over-limit upload gets, naming the limit that applied. */
+export function overLimit(bytes: number, maxBytes: number): string {
+  return (
+    `That upload is ${bytes} bytes, over this deployment's ${maxBytes} byte upload limit. ` +
+    'Send a smaller file, or split it across several uploads.'
+  );
+}
+
+/** What an upload with no body is told. */
+const EMPTY_UPLOAD =
+  'That upload carried no bytes. Send the file as the request body — e.g. ' +
+  '`curl -X POST --data-binary @<file> "<uploadUrl>?filename=<name>"`.';
+
+/** A refusal with the HTTP status the upload route and the apply tool both answer. */
+export class UploadTokenError extends Error {
+  constructor(
+    message: string,
+    readonly status: number,
+  ) {
+    super(message);
+    this.name = 'UploadTokenError';
+  }
+}
+
+/** What `request_file_upload` answers: where to send the bytes, and the terms. */
+export interface IssuedUpload {
+  /** The absolute URL the bytes are POSTed to. Carries the token in its path. */
+  uploadUrl: string;
+  /**
+   * The token itself — what `apply_file_upload` takes, and the credential the
+   * upload route is authenticated by.
+   *
+   * Named separately from `uploadUrl` for two reasons: the apply needs it on
+   * its own, and the route takes it either way round. The bytes can go to
+   * `uploadUrl`, which carries the token in its last path segment, or to that
+   * address WITHOUT that segment with the token in the `x-upload-token`
+   * header — the spelling for a caller that would rather its credential not
+   * land in an access log or a shell history on the way.
+   */
+  token: string;
+  /** ISO-8601 instant after which the token and any bytes sent with it are gone. */
+  expiresAt: string;
+  /** Seconds from now until `expiresAt`, so a caller need not parse the date. */
+  expiresInSeconds: number;
+  /** The largest upload this deployment accepts, in bytes. */
+  maxBytes: number;
+}
+
+/** What the upload route answers, and what `apply_file_upload` reads back. */
+export interface ReceivedUpload {
+  /** The name the sender gave the file. For a single file, the name it lands under. */
+  filename: string;
+  bytes: number;
+  /** `zip` when the name ends in `.zip` and the bytes parse as an archive. */
+  kind: 'file' | 'zip';
+  /** Present for a zip: how many members the archive holds. */
+  entries?: number;
+}
+
+/** A received upload, with the bytes on disk, as `apply_file_upload` claims it. */
+export interface ClaimedUpload extends ReceivedUpload {
+  /** Absolute path of the stored bytes — OUTSIDE every workspace. */
+  absolutePath: string;
+}
+
+interface UploadRecord {
+  /** The file name under the store's root. Never derived from the sender's name. */
+  id: string;
+  userId: string;
+  expiresAt: number;
+  received?: ReceivedUpload;
+  /**
+   * Taken by an upload that is still writing its bytes. Set BEFORE the first
+   * `await` in {@link AgentUploadStore.receive}, so two uploads arriving at
+   * once cannot both pass the one-file-per-token check and then race each
+   * other's bytes onto the same path. Cleared only when a receive fails.
+   */
+  attaching: boolean;
+  /** Held by an apply that is running: a second apply finds the token in use. */
+  claimed: boolean;
+  /**
+   * When the apply holding this record claimed it. An expired record that is
+   * CLAIMED is pinned rather than swept — the apply is reading those bytes —
+   * until the claim itself looks abandoned (see {@link CLAIM_GRACE_MS}).
+   */
+  claimedAt?: number;
+}
+
+export interface AgentUploadStoreOptions {
+  /** Directory the bytes are written to. A sibling of `workspacesRoot`, never inside one. */
+  root: string;
+  /** Absolute base URL of this deployment's API, e.g. `https://core.example.com`. */
+  publicBaseUrl: string;
+  /** Token prefix, so a leaked credential can be recognised by shape. */
+  tokenPrefix?: string;
+  ttlMs?: number;
+  maxBytes?: number;
+  /** How many tokens one user may hold open at once. Defaults to {@link MAX_OPEN_UPLOADS_PER_USER}. */
+  maxOpenPerUser?: number;
+  /**
+   * How the staging root is listed. Test seam — defaults to `fs.readdir`.
+   *
+   * The ONE thing about the sweep a suite needs to control. Whether a token
+   * issued and uploaded while the sweep is running survives depends on the
+   * order of two steps inside {@link AgentUploadStore.sweepNow} — the listing
+   * and the live-id set — and the gap between them is a filesystem round-trip
+   * no test can otherwise sit inside. Given a listing it can hold open, a test
+   * can put a whole upload in that gap and assert the bytes are still there.
+   */
+  listRoot?: (root: string) => Promise<string[]>;
+}
+
+/**
+ * The upload tokens an agent lands files with, and the bytes sent against
+ * them.
+ *
+ * An agent that wants to put 27 files on a branch has no way to do it through
+ * a tool argument: the content would have to pass through the model, which
+ * truncates, mangles escape characters and cannot carry a PNG at all. So the
+ * bytes take a route of their own — this store issues a one-time token, the
+ * upload route attaches bytes to it, and `apply_file_upload` lands them.
+ *
+ * Three properties the whole design rests on, because the upload route is the
+ * one endpoint on this server that authenticates by a token alone:
+ *
+ *  - **Unguessable.** 32 random bytes, base64url. Nothing about the token is
+ *    derived from the user, the time or a counter.
+ *  - **Never stored in the clear.** Records are keyed by the token's SHA-256,
+ *    so the store's own memory (and anything that dumps it) holds no usable
+ *    credential.
+ *  - **Bound and single-use.** A record carries the id of the user it was
+ *    issued to and is consumed by the apply that lands it; a second apply, an
+ *    apply by anyone else, and an apply after the expiry all meet
+ *    {@link UPLOAD_TOKEN_REFUSAL}.
+ *
+ * And the bytes land in `root` — a directory beside the workspaces root, not
+ * inside one. No file tool can reach it: every workspace path is resolved
+ * against a branch's checkout, so there is no spelling of a tool argument that
+ * names a file in here. An upload nobody applies is deleted when its token
+ * expires (see {@link startSweeping}), so an abandoned drop costs disk for at
+ * most one TTL.
+ */
+export class AgentUploadStore {
+  private readonly records = new Map<string, UploadRecord>();
+  private readonly root: string;
+  private readonly publicBaseUrl: string;
+  private readonly tokenPrefix: string;
+  private readonly ttlMs: number;
+  private readonly listRoot: (root: string) => Promise<string[]>;
+  readonly maxBytes: number;
+  private readonly maxOpenPerUser: number;
+  private sweepTimer: ReturnType<typeof setInterval> | null = null;
+  /** The sweep in flight, so a shutdown can wait for it — see {@link drainSweep}. */
+  private sweeping: Promise<void> | null = null;
+  /** Set by {@link stopSweeping}: no further sweep deletes anything. */
+  private stopped = false;
+
+  constructor(options: AgentUploadStoreOptions) {
+    this.root = options.root;
+    this.publicBaseUrl = options.publicBaseUrl.replace(/\/+$/, '');
+    this.tokenPrefix = options.tokenPrefix ?? '';
+    this.ttlMs = options.ttlMs ?? UPLOAD_TOKEN_TTL_MS;
+    this.listRoot = options.listRoot ?? ((root) => fs.readdir(root));
+    this.maxBytes = options.maxBytes ?? MAX_UPLOAD_BYTES;
+    this.maxOpenPerUser = options.maxOpenPerUser ?? MAX_OPEN_UPLOADS_PER_USER;
+  }
+
+  /** The URL a token's bytes are sent to. The one spelling of this route's address. */
+  uploadUrlFor(token: string): string {
+    return `${this.publicBaseUrl}/api/agent/uploads/${encodeURIComponent(token)}`;
+  }
+
+  /**
+   * Mint a token for `user` and answer the terms it is good for. Nothing is
+   * written to disk yet — a token nobody uploads against costs one map entry
+   * until the sweep drops it.
+   */
+  issue(user: { id: string }): IssuedUpload {
+    // Counted over what the user still HOLDS: a token that can still be used,
+    // and one whose expiry has passed but whose work has not ended — an upload
+    // still arriving, an apply still reading. Those keep a connection or a file
+    // open exactly as a live token does, so letting them fall out of the count
+    // at expiry would let a caller keep ten slow uploads going and ask for ten
+    // more. An expired record that is doing nothing is waiting for the sweep,
+    // and is not one the user holds.
+    const now = Date.now();
+    let open = 0;
+    for (const record of this.records.values()) {
+      if (record.userId !== user.id) continue;
+      if (record.expiresAt > now || this.pinned(record, now)) open += 1;
+    }
+    if (open >= this.maxOpenPerUser) {
+      throw new UploadTokenError(
+        `You already hold ${open} upload tokens, the most one user may have open at once. Apply one of them with ` +
+          '`apply_file_upload`, or wait for one to expire, then ask again. One token carries a zip of any number of files.',
+        429,
+      );
+    }
+    const token = this.tokenPrefix + randomBytes(32).toString('base64url');
+    const expiresAt = Date.now() + this.ttlMs;
+    this.records.set(hash(token), {
+      id: `upload-${Date.now()}-${randomBytes(8).toString('hex')}`,
+      userId: user.id,
+      expiresAt,
+      attaching: false,
+      claimed: false,
+    });
+    // AFTER the record exists, not before. The first sweep runs the moment the
+    // timer starts, and started from an empty map it would be a sweep that
+    // believes nothing is live — this token's own id included.
+    this.startSweeping();
+    return {
+      uploadUrl: this.uploadUrlFor(token),
+      token,
+      expiresAt: new Date(expiresAt).toISOString(),
+      expiresInSeconds: Math.round(this.ttlMs / 1000),
+      maxBytes: this.maxBytes,
+    };
+  }
+
+  /**
+   * Store the bytes of `body` against `token` and say what was received. One
+   * file per token: a second upload against the same token is refused, so a
+   * token cannot be used to keep replacing bytes an apply is about to land.
+   *
+   * WRITTEN AS THEY ARRIVE, never gathered first. The body is up to the
+   * deployment's whole upload limit, and this is the one route a caller can
+   * send to without a session: held in memory, a handful of uploads at once
+   * cost the process several times that limit each (the chunks, the buffer
+   * they were joined into, the archive read over it). Streamed, an upload in
+   * flight costs one chunk.
+   *
+   * A name ending in `.zip` is read as an archive HERE, once the bytes are on
+   * disk, so the answer can carry the entry count and so a corrupt archive is
+   * refused while the caller is still holding the file — rather than at apply
+   * time, when its token would already be spent.
+   */
+  async receive(
+    token: string,
+    filename: string,
+    body: AsyncIterable<Buffer | string>,
+  ): Promise<ReceivedUpload> {
+    const record = this.openRecord(token);
+    // RESERVED FIRST, before any `await`: two uploads arriving at once would
+    // otherwise both find the token open, both write, and the apply would land
+    // whichever set of bytes finished last — against a token whose answer
+    // described the other one.
+    record.attaching = true;
+    const stored = path.join(this.root, record.id);
+    let bytes = 0;
+    const received: ReceivedUpload = { filename, bytes: 0, kind: 'file' };
+    try {
+      await fs.mkdir(this.root, { recursive: true });
+      const file = await fs.open(stored, 'w');
+      try {
+        for await (const chunk of body) {
+          const buf = typeof chunk === 'string' ? Buffer.from(chunk) : chunk;
+          bytes += buf.byteLength;
+          // The real total, counted as it arrives: a `content-length` is the
+          // sender's claim, and a chunked body makes none.
+          if (bytes > this.maxBytes) throw new UploadTokenError(overLimit(bytes, this.maxBytes), 413);
+          // To the last byte: one `write` may put down only part of what it
+          // was handed and say so in `bytesWritten`, and a chunk counted as
+          // received while half of it is on disk would land a file shorter
+          // than the size the answer names.
+          for (let written = 0; written < buf.byteLength; ) {
+            written += (await file.write(buf, written, buf.byteLength - written)).bytesWritten;
+          }
+        }
+      } finally {
+        await file.close();
+      }
+      if (bytes === 0) throw new UploadTokenError(EMPTY_UPLOAD, 400);
+      received.bytes = bytes;
+      if (filename.toLowerCase().endsWith('.zip')) {
+        try {
+          received.entries = new AdmZip(stored).getEntries().length;
+        } catch (err) {
+          throw new UploadTokenError(
+            `"${filename}" is not a readable .zip archive: ${err instanceof Error ? err.message : String(err)}`,
+            422,
+          );
+        }
+        received.kind = 'zip';
+      }
+    } catch (err) {
+      // Nothing was received, so the token is open again: the sender may retry
+      // with the same one rather than ask for another. Whatever part of the
+      // body reached the disk goes now, rather than at the next sweep.
+      await this.remove(record.id);
+      record.attaching = false;
+      throw err;
+    }
+    // THE TTL MAY HAVE PASSED while the bytes were arriving — a 40 MB upload
+    // over a slow link takes real time, and the token was issued before it
+    // started. The record was pinned against the sweep throughout (see
+    // {@link pinned}), so neither it nor the file could be deleted under the
+    // write; but an expired token cannot be applied, so answering "received"
+    // would hand the sender a success it can do nothing with. The refusal is
+    // the honest answer, and the bytes go with it rather than waiting for a
+    // sweep to notice.
+    if (record.expiresAt <= Date.now()) {
+      this.records.delete(hash(token));
+      await this.remove(record.id);
+      throw new UploadTokenError(UPLOAD_TOKEN_REFUSAL, 404);
+    }
+    record.received = received;
+    return received;
+  }
+
+  /**
+   * Refuse `token` if it cannot accept bytes — unknown, expired, somebody
+   * else's doing, or already holding a file — and answer its record if it can.
+   *
+   * Exists as its own step so the upload route can ask BEFORE it reads the
+   * body. The route is the one endpoint here authenticated by a token alone,
+   * and buffering up to the deployment's whole upload limit for an invented
+   * token would let a handful of concurrent requests spend the process's
+   * memory on bytes that were never going to be stored.
+   */
+  assertOpen(token: string): void {
+    this.openRecord(token);
+  }
+
+  /** {@link assertOpen}, with the record it found — the store's own view of it. */
+  private openRecord(token: string): UploadRecord {
+    const record = this.find(token);
+    if (record.received !== undefined || record.attaching) {
+      throw new UploadTokenError(UPLOAD_TOKEN_REFUSAL, 404);
+    }
+    return record;
+  }
+
+  /**
+   * Hold `token` for an apply by `userId`, and answer where its bytes are.
+   *
+   * A CLAIM rather than a consume, because an apply can be refused whole —
+   * the destination is on a protected branch the caller may not write, the
+   * archive turned out unreadable — and a token spent on a refusal would make
+   * the agent send the same 40 MB again to try a different destination. The
+   * claim is what keeps it single-use meanwhile: a second apply arriving while
+   * the first runs finds the token in use and is refused. The caller
+   * {@link consume}s it once an answer exists, or {@link release}s it on a
+   * refusal that landed nothing.
+   */
+  claim(token: string, userId: string): ClaimedUpload {
+    const record = this.find(token);
+    if (record.userId !== userId || record.claimed || record.received === undefined) {
+      throw new UploadTokenError(UPLOAD_TOKEN_REFUSAL, 404);
+    }
+    record.claimed = true;
+    record.claimedAt = Date.now();
+    return { ...record.received, absolutePath: path.join(this.root, record.id) };
+  }
+
+  /** Give a claimed token back, unused — the apply refused without landing anything. */
+  release(token: string): void {
+    const record = this.records.get(hash(token));
+    if (record) {
+      record.claimed = false;
+      record.claimedAt = undefined;
+    }
+  }
+
+  /** Spend the token and delete its bytes. Idempotent. */
+  async consume(token: string): Promise<void> {
+    const key = hash(token);
+    const record = this.records.get(key);
+    if (!record) return;
+    this.records.delete(key);
+    await this.remove(record.id);
+  }
+
+  /**
+   * Delete every record whose token has expired, with the bytes it was
+   * holding — then delete any file in the root that no LIVE record claims and
+   * that has outlived every token that could name it.
+   *
+   * The second half is what makes the first one true. A process killed between
+   * the write and the apply leaves a file no map will ever mention again, and
+   * sweeping by what the records DON'T name is the honest reading of the
+   * promise: an upload nobody applied does not stay. By AGE rather than at
+   * once, because a file this map does not name may be another process's (see
+   * the loop below).
+   *
+   * ONE record is kept past its expiry: one an apply has CLAIMED. Those bytes
+   * are being read right now, and a sweep that deleted them would make the
+   * commit land short of what the answer promised. The pin lasts
+   * {@link CLAIM_GRACE_MS}, after which a claim nobody consumed or released
+   * belongs to a dead process and is reclaimed.
+   */
+  async sweepNow(): Promise<void> {
+    if (this.stopped) return;
+    const now = Date.now();
+    for (const [key, record] of [...this.records]) {
+      if (record.expiresAt > now) continue;
+      if (this.pinned(record, now)) continue;
+      this.records.delete(key);
+      await this.remove(record.id);
+    }
+    let names: string[];
+    try {
+      names = await this.listRoot(this.root);
+    } catch {
+      return; // root not created yet, or unreadable — nothing to reclaim
+    }
+    // Every id a live record is holding — including one whose bytes have not
+    // arrived yet, so an upload in flight is never swept out from under itself,
+    // and one an apply has pinned past its expiry.
+    //
+    // Read AFTER the directory, never before. The gap between the two is a
+    // real one — the readdir is a filesystem round-trip, and this process
+    // serves other requests across it — so a token issued and uploaded during
+    // that gap would appear in `names` while a set taken earlier had never
+    // heard of it, and the sweep would delete bytes somebody had just been
+    // told were received. Taken afterwards, the set is a superset of what the
+    // listing could possibly name.
+    const live = new Set([...this.records.values()].map((r) => r.id));
+    for (const name of names) {
+      // Asked again per name: a stop (an evicted tenant, a shutdown) that
+      // landed while this sweep was reading the directory must not go on to
+      // delete files a replacement store may already have issued ids for.
+      if (this.stopped) return;
+      if (live.has(name)) continue;
+      // A file this map does not name is not necessarily nobody's. The records
+      // are ONE PROCESS's memory, and two processes share this directory
+      // whenever a deployment restarts by starting the new one before the old
+      // one has stopped: to each, the other's uploads are files no record
+      // names. So such a file goes only once it is older than any token could
+      // still be good for — its TTL, and the grace an apply holding it is
+      // given. A process that died leaves its files for that long and no
+      // longer; a process that is alive never loses one in use.
+      if (await this.outlivedEveryToken(name, now)) await this.remove(name);
+    }
+  }
+
+  /**
+   * Whether the stored file `name` is older than any token that could still be
+   * naming it, in this process or another. A file that cannot be examined is
+   * left for the next sweep rather than judged.
+   */
+  private async outlivedEveryToken(name: string, now: number): Promise<boolean> {
+    try {
+      const { mtimeMs } = await fs.stat(path.join(this.root, name));
+      return now - mtimeMs > this.ttlMs + CLAIM_GRACE_MS;
+    } catch {
+      return false;
+    }
+  }
+
+  /**
+   * Whether an expired record is held open by work that is still running:
+   * an UPLOAD still writing its bytes, or an APPLY still reading them.
+   *
+   * Both windows can outlast a TTL — a large upload over a slow link, an apply
+   * that clones a branch before it reads — and in both the record's own file is
+   * being written or read right now. A sweep that deleted either would leave a
+   * file no map mentions (the upload writes after the delete) or an apply
+   * reading a path that has gone. The apply's pin has a grace, because a claim
+   * can be abandoned by a process that dies; the upload's needs none, because
+   * `receive` always ends — it clears `attaching` on failure and sets
+   * `received` on success, and one whose TTL passed meanwhile deletes the
+   * record itself rather than leaving it pinned.
+   */
+  private pinned(record: UploadRecord, now: number): boolean {
+    if (record.attaching && record.received === undefined) return true;
+    return record.claimed && now - (record.claimedAt ?? now) < CLAIM_GRACE_MS;
+  }
+
+  /**
+   * Sweep now, and keep sweeping. Started by the first {@link issue} rather
+   * than at boot, so a deployment nobody uploads to runs no timer; idempotent,
+   * and the timer is `unref`'d because nothing here is worth keeping a process
+   * alive for — the records are in memory and go with it.
+   *
+   * The first sweep runs IMMEDIATELY, not one interval later, because the
+   * records are in memory: a process that restarted holds no record of what
+   * the previous one stored, and what that one left behind long enough ago
+   * has no reason to wait a further interval.
+   */
+  startSweeping(intervalMs: number = SWEEP_INTERVAL_MS): void {
+    if (this.sweepTimer) return;
+    this.stopped = false;
+    this.sweep();
+    this.sweepTimer = setInterval(() => this.sweep(), intervalMs);
+    this.sweepTimer.unref?.();
+  }
+
+  /**
+   * Stop sweeping, for good: the timer is cleared AND a sweep already running
+   * abandons the rest of its work.
+   *
+   * Both halves matter when a graph is stopped — a tenant evicted, the process
+   * shutting down. The root is this tenant's, and a reactivation builds a new
+   * store over the same directory with an empty record map: a sweep left
+   * running from the old store would find the new store's files named by no
+   * record of ITS own and delete them, under an apply that is about to read
+   * them. {@link drainSweep} is how a caller waits for the abandonment to
+   * actually have happened.
+   */
+  stopSweeping(): void {
+    this.stopped = true;
+    if (this.sweepTimer) clearInterval(this.sweepTimer);
+    this.sweepTimer = null;
+  }
+
+  /** Wait for the sweep in flight, if any. Pairs with {@link stopSweeping} on shutdown. */
+  async drainSweep(): Promise<void> {
+    await this.sweeping;
+  }
+
+  /** One sweep, with its failure logged and its promise kept for {@link drainSweep}. */
+  private sweep(): void {
+    const running = this.sweepNow()
+      .catch((err: unknown) => {
+        log.warn('could not sweep expired uploads:', { err });
+      })
+      .finally(() => {
+        if (this.sweeping === running) this.sweeping = null;
+      });
+    this.sweeping = running;
+  }
+
+  /** The live record for `token`, or the one refusal. Expiry is judged here. */
+  private find(token: string): UploadRecord {
+    const key = hash(token);
+    const record = this.records.get(key);
+    if (!record) throw new UploadTokenError(UPLOAD_TOKEN_REFUSAL, 404);
+    const now = Date.now();
+    if (record.expiresAt <= now) {
+      // Expired either way — but a record an apply is still reading keeps its
+      // bytes (and its map entry) until that apply ends, for the reason
+      // {@link sweepNow} gives. The refusal is the same; only the deletion waits.
+      if (!this.pinned(record, now)) {
+        this.records.delete(key);
+        void this.remove(record.id);
+      }
+      throw new UploadTokenError(UPLOAD_TOKEN_REFUSAL, 404);
+    }
+    return record;
+  }
+
+  /** Best-effort delete of one stored upload. Never throws. */
+  private async remove(id: string): Promise<void> {
+    try {
+      await fs.rm(path.join(this.root, id), { force: true });
+    } catch (err) {
+      log.warn(`could not delete the stored upload "${id}":`, { err });
+    }
+  }
+}
+
+function hash(token: string): string {
+  return createHash('sha256').update(token).digest('hex');
+}
+
+/**
+ * Refuse a boot whose upload root is not OUTSIDE every workspace.
+ *
+ * The whole safety of this route rests on where the bytes land: they are a
+ * buffer somebody sent, judged by nothing until `apply_file_upload` judges each
+ * path against a branch. A root configured inside `workspacesRoot` would put
+ * that unjudged buffer where the file tools read — `read_file`, `grep`, the
+ * download route — bypassing the access and platform-file rules the apply
+ * exists to apply. The invariant is documented on `AGENT_UPLOADS_ROOT`; this is
+ * it checked, at boot, naming the variable, rather than trusted.
+ *
+ * Both directions and both spellings: equal paths, either containing the
+ * other, and the real paths, so a root that is a LINK into the workspaces tree
+ * is refused too.
+ */
+export async function assertUploadsRootOutsideWorkspaces(
+  uploadsRoot: string,
+  workspacesRoot: string,
+): Promise<void> {
+  for (const [uploads, workspaces] of [
+    [path.resolve(uploadsRoot), path.resolve(workspacesRoot)],
+    [await realBase(uploadsRoot), await realBase(workspacesRoot)],
+  ]) {
+    if (uploads === workspaces || contains(workspaces, uploads) || contains(uploads, workspaces)) {
+      throw new Error(
+        `AGENT_UPLOADS_ROOT ("${uploadsRoot}") must be outside WORKSPACES_ROOT ("${workspacesRoot}"): uploaded ` +
+          'bytes are staged there before any access or platform-file rule has judged them, so a root inside a ' +
+          'workspace would let the file tools read them. Point it at a sibling directory.',
+      );
+    }
+  }
+}
+
+/** Whether `child` is inside `parent`. Paths already resolved. */
+function contains(parent: string, child: string): boolean {
+  return child.startsWith(parent.endsWith(path.sep) ? parent : parent + path.sep);
+}
+
+/**
+ * `dir` with every link on it resolved — as far as it exists. Neither root is
+ * required to exist yet (the store makes its own on the first upload), so the
+ * deepest existing ancestor is resolved and the missing rest joined back on:
+ * a link anywhere along the part that DOES exist is what could redirect the
+ * one into the other.
+ */
+async function realBase(dir: string): Promise<string> {
+  let existing = path.resolve(dir);
+  const rest: string[] = [];
+  for (;;) {
+    try {
+      return path.join(await fs.realpath(existing), ...rest);
+    } catch {
+      const up = path.dirname(existing);
+      if (up === existing) return path.join(existing, ...rest);
+      rest.unshift(path.basename(existing));
+      existing = up;
+    }
+  }
+}