@bevel-software/platform-core-backend 0.15.1 → 0.20.0

package/src/modules/plugins/__tests__/join-request-jobs.service.test.ts

@@ -0,0 +1,608 @@
+import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
+import { DEFAULT_BRANCH, joinBranchFor } from '@bevel-software/platform-shared';
+import {
+  PluginJoinRequestJobs,
+  JoinRequestNotReadyError,
+  type JoinRequestJobsDeps,
+  CLAIM_STALE_AFTER_MS,
+} from '../join-request-jobs.service.js';
+import { workspaceIdForBranch } from '../../../shared/workspace-id.js';
+import { FakeJoinRequestStore } from './fake-join-request-store.js';
+
+/**
+ * The git half of a join request, on its own: what it does to the row it was
+ * handed, and what it refuses to do twice.
+ *
+ * The route's own tests cover the pairing — one click, one row, one change
+ * request. What is here is the part a route cannot reach: two flights racing,
+ * a key nothing answers to any more, and the words a failure leaves behind.
+ */
+
+const ALI = 'ali@bevel.software';
+const ALI_USER = { id: 'u-1', email: ALI, name: 'Ali Baba' };
+const KB = 'knowledge-base';
+const ACCESS_MD = '---\nread:\n  - everyone\n---\nread: []\n';
+
+function harness(over: Partial<JoinRequestJobsDeps> = {}, sharedStore?: FakeJoinRequestStore) {
+  const workflow = {
+    listChangeRequestsAuthoredBy: vi.fn(async () => [] as never[]),
+    createBranch: vi.fn(async () => ({ name: 'x', isDefault: false, isProtected: false })),
+    listBranches: vi.fn(async () => [] as { name: string }[]),
+    commitChanges: vi.fn(async () => null),
+    openChangeRequest: vi.fn(async () => ({ number: 42 })),
+  };
+  const workspaceService = {
+    getOrCreateForBranch: async (branch: string) => ({ id: workspaceIdForBranch(branch) }),
+    readFile: vi.fn(async () => ACCESS_MD),
+    writeFile: vi.fn(async () => undefined),
+  };
+  const store = sharedStore ?? new FakeJoinRequestStore();
+  const jobs = new PluginJoinRequestJobs(store, {
+    workflow: workflow as never,
+    workspaceService: workspaceService as never,
+    kbDirName: KB,
+    target: async () => ({ folder: 'Plugins/Finance', displayName: 'Finance' }),
+    requester: async (mail) => ({ ...ALI_USER, email: mail }),
+    ...over,
+  });
+  return { workflow, workspaceService, store, jobs };
+}
+
+function pendingRow(store: FakeJoinRequestStore) {
+  return store.seed({
+    requesterEmail: ALI,
+    requesterName: 'Ali Baba',
+    pluginKey: 'Finance',
+    status: 'pending',
+    failureReason: null,
+    changeRequestNumber: null,
+    claimedAt: null,
+  });
+}
+
+describe('PluginJoinRequestJobs', () => {
+  // Silenced for the whole file and restored after EVERY test, including one
+  // that fails an assertion partway: a `mockRestore()` on the success path
+  // leaves console.error stubbed for every later test in the file the moment
+  // anything above it throws, which hides exactly the output you need to see.
+  let consoleError: ReturnType<typeof vi.spyOn>;
+  beforeEach(() => {
+    consoleError = vi.spyOn(console, 'error').mockImplementation(() => {});
+  });
+  afterEach(() => {
+    consoleError.mockRestore();
+    // The same discipline, for the same reason. Several tests below install
+    // fake timers and move the system clock minutes forward; a
+    // `useRealTimers()` on the success path leaves both in place for every
+    // later test in the file the moment an assertion above it fails, and a
+    // shifted clock makes the claim-window tests fail in ways that have
+    // nothing to do with what they are testing. Harmless when no test
+    // installed them.
+    vi.useRealTimers();
+  });
+
+  it('runs the same steps the endpoint used to run inline, and records the number', async () => {
+    const h = harness();
+    await h.jobs.start(pendingRow(h.store));
+
+    const branch = joinBranchFor(ALI, 'Finance');
+    expect(h.workflow.createBranch).toHaveBeenCalledWith(
+      workspaceIdForBranch(DEFAULT_BRANCH),
+      branch,
+      DEFAULT_BRANCH,
+    );
+    const write = h.workspaceService.writeFile.mock.calls[0] as unknown as [string, string, string];
+    expect(write[1]).toBe(`${KB}/Plugins/Finance/access.md`);
+    expect(write[2]).toContain('Ali Baba <ali@bevel.software>');
+    expect(h.workflow.commitChanges).toHaveBeenCalled();
+    expect(h.workflow.openChangeRequest).toHaveBeenCalledWith(
+      workspaceIdForBranch(branch),
+      expect.objectContaining({ email: ALI }),
+      expect.objectContaining({ title: 'Join request: Finance', sourceBranch: branch }),
+    );
+    expect(h.store.all()).toMatchObject([{ status: 'opened', changeRequestNumber: 42 }]);
+  });
+
+  it('two starts on one record are one flight, and one change request', async () => {
+    const h = harness();
+    const record = pendingRow(h.store);
+    await Promise.all([h.jobs.start(record), h.jobs.start(record)]);
+    expect(h.workflow.openChangeRequest).toHaveBeenCalledTimes(1);
+  });
+
+  it('a start on a record another flight already opened does nothing at all', async () => {
+    const h = harness();
+    const record = pendingRow(h.store);
+    await h.jobs.start(record);
+    // The SAME stale snapshot a second tab would hold: still `pending` in
+    // this object, `opened` in the store.
+    await h.jobs.start(record);
+    expect(h.workflow.openChangeRequest).toHaveBeenCalledTimes(1);
+  });
+
+  it('leaves the branch alone when the grant is already on it', async () => {
+    const h = harness();
+    h.workspaceService.readFile.mockResolvedValue(
+      `---\nread:\n  - everyone\n---\nread:\n  - Ali Baba <ali@bevel.software>\n`,
+    );
+    await h.jobs.start(pendingRow(h.store));
+    expect(h.workspaceService.writeFile).not.toHaveBeenCalled();
+    expect(h.workflow.commitChanges).not.toHaveBeenCalled();
+    // Still opens the request: the branch carries the proposal either way.
+    expect(h.workflow.openChangeRequest).toHaveBeenCalledTimes(1);
+  });
+
+  it('records a failure in the words the git work used, with credentials stripped', async () => {
+    const h = harness();
+    h.workflow.openChangeRequest.mockRejectedValue(
+      new Error("fatal: unable to access 'https://x-access-token:ghp_secret@github.com/acme/kb'"),
+    );
+    await h.jobs.start(pendingRow(h.store));
+
+    const [row] = h.store.all();
+    expect(row.status).toBe('failed');
+    expect(row.failureReason).toContain('unable to access');
+    expect(row.failureReason).not.toContain('ghp_secret');
+  });
+
+  it('fails the record rather than throwing when the plugin is gone', async () => {
+    const h = harness({ target: async () => null });
+    await expect(h.jobs.start(pendingRow(h.store))).resolves.toBeUndefined();
+    expect(h.store.all()).toMatchObject([
+      { status: 'failed', failureReason: 'the plugin is no longer available' },
+    ]);
+    expect(h.workflow.createBranch).not.toHaveBeenCalled();
+  });
+
+  it('fails the record when no account answers to the address that asked', async () => {
+    const h = harness({ requester: async () => null });
+    await h.jobs.start(pendingRow(h.store));
+    expect(h.store.all()).toMatchObject([
+      { status: 'failed', failureReason: 'the account that asked no longer exists' },
+    ]);
+  });
+
+  it('opens nothing for an account erased and made again under the same address', async () => {
+    let asked = 0;
+    const h = harness({
+      requester: async (mail) => ({ ...ALI_USER, id: asked++ === 0 ? 'u-1' : 'u-2', email: mail }),
+    });
+    await h.jobs.start(pendingRow(h.store));
+    expect(h.workflow.openChangeRequest).not.toHaveBeenCalled();
+    expect(h.store.all()).toMatchObject([{ status: 'failed' }]);
+  });
+
+  it('starts nothing once the sweep is stopped — the row waits for the next boot', async () => {
+    const h = harness();
+    h.jobs.stopSweeping();
+    await h.jobs.start(pendingRow(h.store));
+    expect(h.workflow.createBranch).not.toHaveBeenCalled();
+    expect(h.store.all()).toMatchObject([{ status: 'pending' }]);
+  });
+
+  it('opens nothing for an account erased while the branch was being pushed', async () => {
+    // The account answers when the job starts and is gone by the time the
+    // change request would open — the erasure landed during the git work.
+    let asked = 0;
+    const h = harness({
+      requester: async (mail) => (asked++ === 0 ? { ...ALI_USER, email: mail } : null),
+    });
+    await h.jobs.start(pendingRow(h.store));
+    expect(h.workflow.commitChanges).toHaveBeenCalled();
+    expect(h.workflow.openChangeRequest).not.toHaveBeenCalled();
+    expect(h.store.all()).toMatchObject([
+      { status: 'failed', failureReason: 'the account that asked no longer exists' },
+    ]);
+  });
+
+  it('sweeps only the pending rows, and leaves the settled ones alone', async () => {
+    const h = harness();
+    pendingRow(h.store);
+    h.store.seed({
+      requesterEmail: 'mia@bevel.software',
+      requesterName: 'Mia',
+      pluginKey: 'GTM',
+      status: 'opened',
+      failureReason: null,
+      changeRequestNumber: 7,
+      claimedAt: null,
+    });
+    h.store.seed({
+      requesterEmail: 'mia@bevel.software',
+      requesterName: 'Mia',
+      pluginKey: 'Ops',
+      status: 'failed',
+      failureReason: 'the remote refused the push',
+      changeRequestNumber: null,
+      claimedAt: null,
+    });
+
+    await h.jobs.sweep();
+    await h.jobs.drain();
+
+    expect(h.workflow.openChangeRequest).toHaveBeenCalledTimes(1);
+    expect(h.store.all().map((r) => r.status)).toEqual(['opened', 'opened', 'failed']);
+  });
+
+  it('the sweep returns without waiting for the git it started', async () => {
+    const h = harness();
+    // A first request from a person is a full clone; the boot must not queue
+    // behind one, so `sweep` resolves once the work is under way.
+    let release = () => {};
+    h.workflow.createBranch.mockReturnValue(
+      new Promise((resolve) => {
+        release = () => resolve({ name: 'x', isDefault: false, isProtected: false });
+      }) as never,
+    );
+    pendingRow(h.store);
+
+    await h.jobs.sweep();
+    expect(h.store.all()).toMatchObject([{ status: 'pending' }]);
+
+    release();
+    await h.jobs.drain();
+    expect(h.store.all()).toMatchObject([{ status: 'opened' }]);
+  });
+
+  it('does not run a record another process holds — the redeploy overlap', async () => {
+    const h = harness();
+    const record = pendingRow(h.store);
+    // What the OTHER server in a redeploy window leaves behind: the row is
+    // still pending, and its claim is live (it is beating).
+    await h.store.claim(record.id, CLAIM_STALE_AFTER_MS);
+
+    await h.jobs.start({ ...record });
+
+    // Not "opened twice" — not started at all. Two servers cloning and
+    // pushing the same branch into one shared workspace is the thing the
+    // claim exists to prevent, and neither one's single-flight map can see
+    // the other.
+    expect(h.workflow.createBranch).not.toHaveBeenCalled();
+    expect(h.workflow.openChangeRequest).not.toHaveBeenCalled();
+    expect(h.store.all()[0].status).toBe('pending');
+  });
+
+  it('takes over a claim left by a process that died holding it', async () => {
+    const h = harness();
+    const record = pendingRow(h.store);
+    await h.store.claim(record.id, CLAIM_STALE_AFTER_MS);
+    // The dead process stopped beating; the window lapses. `setSystemTime`
+    // rather than a `Date.now` spy, because the store stamps claims with
+    // `new Date()` and a spy on `now` leaves that reading the real clock.
+    vi.useFakeTimers();
+    vi.setSystemTime(Date.now() + CLAIM_STALE_AFTER_MS + 1000);
+
+    await h.jobs.start({ ...record });
+
+    expect(h.store.all()[0]).toMatchObject({ status: 'opened', changeRequestNumber: 42 });
+  });
+
+  it('a request claimed the instant before a crash is finished by the next boot', async () => {
+    // THE AC5 REGRESSION, exactly as local testing reproduced it: the click is
+    // answered, `run` claims the row, and the process is killed a moment
+    // later. The row is left pending with a claim seconds old.
+    const h = harness();
+    const record = pendingRow(h.store);
+    await h.store.claim(record.id, CLAIM_STALE_AFTER_MS);
+    expect(h.store.all()[0]).toMatchObject({ status: 'pending', changeRequestNumber: null });
+
+    // A boot takes a couple of minutes. The claim was sized to cover a clone
+    // — fifteen minutes — so the boot sweep used to be refused this very row
+    // and return silently, leaving the request owed forever. The window is
+    // three missed heartbeats now, and a dead process misses all of them.
+    vi.useFakeTimers();
+    vi.setSystemTime(Date.now() + 2 * 60 * 1000);
+
+    const fresh = harness({}, h.store);
+    await fresh.jobs.sweep();
+    await fresh.jobs.drain();
+
+    expect(h.store.all()[0]).toMatchObject({ status: 'opened', changeRequestNumber: 42 });
+    expect(fresh.workflow.openChangeRequest).toHaveBeenCalledTimes(1);
+  });
+
+  it('a live process keeps its claim however long the work takes, by its own heartbeat', async () => {
+    // The other half of the window, and the half that has to be tested
+    // through the SERVICE. Beating the claim by calling `store.heartbeat`
+    // from the test would prove only that the fake store refreshes a
+    // timestamp; the thing that can actually regress is the interval `run`
+    // installs. Nothing below touches the store by hand — the claim survives
+    // only if the service is beating it.
+    vi.useFakeTimers();
+    const h = harness();
+    const record = pendingRow(h.store);
+
+    // A first-ever request is a full clone. This one hangs for well over the
+    // stale window, which is the case the heartbeat exists for: without it
+    // the work would be stolen out from under the process doing it.
+    let finishClone = (): void => {};
+    h.workflow.createBranch.mockReturnValue(
+      new Promise((resolve) => {
+        finishClone = () => resolve({ name: 'x', isDefault: false, isProtected: false });
+      }) as never,
+    );
+
+    const flight = h.jobs.start(record);
+    await vi.advanceTimersByTimeAsync(0); // the claim lands
+    const beats = 4 * CLAIM_STALE_AFTER_MS;
+    await vi.advanceTimersByTimeAsync(beats);
+
+    // Four windows in, and still held — by the interval, and nothing else.
+    const other = harness({}, h.store);
+    await other.jobs.start({ ...record });
+    expect(other.workflow.createBranch).not.toHaveBeenCalled();
+    expect(h.store.all()[0].status).toBe('pending');
+
+    // And it still completes: the beats did not disturb the work.
+    finishClone();
+    await flight;
+    expect(h.store.all()[0]).toMatchObject({ status: 'opened', changeRequestNumber: 42 });
+    expect(h.workflow.openChangeRequest).toHaveBeenCalledTimes(1);
+  });
+
+  it('stops before the change request when the row was erased under it', async () => {
+    // Account erasure deletes a requester's recorded rows in a transaction
+    // this job has no part in and cannot be told about. It lands here between
+    // the commit and the change request — the worst moment, because what
+    // comes next is the one step that leaves something a manager sees.
+    const h = harness();
+    const record = pendingRow(h.store);
+    h.workflow.commitChanges.mockImplementation(async () => {
+      h.store.deleteFor(ALI, 'Finance');
+      return null;
+    });
+
+    await h.jobs.start(record);
+
+    // No change request in a deleted person's name — and no row written back,
+    // which would resurrect the ask the erasure just removed.
+    expect(h.workflow.openChangeRequest).not.toHaveBeenCalled();
+    expect(h.store.all()).toEqual([]);
+  });
+
+  it('settles nothing when its claim was taken over mid-work', async () => {
+    // The straggler. This worker stalled past the window — a frozen host, a
+    // partition that outlived three beats — and another process took the row
+    // over and is doing the work now. Coming back to life, it must not open a
+    // second change request, and must not stamp its own ending on the attempt
+    // that replaced it.
+    const h = harness();
+    const record = pendingRow(h.store);
+    h.workflow.commitChanges.mockImplementation(async () => {
+      // A zero window is "the claim has lapsed", without minutes of fake time.
+      await h.store.claim(record.id, 0);
+      return null;
+    });
+
+    await h.jobs.start(record);
+
+    expect(h.workflow.openChangeRequest).not.toHaveBeenCalled();
+    // Still pending, under the NEW claim: not `failed`, which is what an
+    // unfenced worker would have written over a run that is going fine.
+    expect(h.store.all()[0]).toMatchObject({ status: 'pending', failureReason: null });
+  });
+
+  it('will not push or open anything while ownership is unknown', async () => {
+    // The beat THREW rather than answering. That is not "still mine" — the
+    // database a replacement would have claimed through is exactly the
+    // database that just refused to answer — so in front of an irreversible
+    // step it has to read as "not mine".
+    const h = harness();
+    const record = pendingRow(h.store);
+    const realHeartbeat = h.store.heartbeat.bind(h.store);
+    let beats = 0;
+    h.store.heartbeat = async (id: string, token: string) => {
+      // Healthy up to the gate that stands in front of the git, then down.
+      if (++beats > 1) throw new Error('the database went away');
+      return realHeartbeat(id, token);
+    };
+
+    await h.jobs.start(record);
+
+    // Nothing was written to the branch and nothing was opened — and the row
+    // is left exactly as it was, pending, for the next sweep to redo.
+    expect(h.workspaceService.writeFile).not.toHaveBeenCalled();
+    expect(h.workflow.commitChanges).not.toHaveBeenCalled();
+    expect(h.workflow.openChangeRequest).not.toHaveBeenCalled();
+    expect(h.store.all()[0]).toMatchObject({ status: 'pending', failureReason: null });
+  });
+
+  it('a database blip does not abandon work that is going fine', async () => {
+    // The other direction, and the reason `unknown` is a third case rather
+    // than a synonym for `lost`. The PERIODIC beat takes no action on its
+    // answer, and the window is three beats wide exactly so one failed UPDATE
+    // costs a beat rather than a clone that is running perfectly well.
+    vi.useFakeTimers();
+    const h = harness();
+    const record = pendingRow(h.store);
+    const realHeartbeat = h.store.heartbeat.bind(h.store);
+    // Every beat from the interval fails; the gates' beats still succeed.
+    let inGit = false;
+    h.store.heartbeat = async (id: string, token: string) => {
+      if (inGit) throw new Error('the database blipped');
+      return realHeartbeat(id, token);
+    };
+
+    let finishClone = (): void => {};
+    h.workflow.createBranch.mockReturnValue(
+      new Promise((resolve) => {
+        finishClone = () => resolve({ name: 'x', isDefault: false, isProtected: false });
+      }) as never,
+    );
+
+    const flight = h.jobs.start(record);
+    await vi.advanceTimersByTimeAsync(0);
+    inGit = true;
+    await vi.advanceTimersByTimeAsync(3 * CLAIM_STALE_AFTER_MS);
+    inGit = false;
+
+    // The interval swallowed every failure rather than tearing the job down.
+    finishClone();
+    await flight;
+    expect(h.store.all()[0]).toMatchObject({ status: 'opened', changeRequestNumber: 42 });
+  });
+
+  it('checks the claim again after the clone, before it writes or pushes', async () => {
+    // The gate before the git says nothing about who owns the row by the time
+    // the clone returns, and a first-ever request clones the whole plugins
+    // repository. So the row is taken over DURING the clone here: the write
+    // and the push must not happen against a shared workspace this process no
+    // longer has the right to touch.
+    const h = harness();
+    const record = pendingRow(h.store);
+    h.workspaceService.getOrCreateForBranch = async (branch: string) => {
+      await h.store.claim(record.id, 0); // the window lapsed; another worker took it
+      return { id: workspaceIdForBranch(branch) };
+    };
+
+    await h.jobs.start(record);
+
+    // The branch was created under a claim this process did hold — that is
+    // idempotent and harmless — but nothing after the clone ran.
+    expect(h.workspaceService.writeFile).not.toHaveBeenCalled();
+    expect(h.workflow.commitChanges).not.toHaveBeenCalled();
+    expect(h.workflow.openChangeRequest).not.toHaveBeenCalled();
+  });
+
+  it('a superseded token can neither beat nor decide the row', async () => {
+    // The fence itself, at the store: the contract `DbJoinRequestStore` puts
+    // in SQL (`where claim_token = $token`) and this fake mirrors.
+    const h = harness();
+    const record = pendingRow(h.store);
+    const first = await h.store.claim(record.id, CLAIM_STALE_AFTER_MS);
+    const second = await h.store.claim(record.id, 0);
+    expect(first?.claimToken).toBeTruthy();
+    expect(second?.claimToken).not.toBe(first?.claimToken);
+
+    await h.store.markOpened(record.id, second!.claimToken, 7);
+
+    // Everything the superseded worker might still do, with the token it
+    // believes it holds. All of it must be a no-op.
+    expect(await h.store.heartbeat(record.id, first!.claimToken)).toBe(false);
+    await h.store.markFailed(record.id, first!.claimToken, 'the git host went away');
+    await h.store.markOpened(record.id, first!.claimToken, 99);
+    await h.store.release(record.id, first!.claimToken);
+
+    expect(h.store.all()[0]).toMatchObject({ status: 'opened', changeRequestNumber: 7 });
+  });
+
+  it('keeps sweeping on a tick, so a row nobody is working is picked up without a restart', async () => {
+    vi.useFakeTimers();
+    const h = harness();
+    const record = pendingRow(h.store);
+    // Held by a process that then died: the first sweep is refused it.
+    await h.store.claim(record.id, CLAIM_STALE_AFTER_MS);
+
+    h.jobs.startSweeping(1000);
+    await vi.advanceTimersByTimeAsync(0);
+    await h.jobs.drain();
+    expect(h.store.all()[0].status).toBe('pending');
+
+    // The claim goes stale, and the NEXT tick takes it — no restart involved.
+    // Without the tick this row would wait for the next boot, which on a
+    // healthy deployment may be days.
+    vi.setSystemTime(Date.now() + CLAIM_STALE_AFTER_MS + 1000);
+    await vi.advanceTimersByTimeAsync(1000);
+    await h.jobs.drain();
+
+    expect(h.store.all()[0]).toMatchObject({ status: 'opened', changeRequestNumber: 42 });
+    h.jobs.stopSweeping();
+  });
+
+  it('stopSweeping ends the tick', async () => {
+    vi.useFakeTimers();
+    const h = harness();
+    h.jobs.startSweeping(1000);
+    await vi.advanceTimersByTimeAsync(0);
+    h.jobs.stopSweeping();
+
+    pendingRow(h.store);
+    await vi.advanceTimersByTimeAsync(5000);
+    await h.jobs.drain();
+
+    expect(h.workflow.openChangeRequest).not.toHaveBeenCalled();
+  });
+
+  it('proceeds past a createBranch failure only when the branch is really there', async () => {
+    const h = harness();
+    const branch = joinBranchFor(ALI, 'Finance');
+    h.workflow.createBranch.mockRejectedValue(new Error('a branch named that already exists'));
+    h.workflow.listBranches.mockResolvedValue([{ name: branch }]);
+
+    await h.jobs.start(pendingRow(h.store));
+
+    expect(h.workflow.listBranches).toHaveBeenCalledWith(expect.any(String), {
+      freshFetch: true,
+      strictFetch: true,
+    });
+    expect(h.store.all()[0]).toMatchObject({ status: 'opened' });
+  });
+
+  it('fails the record when createBranch refuses and no branch exists to work against', async () => {
+    const h = harness();
+    h.workflow.createBranch.mockRejectedValue(new Error('remote: permission denied'));
+    h.workflow.listBranches.mockResolvedValue([]);
+
+    await h.jobs.start(pendingRow(h.store));
+
+    // The old bare `catch {}` swallowed this and let every later step run
+    // against a branch that was never created, so the requester read whatever
+    // confused error fell out downstream instead of the real one.
+    expect(h.store.all()[0]).toMatchObject({
+      status: 'failed',
+      failureReason: 'remote: permission denied',
+    });
+    expect(h.workflow.openChangeRequest).not.toHaveBeenCalled();
+  });
+
+  it('refuses to splice when access.md is unreadable for any reason but absence', async () => {
+    const h = harness();
+    const denied = Object.assign(new Error('permission denied'), { code: 'EACCES' });
+    h.workspaceService.readFile.mockRejectedValue(denied);
+
+    await h.jobs.start(pendingRow(h.store));
+
+    // Treating it as empty would have committed a file holding nothing but
+    // this one request, wiping the plugin's real rules.
+    expect(h.workspaceService.writeFile).not.toHaveBeenCalled();
+    expect(h.store.all()[0]).toMatchObject({ status: 'failed' });
+  });
+
+  it('splices from nothing when access.md is merely absent', async () => {
+    const h = harness();
+    const missing = Object.assign(new Error('no such file'), { code: 'ENOENT' });
+    h.workspaceService.readFile.mockRejectedValue(missing);
+
+    await h.jobs.start(pendingRow(h.store));
+
+    expect(h.workspaceService.writeFile).toHaveBeenCalled();
+    expect(h.store.all()[0]).toMatchObject({ status: 'opened' });
+  });
+
+  it('keeps the request pending, and unclaimed, when the platform is not ready yet', async () => {
+    const h = harness({
+      target: async () => {
+        throw new JoinRequestNotReadyError('the plugin catalog is not available yet');
+      },
+    });
+    await h.jobs.start(pendingRow(h.store));
+
+    // A knowledge base that has not finished cloning must not tell everyone
+    // with a request outstanding that their plugin was deleted.
+    const [row] = h.store.all();
+    expect(row).toMatchObject({ status: 'pending', failureReason: null });
+    // And the claim went back, so the next sweep picks it up at once rather
+    // than waiting out a claim nothing is working behind.
+    expect(row.claimedAt).toBeNull();
+  });
+
+  it('asks for a FRESH change-request listing before it opens a second one', async () => {
+    const h = harness();
+    await h.jobs.start(pendingRow(h.store));
+    // A cached listing taken before an earlier attempt opened its change
+    // request shows nothing to adopt, which is how a retry becomes a
+    // duplicate.
+    expect(h.workflow.listChangeRequestsAuthoredBy).toHaveBeenCalledWith(ALI, { fresh: true });
+  });
+
+});