@1agh/maude 0.56.0 → 0.57.0

package/apps/studio/sync/remote-docs.ts

@@ -0,0 +1,191 @@
+// What the PROJECT holds, versus what this machine happens to carry.
+//
+// The sync runtime opens one provider per canvas found on LOCAL disk, and Yjs
+// has no enumeration — so a document that exists only on the hub is invisible
+// to a peer by construction. A desktop with 72 of a project's 75 canvases syncs
+// 72, reports "72/72 synced", and is accurate about the wrong universe. The
+// user-visible form of that was "I pressed Open in Maude and nothing happened":
+// three real canvases (welcome, how-to-use, how-to-make-video) lived only in
+// the cloud and could never arrive.
+//
+// This module asks the hub what it has (`GET /api/documents`, scope-bound to
+// the same token the sync uses) and diffs it against the local set, so the
+// runtime can say which side is missing what.
+//
+// SYNC IS BIDIRECTIONAL AND COMPLETE. A project you have been granted access to
+// is a project you get — all of it, both directions. Local canvases go up (they
+// always did); hub-only documents now come DOWN, materialised as real files.
+//
+// What that trades, stated plainly: a hub can create files in your design root,
+// where before it could only update canvases you already had (DDR-054 treats
+// hub-pushed content as untrusted). That is accepted deliberately — the hub is
+// the project, the caller is authenticated, and a partial project is not a
+// project. The one guard kept is containment: a document NAME can never place a
+// file outside the design root. That is not policy filtering, it is the
+// difference between writing your project and writing your filesystem.
+
+/** One document as the hub reports it. */
+export interface RemoteDoc {
+  name: string;
+  bytes: number;
+}
+
+export interface RemoteDocDiff {
+  /** Documents on both sides — the ones actually syncing. */
+  shared: string[];
+  /** On the hub, absent here. The project has them; this machine cannot get them. */
+  hubOnly: RemoteDoc[];
+  /** Here, not on the hub yet — normal for a canvas this peer just created. */
+  localOnly: string[];
+  /** Null when the hub could not be asked (old hub, offline, refused). */
+  reachable: boolean;
+}
+
+/** How long to wait for the listing. Never blocks a sync — see `fetchRemoteDocs`. */
+const LIST_TIMEOUT_MS = 6000;
+
+/**
+ * Ask the hub which documents this token may open.
+ *
+ * Returns null on ANY failure — an old hub without the route, a refused token,
+ * a network blip. A peer that cannot get the listing must still sync: this is a
+ * reporting improvement, and making it load-bearing would trade a real feature
+ * for a nicer message.
+ */
+export async function fetchRemoteDocs(
+  hubUrl: string,
+  token: string,
+  fetchImpl: typeof fetch = fetch
+): Promise<RemoteDoc[] | null> {
+  try {
+    const base = hubUrl.replace(/\/+$/, '');
+    const res = await fetchImpl(`${base}/api/documents`, {
+      headers: { authorization: `Bearer ${token}` },
+      signal: AbortSignal.timeout(LIST_TIMEOUT_MS),
+    });
+    if (!res.ok) return null;
+    const body = (await res.json()) as { documents?: unknown };
+    if (!Array.isArray(body?.documents)) return null;
+    return body.documents
+      .filter(
+        (d): d is RemoteDoc =>
+          !!d && typeof (d as RemoteDoc).name === 'string' && (d as RemoteDoc).name.length > 0
+      )
+      .map((d) => ({ name: d.name, bytes: Number(d.bytes) || 0 }));
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Diff the hub's documents against the ones this peer syncs.
+ *
+ * `localDocNames` must already be in WIRE form (`docNameFor(slug)`), not raw
+ * slugs — comparing a namespaced hub against flat local names would report
+ * every document as missing on both sides, which is the exact failure the
+ * namespace exists to prevent, arrived at from the reporting side.
+ */
+export function diffRemoteDocs(
+  localDocNames: readonly string[],
+  remote: RemoteDoc[] | null
+): RemoteDocDiff {
+  if (remote === null) {
+    return { shared: [], hubOnly: [], localOnly: [...localDocNames], reachable: false };
+  }
+  const local = new Set(localDocNames);
+  const remoteNames = new Set(remote.map((d) => d.name));
+  return {
+    shared: [...local].filter((n) => remoteNames.has(n)).sort(),
+    hubOnly: remote.filter((d) => !local.has(d.name)).sort((a, b) => a.name.localeCompare(b.name)),
+    localOnly: [...local].filter((n) => !remoteNames.has(n)).sort(),
+    reachable: true,
+  };
+}
+
+/**
+ * One line a person can act on.
+ *
+ * Deliberately names the documents being PULLED rather than only what already
+ * matched: "72 synced" was true and useless, because the number is drawn from
+ * the local file set and can never express what the project holds beyond it.
+ */
+export function describeRemoteDiff(diff: RemoteDocDiff): string | null {
+  if (!diff.reachable) return null;
+  if (diff.hubOnly.length === 0) return null;
+  const n = diff.hubOnly.length;
+  const names = diff.hubOnly
+    .slice(0, 3)
+    .map((d) => d.name)
+    .join(', ');
+  return `pulling ${n} canvas${n === 1 ? '' : 'es'} down from the project (${names}${n > 3 ? ', …' : ''}).`;
+}
+
+/**
+ * The local slug a hub document name maps to.
+ *
+ * Namespaced names (`ws/<workspace>/<branch>/<slug>` — DDR-192 §5) carry the
+ * slug in the last segment; a legacy flat name IS the slug. Returns null for
+ * anything that does not survive the charset the hub itself enforces on
+ * `documentName`, so a crafted name cannot become a path component.
+ */
+export function slugFromDocName(docName: string): string | null {
+  const raw = String(docName ?? '');
+  // Validate the WHOLE name against the two shapes a hub may legitimately use,
+  // never just its last segment. Taking the tail of an arbitrary string would
+  // accept `../../etc/passwd` as `passwd`: contained by the check below, and
+  // still a file this project never asked for, created from a name that should
+  // have been refused outright. A component is `[A-Za-z0-9_-]` — no dots (so no
+  // traversal and no extension smuggling), no spaces.
+  const COMPONENT = '[A-Za-z0-9_-]{1,120}';
+  const FLAT = new RegExp(`^${COMPONENT}$`);
+  const NAMESPACED = new RegExp(`^ws/${COMPONENT}/${COMPONENT}/(${COMPONENT})$`);
+  if (FLAT.test(raw)) return raw.toLowerCase();
+  const ns = NAMESPACED.exec(raw);
+  return ns ? ns[1].toLowerCase() : null;
+}
+
+export interface PullTarget {
+  slug: string;
+  docName: string;
+  /** Absolute path the body will be written to, flat under the design root. */
+  bodyAbs: string;
+}
+
+/**
+ * Where a hub-only document lands on disk.
+ *
+ * FLAT, directly under the design root — the same convention the hub applies in
+ * the mirror direction (`workspace-files.mjs defaultBodyPath`) and for the same
+ * reason: a slug is lossy, `ui-card` cannot be un-flattened into `ui/Card.tsx`
+ * without guessing, and guessing wrong scatters files into directories the user
+ * never made. A flat file is trivially moved; an invented tree is not. Moving it
+ * later does not break sync — both paths slug to the same document.
+ *
+ * Returns only targets that resolve INSIDE the design root. A document name is
+ * hub-controlled input, and this is the last point before a create.
+ */
+export function pullTargets(
+  hubOnly: readonly RemoteDoc[],
+  designRoot: string,
+  join: (...parts: string[]) => string,
+  resolve: (p: string) => string,
+  sep: string
+): PullTarget[] {
+  const rootResolved = resolve(designRoot);
+  const out: PullTarget[] = [];
+  // One slug, one target. `slugFromDocName` lowercases, so a hub advertising
+  // `Foo`, `foo` and `ws/w/main/foo` yields three targets for ONE file — three
+  // providers on the same path, and a `pulled` count that overstates what
+  // arrived by a factor the hub chooses.
+  const seen = new Set<string>();
+  for (const doc of hubOnly) {
+    const slug = slugFromDocName(doc.name);
+    if (!slug || seen.has(slug)) continue;
+    seen.add(slug);
+    const bodyAbs = join(designRoot, `${slug}.tsx`);
+    const target = resolve(bodyAbs);
+    if (target !== rootResolved && !target.startsWith(rootResolved + sep)) continue;
+    out.push({ slug, docName: doc.name, bodyAbs });
+  }
+  return out;
+}

package/apps/studio/sync/status.ts

@@ -79,7 +79,10 @@ export function createSyncStatusStore(opts: SyncStatusStoreOptions): SyncStatusS
   const conflicts: SyncConflict[] = [];
 
   let snapshot: SyncStatusSnapshot = {
-    state: 'online',
+    // Same reason the monitor is seeded `connecting` (see connection-state.ts):
+    // this is the payload a reader gets BEFORE the first monitor snapshot lands,
+    // so seeding it `online` re-introduced the born-connected lie one layer out.
+    state: 'connecting',
     queuedOps: 0,
     lastSyncAt: null,
     offlineSince: null,

package/apps/studio/test/acp-activity-endpoint.test.ts

@@ -0,0 +1,183 @@
+// feature-acp-turn-notifications Task 2/3 — the activity snapshot the native
+// shell's poller reads (notify.rs Task 5) to decide whether to fire an OS
+// notification for a project that isn't the visible one. Covers:
+//   • activitySnapshot() shape + monotonic seq (bumps only on real change).
+//   • AcpBridge.awaitingInputCount reflects pending permission requests.
+//   • The Acp manager's registered probe reports 'awaiting-input', not
+//     'running', for a chat blocked on a permission prompt — the ordering
+//     Task 2's gotcha calls out explicitly.
+//   • /_api/acp/activity is absent from both canvas-origin allowlists
+//     (Decision: NEITHER, like /_api/acp/running today).
+
+import { afterEach, describe, expect, test } from 'bun:test';
+import { readFileSync } from 'node:fs';
+import { join } from 'node:path';
+
+import type { ServerWebSocket } from 'bun';
+
+import { AcpBridge } from '../acp/bridge.ts';
+import { createAcp } from '../acp/index.ts';
+import { activitySnapshot, registerActivityProbe } from '../acp/running.ts';
+import type { Context } from '../context.ts';
+import type { WsData } from '../ws.ts';
+
+const FIXTURE = join(import.meta.dir, 'fixtures', 'mock-acp-agent-permission.mjs');
+const TEST_ENV_KEYS = ['MAUDE_ACP_ADAPTER_ENTRY', 'MAUDE_ACP_RUNTIME', 'MAUDE_CLAUDE_BIN'];
+
+function useMockAgent() {
+  process.env.MAUDE_ACP_ADAPTER_ENTRY = FIXTURE;
+  process.env.MAUDE_ACP_RUNTIME = process.execPath;
+  process.env.MAUDE_CLAUDE_BIN = process.execPath;
+}
+
+afterEach(() => {
+  for (const key of TEST_ENV_KEYS) delete process.env[key];
+  // Restore a no-op probe so a later file's `registerRunningChatsProbe`-style
+  // module-level state doesn't leak a stale closure across test files.
+  registerActivityProbe(() => []);
+});
+
+async function until<T>(fn: () => T | undefined, timeoutMs = 12000): Promise<T> {
+  const start = performance.now();
+  for (;;) {
+    const v = fn();
+    if (v !== undefined) return v;
+    if (performance.now() - start > timeoutMs) throw new Error('timeout waiting for condition');
+    await new Promise((r) => setTimeout(r, 25));
+  }
+}
+
+function fakeWs(id: string) {
+  const frames: Array<Record<string, unknown>> = [];
+  const ws = {
+    data: { id } as WsData,
+    send: (raw: string) => {
+      frames.push(JSON.parse(raw));
+    },
+  } as unknown as ServerWebSocket<WsData>;
+  return { ws, frames };
+}
+
+describe('activitySnapshot (running.ts)', () => {
+  test('reflects the registered probe and never throws on a bad one', () => {
+    registerActivityProbe(() => [{ chatId: 'a', state: 'running' }]);
+    expect(activitySnapshot().chats).toEqual([{ chatId: 'a', state: 'running' }]);
+
+    registerActivityProbe(() => {
+      throw new Error('boom');
+    });
+    expect(activitySnapshot().chats).toEqual([]);
+  });
+
+  test('seq bumps only when the (chatId, state) set actually changes', () => {
+    registerActivityProbe(() => [{ chatId: 'a', state: 'idle' }]);
+    const first = activitySnapshot().seq;
+    const second = activitySnapshot().seq; // same state — no bump
+    expect(second).toBe(first);
+
+    registerActivityProbe(() => [{ chatId: 'a', state: 'running' }]);
+    const third = activitySnapshot().seq; // real transition — bumps
+    expect(third).toBe(first + 1);
+
+    // Enumeration-order shuffle with the SAME states must NOT bump.
+    registerActivityProbe(() => [
+      { chatId: 'b', state: 'idle' },
+      { chatId: 'a', state: 'running' },
+    ]);
+    activitySnapshot();
+    registerActivityProbe(() => [
+      { chatId: 'a', state: 'running' },
+      { chatId: 'b', state: 'idle' },
+    ]);
+    const fourth = activitySnapshot().seq;
+    const fifth = activitySnapshot().seq;
+    expect(fifth).toBe(fourth);
+  });
+});
+
+describe('AcpBridge.awaitingInputCount + the manager probe (Task 2 gotcha: awaiting-input wins over running)', () => {
+  test('a pending permission request reports awaiting-input, not running', async () => {
+    useMockAgent();
+    const ctx = {
+      paths: { repoRoot: process.cwd(), designRoot: '/tmp/does-not-matter-activity' },
+    } as unknown as Context;
+    const acp = createAcp(ctx);
+
+    const a = fakeWs('activity-ws-a');
+    try {
+      acp.onOpen(a.ws);
+      acp.onMessage(a.ws, JSON.stringify({ t: 'prompt', text: 'hi', chat: 'activity-chat' }));
+      const reqFrame = await until(() => a.frames.find((f) => f.t === 'permission-request'));
+
+      const snap = activitySnapshot();
+      expect(snap.chats).toContainEqual({ chatId: 'activity-chat', state: 'awaiting-input' });
+
+      acp.onMessage(
+        a.ws,
+        JSON.stringify({ t: 'permission-response', id: reqFrame.id, decision: 'allow-once' })
+      );
+      await until(() => a.frames.find((f) => f.t === 'turn-end'));
+
+      // Turn settled — no longer awaiting-input (running or idle, never stuck).
+      const after = activitySnapshot().chats.find((c) => c.chatId === 'activity-chat');
+      expect(after?.state).not.toBe('awaiting-input');
+    } finally {
+      acp.onClose(a.ws);
+      acp.stopAll();
+      await new Promise((r) => setTimeout(r, 50));
+    }
+  }, 20000);
+});
+
+describe('AcpBridge.awaitingInputCount — direct unit', () => {
+  test('starts at 0 and counts a pending permission before resolution', async () => {
+    useMockAgent();
+    const bridge = new AcpBridge({ repoRoot: process.cwd(), onUpdate: () => {} });
+    expect(bridge.awaitingInputCount).toBe(0);
+    try {
+      const promptPromise = bridge.prompt('hi', 'c1');
+      await until(() => (bridge.awaitingInputCount > 0 ? true : undefined));
+      expect(bridge.awaitingInputCount).toBe(1);
+      await bridge.cancel(); // denies the pending request rather than hanging
+      await promptPromise;
+      expect(bridge.awaitingInputCount).toBe(0);
+    } finally {
+      await bridge.stop();
+    }
+  }, 15000);
+});
+
+describe('/_api/acp/activity — NEITHER canvas-origin allowlist (Decision D + dual-allowlist rule)', () => {
+  const httpSrc = readFileSync(join(import.meta.dir, '..', 'http.ts'), 'utf8');
+  const serverSrc = readFileSync(join(import.meta.dir, '..', 'server.ts'), 'utf8');
+
+  test('the route exists on the main-origin route table', () => {
+    expect(httpSrc).toContain("'/_api/acp/activity'");
+  });
+
+  test('is absent from CANVAS_SAFE_API', () => {
+    const start = httpSrc.indexOf('CANVAS_SAFE_API = new Set([');
+    const end = httpSrc.indexOf(']);', start);
+    expect(start).toBeGreaterThan(-1);
+    const block = httpSrc.slice(start, end);
+    expect(block).not.toContain('/_api/acp/activity');
+  });
+
+  test("is absent from startCanvasServer's routes map", () => {
+    const start = serverSrc.indexOf('function startCanvasServer');
+    const routesEnd = serverSrc.indexOf('async fetch(req, srv)', start);
+    expect(start).toBeGreaterThan(-1);
+    const block = serverSrc.slice(start, routesEnd);
+    expect(block).not.toContain('/_api/acp/activity');
+  });
+
+  test('the payload never carries a chat title, message text, or transcript field name', () => {
+    // Grep the route handler body itself (not the whole file) for the fields
+    // Decision D forbids — a title/text field appearing HERE would mean
+    // someone widened the payload without reading the comment.
+    const start = httpSrc.indexOf("'/_api/acp/activity':");
+    const end = httpSrc.indexOf('},', start);
+    const block = httpSrc.slice(start, end);
+    expect(block).not.toMatch(/title|message|transcript|content/i);
+  });
+});

package/apps/studio/test/acp-branch-guard.test.ts

@@ -0,0 +1,123 @@
+// feature-acp-write-path-scope Addendum, Task 9 — the branch-switch warning.
+//
+// `RepoBranchSwitcher.jsx` has THREE `window.location.reload()` call sites
+// (`switchDraft`, `createDraft`, and the local-merge fold in `foldDraft`). All
+// three move the worktree, so all three need the gate; covering one and missing
+// two is the easy mistake, which is why the source-shape assertion below exists
+// alongside the copy tests.
+//
+// Why the wording matters enough to test: this is a safety prompt, and Task 8
+// changed what is TRUE about it. Before, the reload killed the running turn, so
+// "this will stop it" was accurate. Now the bridge outlives its socket, so the
+// turn keeps going — into a worktree that has moved under it. "Will stop it"
+// would now be false in the reassuring direction.
+
+import { describe, expect, test } from 'bun:test';
+import { readFileSync } from 'node:fs';
+import { join } from 'node:path';
+
+import { chatGuardCopy } from '../client/panels/RepoBranchSwitcher.jsx';
+
+describe('chatGuardCopy', () => {
+  const one = chatGuardCopy({ count: 1, branch: 'draft-a', verb: 'switch to “main”' });
+
+  test('names the branch, the action, and the real consequence', () => {
+    expect(one.title).toBe('A chat is working right now');
+    expect(one.body).toContain('draft-a');
+    expect(one.body).toContain('switch to “main”');
+    expect(one.body).toContain('wrong branch');
+  });
+
+  test('does NOT claim the switch stops the chat (it no longer does — Task 8)', () => {
+    const all = `${one.title} ${one.body} ${one.meta}`.toLowerCase();
+    expect(all).not.toContain('will stop it');
+    // …and it does say the thing that IS true: the agent is actively editing.
+    expect(one.body).toContain('editing files on');
+  });
+
+  test('is honest that History cannot undo this', () => {
+    // The `_history/` snapshot stack is per-canvas-slug with NO branch
+    // awareness, so the "it's reversible" reassurance this whole feature leans
+    // on genuinely does not hold across a checkout. Promising rollback here
+    // would be the one lie that makes the confirm worse than no confirm.
+    expect(one.body).toContain("History can't undo");
+  });
+
+  test('pluralizes when several chats are running', () => {
+    expect(chatGuardCopy({ count: 3, branch: 'x', verb: 'y' }).title).toBe(
+      '3 chats are working right now'
+    );
+  });
+
+  test('the confirm is not the safe default', () => {
+    // "Wait" is the cancel; "Do it anyway" is the deliberate, phrased-as-a-risk
+    // confirm. A neutral "OK/Cancel" pair would read as routine.
+    expect(one.cancel).toBe('Wait');
+    expect(one.confirm).toBe('Do it anyway');
+  });
+});
+
+describe('all THREE reload call sites are gated', () => {
+  const src = readFileSync(
+    join(import.meta.dir, '..', 'client', 'panels', 'RepoBranchSwitcher.jsx'),
+    'utf8'
+  );
+
+  test('every reload lives in a `do*` action reached only through guardedByChat', () => {
+    // Source-shape guard rather than a render test: the failure mode is a
+    // FOURTH reload being added later without a gate, which no render test of
+    // the existing three would catch.
+    const guarded = src.match(/guardedByChat\(/g) ?? [];
+    // One definition + three call sites.
+    expect(guarded.length).toBeGreaterThanOrEqual(4);
+    for (const fn of ['doSwitchDraft', 'doCreateDraft', 'doFoldDraft']) {
+      expect(src).toContain(`function ${fn}`);
+    }
+  });
+
+  test('the gate asks the SERVER, not local chat state', () => {
+    // A bridge can be running DETACHED (post-reload, pre-re-attach) with no
+    // client that knows about it — precisely the case the warning is for. Local
+    // busy state would answer "nothing is running" exactly when it matters.
+    expect(src).toContain('/_api/acp/running');
+  });
+
+  test('a failed probe never BLOCKS a switch', () => {
+    // A switch the user asked for must not be hostage to an unrelated fetch
+    // error. The catch returns 0 (⇒ proceed), not a thrown error.
+    expect(src).toMatch(/catch\s*{[\s\S]*?return 0;/);
+  });
+});
+
+// SECURITY (security-auditor A6) — cross-PROJECT loopback adjacency.
+//
+// Before the sidecar pool (Addendum Task 10) at most one dev-server ran at a
+// time, so "another Maude instance on another loopback port" was not a state the
+// product produced. The pool makes concurrency DESIGNED-IN: projects A and B now
+// routinely run side by side on different ports, and A's canvas iframe is
+// untrusted content (DDR-054). So the question "can A's canvas reach B's
+// privileged main origin?" went from hypothetical to worth an explicit test.
+//
+// The defence is `sameOriginWrite`'s host+port compare — different instance,
+// different port, different `host`, rejected. Asserting it here so the property
+// is pinned to the pool's introduction rather than left as an inference.
+describe('cross-project origin adjacency (A6)', () => {
+  test('a request from ANOTHER instance’s origin is not same-origin', async () => {
+    const { sameOriginWrite } = await import('../http.ts');
+    const projectB = new Request('http://localhost:4400/_api/acp/running', {
+      method: 'POST',
+      // Project A's canvas/main origin, aimed at project B's server.
+      headers: { origin: 'http://localhost:4399' },
+    });
+    expect(sameOriginWrite(projectB)).toBe(false);
+  });
+
+  test('a request from the SAME instance is allowed (the pool must not break normal use)', async () => {
+    const { sameOriginWrite } = await import('../http.ts');
+    const own = new Request('http://localhost:4400/_api/acp/running', {
+      method: 'POST',
+      headers: { origin: 'http://localhost:4400' },
+    });
+    expect(sameOriginWrite(own)).toBe(true);
+  });
+});