switchroom 0.15.44 → 0.16.4

package/telegram-plugin/uat/scenarios/jtbd-reply-is-last-dm.test.ts

@@ -0,0 +1,202 @@
+/**
+ * JTBD: "the reply is last" + "a conversational turn opens no card" — the
+ * CI-enforced form of the deterministic emission invariants (design
+ * `docs/message-emission-determinism.md` §11; #2556).
+ *
+ * Four cases, each pulled from server SEND-ORDER history (`driver.getHistory`)
+ * so a post-reply card that landed before any live observer started is still
+ * caught. The ordering assertion is the SCOPED one (§6): within a single
+ * foreground turn, no activity-card / worker-feed surface opens after that
+ * turn's reply — NOT a naive "answer has the max message_id" (that would
+ * false-positive on a legitimate later background / represent / error surface).
+ * `assertReplyIsLast` filters to the activity/answer lanes of the SAME turn.
+ *
+ *   1. Conversational, zero-tool ("Reply with only: pong") — NO activity card
+ *      opens in this turn at all (lever 5 base case / G1, the triplication).
+ *   2. Tool-heavy (a REAL_WORK activity-surface prompt) — a card opened AND no
+ *      card for this turn sits below the substantive reply (lever 1 / races
+ *      A/B/E).
+ *   3. Short-pinging final ("Reply 'Done!' then write one memory") — the
+ *      currently-reordering case; green only once lever 2 lands (G5).
+ *   4. Two-turn backstop — a prompt that ends a turn without a qualifying reply
+ *      (forcing the silent-end re-prompt); no card opens below the final answer
+ *      across the re-prompt boundary (G3/C). Needs `getHistory`.
+ *
+ * Runs under CI `uat-gate`; the full live MTProto run needs the test-harness
+ * agent + a vault session, so locally this self-skips green (no driver).
+ */
+import { describe, it, expect, beforeAll } from "vitest";
+import { spinUp, type Scenario } from "../harness.js";
+import {
+  assertReplyIsLast,
+  isAnswer,
+  isActivityFeedMessage,
+  isWorkerFeedMessage,
+} from "../assertions.js";
+import { collectTurn } from "../real-work-prompts.js";
+import type { ObservedMessage } from "../driver.js";
+
+/** Per-case overall budget. */
+const TURN_BUDGET_MS = 130_000;
+/** History pull depth — comfortably covers a multi-surface two-turn exchange. */
+const HISTORY_LIMIT = 80;
+
+describe("uat: reply-is-last + conversational-turn-opens-no-card (DM)", () => {
+  let sc: Scenario | null = null;
+
+  beforeAll(async () => {
+    try {
+      sc = await spinUp({ agent: "test-harness" });
+      await sc.driver.primeDialogs();
+    } catch (err) {
+      console.warn(
+        `[reply-is-last] no live driver — self-skipping green: ${(err as Error).message}`,
+      );
+      sc = null;
+    }
+  });
+
+  // Case 1 — conversational, zero-tool: NO activity card opens at all.
+  it(
+    "case 1: a conversational 0-tool turn opens NO activity card (lever 5 / G1)",
+    async () => {
+      if (sc == null) return; // self-skip green
+      const { driver, botUserId, driverUserId } = sc;
+
+      const obs = await collectTurn(
+        driver,
+        botUserId,
+        driverUserId,
+        "Reply with only this exact word and nothing else, using no tools at all: pong",
+        { timeoutMs: TURN_BUDGET_MS, minAnswerChars: 1, settleMs: 8_000 },
+      );
+
+      expect(obs.answer, "the pong answer must land").not.toBeNull();
+
+      // Pull send-order history and confirm: no activity-card surface opened in
+      // this turn. We scope to surfaces at/after the answer's turn by reusing
+      // assertReplyIsLast, AND additionally assert the live collector saw no
+      // activity feed for this minimal turn.
+      const history = await driver.getHistory(botUserId, HISTORY_LIMIT);
+      assertReplyIsLast(history, driverUserId, { turn: obs.answer! });
+
+      // The strong G1 assertion: a 0-tool conversational turn must produce no
+      // activity feed at all (neither open nor edit).
+      expect(
+        obs.sawActivityFeed,
+        "a 0-tool conversational turn must not open an activity card (the triplication)",
+      ).toBe(false);
+    },
+    TURN_BUDGET_MS + 30_000,
+  );
+
+  // Case 2 — tool-heavy: a card opens, but none below the substantive reply.
+  it(
+    "case 2: a tool-heavy turn opens a card but none below the reply (lever 1 / races A/B/E)",
+    async () => {
+      if (sc == null) return; // self-skip green
+      const { driver, botUserId, driverUserId } = sc;
+
+      const obs = await collectTurn(
+        driver,
+        botUserId,
+        driverUserId,
+        "Use your Bash tool to run `uname -a`, then tell me in one sentence what " +
+          "operating system this machine is running. Keep the answer substantive " +
+          "(a few sentences explaining what the output means).",
+        { timeoutMs: TURN_BUDGET_MS, minAnswerChars: 60, settleMs: 8_000 },
+      );
+
+      if (obs.answer == null) {
+        console.warn("[reply-is-last] case 2 INCONCLUSIVE — no answer landed in budget.");
+        return;
+      }
+      if (!obs.sawActivityFeed && !obs.sawWorkerFeed) {
+        console.warn(
+          "[reply-is-last] case 2 INCONCLUSIVE — the agent answered without a " +
+            "tool feed; the reorder vector this guards was not exercised.",
+        );
+        return;
+      }
+
+      const history = await driver.getHistory(botUserId, HISTORY_LIMIT);
+      // The scoped invariant: no activity/worker-feed surface for this turn
+      // lands below the substantive reply.
+      assertReplyIsLast(history, driverUserId, { turn: obs.answer });
+    },
+    TURN_BUDGET_MS + 30_000,
+  );
+
+  // Case 3 — short-pinging final: the case the lever-2 ordering fix makes pass.
+  it(
+    "case 3: a short-pinging final stays last even with post-reply tool work (lever 2 / G5)",
+    async () => {
+      if (sc == null) return; // self-skip green
+      const { driver, botUserId, driverUserId } = sc;
+
+      const obs = await collectTurn(
+        driver,
+        botUserId,
+        driverUserId,
+        "Reply with only the single word 'Done!' (with the exclamation mark) — " +
+          "then, AFTER that reply, save a one-line memory noting you completed " +
+          "this test. The short reply is your final answer.",
+        // The "Done!" reply is short; accept it as the answer.
+        { timeoutMs: TURN_BUDGET_MS, minAnswerChars: 1, settleMs: 10_000 },
+      );
+
+      if (obs.answer == null) {
+        console.warn("[reply-is-last] case 3 INCONCLUSIVE — no answer landed in budget.");
+        return;
+      }
+
+      const history = await driver.getHistory(botUserId, HISTORY_LIMIT);
+      // The post-'Done!' memory write must NOT have reopened a card BELOW the
+      // reply. Before lever 2 this reorders (the named G5 residual); after, the
+      // card is finalized before the send and stays above.
+      assertReplyIsLast(history, driverUserId, { turn: obs.answer });
+    },
+    TURN_BUDGET_MS + 30_000,
+  );
+
+  // Case 4 — two-turn backstop: no card below the final answer across the
+  // silent-end re-prompt boundary.
+  it(
+    "case 4: no card opens below the final answer across a re-prompt boundary (G3/C)",
+    async () => {
+      if (sc == null) return; // self-skip green
+      const { driver, botUserId, driverUserId } = sc;
+
+      // A prompt that nudges the model to write its answer as prose first
+      // (no reply tool) — forcing the silent-end re-prompt, then a real answer
+      // on the re-prompted turn. The model's exact path isn't forceable, so the
+      // ordering assertion is the durable part; the re-prompt is best-effort.
+      const obs = await collectTurn(
+        driver,
+        botUserId,
+        driverUserId,
+        "Think out loud briefly, then give me a thorough multi-sentence answer " +
+          "(at least 220 characters) explaining what a Telegram supergroup is and " +
+          "how forum topics work inside one.",
+        { timeoutMs: TURN_BUDGET_MS, minAnswerChars: 200, settleMs: 12_000 },
+      );
+
+      if (obs.answer == null) {
+        console.warn("[reply-is-last] case 4 INCONCLUSIVE — no answer landed in budget.");
+        return;
+      }
+
+      // Pull full send-order history (the re-prompt may have produced a second
+      // card before the live observer in collectTurn caught it) and assert the
+      // final answer's turn has no feed surface below it.
+      const history = await driver.getHistory(botUserId, HISTORY_LIMIT);
+      assertReplyIsLast(history, driverUserId, { turn: obs.answer });
+
+      // Sanity: the answer is a genuine answer-lane message (not a feed).
+      expect(isAnswer(obs.answer, driverUserId)).toBe(true);
+      expect(isActivityFeedMessage(obs.answer)).toBe(false);
+      expect(isWorkerFeedMessage(obs.answer)).toBe(false);
+    },
+    TURN_BUDGET_MS + 30_000,
+  );
+});

package/telegram-plugin/uat/scenarios/reactions-dm.test.ts

@@ -10,36 +10,27 @@
  * of the bot's final reply — otherwise the user looks at their
  * inbound, sees it still wearing 🤔, and asks "you done?").
  *
- * History: this scenario was previously `describe.skip` with a
- * rationale that the pinned progress card "renders INSTEAD of
- * reactions". The card was retired in #1126; the card-vs-reaction
- * branch in the gateway is dead. We can now exercise the full
- * lifecycle end-to-end without the two-agent split.
- *
  * What we assert (in priority order):
  *
- *  1. Within the turn, the driver sees AT LEAST ONE `+` reaction
- *     op (the L1 "I'm alive" signal). Fast turns may collapse
- *     intermediate states, so we only require *one* add, not a
- *     specific emoji.
+ *  1. Within the turn, the bot places AT LEAST ONE reaction on the
+ *     inbound message (the L1 "I'm alive" signal). We poll via
+ *     `driver.pollReactions()` rather than subscribing to push
+ *     events — Telegram does not deliver `updateMessageReactions`
+ *     push events to the human account when a bot sets a reaction
+ *     in a DM (fixes #2502).
  *  2. By the time the bot has sent a final reply (+ a short tail
- *     for Telegram to deliver the terminal-emoji replace), the
- *     LAST observed `+` op is in the `done` set (`👍 / 💯 / 🎉`).
- *
- * Why "last `+` op wins" rather than `expectReaction(['👍'])` with
- * a literal sequence: `setMessageReaction` REPLACES the prior emoji
- * atomically. mtcute's update stream can deliver the replace as a
- * `-prev` followed by a `+next`, or as a single coalesced event,
- * depending on server batching. The "last add wins" shape matches
- * the production semantics — whatever's *currently* on the message
- * is what the user actually sees.
+ *     for Telegram to apply the terminal-emoji replace), the reaction
+ *     on the inbound message is in the `done` set (`👍 / 💯 / 🎉`).
  *
- * The observer must be attached BEFORE the reply lands so we
- * capture the queued / working reactions, not just the terminal
- * one. Pattern: `observeReactions` immediately after `sendDM`
- * returns the messageId, drain into a trail array while we wait
- * for the reply, then run a short tail to catch the terminal
- * after the reply.
+ * Polling strategy:
+ *   - Poll every `POLL_INTERVAL_MS` until a terminal-done emoji
+ *     appears OR the bot has replied AND `TAIL_AFTER_REPLY_MS` has
+ *     elapsed. Bail immediately on reply-timeout so CI doesn't burn
+ *     the full 90s safety ceiling.
+ *   - After the reply arrives, keep polling through the tail window
+ *     so the terminal emoji (👍) has time to replace the working
+ *     emoji (👀/🤔). In practice the replace happens within 1-2s
+ *     of the reply on a healthy bot; the 8s ceiling absorbs jitter.
  *
  * Requires the same env as `smoke-dm-reply.test.ts` (see
  * `uat/SETUP.md` §6).
@@ -49,16 +40,11 @@ import { describe, expect, it } from "vitest";
 import { spinUp } from "../harness.js";
 
 const TERMINAL_DONE_EMOJI = new Set(["👍", "💯", "🎉"]);
+const POLL_INTERVAL_MS = 1_000;
 const TAIL_AFTER_REPLY_MS = 8_000;
 
 const INBOUND = (): string => `uat-reactions ${new Date().toISOString()}`;
 
-interface ObservedOp {
-  emoji: string;
-  op: "+" | "-";
-  at: number;
-}
-
 describe("uat: reaction lifecycle on driver DM", () => {
   it(
     "driver sees an alive reaction, then a terminal-done emoji by reply tail",
@@ -67,71 +53,91 @@ describe("uat: reaction lifecycle on driver DM", () => {
       try {
         const sent = await sc.sendDM(INBOUND());
 
-        // Attach the observer immediately so the queued (👀) and
-        // working reactions don't fire before the listener exists.
-        const trail: ObservedOp[] = [];
-        const iter = sc.driver
-          .observeReactions(sc.botUserId, { messageId: sent.messageId })
-          [Symbol.asyncIterator]();
-        let pump: Promise<void> | null = null;
-        let stopPump = false;
-        pump = (async () => {
-          while (!stopPump) {
-            const next = await iter.next();
-            if (next.done === true) return;
-            trail.push({
-              emoji: next.value.emoji,
-              op: next.value.op,
-              at: Date.now(),
-            });
-          }
-        })();
+        // Poll the reaction state on the sent message. We use polling
+        // rather than `observeReactions` because Telegram does not
+        // deliver `updateMessageReactions` push updates to user accounts
+        // when a bot sets a reaction in a DM — see module docblock.
+        const reactionHistory: string[][] = [];
+        let replyReceived = false;
+        let replyReceivedAt = 0;
 
-        try {
-          // Wait for the bot's reply (any content). Gives the L1
-          // lifecycle time to traverse queued → working → done.
-          const reply = await sc.expectMessage(/\S/, {
-            from: "bot",
-            timeout: 60_000,
+        // Wait for the bot's reply (any content). We start polling
+        // concurrently so we capture intermediate reactions during
+        // the turn.
+        let replyTimedOut = false;
+        const replyPromise = sc
+          .expectMessage(/\S/, { from: "bot", timeout: 60_000 })
+          .then((reply) => {
+            expect(reply.text.length).toBeGreaterThan(0);
+            replyReceived = true;
+            replyReceivedAt = Date.now();
+            return reply;
+          })
+          .catch((err: unknown) => {
+            // expectMessage timeout → mark so the poll loop exits immediately
+            // instead of burning the full 90s safety ceiling on CI failure.
+            replyTimedOut = true;
+            throw err;
           });
-          expect(reply.text.length).toBeGreaterThan(0);
 
-          // Tail after the reply for Telegram to deliver the
-          // terminal-emoji replace. In practice <1s on a healthy bot;
-          // 8s ceiling absorbs server batching jitter.
-          await new Promise((resolve) =>
-            setTimeout(resolve, TAIL_AFTER_REPLY_MS),
-          );
-        } finally {
-          stopPump = true;
-          await iter.return?.();
-          if (pump) {
-            await pump.catch(() => {
-              /* generator return triggers rejection on pending iter.next() — ignore */
-            });
+        // Polling loop: sample reactions while waiting for the reply,
+        // then continue for TAIL_AFTER_REPLY_MS after the reply lands.
+        const poll = async (): Promise<void> => {
+          const deadline = Date.now() + 90_000; // safety ceiling
+          while (Date.now() < deadline) {
+            // Bail early if the reply timed out — no point polling further.
+            if (replyTimedOut) break;
+            const emojis = await sc.driver.pollReactions(
+              sc.botUserId,
+              sent.messageId,
+            );
+            if (emojis.length > 0) {
+              reactionHistory.push([...emojis]);
+            }
+            if (
+              replyReceived &&
+              Date.now() - replyReceivedAt >= TAIL_AFTER_REPLY_MS
+            ) {
+              break;
+            }
+            await new Promise((resolve) => setTimeout(resolve, POLL_INTERVAL_MS));
           }
-        }
+        };
+
+        // Run both concurrently; wait for both to settle.
+        await Promise.all([replyPromise, poll()]);
 
-        // L1 alive signal: at least one `+` op landed during the turn.
-        const adds = trail.filter((o) => o.op === "+");
+        // L1 alive signal: at least one non-empty reaction set was
+        // observed during the turn.
+        const allSeen = reactionHistory.flat();
+        const uniqueSeen = [...new Set(allSeen)];
         expect(
-          adds.length,
-          `expected at least one reaction-add during the turn, got 0. ` +
-            `Full trail: ${trail.map((o) => `${o.op}${o.emoji}`).join(" ") || "(empty)"}`,
+          reactionHistory.length,
+          `expected at least one reaction poll to show a reaction during the ` +
+            `turn, but all ${reactionHistory.length > 0 ? "polls returned nothing with emojis" : "polls returned empty"}. ` +
+            `History snapshots: ${reactionHistory.map((s) => `[${s.join(",")}]`).join(" ") || "(none)"}`,
         ).toBeGreaterThan(0);
 
-        // L1 terminal: the LAST `+` op should be a terminal-done emoji.
-        // Extra `-` ops after the final `+` are tolerated (Telegram
-        // sometimes emits a bare clean-up `-`); the last `+` is what
-        // the user actually sees.
-        const lastAdd = adds[adds.length - 1];
+        // L1 terminal: the LAST non-empty snapshot should contain a
+        // terminal-done emoji. `setMessageReaction` replaces atomically,
+        // so the last snapshot holds whatever is currently on the message.
+        const lastSnapshot = reactionHistory[reactionHistory.length - 1];
+        // The bot uses setMessageReaction (replace, not append) — exactly one
+        // emoji should be set at any time. Assert the invariant so we catch
+        // accidental multi-emoji states, then check the terminal-done value.
+        expect(
+          lastSnapshot.length,
+          `expected exactly 1 reaction in the final snapshot, got [${lastSnapshot.join(",")}]. ` +
+            `All snapshots: ${reactionHistory.map((s) => `[${s.join(",")}]`).join(" ")}`,
+        ).toBe(1);
+        const lastEmoji = lastSnapshot[0];
         expect(
-          TERMINAL_DONE_EMOJI.has(lastAdd.emoji),
-          `expected last reaction-add to be one of ${[
+          TERMINAL_DONE_EMOJI.has(lastEmoji),
+          `expected last reaction to be one of ${[
             ...TERMINAL_DONE_EMOJI,
-          ].join(", ")}, got ${lastAdd.emoji}. Full trail: ${trail
-            .map((o) => `${o.op}${o.emoji}`)
-            .join(" ")}`,
+          ].join(", ")}, got ${lastEmoji}. ` +
+            `All snapshots: ${reactionHistory.map((s) => `[${s.join(",")}]`).join(" ")}. ` +
+            `Unique emojis seen: ${uniqueSeen.join(", ") || "(none)"}`,
         ).toBe(true);
       } finally {
         await sc.tearDown();

package/telegram-plugin/welcome-text.ts

@@ -11,6 +11,8 @@
  * monospace inline and avoid Telegram treating them as markdown.
  */
 
+import { maskUsername } from "./demo-mask.js";
+
 export type AuthSummary = {
   authenticated: boolean;
   subscription_type: string | null;
@@ -198,10 +200,19 @@ const STATUS_DOT: Record<StatusProbeRow['status'], string> = {
 export function statusPairedText(params: {
   user: string;
   meta: AgentMetadata;
+  /**
+   * Demo mode (the `/status demo` suffix). When true the paired-user tag
+   * (`@handle` or numeric sender id) is run through `maskUsername` so a
+   * screen recording shows a stable fake `@demo_user…` handle instead of
+   * the operator's real Telegram identity. Off by default — the agent /
+   * model / health / audit topology below is NOT masked (out of scope).
+   */
+  demo?: boolean;
 }): string {
   const { user, meta } = params;
+  const shownUser = params.demo ? maskUsername(user) : user;
   const lines = [
-    `Paired as ${escapeHtml(user)}.`,
+    `Paired as ${escapeHtml(shownUser)}.`,
     ``,
     `Agent: ${formatAgentLine(meta)}`,
     `Auth: ${formatAuthLine(meta.auth)}`,
@@ -327,6 +338,7 @@ export const TELEGRAM_MENU_COMMANDS = [
   { command: "effort", description: "Show or switch the reasoning effort" },
   { command: "doctor", description: "Health check (deps, services, MCP)" },
   { command: "usage", description: "Pro/Max plan quota (5h + 7d windows)" },
+  { command: "whoami", description: "This agent's sandbox: tools, MCP, vault key-names" },
   // Vault — secrets + capability grants. /vault is a top-level command
   // dispatching subcommands (list, get, set, delete, status, unlock, lock,
   // grant, grants). Surfaced in the menu so mobile users can tap-to-pick