switchroom 0.15.45 → 0.16.4

package/telegram-plugin/uat/scenarios/jtbd-foreground-feed-visibility-dm.test.ts

@@ -0,0 +1,396 @@
+/**
+ * Foreground activity-feed visibility across the silence-fallback threshold.
+ *
+ * ## Cause class
+ *
+ * A foreground turn that does REAL sequential work (several tool calls, each
+ * followed by a model-thinking gap) can trip the silence-fallback timer
+ * (`SILENCE_FALLBACK_MS`, default 300 000 ms) even while it is visibly
+ * progressing. The silence clock is reset by:
+ *
+ *   - A fresh `reply` or `stream_reply` first-emit (any real user-visible send).
+ *   - `SILENCE_LIVENESS_PRODUCTION` ON (the default): a new tool-activity label
+ *     appearing on the feed, or an answer-stream draft update.
+ *
+ * Crucially, every render resets the clock to ZERO — so the failure is NOT
+ * cumulative across many short gaps. The feed only darkens on a SINGLE
+ * continuous no-render window longer than the threshold. Heartbeat edits keep
+ * the message visually advancing but do NOT count as liveness, so they do not
+ * reset the clock; only a real tool-label render or an answer-stream draft does.
+ *
+ * This file is the DETERMINISTIC guard half of the pair. The fleet runs
+ * `SWITCHROOM_SILENCE_DEFER_INFLIGHT_TOOLS=1` (set in defaults.env — it is the
+ * fleet default, despite a stale gateway.ts comment that still says "OFF, canary
+ * on marko"). With the defer ON, a single long IN-FLIGHT tool does NOT trip the
+ * base fallback: the defer holds it back while the tool runs (up to the hard
+ * ceiling, default 15 min) and the feed heartbeat keeps editing the live
+ * message. So the CORRECT behaviour for a long in-flight turn is: the feed stays
+ * lit. This guard pins exactly that. A regression that breaks the defer, stops
+ * the heartbeat, or nulls `currentTurn` for an in-flight turn darkens the feed,
+ * and this test catches it.
+ *
+ * The workload is one ~35 s no-output command (`sleep 35`) — well under the
+ * default 15 min hard ceiling — so under prod config the feed must stay live
+ * across the shrunk 20 s base fallback. A prompt of several FAST steps would not
+ * exercise the silence window at all (each tool start re-renders and resets the
+ * clock); one long stretch is what holds the clock open.
+ *
+ * This guard does NOT reproduce #680's exact trigger — silent model thinking
+ * BETWEEN tools, with no tool in-flight. The defer does not cover that vector
+ * and it cannot be forced deterministically; it lives in the sibling best-effort
+ * scenario `jtbd-foreground-feed-thinkgap-dm.test.ts`. Together the pair covers
+ * the feed-visibility invariant deterministically (here) and the true #680
+ * vector best-effort (there).
+ *
+ * This scenario shrinks the base fallback to 20 s via
+ * `SWITCHROOM_SILENCE_FALLBACK_MS=20000` on the test-harness agent so the
+ * window is exercised within a test budget instead of 5 minutes.
+ *
+ * ## Required env precondition (operator must set on test-harness agent)
+ *
+ *   SWITCHROOM_SILENCE_FALLBACK_MS=20000
+ *
+ * Set this under the `test-harness` agent's `env:` block in
+ * `~/.switchroom/switchroom.yaml`, then restart the agent
+ * (`switchroom agent restart test-harness --wait --force`) before running
+ * this scenario. The scenario detects whether the threshold is plausibly
+ * shrunk by reading `SWITCHROOM_UAT_SILENCE_FALLBACK_MS` from the test
+ * env (a parallel knob populated by the UAT `.env` file) and
+ * skip-with-message if it is not set to ≤ 30 000.
+ *
+ * Without the shrunk threshold the scenario still runs but the timing
+ * assertions become vacuous (the default 300 s threshold far exceeds the
+ * test budget) and the test exits before the scenario would have had a
+ * chance to catch a regression. The skip keeps the failure signal honest.
+ *
+ * ## What it asserts (the gap no existing scenario covers)
+ *
+ * 1. **Feed opened.** An activity-feed message (`→`/`✓` lines) appears in the
+ *    DM at some point during the turn — the agent started reporting progress.
+ *
+ * 2. **Feed survived the fallback window.** After the silence-fallback interval
+ *    has elapsed from the point the feed was first observed, the feed message
+ *    is still present and carries at least one more edit that arrived AFTER
+ *    the threshold mark. If `currentTurn` was nulled mid-turn, the gateway
+ *    stops sending activity-feed edits and the message goes stale or disappears
+ *    — this assertion catches that.
+ *
+ * 3. **Final answer arrives.** A substantive reply (≥ 150 chars) eventually
+ *    lands, confirming the turn completed rather than being wedged.
+ *
+ * ## Failure shapes
+ *
+ *   (a) Feed never opened — the activity feed did not paint at all. Either the
+ *       agent never used tools or the very first drainActivitySummary call
+ *       failed. Distinct from the regression; both are failures.
+ *
+ *   (b) Feed went dark — the feed message was present before the fallback mark
+ *       but received no fresh edit after the mark. This IS the regression this
+ *       test exists to catch: `currentTurn` was nulled mid-turn, silencing the
+ *       live feed while the agent was still working.
+ *
+ *   (c) No final answer — the turn never produced a substantive reply. Possibly
+ *       the fallback also dropped the answer path (compound regression), or the
+ *       prompt was too slow for the overall test budget.
+ *
+ * ## Tolerances
+ *
+ * The feed edit observation is polled via `driver.getMessage` (the same
+ * technique used by `jtbd-worker-activity-feed-dm`). Because mtcute's live
+ * `observeMessages` may miss edits that arrive before the observer is attached,
+ * we cross-check by comparing the snapshot taken just before the fallback mark
+ * against a fresh fetch just after it. A changed body confirms a live edit
+ * occurred across the threshold; an UNCHANGED body with no subsequent new edit
+ * in the live stream is the regression signal.
+ *
+ * The prompt is engineered so the model does at least 4 sequential tool calls
+ * with brief thinking gaps between them, each step taking ~4–6 s, giving a
+ * natural total span of ~25–35 s that straddles the shrunk 20 s fallback.
+ */
+
+import { describe, expect, it } from "vitest";
+import { spinUp } from "../harness.js";
+import { isActivityFeedMessage } from "../assertions.js";
+import type { ObservedMessage } from "../driver.js";
+
+/**
+ * The shrunk fallback threshold the operator must set on the test-harness
+ * agent. The test reads a parallel UAT-env knob so we can detect whether
+ * the precondition is satisfied without reaching into the agent's process env.
+ *
+ * Set `SWITCHROOM_UAT_SILENCE_FALLBACK_MS=20000` in the repo-root `.env`
+ * alongside `SWITCHROOM_SILENCE_FALLBACK_MS=20000` in the agent's env block.
+ */
+const PRECONDITION_FALLBACK_MS = Number.parseInt(
+  process.env.SWITCHROOM_UAT_SILENCE_FALLBACK_MS ?? "",
+  10,
+);
+
+/**
+ * How long to wait after the feed first appears before taking the
+ * "before-mark" snapshot. We want the fallback timer to have clearly
+ * elapsed by the time we take the "after-mark" snapshot. Chosen to be
+ * safely above the shrunk 20 s fallback while staying comfortably within
+ * the test budget.
+ */
+const FALLBACK_WINDOW_MS = 25_000;
+
+/**
+ * How long to poll after the fallback window for a fresh feed edit. Short
+ * enough not to waste budget but long enough for one more drainActivitySummary
+ * cycle to land (the feed heartbeat fires roughly every 5–8 s).
+ */
+const POST_MARK_EDIT_WAIT_MS = 15_000;
+
+/**
+ * A substantive answer is at least this many characters. Avoids latching
+ * onto a brief "on it" ack or a stub.
+ */
+const MIN_ANSWER_CHARS = 150;
+
+/**
+ * Overall test budget. Includes:
+ *   - spinUp settle: ~8 s
+ *   - turn onset (first tool + first feed paint): ~20 s
+ *   - FALLBACK_WINDOW_MS: 25 s
+ *   - POST_MARK_EDIT_WAIT_MS: 15 s
+ *   - final-answer wait: ~30 s
+ *   - headroom: ~20 s
+ */
+const OVERALL_BUDGET_MS = 150_000;
+
+/**
+ * Workload prompt: one ~33 s no-output command, `timeout 33 tail -f /dev/null`.
+ * NOTE: standalone `sleep` is blocked by the Claude Code harness ("foreground
+ * sleep is blocked"), so this is the hook-safe equivalent of a long in-flight
+ * no-op. The tool's label renders once at its start (resetting the clock to
+ * zero), then nothing renders for ~33 s while the tool is in-flight. Under prod
+ * config (defer ON) the base fallback is HELD during that in-flight stretch, so
+ * the feed must stay live and the heartbeat keeps editing it. That is the
+ * invariant this guard asserts; a regression that lets `currentTurn` get nulled
+ * mid-stretch breaks it.
+ *
+ * The prompt explicitly asks the model to give the Bash call a short
+ * DESCRIPTION. That matters: an empty-label tool is dropped from the activity
+ * feed (it never opens), which would fail the test on "feed never appeared"
+ * (shape a) for the wrong reason. A labelled tool opens the feed reliably.
+ *
+ * Why not several fast steps: each fast tool start emits a fresh label that
+ * resets the clock, so the silence window never opens at all and the test would
+ * pass vacuously (a false green). One long stretch is what holds it open.
+ *
+ * We do NOT use run_in_background — this must be a FOREGROUND turn so
+ * currentTurn stays in place and the silence clock applies to it directly.
+ */
+const SEQUENTIAL_WORK_PROMPT =
+  "Use the Bash tool to run EXACTLY this command — and give the tool call a " +
+  'short description such as "long-running wait" so it is clearly labelled: ' +
+  "`timeout 33 tail -f /dev/null`. It runs for about 33 seconds and then exits " +
+  "on its own (a non-zero timeout exit code is expected and fine). Do not run " +
+  "any other tool while it is running. After it finishes, reply with a short " +
+  "paragraph (a few sentences) telling me it completed and that you waited " +
+  "about 33 seconds.";
+
+describe("uat: foreground activity-feed visibility across silence-fallback threshold", () => {
+  it(
+    "feed remains live and receives edits after the shrunk fallback window",
+    async () => {
+      // Precondition guard: if the operator hasn't shrunk the silence
+      // fallback to ≤ 30 000 ms on the test-harness agent (and mirrored
+      // it into the UAT env), the timing assertions are vacuous. Skip
+      // with a clear message rather than silently producing a false green.
+      if (!Number.isFinite(PRECONDITION_FALLBACK_MS) || PRECONDITION_FALLBACK_MS > 30_000) {
+        console.warn(
+          "[uat/foreground-feed-visibility] SKIPPED — precondition not met.\n" +
+            "  This scenario requires SWITCHROOM_SILENCE_FALLBACK_MS=20000 set on\n" +
+            "  the test-harness agent AND SWITCHROOM_UAT_SILENCE_FALLBACK_MS=20000\n" +
+            "  in the repo-root .env. Without it the silence fallback does not fire\n" +
+            "  within the test window and the regression cannot be detected.\n" +
+            "  See the header doc comment in this file for setup instructions.",
+        );
+        return;
+      }
+
+      const sc = await spinUp({ agent: "test-harness" });
+      try {
+        // Start observing BEFORE sending so no activity-feed messages are missed.
+        const iter = sc.driver
+          .observeMessages(sc.botUserId)
+          [Symbol.asyncIterator]();
+
+        await sc.sendDM(SEQUENTIAL_WORK_PROMPT);
+
+        console.log("[foreground-feed] prompt sent; watching for activity-feed message…");
+
+        // ── Assertion 1: feed opened ─────────────────────────────────────────
+        // Drain the live message stream until we see an activity-feed message
+        // (lines matching `→ …` or `✓ …`). Give generous budget for the agent to
+        // start tools and open the feed.
+        let feedMsg: ObservedMessage | null = null;
+        const feedDeadline = Date.now() + 90_000;
+
+        while (Date.now() < feedDeadline) {
+          const remaining = feedDeadline - Date.now();
+          const next = await Promise.race([
+            iter.next(),
+            new Promise<{ done: true; value: undefined }>((r) =>
+              setTimeout(() => r({ done: true, value: undefined }), Math.max(0, remaining)),
+            ),
+          ]);
+          if (next.done || next.value == null) break;
+          const m = next.value as ObservedMessage;
+          // Skip our own echo and worker-feed messages; we want the foreground
+          // activity feed.
+          if (m.senderUserId === sc.driverUserId) continue;
+          if (isActivityFeedMessage(m)) {
+            feedMsg = m;
+            break;
+          }
+        }
+
+        expect(
+          feedMsg,
+          "Failure shape (a): the foreground activity-feed message never appeared. " +
+            "Either the agent did not use tools, the prompt was too fast for the " +
+            "feed to paint, or drainActivitySummary failed on every attempt this turn.",
+        ).not.toBeNull();
+
+        // feedMsg is confirmed non-null beyond this point.
+        const { messageId: feedId } = feedMsg!;
+        console.log(
+          `[foreground-feed] feed opened (id=${feedId}): ` +
+            JSON.stringify(feedMsg!.text.slice(0, 120)),
+        );
+
+        // ── Snapshot before the fallback mark ───────────────────────────────
+        // Record the feed body and clock, then wait FALLBACK_WINDOW_MS so the
+        // shrunk base fallback has definitely elapsed. During this wait the
+        // single in-flight stretch (`sleep 35`) emits no renders, but the defer
+        // (prod default ON) HOLDS the base fallback while the tool is in-flight,
+        // so currentTurn survives and the feed heartbeat keeps editing the
+        // message. We assert those heartbeat edits keep landing after the mark;
+        // if a regression nulls currentTurn mid-stretch, the edits stop and we
+        // catch it.
+        const beforeMarkText = feedMsg!.text;
+        const markAt = Date.now() + FALLBACK_WINDOW_MS;
+
+        console.log(
+          `[foreground-feed] waiting ${FALLBACK_WINDOW_MS}ms for fallback window to elapse…`,
+        );
+        await new Promise((r) => setTimeout(r, FALLBACK_WINDOW_MS));
+
+        // ── Assertion 2: feed survived the fallback window ───────────────────
+        // Fetch the feed message directly. If currentTurn was nulled mid-turn,
+        // the gateway either stopped editing (stale text) or the message may have
+        // been deleted by clearActivitySummary (null). Either condition is the
+        // regression.
+        const afterMark = await sc.driver.getMessage(sc.botUserId, feedId);
+
+        console.log(
+          `[foreground-feed] feed state after ${FALLBACK_WINDOW_MS}ms mark ` +
+            `(id=${feedId}): ` +
+            JSON.stringify(afterMark?.text?.slice(0, 120) ?? null),
+        );
+
+        expect(
+          afterMark,
+          "Failure shape (b): the activity-feed message was deleted after the " +
+            `silence-fallback window (${FALLBACK_WINDOW_MS}ms). This means ` +
+            "currentTurn was nulled mid-turn by the silence-fallback handler, " +
+            "which then triggered clearActivitySummary and removed the live feed " +
+            "message. The regression: a productive foreground turn went dark " +
+            "because a continuous no-render window exceeded the " +
+            `shrunk threshold (SWITCHROOM_SILENCE_FALLBACK_MS=${PRECONDITION_FALLBACK_MS}).`,
+        ).not.toBeNull();
+
+        // The body should have changed — i.e. the feed received at least one
+        // edit after the fallback mark — proving it is still alive. We check
+        // both the polled snapshot and the live-stream edits we may have
+        // collected during the wait.
+        const bodyChangedAfterMark = afterMark!.text !== beforeMarkText;
+
+        // Also drain any edits that arrived during the POST_MARK_EDIT_WAIT_MS
+        // window to catch the next drainActivitySummary cycle if the polled
+        // snapshot was taken slightly before the edit landed.
+        let sawFeedEditAfterMark = bodyChangedAfterMark;
+        const postMarkDeadline = Date.now() + POST_MARK_EDIT_WAIT_MS;
+
+        while (!sawFeedEditAfterMark && Date.now() < postMarkDeadline) {
+          const remaining = postMarkDeadline - Date.now();
+          const next = await Promise.race([
+            iter.next(),
+            new Promise<{ done: true; value: undefined }>((r) =>
+              setTimeout(() => r({ done: true, value: undefined }), Math.max(0, remaining)),
+            ),
+          ]);
+          if (next.done || next.value == null) break;
+          const m = next.value as ObservedMessage;
+          if (m.senderUserId === sc.driverUserId) continue;
+          // An edit of the feed message that arrived after the mark
+          if (m.edited && m.messageId === feedId && m.date.getTime() >= markAt) {
+            sawFeedEditAfterMark = true;
+            console.log(
+              `[foreground-feed] feed edit confirmed after mark (id=${feedId}): ` +
+                JSON.stringify(m.text.slice(0, 120)),
+            );
+          }
+        }
+
+        expect(
+          sawFeedEditAfterMark,
+          "Failure shape (b): the activity-feed message still exists but received " +
+            `no edit after the ${FALLBACK_WINDOW_MS}ms fallback window. This is the ` +
+            "feed-went-dark regression: currentTurn was nulled mid-turn so no further " +
+            "drainActivitySummary calls fired. The feed body was frozen at " +
+            `${JSON.stringify(beforeMarkText.slice(0, 80))} and did not advance.`,
+        ).toBe(true);
+
+        // ── Assertion 3: final answer lands ─────────────────────────────────
+        // Collect from the live stream until a substantive bot reply arrives.
+        // This confirms the turn completed — the fallback did not also drop
+        // the answer path.
+        let finalAnswer: ObservedMessage | null = null;
+        const answerDeadline = Date.now() + 60_000;
+
+        while (Date.now() < answerDeadline) {
+          const remaining = answerDeadline - Date.now();
+          const next = await Promise.race([
+            iter.next(),
+            new Promise<{ done: true; value: undefined }>((r) =>
+              setTimeout(() => r({ done: true, value: undefined }), Math.max(0, remaining)),
+            ),
+          ]);
+          if (next.done || next.value == null) break;
+          const m = next.value as ObservedMessage;
+          if (m.senderUserId === sc.driverUserId) continue;
+          if (m.edited) continue; // edits are feed updates, not the answer
+          if (isActivityFeedMessage(m)) continue; // skip feed-only sends
+          if (m.text.trim().length >= MIN_ANSWER_CHARS) {
+            finalAnswer = m;
+            break;
+          }
+        }
+
+        console.log(
+          `[foreground-feed] final answer (id=${finalAnswer?.messageId ?? "NONE"}): ` +
+            JSON.stringify(finalAnswer?.text?.slice(0, 180) ?? null),
+        );
+
+        expect(
+          finalAnswer,
+          "Failure shape (c): no final answer arrived after the silence-fallback " +
+            `window. The turn did not produce a substantive reply (≥${MIN_ANSWER_CHARS} chars). ` +
+            "If the feed-gone-dark assertion also failed, the fallback may have " +
+            "suppressed the entire turn's output. If only this assertion failed, " +
+            "the turn is still in flight past the test budget — increase the prompt " +
+            "timeout or check that the agent is not wedged.",
+        ).not.toBeNull();
+
+        await iter.return?.();
+      } finally {
+        await sc.tearDown();
+      }
+    },
+    OVERALL_BUDGET_MS,
+  );
+});

package/telegram-plugin/uat/scenarios/jtbd-liveness-feed-open-dm.test.ts

@@ -0,0 +1,202 @@
+/**
+ * Liveness-driven feed open — a thinking-only turn still surfaces a live feed.
+ *
+ * ## Cause class (the #680 dark-turn, true vector)
+ *
+ * The activity feed is TOOL-driven: it opens only when a tool emits a non-null
+ * label (`drainActivitySummary`). A turn dominated by model thinking, or by
+ * suppressed-by-design tools (typing / memory recall / reply), emits no label —
+ * so the feed never opens and a long turn reads as pure silence until the 300s
+ * silence-poke. Turn #680 was exactly this: 335 s alive, `tools=4`, yet
+ * `feedOpened=false / activityMsgId=none` the entire time.
+ *
+ * The fix (gateway `feedHeartbeatTick`): once a turn has been alive for
+ * `FEED_LIVENESS_OPEN_MS` with no labelled tool yet, open a minimal `Working…`
+ * feed and let the existing 6 s heartbeat climb its elapsed. The first real
+ * tool label takes over and its edit replaces the placeholder; a pure-thinking
+ * turn finalizes to `✓ Working…` rather than freezing on the live line.
+ *
+ * ## Precondition (set on the test-harness agent for this run)
+ *
+ *   SWITCHROOM_FEED_LIVENESS_OPEN_MS=6000   (shrinks the default 12 s so the
+ *                                            window is exercised within budget)
+ *
+ * The scenario does not hard-require it — at the 12 s default the feed still
+ * opens, just later; the deadlines below tolerate either.
+ *
+ * ## What it asserts (asymmetric — non-determinism is handled, not faked)
+ *
+ * The trigger is a "think, then answer at length, use NO tools" prompt. The
+ * model's exact behaviour is not fully forceable, so the branches are:
+ *
+ *   PASS         — a `Working…` activity-feed message appeared. ONLY the
+ *                  liveness path produces a bare "Working…" feed (a tool would
+ *                  carry a tool label), so this is positive proof the timer
+ *                  opened the feed on a tool-less turn.
+ *   INCONCLUSIVE — the agent used a tool anyway (feed opened with a tool label,
+ *                  no "Working…"). The liveness path was not exercised; not a
+ *                  failure of the fix. Warn + pass.
+ *   INCONCLUSIVE — the turn was too short (answer landed before the threshold).
+ *                  Warn + pass.
+ *   HARD FAIL    — the turn ran clearly longer than the threshold with NO feed
+ *                  of any kind, yet produced an answer. Liveness should have
+ *                  opened a feed and did not — the regression this guard exists
+ *                  to catch.
+ *   FAIL         — no answer at all within budget (wedged).
+ */
+
+import { describe, expect, it } from "vitest";
+import { spinUp } from "../harness.js";
+import { isActivityFeedMessage } from "../assertions.js";
+import type { ObservedMessage } from "../driver.js";
+
+/** A substantive answer is at least this many characters (skips a brief ack). */
+const MIN_ANSWER_CHARS = 200;
+
+/** Overall test budget. */
+const OVERALL_BUDGET_MS = 150_000;
+
+/**
+ * Workload: think, then answer at length, with NO tools. The long answer
+ * generation holds the turn open past the liveness threshold with no tool
+ * label, which is exactly the condition that should open the `Working…` feed.
+ */
+const THINKING_WORKLOAD_PROMPT =
+  "Do NOT use any tools at all for this — no Bash, no Read, no memory search, " +
+  "nothing. Just think carefully and then write me a thorough, detailed " +
+  "explanation (at least 450 words) of how the TCP three-way handshake works, " +
+  "including SYN, SYN-ACK, ACK, sequence numbers, and what happens if the final " +
+  "ACK is lost. Take your time getting it right, then reply with the full essay " +
+  "in one message.";
+
+describe("uat: liveness-driven feed open (thinking-only turn stays visible)", () => {
+  it(
+    "opens a 'Working…' feed for a turn that emits no tool label",
+    async () => {
+      const sc = await spinUp({ agent: "test-harness" });
+      try {
+        const iter = sc.driver
+          .observeMessages(sc.botUserId)
+          [Symbol.asyncIterator]();
+
+        await sc.sendDM(THINKING_WORKLOAD_PROMPT);
+        const sentAt = Date.now();
+        console.log("[liveness-feed] prompt sent; watching for feed + answer…");
+
+        // Drain the stream until EITHER a feed message appears OR a substantive
+        // answer lands. Track which kind of feed (liveness vs tool) we saw.
+        let livenessFeed: ObservedMessage | null = null;
+        let toolFeed: ObservedMessage | null = null;
+        let answer: ObservedMessage | null = null;
+        let answerAt = 0;
+
+        const deadline = Date.now() + 110_000;
+        while (Date.now() < deadline) {
+          if (livenessFeed && answer) break;
+          const remaining = deadline - Date.now();
+          const next = await Promise.race([
+            iter.next(),
+            new Promise<{ done: true; value: undefined }>((r) =>
+              setTimeout(
+                () => r({ done: true, value: undefined }),
+                Math.max(0, remaining),
+              ),
+            ),
+          ]);
+          if (next.done || next.value == null) break;
+          const m = next.value as ObservedMessage;
+          if (m.senderUserId === sc.driverUserId) continue;
+
+          if (isActivityFeedMessage(m)) {
+            // A "Working…" feed body is the liveness placeholder; anything else
+            // is a tool-label feed.
+            if (/Working/.test(m.text)) {
+              if (!livenessFeed) {
+                livenessFeed = m;
+                console.log(
+                  `[liveness-feed] LIVENESS feed opened at +${Date.now() - sentAt}ms: ` +
+                    JSON.stringify(m.text.slice(0, 120)),
+                );
+              }
+            } else if (!toolFeed) {
+              toolFeed = m;
+              console.log(
+                `[liveness-feed] tool-label feed opened at +${Date.now() - sentAt}ms: ` +
+                  JSON.stringify(m.text.slice(0, 120)),
+              );
+            }
+            continue;
+          }
+          if (m.edited) continue;
+          if (m.text.trim().length >= MIN_ANSWER_CHARS && !answer) {
+            answer = m;
+            answerAt = Date.now();
+            console.log(
+              `[liveness-feed] answer landed at +${answerAt - sentAt}ms (len=${m.text.trim().length}).`,
+            );
+          }
+        }
+
+        const turnSpanMs = (answerAt || Date.now()) - sentAt;
+
+        // ── Branch resolution ───────────────────────────────────────────────
+        if (livenessFeed) {
+          // PASS: the liveness timer opened a feed on a tool-less turn.
+          expect(
+            livenessFeed,
+            "liveness feed should be present in the PASS branch",
+          ).not.toBeNull();
+          // Confirm it carries the in-progress placeholder shape.
+          expect(livenessFeed!.text).toMatch(/Working/);
+          console.log(
+            "[liveness-feed] PASS — liveness-driven feed open confirmed.",
+          );
+          return;
+        }
+
+        if (toolFeed) {
+          console.warn(
+            "[liveness-feed] INCONCLUSIVE — the agent used a tool, so the feed " +
+              "opened via the normal tool-label path and the liveness timer was " +
+              "not exercised. Not a failure of the fix.",
+          );
+          // Still require the turn completed.
+          expect(
+            answer,
+            "even in the tool-feed branch the turn must complete with an answer",
+          ).not.toBeNull();
+          return;
+        }
+
+        // No feed of any kind appeared.
+        if (answer && turnSpanMs < 8_000) {
+          console.warn(
+            `[liveness-feed] INCONCLUSIVE — turn completed in ${turnSpanMs}ms, ` +
+              "below the liveness threshold; no feed was expected.",
+          );
+          return;
+        }
+
+        // No feed, and the turn ran long enough that liveness SHOULD have opened.
+        expect(
+          answer,
+          "FAIL — no answer arrived within budget; the turn may be wedged.",
+        ).not.toBeNull();
+
+        expect(
+          livenessFeed,
+          "HARD FAIL — the turn ran for " +
+            `${turnSpanMs}ms (well past the liveness threshold) with NO activity ` +
+            "feed of any kind, yet produced an answer. The liveness timer should " +
+            "have opened a 'Working…' feed and did not — this is the #680 " +
+            "dark-turn regression.",
+        ).not.toBeNull();
+
+        await iter.return?.();
+      } finally {
+        await sc.tearDown();
+      }
+    },
+    OVERALL_BUDGET_MS,
+  );
+});