From eea5ea3bd64d03401eb0b21b9ac9eddf9092c834 Mon Sep 17 00:00:00 2001
From: ciregenz <ciregenz@berkeley.edu>
Date: Sun, 26 Jul 2026 22:53:41 -0700
Subject: [PATCH] [eric] browser: seal today's three bug classes (mute aux
 lane, finder budget, soft signed-out)

---
 backend/tests/test_browser_classifier_lane.py | 43 ++++++++++
 backend/tests/test_browser_signed_out.py      | 82 +++++++++++++++++++
 2 files changed, 125 insertions(+)
 create mode 100644 backend/tests/test_browser_classifier_lane.py
 create mode 100644 backend/tests/test_browser_signed_out.py

diff --git a/backend/tests/test_browser_classifier_lane.py b/backend/tests/test_browser_classifier_lane.py
new file mode 100644
index 00000000..fc797e46
--- /dev/null
+++ b/backend/tests/test_browser_classifier_lane.py
@@ -0,0 +1,43 @@
+"""The fast-path classifier must not let a mute provider disable the browser fast path.
+
+Measured live 2026-07-26: on the codex lane the aux resolves to cx/gpt-5.4-mini, which returns an
+EMPTY body for this call. Empty parsed to verdict 'no', which reads identically to "this is not a
+browser task", so the entire browser fast path silently switched off for every GPT user with no
+error to show for it (gpt-5.4 filled 0/3; after the empty-body fallback, 2/2). These tests pin both
+halves of the contract: the parser may call empty 'no', and the caller must not accept that as a
+verdict when a provider was pinned.
+"""
+from backend.apps.agents.browser import browser_fast_path as fp
+
+CLAUDE_REPLY = "ACT\n\nENTRY: https://x.com/home\n\n1. Navigate to X.\n2. Click the composer."
+
+
+def test_empty_body_parses_as_no() -> None:
+    """The parser is allowed to say 'no' on empty; the BUG was the caller treating that as a real
+    verdict instead of a mute lane."""
+    assert fp.parse_verdict_and_brief("") == ("no", "")
+
+
+def test_whitespace_only_body_is_treated_as_empty() -> None:
+    """A lane answering with only whitespace is exactly as mute as one answering ''."""
+    assert fp.parse_verdict_and_brief("   \n\t \n") == ("no", "")
+
+
+def test_real_verdict_still_parses() -> None:
+    verdict, brief = fp.parse_verdict_and_brief(CLAUDE_REPLY)
+    assert verdict == "act"
+    assert "ENTRY: https://x.com/home" in brief
+
+
+def test_read_verdict_still_parses() -> None:
+    verdict, _ = fp.parse_verdict_and_brief("READ\n\nENTRY: https://example.com")
+    assert verdict == "read"
+
+
+def test_caller_retries_provider_agnostic_on_a_mute_lane() -> None:
+    """The fix itself: classify_and_brief must re-ask WITHOUT the provider pin when the pinned lane
+    returns nothing. Pinned by source so the retry cannot be quietly deleted."""
+    import inspect
+    src = inspect.getsource(fp.classify_and_brief)
+    assert "if not text.strip() and primary_api:" in src, "empty-body fallback missing"
+    assert "p_ask(None)" in src, "fallback must drop the provider pin"
diff --git a/backend/tests/test_browser_signed_out.py b/backend/tests/test_browser_signed_out.py
new file mode 100644
index 00000000..7bc3c27a
--- /dev/null
+++ b/backend/tests/test_browser_signed_out.py
@@ -0,0 +1,82 @@
+"""Soft signed-out detection, plus the two invariants behind the 2026-07-26 bug hunt.
+
+Each test here corresponds to a bug that actually escaped to a live run, so the point is to make
+that whole CLASS unwritable rather than to re-check a line:
+
+  1. Sites that browse fine while withholding the composer (bsky/stackoverflow/tiktok/threads) hit
+     no login URL and show no password field, so the hard-wall gate saw nothing and the run
+     reported "couldn't find the compose box" when the truth was "you are not signed in".
+  2. The composer finder budgeted ~24s of reveal work into a 15s command timeout, so heavy pages
+     were killed mid-ladder and threw away everything, and the last two tiers were unreachable.
+"""
+import os
+import re
+
+from backend.apps.agents.browser import browser_send_parse as sp
+from backend.apps.agents.browser import browser_login_handoff as lh
+from backend.apps.agents.core.ws_manager import BROWSER_CMD_TIMEOUTS, BROWSER_CMD_TIMEOUT_DEFAULT
+
+P_REPO_ROOT = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+P_HANDLER_TS = os.path.join(P_REPO_ROOT, "frontend", "src", "shared", "browserCommandHandler.ts")
+
+# Page-shaped perceptions, in the interactives format the agent actually sees.
+BSKY_SIGNED_OUT = '[1]<link "Sign in">\n[2]<button "Create account">\n[3]<heading "Discover">'
+SO_SIGNED_OUT = '[4]<link "Log in">\n[5]<link "Sign up">\n[6]<heading "Questions">'
+X_SIGNED_IN = '[1]<button "Profile">\n[2]<textbox "What is happening?">\n[9]<button "Post">'
+# A signed-IN page that still advertises a sign-up somewhere: the veto must win, or we would tell
+# an authenticated user to log in again.
+SIGNED_IN_WITH_PROMO = '[1]<link "Sign up">\n[2]<button "Sign out">\n[3]<textbox "Post text">'
+LINKEDIN_FEED = '[1]<button "Start a post">\n[2]<link "Notifications">\n[3]<link "My Network">'
+
+
+def test_soft_signed_out_pages_are_detected():
+    assert sp.looks_signed_out(BSKY_SIGNED_OUT)
+    assert sp.looks_signed_out(SO_SIGNED_OUT)
+
+
+def test_signed_in_pages_are_never_called_signed_out():
+    for state in (X_SIGNED_IN, SIGNED_IN_WITH_PROMO, LINKEDIN_FEED, ""):
+        assert not sp.looks_signed_out(state), state[:40]
+
+
+def test_signed_in_marker_vetoes_a_sign_in_link():
+    """The whole point of the veto: 'Sign up' present AND 'Sign out' present means signed IN."""
+    assert not sp.looks_signed_out(SIGNED_IN_WITH_PROMO)
+
+
+def test_soft_detection_is_separate_from_the_hard_wall():
+    """A soft page must NOT read as a hard login wall: the hard gate declines the send outright,
+    and mislabelling here would change behaviour on pages we can still act on."""
+    assert not sp.looks_like_login_wall("https://bsky.app/", BSKY_SIGNED_OUT)
+    assert sp.looks_signed_out(BSKY_SIGNED_OUT)
+
+
+def test_handoff_ignores_soft_pages_unless_explicitly_allowed():
+    """allow_soft is off by default because the pause interrupts the user; the caller turns it on
+    only once the agent is demonstrably stuck."""
+    assert lh.login_wall_domain("https://bsky.app/", BSKY_SIGNED_OUT) is None
+    assert lh.login_wall_domain("https://bsky.app/", BSKY_SIGNED_OUT, allow_soft=True) == "bsky.app"
+
+
+def test_handoff_still_catches_hard_walls_without_soft():
+    assert lh.login_wall_domain("https://www.instagram.com/accounts/login/", "") == "instagram.com"
+
+
+def test_find_composer_timeout_exceeds_its_own_in_page_budget():
+    """INVARIANT that seals the linkedin bug: the reveal ladder self-caps at a deadline, and the
+    command carrying it must outlast that deadline. If someone lowers the timeout (or raises the
+    in-page budget) past each other, heavy pages silently die mid-ladder again."""
+    budget_ms = p_in_page_budget_ms()
+    timeout_s = BROWSER_CMD_TIMEOUTS.get("find_composer", BROWSER_CMD_TIMEOUT_DEFAULT)
+    assert timeout_s * 1000 > budget_ms, (
+        f"find_composer timeout {timeout_s}s must exceed the in-page deadline {budget_ms}ms")
+
+
+def p_in_page_budget_ms() -> int:
+    """The deadline the in-page finder gives itself, read from the renderer source so the two sides
+    of the invariant can never drift apart silently."""
+    with open(P_HANDLER_TS, encoding="utf-8") as fh:
+        src = fh.read()
+    m = re.search(r"const DEADLINE = Date\.now\(\) \+ (\d+)", src)
+    assert m, "in-page finder deadline not found; did the reveal ladder change?"
+    return int(m.group(1))