From 78843dba2f1da55125089c0af7afc3e397340813 Mon Sep 17 00:00:00 2001
From: ciregenz <ciregenz@berkeley.edu>
Date: Thu, 30 Jul 2026 17:31:16 -0700
Subject: [PATCH] [eric] browser: a prose decline from the read script is a
 decline, not an answer

---
 .../agents/browser/browser_read_script.py     | 18 +++++++++++
 backend/tests/test_browser_read_script.py     | 31 +++++++++++++++++++
 2 files changed, 49 insertions(+)

diff --git a/backend/apps/agents/browser/browser_read_script.py b/backend/apps/agents/browser/browser_read_script.py
index 23d8ff6d..937c0967 100644
--- a/backend/apps/agents/browser/browser_read_script.py
+++ b/backend/apps/agents/browser/browser_read_script.py
@@ -8,6 +8,7 @@ Fail-open everywhere: thin page, decline, error = the loop runs exactly as today
 import asyncio
 import logging
 import os
+import re
 import time
 from typing import Awaitable, Callable, Dict, Optional
 
@@ -57,12 +58,29 @@ def read_script_enabled() -> bool:
     return os.environ.get("OSW_READ_SCRIPT", "1") != "0"
 
 
+# A decline written as prose instead of the token. The aux is asked to say INSUFFICIENT when the
+# answer lives behind a click, and it usually does, but on a page that is ALMOST right it explains
+# itself instead: "I can see the search results, but I cannot access the individual product page."
+# That reply used to be accepted as the answer, which ended the run at the wrong page and made a
+# perfectly good multi-step task fail 3 times out of 3 on amazon, deterministically. It is a decline,
+# so it has to be read as one. Anchored on the aux declaring it cannot REACH somewhere, never on a
+# page merely lacking a field, because "the page does not show a price" IS a legitimate answer here.
+P_PROSE_DECLINE_RE = re.compile(
+    r"\b(?:cannot|can'?t|unable to|don'?t have (?:the )?ability to)\s+"
+    r"(?:\w+\s+){0,3}?(?:access|open|reach|navigate(?:\s+to)?|visit|click(?:\s+(?:on|into))?|load)\b"
+    r"|\b(?:would|will|you)\s+need\s+to\s+(?:\w+\s+){0,2}?"
+    r"(?:open|click|visit|navigate|go\s+to)\b",
+    re.I)
+
+
 def is_answer(reply: str) -> Optional[str]:
     """The usable answer text, or None. Declines, empties, and hedge-shaped replies
     all fail closed to the loop, so a thin extraction can never become a wrong answer."""
     answer = (reply or "").strip()
     if not answer or answer.upper().startswith("INSUFFICIENT"):
         return None
+    if P_PROSE_DECLINE_RE.search(answer):
+        return None
     return answer
 
 
diff --git a/backend/tests/test_browser_read_script.py b/backend/tests/test_browser_read_script.py
index 242e8001..4ce8cff3 100644
--- a/backend/tests/test_browser_read_script.py
+++ b/backend/tests/test_browser_read_script.py
@@ -132,3 +132,34 @@ def test_no_aux_client_and_tool_error_both_fail_open():
         return {"error": "card is gone"}
     assert asyncio.run(rs.run_read_script(
         Aux("answer"), "m", "t", "b1", "t1", broken_tool)) is None
+
+
+def test_a_prose_decline_is_not_an_answer():
+    """The exact reply that made amazon fail 3/3: the aux explains it cannot reach the page it
+    needs instead of emitting INSUFFICIENT. Accepting it ended the run on the search results and
+    the model loop, which could have clicked through, never got to run."""
+    verbatim = ("I can see this is a search results page for \"usb c cable\" on Amazon, but I "
+                "cannot access the individual product page for the first result. The page shows "
+                "search results with multiple products listed, but to get the exact price, star "
+                "rating, and number of ratings I would need to open the product page.")
+    assert rs.is_answer(verbatim) is None
+
+    for decline in (
+        "I can't open the product page from here.",
+        "I am unable to navigate to that profile.",
+        "You would need to click into the post to see the replies.",
+        "I cannot reach the comments section on this page.",
+    ):
+        assert rs.is_answer(decline) is None, decline
+
+
+def test_a_page_that_simply_lacks_the_field_is_still_an_answer():
+    """The guard must not eat real answers. Per the read-script contract, a page that visibly
+    lacks the field IS the answer, and so is any answer that happens to contain the word 'open'."""
+    for answer in (
+        "The price is $12.99 and it has 4.5 stars from 8,214 ratings.",
+        "The profile does not show a headline; the name is Eric Zeng.",
+        "The first post is by @someone and the shop is open until 5pm.",
+        "Title: 'How to open a jar'. Channel: Kitchen Tips. 1.2M views.",
+    ):
+        assert rs.is_answer(answer) == answer, answer