From bd24d0bf6c016c8fceb27896bb7a46db2d284123 Mon Sep 17 00:00:00 2001
From: ciregenz <ciregenz@berkeley.edu>
Date: Tue, 4 Aug 2026 16:11:51 -0700
Subject: [PATCH] [eric] browser: a comment task must not click Create-new-post

---
 .../apps/agents/browser/browser_send_parse.py | 32 +++++++++++++++++--
 .../agents/browser/browser_send_script.py     |  2 +-
 backend/tests/test_browser_opener_match.py    | 25 +++++++++++++++
 3 files changed, 56 insertions(+), 3 deletions(-)

diff --git a/backend/apps/agents/browser/browser_send_parse.py b/backend/apps/agents/browser/browser_send_parse.py
index 6d4f9fd1..6c7db5ff 100644
--- a/backend/apps/agents/browser/browser_send_parse.py
+++ b/backend/apps/agents/browser/browser_send_parse.py
@@ -9,6 +9,7 @@ imports from here, never the reverse.
 """
 
 import re
+from typing import Optional
 
 # Double quotes are unambiguous. Single quotes only delimit when the opener is at a word boundary (start/space/colon), so an in-word apostrophe like "chen's" is never mistaken for a payload quote, that mispairing was silently corrupting the canonical "text him '...'" errand.
 P_QUOTED_DQ_RE = re.compile(r'"([^"]{4,300})"')
@@ -249,21 +250,48 @@ def quoted_payload(task: str) -> str:
     return sq.pop() if len(sq) == 1 else ""
 
 
-def opener_index_in_state(state_text: str):
+def opener_index_in_state(state_text: str, task: Optional[str] = None):
     """(index, name) of the single composer OPENER, or None.
 
     An exact name, or a verb+noun compose phrase anywhere in a longer label. The second half is
     what reaches the openers whose label is a whole sentence, and it keeps the exactness the first
     half was buying: an upsell ('Send InMail') has the wrong noun and a count ('526 comments') has
-    no verb. Still a SINGLETON, so two candidates stay the model's problem, not a coin flip."""
+    no verb. Still a SINGLETON, so two candidates stay the model's problem, not a coin flip.
+
+    Pass `task` to drop openers that contradict it, which is what keeps a comment task off the
+    'New post' upload button. Optional so existing read-only callers are unaffected."""
     hits = [(int(m.group(1)), m.group(2)) for m in P_OPENER_ROW_RE.finditer(state_text or "")]
     if not hits:
         hits = [(int(m.group(1)), m.group(2))
                 for m in P_CONTROL_ROW_RE.finditer(state_text or "")
                 if P_OPENER_PHRASE_RE.search(m.group(2) or "")]
+    if task is not None:
+        hits = [h for h in hits if not opener_contradicts_task(task, h[1])]
     return hits[0] if len(hits) == 1 else None
 
 
+# "New post", "Create", "Compose" open a BLANK thing. Fine when the task is to write something new;
+# wrong when the task is to respond to something that already exists.
+P_CREATE_OPENER_RE = re.compile(r"\b(new|create|compose|start a)\b", re.I)
+
+
+def opener_contradicts_task(task: str, opener_name: str) -> bool:
+    """True when this opener would take us somewhere the task did not ask to go.
+
+    Measured on instagram 2026-08-04: a "write a comment on the first post" task matched the opener
+    `New post Create`, which is the UPLOAD flow, so the run left the feed for a file picker and
+    never saw a post. Nothing downstream could catch it, because by then the only evidence left was
+    a page with no composer on it.
+
+    Only the create-vs-respond direction, because that is the one with a wrong destination. A
+    respond-shaped opener on a create task is left alone: some sites really do route a new post
+    through a control labelled "Write".
+    """
+    t, name = task or "", opener_name or ""
+    responding = bool(P_COMMENT_INTENT_RE.search(t))
+    return responding and bool(P_CREATE_OPENER_RE.search(name))
+
+
 def composer_index_in_state(state_text: str):
     """(index, name) of the single compose-shaped textbox, or None. Two
     candidates = ambiguous = model's problem."""
diff --git a/backend/apps/agents/browser/browser_send_script.py b/backend/apps/agents/browser/browser_send_script.py
index e0921d53..33fd814e 100644
--- a/backend/apps/agents/browser/browser_send_script.py
+++ b/backend/apps/agents/browser/browser_send_script.py
@@ -293,7 +293,7 @@ async def run_send_script(
     p_struct_selector: str = ""
     if not composer:
         # Reversible-opener hop: prestage often stops on the profile with the "Message" opener visible (its settle raced the overlay). Opening a composer is the allowed opener class; the irreversible bar is unchanged.
-        opener = browser_send_parse.opener_index_in_state(state_text)
+        opener = browser_send_parse.opener_index_in_state(state_text, task_sans_brief)
         if opener and browser_send_parse.surface_mismatch(task_sans_brief, opener[1]):
             # The same wrong-surface rule the composer already enforces, applied one step earlier.
             # Measured on linkedin.com with "start a post": the only opener listed was 'Comment', so
diff --git a/backend/tests/test_browser_opener_match.py b/backend/tests/test_browser_opener_match.py
index b9375aa7..c0250658 100644
--- a/backend/tests/test_browser_opener_match.py
+++ b/backend/tests/test_browser_opener_match.py
@@ -57,3 +57,28 @@ def test_the_exact_rule_is_preferred_over_the_phrase_rule():
     hit = sp.opener_index_in_state(
         '[3]<button "Comment">\n[9]<button "Read or add comments 526 comments">')
     assert hit is not None and hit[0] == 3
+
+
+def test_a_comment_task_must_not_click_create_new_post():
+    """Measured live on instagram 2026-08-04, and caused by widening the opener match above.
+
+    "New post Create" contains the phrase "new post", so a task that says "write a comment on the
+    first post" matched Instagram's UPLOAD button. The run left the feed for a file picker and never
+    saw a post, and nothing downstream could catch it: by then the only evidence left was a page
+    with no composer on it, which reads as a missing composer rather than a wrong destination.
+
+    Only the create-vs-respond direction is guarded. A respond-shaped label on a create task is left
+    alone, because some sites really do route a new post through a control labelled "Write"."""
+    ig = '[11]<button "New post Create">'
+    assert sp.opener_index_in_state(ig, "write a comment on the first post") is None
+    # the same button is exactly right when the task IS to create something new
+    assert sp.opener_index_in_state(ig, "create a new post saying hi") is not None
+    # and a genuine comment opener still resolves for the comment task
+    assert sp.opener_index_in_state(
+        '[11]<button "Read or add comments 526 comments">', "write a comment on the first post")
+
+
+def test_omitting_the_task_leaves_opener_matching_exactly_as_it_was():
+    """The task argument is optional so read-only callers (the perception summary line) keep their
+    behaviour; a guard that changed what `opener=` reports would corrupt the dryrun report."""
+    assert sp.opener_index_in_state('[11]<button "New post Create">') is not None