[eric] browser: a post or comment task can never land in a DM box, at composer or opener

backend/apps/agents/browser/browser_send_parse.py

@@ -138,15 +138,25 @@ def looks_signed_out(state_text: str) -> bool:
 P_POST_INTENT_RE = re.compile(r"\b(post|tweet|publish|share)\b", re.I)
 P_COMMENT_INTENT_RE = re.compile(r"\b(comment|reply|respond)\b", re.I)
 P_COMMENT_SURFACE_RE = re.compile(r"\b(comment|reply)\b", re.I)
+# The private-message box is the worst wrong surface there is, because getting it wrong is not a
+# failed action, it is the user's words delivered privately to a named stranger. Live sweep, dry
+# run: 'write a comment on the first post' walked to instagram.com/<someone>/, took that profile's
+# 'Message' opener and filled 'Message...'. Armed, it would have DM'd them.
+P_DM_SURFACE_RE = re.compile(r"\b(message|messages|dm)\b", re.I)
+P_DM_INTENT_RE = re.compile(r"\b(dm|dms|message|messages|email|e-mail|mail|inbox|chat)\b", re.I)
 
 
 def surface_mismatch(task: str, composer_name: str) -> bool:
-    """True when the task asks to create a POST but the composer found is a comment/reply box.
+    """True when the composer we found contradicts what the task actually asked for.
 
-    Deliberately one-directional: a task that mentions commenting is left alone, so this can only
-    ever reject a comment box for a post task, never the reverse. A rejection is cheap (the
-    structural finder, which does find LinkedIn's real composer, gets its turn instead)."""
+    A public ask (post, comment) is contradicted by a DM box, and a post ask is contradicted by a
+    comment box. A task that asked for neither is left alone entirely, so 'text tyler hello' still
+    gets its message box. Rejecting only ever costs a turn: the structural finder, which does find
+    LinkedIn's real composer, gets its go instead."""
     t, name = task or "", composer_name or ""
+    p_public = bool(P_POST_INTENT_RE.search(t) or P_COMMENT_INTENT_RE.search(t))
+    if p_public and P_DM_SURFACE_RE.search(name) and not P_DM_INTENT_RE.search(t):
+        return True
     if not P_POST_INTENT_RE.search(t) or P_COMMENT_INTENT_RE.search(t):
         return False
     return bool(P_COMMENT_SURFACE_RE.search(name))

backend/apps/agents/browser/browser_send_script.py

@@ -259,10 +259,10 @@ async def run_send_script(
 
     composer = browser_send_parse.composer_index_in_state(state_text)
     if composer and browser_send_parse.surface_mismatch(task_sans_brief, composer[1]):
-        # Asked to POST, found a COMMENT box: that is someone else's content, not a slower route to
-        # ours. Drop it and let the tiers below (opener, then the structural finder, which does find
-        # LinkedIn's real composer) look properly.
-        logger.info(f"[browser-sendscript] ignoring {composer[1]!r}: a comment box is not where a post goes")
+        # Asked to POST and found a COMMENT box, or found a DM box nobody asked for: either way it
+        # is someone else's surface, not a slower route to ours. Drop it and let the tiers below
+        # (opener, then the structural finder, which does find LinkedIn's real composer) look properly.
+        logger.info(f"[browser-sendscript] ignoring composer {composer[1]!r}: wrong surface for this task")
         composer = None
     if not composer:
         # The staged snapshot is prestage's, frozen the instant it clicked Message; the overlay composer lazy-renders a beat later (r263/r269 declined on exactly this, prestage's LAST step was the Message click). Poll a short window so the overlay has time to appear before we fall back to the opener.
@@ -278,13 +278,14 @@ async def run_send_script(
         # 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)
         if opener and browser_send_parse.surface_mismatch(task_sans_brief, opener[1]):
-            # The same post-is-not-a-comment rule the composer already enforces, applied one step
-            # earlier. Measured on linkedin.com with the task "start a post": the only opener listed
-            # was 'Comment', so the script opened a stranger's comment box, found no post composer
-            # inside it, and declined. Opening the wrong surface is not a slower route to the right
-            # one, and here it also burns the reversible-opener hop we only get once.
-            logger.info(f"[browser-sendscript] ignoring opener {opener[1]!r}: a comment box is not "
-                        f"where a post goes")
+            # 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
+            # the script opened a stranger's comment box, found no post composer inside it, and
+            # declined. Measured on instagram.com with "write a comment on the first post": the
+            # opener taken was the profile's 'Message', which is a DM to that person. Opening the
+            # wrong surface is not a slower route to the right one, and it burns the one reversible
+            # opener hop we get.
+            logger.info(f"[browser-sendscript] ignoring opener {opener[1]!r}: wrong surface for this task")
             opener = None
         if opener:
             logger.info(f"[browser-sendscript] firing via opener {opener[1]!r} [{opener[0]}]")

backend/tests/test_browser_readonly_guard.py

@@ -117,7 +117,10 @@ def test_a_comment_task_keeps_its_comment_opener():
 
 
 def test_a_post_task_keeps_a_real_post_opener():
-    for opener in ("Post", "Compose", "New message", "Message"):
+    # "New message" and "Message" used to sit in this list as merely not-a-comment. They are DM
+    # openers on every site we drive, and taking one for a post task is how instagram came within
+    # a dry-run flag of DMing a stranger. See test_browser_wrong_surface.py.
+    for opener in ("Post", "Compose", "Create post", "Write"):
         assert not sp.surface_mismatch('post this, exactly: "hi"', opener), opener
 
 

backend/tests/test_browser_wrong_surface.py

@@ -0,0 +1,69 @@
+"""A send must land on the surface the task asked for, or not at all.
+
+Getting this wrong is not a failed action. Measured live 2026-07-31 in a dry-run sweep: the task
+"write a comment on the first post" walked to instagram.com/<someone>/, took that profile's
+"Message" opener and filled "Message...". The only reason it did not DM a stranger the text is that
+the run was armed with OSW_SENDSCRIPT_DRYRUN=1.
+
+So the real phrasings from that sweep are pinned here, both the ones that must be refused and the
+ones that must keep working, because a guard this blunt breaks whole sites when it drifts.
+"""
+
+from backend.apps.agents.browser import browser_send_parse as sp
+
+# Verbatim from the coverage harness, which is what actually reaches the send script.
+X = 'Go to x.com and post this tweet, exactly: "coverage probe alpha"'
+LINKEDIN = 'Go to linkedin.com and create a post with exactly this text: "coverage probe alpha"'
+INSTAGRAM = 'Go to instagram.com and write a comment "coverage probe alpha" on the first post'
+GMAIL = 'Go to gmail and start a new email with the body "coverage probe alpha"'
+TWITCH = 'Go to twitch.tv, open the first live channel, and write "coverage probe alpha" in chat'
+YOUTUBE = 'Go to youtube.com, open the first video, and write the comment "coverage probe alpha"'
+
+
+def test_the_instagram_dm_that_started_this():
+    assert sp.surface_mismatch(INSTAGRAM, "Message") is True
+    assert sp.surface_mismatch(INSTAGRAM, "Message...") is True
+
+
+def test_no_task_that_never_mentions_messaging_may_reach_a_dm_box():
+    for task in (X, LINKEDIN, INSTAGRAM, YOUTUBE):
+        for surface in ("Message", "Message...", "Direct message", "Send a DM", "New message"):
+            assert sp.surface_mismatch(task, surface) is True, (task, surface)
+
+
+def test_a_task_that_does_ask_to_message_still_gets_its_composer():
+    # Gmail's box is literally named "Message Body" and twitch's chat input is "Send a message";
+    # refusing those would take both sites out of the write path entirely.
+    assert sp.surface_mismatch(GMAIL, "Message Body") is False
+    assert sp.surface_mismatch(TWITCH, "Send a message") is False
+    assert sp.surface_mismatch("DM @someone \"hi\"", "Message...") is False
+    assert sp.surface_mismatch("text tyler hello", "Write a message") is False
+    # "reply" reads as public, so this one needs the word inbox to survive.
+    assert sp.surface_mismatch("Reply to her inbox thread with \"hi\"", "Message") is False
+
+
+def test_the_original_post_versus_comment_rule_is_untouched():
+    assert sp.surface_mismatch(LINKEDIN, "Text editor for creating comment") is True
+    assert sp.surface_mismatch(X, "Add a comment") is True
+    # A task that ASKS to comment keeps its comment box.
+    assert sp.surface_mismatch(YOUTUBE, "Add a comment") is False
+    assert sp.surface_mismatch(INSTAGRAM, "Add a comment...") is False
+
+
+def test_the_composers_we_actually_want_are_never_refused():
+    assert sp.surface_mismatch(X, "Post text") is False
+    assert sp.surface_mismatch(LINKEDIN, "Share") is False
+    assert sp.surface_mismatch(LINKEDIN, "start a post") is False
+    assert sp.surface_mismatch(TWITCH, "Chat") is False
+
+
+def test_a_task_asking_for_neither_a_post_nor_a_comment_is_left_alone():
+    # Nothing to contradict, so the guard stays out of it rather than guessing.
+    for surface in ("Message", "Message...", "New message"):
+        assert sp.surface_mismatch("send tyler hello", surface) is False, surface
+
+
+def test_empty_inputs_never_refuse():
+    assert sp.surface_mismatch("", "") is False
+    assert sp.surface_mismatch(X, "") is False
+    assert sp.surface_mismatch("", "Message") is False