From bdb30c26b084c7b5a7cc4965737f9b8d6bb50c07 Mon Sep 17 00:00:00 2001
From: ciregenz <ciregenz@berkeley.edu>
Date: Tue, 7 Jul 2026 00:22:34 -0700
Subject: [PATCH] [eric] browser: fix quoted_payload apostrophe bug (chen's
 mispaired with the payload quote, silently broke the canonical send errand) +
 6 mechanism tests (opener-hop/fill/verify/send/receipt + all abort branches)

---
 .../agents/browser/browser_send_script.py     | 15 ++-
 backend/tests/test_browser_send_script.py     | 98 +++++++++++++++++++
 2 files changed, 109 insertions(+), 4 deletions(-)
 create mode 100644 backend/tests/test_browser_send_script.py

diff --git a/backend/apps/agents/browser/browser_send_script.py b/backend/apps/agents/browser/browser_send_script.py
index 3129d2f2..c54b48ad 100644
--- a/backend/apps/agents/browser/browser_send_script.py
+++ b/backend/apps/agents/browser/browser_send_script.py
@@ -21,7 +21,9 @@ from typing import Awaitable, Callable
 
 logger = logging.getLogger(__name__)
 
-P_QUOTED_RE = re.compile(r"['\"]([^'\"]{4,300})['\"]")
+# 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})"')
+P_QUOTED_SQ_RE = re.compile(r"(?:^|[\s:>])'([^']{4,300})'")
 P_COMPOSER_ROW_RE = re.compile(r"\[(\d+)\]\*?<\s*textbox\s+\"([^\"]*)\"", re.I)
 P_COMPOSER_NAME_RE = re.compile(r"write|message|compose|reply", re.I)
 
@@ -34,9 +36,14 @@ def script_enabled() -> bool:
 
 def quoted_payload(task: str) -> str:
     """The exact text the user quoted, only when it's unambiguous: exactly one
-    distinct quoted span in the task. Anything else is the model's judgment call."""
-    spans = {m.group(1).strip() for m in P_QUOTED_RE.finditer(task or "") if m.group(1).strip()}
-    return spans.pop() if len(spans) == 1 else ""
+    distinct quoted span in the task. Anything else is the model's judgment call.
+    Double quotes win outright; single quotes must be word-boundary-delimited so
+    an apostrophe inside a name can't hijack the match."""
+    dq = {m.group(1).strip() for m in P_QUOTED_DQ_RE.finditer(task or "") if m.group(1).strip()}
+    if dq:
+        return dq.pop() if len(dq) == 1 else ""
+    sq = {m.group(1).strip() for m in P_QUOTED_SQ_RE.finditer(task or "") if m.group(1).strip()}
+    return sq.pop() if len(sq) == 1 else ""
 
 
 P_OPENER_ROW_RE = re.compile(r"\[(\d+)\]\*?<\s*(?:link|button)\s+\"(Message|Reply|Compose|New message)\"", re.I)
diff --git a/backend/tests/test_browser_send_script.py b/backend/tests/test_browser_send_script.py
new file mode 100644
index 00000000..5bbcfdbc
--- /dev/null
+++ b/backend/tests/test_browser_send_script.py
@@ -0,0 +1,98 @@
+"""Send-script mechanism, verified without a live webview: real LinkedIn-shaped
+interactives fixtures driven through a mock executor exercise the exact path the
+live rig would (opener -> composer -> fill -> commit-check -> send -> clear-check),
+plus every abort/honesty branch. The wall-clock a live run measures is not here;
+the correctness the live run would prove is."""
+import pytest
+
+from backend.apps.agents.browser import browser_send_script as ss
+from backend.apps.agents.browser.browser_agent import send_index_in_state, payload_in_textbox
+
+PROFILE = '[22]*<link "Tyler Chen Premium 1st">\n[50]*<link "Message">\n[51]<button "Follow">'
+COMPOSER_EMPTY = '[2]<textbox "Write a message">\n[9]<button "Attach">'
+COMPOSER_FILLED = '[2]<textbox "Write a message" value="[test] hello world r9-os">\n[14]<button "Send">'
+COMPOSER_SENT = '[2]<textbox "Write a message">\n[9]<button "Attach">'  # cleared, Send gone
+
+TASK = "go to tyler chen's linkedin hes in entrepreneurs first and text him '[test] hello world r9-os'"
+
+
+def make_exec(list_script):
+    """execute_tool mock: each BrowserListInteractives call returns the next
+    scripted state; clicks/fills succeed and are recorded."""
+    calls = {"list": 0, "clicks": []}
+    states = list(list_script)
+
+    async def execute(tool, params, bid, tid):
+        if tool == "BrowserListInteractives":
+            i = min(calls["list"], len(states) - 1)
+            calls["list"] += 1
+            return {"text": states[i]}
+        calls["clicks"].append((tool, params))
+        return {"ok": True}
+    return execute, calls
+
+
+async def run(task, state0, list_script):
+    ex, calls = make_exec(list_script)
+    r = await ss.run_send_script(task, "b1", "", state0, ex, send_index_in_state, payload_in_textbox)
+    return r, calls
+
+
+@pytest.mark.asyncio
+async def test_opener_hop_full_success():
+    """Prestage stopped on the profile: script opens the composer, fills,
+    sees it commit, finds the late Send, clicks, sees it clear -> receipt passes."""
+    r, calls = await run(TASK, PROFILE, [COMPOSER_EMPTY, COMPOSER_FILLED, COMPOSER_FILLED, COMPOSER_SENT])
+    assert r is not None and r["sent"] is True
+    assert r["payload"] == "[test] hello world r9-os"
+    # opener click (50), fill into composer (2 w/ text), solo send click (14)
+    idxs = [c[1].get("index") for c in calls["clicks"]]
+    assert 50 in idxs and 2 in idxs and 14 in idxs
+    fill = next(c for c in calls["clicks"] if c[1].get("text"))
+    assert fill[1]["text"] == "[test] hello world r9-os"
+
+
+@pytest.mark.asyncio
+async def test_composer_already_open_skips_opener():
+    """Prestage left the composer open: no opener click, straight to fill+send."""
+    r, calls = await run(TASK, COMPOSER_EMPTY, [COMPOSER_FILLED, COMPOSER_FILLED, COMPOSER_SENT])
+    assert r is not None and r["sent"] is True
+    assert 50 not in [c[1].get("index") for c in calls["clicks"]]  # never clicked an opener
+
+
+@pytest.mark.asyncio
+async def test_abort_when_no_payload():
+    """No quoted payload = the model's judgment call, never the script's."""
+    r, _ = await run("open tyler chen's linkedin and message him something nice", PROFILE,
+               [COMPOSER_EMPTY])
+    assert r is None
+
+
+@pytest.mark.asyncio
+async def test_abort_when_fill_not_seen_committed():
+    """Fill click ran but the textbox never shows the payload -> abort PRE-click,
+    the irreversible send never fires (no false 'sent')."""
+    r, calls = await run(TASK, COMPOSER_EMPTY, [COMPOSER_EMPTY, COMPOSER_EMPTY, COMPOSER_EMPTY])
+    assert r is None
+    # a Send-class click (index 14) must NEVER have been issued
+    assert all(c[1].get("index") != 14 for c in calls["clicks"])
+
+
+@pytest.mark.asyncio
+async def test_post_click_unverified_yields_honest_note_not_resend():
+    """Send clicked but the composer never verifiably clears -> the run returns
+    sent=False with a do-not-resend note, never a silent retry."""
+    r, calls = await run(TASK, COMPOSER_EMPTY, [COMPOSER_FILLED, COMPOSER_FILLED,
+                                          COMPOSER_FILLED, COMPOSER_FILLED, COMPOSER_FILLED])
+    assert r is not None and r["sent"] is False
+    assert "do NOT send again" in r["note"] or "not send again" in r["note"].lower()
+    # exactly one send-class click was issued (no blind re-fire)
+    assert sum(1 for c in calls["clicks"] if c[1].get("index") == 14) == 1
+
+
+@pytest.mark.asyncio
+async def test_ambiguous_opener_aborts():
+    """Two 'Message' openers = ambiguous = hands to the model, no guess."""
+    two = '[50]*<link "Message">\n[70]*<link "Message">'
+    r, _ = await run(TASK, two, [COMPOSER_EMPTY])
+    assert r is None