From 50355206292875725aa228cbca3e044f98916193 Mon Sep 17 00:00:00 2001
From: ciregenz <ciregenz@berkeley.edu>
Date: Tue, 4 Aug 2026 17:40:39 -0700
Subject: [PATCH] [eric] browser: prove a send by the posted content, not only
 by an emptied box

---
 .../agents/browser/browser_send_script.py     | 20 +++++++++++++++++++
 scripts/browser_canary.py                     |  8 ++++++--
 2 files changed, 26 insertions(+), 2 deletions(-)

diff --git a/backend/apps/agents/browser/browser_send_script.py b/backend/apps/agents/browser/browser_send_script.py
index 5622d7d6..83d77b41 100644
--- a/backend/apps/agents/browser/browser_send_script.py
+++ b/backend/apps/agents/browser/browser_send_script.py
@@ -148,6 +148,26 @@ async def complete_send(
             sent = True
             break
         p_why = f"payload-still-in-a-textbox (textbox rows={sum(1 for x in state3.splitlines() if '<textbox' in x)})"
+    if not sent and p_why.startswith("payload-still"):
+        # LinkedIn keeps the text in its composer AFTER a successful post: measured at 2.6s, then
+        # again at 7.6s, on every round of a 3-round canary whose posts an independent feed read
+        # found PRESENT. So a cleared composer is a fine signal on X and simply false here, and
+        # waiting longer only buys a slower wrong answer.
+        #
+        # Ask the opposite question instead. A cleared box is ABSENCE evidence; the payload showing
+        # up in the page as CONTENT is presence evidence, and presence is what "it posted" means.
+        # One extra read, only on the path that was already about to report a failure.
+        try:
+            r_txt = await asyncio.wait_for(
+                execute_tool("BrowserGetText", {}, browser_id, tab_id), timeout=6.0)
+            p_page = str(r_txt.get("text") or "") if isinstance(r_txt, dict) else ""
+            p_boxes = "\n".join(x for x in (await fresh_list()).splitlines() if "<textbox" in x)
+            if payload and payload in p_page and payload not in p_boxes:
+                sent = True
+                logger.info("[browser-sendscript] receipt via rendered content: the payload is on the "
+                            "page outside any textbox, so it posted even though the composer kept it")
+        except Exception:
+            pass
     if not sent:
         logger.info(f"[browser-sendscript] receipt withheld after {sum(p_waits):.1f}s of polling: {p_why}")
     # A cleared composer is proof of delivery everywhere EXCEPT the ghost-drop hosts, which clear
diff --git a/scripts/browser_canary.py b/scripts/browser_canary.py
index d7034516..86adcb6c 100644
--- a/scripts/browser_canary.py
+++ b/scripts/browser_canary.py
@@ -146,8 +146,12 @@ def marker_in_page(cfg: Dict[str, str], marker: str, handle: str, site: str) ->
     # the post was really there (the cleanup that followed deleted it), the read just never
     # surfaced its text. Absence only counts as evidence once the read itself is known good, so
     # require proof we saw the destination at all before believing what we did not see on it.
-    saw_page = any(k in log for k in ("[browser-action] BrowserGetText",
-                                      "[browser-action] BrowserListInteractives"))
+    # Any evidence the audit run actually looked at a page. The first version of this listed two
+    # exact "[browser-action] X" strings that the log does not emit for reads, so `saw_page` was
+    # always False and every audit came back "unprovable" no matter what it saw. An instrument that
+    # can only ever return "don't know" is worse than none, because it looks like data.
+    saw_page = any(k in log for k in ("BrowserGetText", "BrowserListInteractives",
+                                      "dryrun-report", "browser-action", "browser-time"))
     hit = any(marker in line for line in log.splitlines()
               if "browser-action" not in line and "prompt" not in line.lower()
               and "canary-audit" not in line)