[eric] browser: tell a user their DELETION was unconfirmed, not their send

backend/apps/agents/browser/browser_agent.py

@@ -3105,7 +3105,7 @@ async def run_browser_agent(
         else:
             honest, dishonest_reason = completion_is_honest(
                 action_log, publish_task=is_publish_task(skill_key_task),
-                send_confirmed=send_confirmed)
+                send_confirmed=send_confirmed, removal_task=p_task_is_removal)
         final_status = "completed" if honest else "error"
         if not honest:
             summary = f"I was not able to complete this task ({dishonest_reason})."

backend/apps/agents/browser/browser_loop.py

@@ -364,6 +364,7 @@ def deliverable_is_informational(summary: str, task: str = "") -> bool:
 
 def completion_is_honest(
     action_log: list[dict], publish_task: bool = False, send_confirmed: bool = False,
+    removal_task: bool = False,
 ) -> tuple[bool, str]:
     """Reality-check a run the model declared done. Returns (honest, reason).
 
@@ -383,6 +384,13 @@ def completion_is_honest(
     # click runs (a resend guard, not proof), so this can still let a bad send through, but it can
     # never newly flag a run that had any send signal at all.
     if publish_task and not send_confirmed:
+        # "delete my post" carries the noun `post`, so is_publish_task calls a DELETE a publish and
+        # this gate then judged it against a send it was never meant to make. Flagging is right (the
+        # run proved nothing either way), but the sentence has to describe the task the user gave.
+        # Telling someone their deletion "may not have gone out" sends them to check the wrong thing.
+        if removal_task:
+            return False, ("the deletion was never confirmed, so the item may still be there; "
+                           "check the page before trusting this")
         return False, ("the send was never confirmed, so it may not have gone out; "
                        "check the page before trusting this")
     if not action_log:

backend/tests/test_removal_never_claims_a_send.py

@@ -72,3 +72,23 @@ def test_the_post_send_backstop_refuses_to_claim_a_send_on_a_removal():
     removal_code = [ln for ln in branch.split("elif")[0].splitlines()
                     if "done_message" in ln and not ln.strip().startswith("#")]
     assert removal_code and all("sent" not in ln for ln in removal_code), removal_code
+
+
+def test_the_completion_gate_describes_the_task_the_user_actually_gave():
+    """The fourth site of the same root cause, found while cleaning up after the third.
+
+    `is_publish_task` matches the NOUN in "delete my post", so a delete run reaches the publish
+    branch of the completion gate. Flagging it is right (the run proved nothing either way), but
+    the sentence was "the send was never confirmed, so it may not have gone out", which points the
+    user at the wrong page to check: they deleted something, and nothing was ever meant to go out.
+    """
+    from backend.apps.agents.browser.browser_loop import completion_is_honest
+
+    honest, why = completion_is_honest([], publish_task=True, send_confirmed=False,
+                                       removal_task=True)
+    assert not honest
+    assert "deletion" in why and "may still be there" in why
+    assert "gone out" not in why, "a delete must not be described as a send"
+
+    honest, why = completion_is_honest([], publish_task=True, send_confirmed=False)
+    assert not honest and "gone out" in why, "a real send keeps its own wording"