From 706a93e0e90a8d7ab5649a02e671b047fe93ca82 Mon Sep 17 00:00:00 2001
From: ciregenz <ciregenz@berkeley.edu>
Date: Thu, 16 Jul 2026 16:02:41 -0700
Subject: [PATCH] [eric] browser: prefill [ so the plan aux can't waffle in
 prose (caught live) + read-script skips the settle-retry on results lists

---
 backend/apps/agents/browser/browser_agent.py         |  1 +
 backend/apps/agents/browser/browser_plan_dispatch.py | 10 +++++++---
 backend/apps/agents/browser/browser_read_script.py   | 10 +++++++---
 3 files changed, 15 insertions(+), 6 deletions(-)

diff --git a/backend/apps/agents/browser/browser_agent.py b/backend/apps/agents/browser/browser_agent.py
index 0ff980b0..fcb2bd45 100644
--- a/backend/apps/agents/browser/browser_agent.py
+++ b/backend/apps/agents/browser/browser_agent.py
@@ -1430,6 +1430,7 @@ async def run_browser_agent(
             p_aux_c, p_aux_m = await p_get_aux_client()
             p_read_answer = await asyncio.wait_for(browser_read_script.run_read_script(
                 p_aux_c, p_aux_m, task, browser_id, tab_id, execute_browser_tool,
+                current_url=current_url,
             ), timeout=25.0)
         except Exception as p_re:
             logger.info(f"[browser-readscript] outer skip ({p_re})")
diff --git a/backend/apps/agents/browser/browser_plan_dispatch.py b/backend/apps/agents/browser/browser_plan_dispatch.py
index 4fcfa790..0f3e411d 100644
--- a/backend/apps/agents/browser/browser_plan_dispatch.py
+++ b/backend/apps/agents/browser/browser_plan_dispatch.py
@@ -97,11 +97,15 @@ async def run_plan_dispatch(
 
         aux_model, _ = await resolve_aux_model(settings, preferred_tier="haiku", primary_api=primary_api)
         client = get_anthropic_client_for_model(settings, aux_model)
-        reply = safe_resp_text(await asyncio.wait_for(
+        # Assistant prefill "[" makes prose unwritable: the aux was narrating the ambiguity instead of emitting the chosen click (caught live via the empty-plan reply log).
+        reply = "[" + safe_resp_text(await asyncio.wait_for(
             client.messages.create(
                 model=aux_model, max_tokens=600, temperature=0, system=P_SYSTEM,
-                messages=[{"role": "user", "content": (
-                    f"Task: {task[:1200]}\n\nInteractive elements:\n{state_text[:P_STATE_CAP]}")}],
+                messages=[
+                    {"role": "user", "content": (
+                        f"Task: {task[:1200]}\n\nInteractive elements:\n{state_text[:P_STATE_CAP]}")},
+                    {"role": "assistant", "content": "["},
+                ],
             ), timeout=P_AUX_TIMEOUT_S))
         steps = parse_plan(reply)
         if not steps:
diff --git a/backend/apps/agents/browser/browser_read_script.py b/backend/apps/agents/browser/browser_read_script.py
index e802f4bb..9945c89a 100644
--- a/backend/apps/agents/browser/browser_read_script.py
+++ b/backend/apps/agents/browser/browser_read_script.py
@@ -11,6 +11,8 @@ import os
 import time
 from typing import Awaitable, Callable, Dict, Optional
 
+from backend.apps.agents.browser.browser_prestage import RESULTS_URL_RE
+
 logger = logging.getLogger(__name__)
 
 ToolRunner = Callable[[str, Dict, str, str], Awaitable[Dict]]
@@ -53,13 +55,15 @@ def is_answer(reply: str) -> Optional[str]:
 
 async def run_read_script(
     aux_client, aux_model, task: str, browser_id: str, tab_id: str,
-    execute_tool: ToolRunner,
+    execute_tool: ToolRunner, current_url: str = "",
 ) -> Optional[str]:
     """The answer to a read task from the staged page, or None (= run the loop).
     Never raises; never acts on the page beyond reading it."""
     t0 = time.monotonic()
     if aux_client is None or not aux_model:
         return None
+    # On a results LIST the miss is structural (the answer lives one click deeper), not hydration; the settle-retry would just re-decline ~3s later.
+    p_retries = 0 if RESULTS_URL_RE.search(current_url or "") else P_INSUFFICIENT_RETRIES
     try:
         from backend.apps.agents.core.aux_llm import safe_resp_text
 
@@ -73,7 +77,7 @@ async def run_read_script(
                 await asyncio.sleep(P_THIN_SETTLE_S)
             return ""
 
-        for ask in range(1 + P_INSUFFICIENT_RETRIES):
+        for ask in range(1 + p_retries):
             page = await p_page_text()
             if len(page) < P_MIN_PAGE_CHARS:
                 logger.info(f"[browser-readscript] page too thin ({len(page)} chars); loop runs")
@@ -89,7 +93,7 @@ async def run_read_script(
             if answer is not None:
                 logger.info(f"[browser-readscript] answered from the staged page in {ms}ms (ask {ask + 1})")
                 return answer
-            if ask < P_INSUFFICIENT_RETRIES:
+            if ask < p_retries:
                 await asyncio.sleep(P_INSUFFICIENT_SETTLE_S)
         logger.info(f"[browser-readscript] insufficient in {int((time.monotonic() - t0) * 1000)}ms; loop runs")
         return None