From e46013f2348fc62d75e58d35a78a5f36e762d1c2 Mon Sep 17 00:00:00 2001
From: ciregenz <ciregenz@berkeley.edu>
Date: Sun, 5 Jul 2026 17:52:53 -0700
Subject: [PATCH] [eric] browser-agent: cap result summary+action-log (CLI
 hard-rejects >25K-token MCP results)

---
 .../apps/agents/browser_agent_mcp_server.py   | 20 +++++-
 .../tests/test_browser_agent_mcp_format.py    | 67 +++++++++++++++++++
 2 files changed, 85 insertions(+), 2 deletions(-)
 create mode 100644 backend/tests/test_browser_agent_mcp_format.py

diff --git a/backend/apps/agents/browser_agent_mcp_server.py b/backend/apps/agents/browser_agent_mcp_server.py
index 6e3b71a5..d2c6df06 100644
--- a/backend/apps/agents/browser_agent_mcp_server.py
+++ b/backend/apps/agents/browser_agent_mcp_server.py
@@ -190,6 +190,18 @@ def call_backend(tasks: list[dict]) -> dict:
 
 
 MAX_IMAGE_B64_BYTES = 400_000
+MAX_SUMMARY_CHARS = 16_000
+MAX_ACTION_LOG_ENTRIES = 40
+
+
+def p_cap_summary(text: str) -> str:
+    """Head+tail split: the CLI hard-rejects tool results past ~25K tokens, and a vanished report is worse than a trimmed one."""
+    if len(text) <= MAX_SUMMARY_CHARS:
+        return text
+    head = text[: MAX_SUMMARY_CHARS - 4_000]
+    tail = text[-3_500:]
+    omitted = len(text) - len(head) - len(tail)
+    return f"{head}\n\n[... {omitted} chars of the report omitted ...]\n\n{tail}"
 
 
 def p_sniff_image_mime(b64: str) -> str:
@@ -233,12 +245,16 @@ def format_result(result: dict) -> dict:
     action_log = result.get("action_log", [])
 
     lines = [f"**Browser Agent Result** (browser: {browser_id}, session: {session_id})", ""]
-    lines.append(f"**Summary:** {summary}")
+    lines.append(f"**Summary:** {p_cap_summary(summary)}")
 
     if action_log:
         lines.append("")
         lines.append("**Actions taken:**")
-        for i, entry in enumerate(action_log, 1):
+        entries = action_log[-MAX_ACTION_LOG_ENTRIES:]
+        omitted = len(action_log) - len(entries)
+        if omitted > 0:
+            lines.append(f"  (... {omitted} earlier actions omitted ...)")
+        for i, entry in enumerate(entries, omitted + 1):
             tool = entry.get("tool", "?")
             inp = entry.get("input", {})
             ms = entry.get("elapsed_ms", 0)
diff --git a/backend/tests/test_browser_agent_mcp_format.py b/backend/tests/test_browser_agent_mcp_format.py
new file mode 100644
index 00000000..6800c577
--- /dev/null
+++ b/backend/tests/test_browser_agent_mcp_format.py
@@ -0,0 +1,67 @@
+"""Browser-agent MCP result payload caps.
+
+The bug: format_result forwarded the sub-agent's summary and full action log
+uncapped. The bundled Claude CLI rejects any MCP tool result past ~25K tokens
+(the model never sees the report at all), and repeated near-cap results were
+the refill mass behind the CLI's "Autocompact is thrashing" turn-killer seen
+on 1.5.4 installs.
+
+The seal: summary is head+tail capped at MAX_SUMMARY_CHARS and the action log
+keeps only the last MAX_ACTION_LOG_ENTRIES entries, so one delegation result
+can never approach the CLI rejection threshold on the text side.
+"""
+
+from backend.apps.agents.browser_agent_mcp_server import (
+    MAX_ACTION_LOG_ENTRIES,
+    MAX_SUMMARY_CHARS,
+    format_result,
+)
+
+
+def result_text(result: dict) -> str:
+    blocks = [b for b in result["content"] if b.get("type") == "text"]
+    return "\n".join(b["text"] for b in blocks)
+
+
+def test_small_summary_passes_through_unchanged() -> None:
+    text = result_text(format_result({"summary": "all done"}))
+    assert "**Summary:** all done" in text
+    assert "omitted" not in text
+
+
+def test_giant_summary_keeps_head_and_tail() -> None:
+    summary = "HEADSTART " + ("x" * 60_000) + " TAILEND"
+    text = result_text(format_result({"summary": summary}))
+    assert len(text) < MAX_SUMMARY_CHARS + 300
+    assert "HEADSTART" in text
+    assert text.endswith("TAILEND")
+    assert "omitted" in text
+
+
+def test_action_log_keeps_last_entries_with_original_numbering() -> None:
+    log = [{"tool": f"Act{i}", "input": {}, "elapsed_ms": i} for i in range(100)]
+    text = result_text(format_result({"summary": "ok", "action_log": log}))
+    assert "(... 60 earlier actions omitted ...)" in text
+    assert "Act59" not in text
+    assert "61. Act60(" in text
+    assert "100. Act99(" in text
+
+
+def test_short_action_log_has_no_omission_line() -> None:
+    log = [{"tool": "Click", "input": {"x": 1}, "elapsed_ms": 5}]
+    text = result_text(format_result({"summary": "ok", "action_log": log}))
+    assert "omitted" not in text
+    assert "1. Click(" in text
+
+
+def test_pathological_result_stays_far_under_cli_rejection_cap() -> None:
+    log = [{"tool": "T", "input": {"v": "y" * 500}, "elapsed_ms": 1} for i in range(500)]
+    out = format_result({"summary": "z" * 200_000, "action_log": log})
+    total = len(result_text(out))
+    assert total < MAX_SUMMARY_CHARS + MAX_ACTION_LOG_ENTRIES * 160 + 500
+
+
+def test_error_result_untouched() -> None:
+    out = format_result({"error": "boom"})
+    assert out["isError"] is True
+    assert "boom" in result_text(out)