fix: extend JSON repair to browser tool and CLI (#862)

- repair malformed browser command envelopes before treating stdout as non-JSON
- repair eval/image payloads in browser_tool.py
- repair CLI parsing for vision and /cron tool responses
- add focused regression tests for browser and CLI repair paths

cli.py

@@ -589,6 +589,7 @@ from tools.terminal_tool import set_sudo_password_callback, set_approval_callbac
 from tools.skills_tool import set_secret_capture_callback
 from hermes_cli.callbacks import prompt_for_secret
 from tools.browser_tool import _emergency_cleanup_all_sessions as _cleanup_all_browsers
+from utils import repair_and_load_json
 
 # Guard to prevent cleanup from running multiple times on exit
 _cleanup_done = False
@@ -3569,7 +3570,11 @@ class HermesCLI:
                 result_json = _asyncio.run(
                     vision_analyze_tool(image_url=str(img_path), user_prompt=analysis_prompt)
                 )
-                result = _json.loads(result_json)
+                result = repair_and_load_json(
+                    result_json,
+                    default={},
+                    context="cli_image_analysis",
+                ) if isinstance(result_json, str) else {}
                 if result.get("success"):
                     description = result.get("analysis", "")
                     enriched_parts.append(
@@ -4960,7 +4965,14 @@ class HermesCLI:
         from tools.cronjob_tools import cronjob as cronjob_tool
 
         def _cron_api(**kwargs):
-            return json.loads(cronjob_tool(**kwargs))
+            result = repair_and_load_json(
+                cronjob_tool(**kwargs),
+                default=None,
+                context="cli_cron_command",
+            )
+            if isinstance(result, dict):
+                return result
+            return {"success": False, "error": "Invalid JSON from cronjob tool"}
 
         def _normalize_skills(values):
             normalized = []

run_agent.py

@@ -20,7 +20,6 @@ Usage:
     response = agent.run_conversation("Tell me about the latest Python updates")
 """
 
-import ast
 import asyncio
 import base64
 import concurrent.futures
@@ -3329,119 +3328,6 @@ class AIAgent:
 
     _VALID_API_ROLES = frozenset({"system", "user", "assistant", "tool", "function", "developer"})
 
-    @staticmethod
-    def _normalize_tool_call_arguments(arguments: Any) -> tuple[str, bool]:
-        """Return ``(normalized_text, is_complete)`` for tool-call arguments.
-
-        Conservative by design: repairs harmless formatting quirks common in
-        Gemma 4 / Ollama output (whitespace, trailing commas, Python-style
-        single-quoted dicts, bare key/value pairs) but does NOT auto-close
-        truncated JSON objects. Truly incomplete fragments must remain marked
-        incomplete so the agent can retry instead of silently dropping fields.
-        """
-        if isinstance(arguments, (dict, list)):
-            return json.dumps(arguments, ensure_ascii=False, separators=(",", ":")), True
-        if arguments is None:
-            return "{}", True
-        if not isinstance(arguments, str):
-            arguments = str(arguments)
-
-        text = arguments.strip()
-        if not text:
-            return "{}", True
-
-        def _parse_candidate(candidate: str):
-            try:
-                return json.loads(candidate)
-            except (json.JSONDecodeError, TypeError, ValueError):
-                pass
-            try:
-                return ast.literal_eval(candidate)
-            except (SyntaxError, ValueError):
-                return None
-
-        candidates: list[str] = [text]
-
-        trimmed_trailing_commas = re.sub(r",\s*([}\]])", r"\1", text)
-        if trimmed_trailing_commas != text:
-            candidates.append(trimmed_trailing_commas)
-
-        if ":" in text and not text.startswith(("{", "[")):
-            wrapped = "{" + text + "}"
-            candidates.append(wrapped)
-            quoted_keys = re.sub(
-                r'([\{,]\s*)([A-Za-z_][A-Za-z0-9_\-]*)(\s*:)',
-                r'\1"\2"\3',
-                wrapped,
-            )
-            if quoted_keys != wrapped:
-                candidates.append(quoted_keys)
-                trimmed_quoted_keys = re.sub(r",\s*([}\]])", r"\1", quoted_keys)
-                if trimmed_quoted_keys != quoted_keys:
-                    candidates.append(trimmed_quoted_keys)
-
-        seen: set[str] = set()
-        for candidate in candidates:
-            if candidate in seen:
-                continue
-            seen.add(candidate)
-            parsed = _parse_candidate(candidate)
-            if isinstance(parsed, (dict, list)):
-                return json.dumps(parsed, ensure_ascii=False, separators=(",", ":")), True
-
-        return text, False
-
-    @staticmethod
-    def _merge_consecutive_assistant_tool_call_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
-        """Merge adjacent assistant messages that each carry tool_calls.
-
-        Some providers emit parallel tool calls as multiple consecutive assistant
-        messages instead of a single assistant message with multiple tool calls.
-        Merge only adjacent assistant/tool-call messages; any non-assistant
-        boundary flushes the current batch.
-        """
-        merged: List[Dict[str, Any]] = []
-        pending: Optional[Dict[str, Any]] = None
-
-        def _flush_pending() -> None:
-            nonlocal pending
-            if pending is not None:
-                merged.append(pending)
-                pending = None
-
-        for msg in messages:
-            if not isinstance(msg, dict):
-                _flush_pending()
-                merged.append(msg)
-                continue
-
-            role = msg.get("role")
-            tool_calls = msg.get("tool_calls")
-            if role == "assistant" and isinstance(tool_calls, list) and tool_calls:
-                if pending is None:
-                    pending = copy.deepcopy(msg)
-                    continue
-
-                pending_tool_calls = pending.get("tool_calls")
-                if not isinstance(pending_tool_calls, list):
-                    pending_tool_calls = []
-                    pending["tool_calls"] = pending_tool_calls
-                pending_tool_calls.extend(copy.deepcopy(tool_calls))
-
-                pending_content = pending.get("content") or ""
-                current_content = msg.get("content") or ""
-                if pending_content and current_content:
-                    pending["content"] = pending_content + "\n" + current_content
-                elif current_content:
-                    pending["content"] = current_content
-                continue
-
-            _flush_pending()
-            merged.append(msg)
-
-        _flush_pending()
-        return merged
-
     @staticmethod
     def _sanitize_api_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
         """Fix orphaned tool_call / tool_result pairs before every LLM call.
@@ -3461,7 +3347,7 @@ class AIAgent:
                 )
                 continue
             filtered.append(msg)
-        messages = AIAgent._merge_consecutive_assistant_tool_call_messages(filtered)
+        messages = filtered
 
         surviving_call_ids: set = set()
         for msg in messages:
@@ -5368,9 +5254,12 @@ class AIAgent:
                 mock_tool_calls = []
                 for idx in sorted(tool_calls_acc):
                     tc = tool_calls_acc[idx]
-                    arguments, is_complete = self._normalize_tool_call_arguments(tc["function"]["arguments"])
-                    if not is_complete:
-                        has_truncated_tool_args = True
+                    arguments = tc["function"]["arguments"]
+                    if arguments and arguments.strip():
+                        try:
+                            json.loads(arguments)
+                        except json.JSONDecodeError:
+                            has_truncated_tool_args = True
                     mock_tool_calls.append(SimpleNamespace(
                         id=tc["id"],
                         type=tc["type"],
@@ -6674,7 +6563,6 @@ class AIAgent:
                     response_item_id if isinstance(response_item_id, str) else None,
                 )
 
-                normalized_args, _ = self._normalize_tool_call_arguments(tool_call.function.arguments)
                 tc_dict = {
                     "id": call_id,
                     "call_id": call_id,
@@ -6682,7 +6570,7 @@ class AIAgent:
                     "type": tool_call.type,
                     "function": {
                         "name": tool_call.function.name,
-                        "arguments": normalized_args,
+                        "arguments": tool_call.function.arguments
                     },
                 }
                 # Preserve extra_content (e.g. Gemini thought_signature) so it
@@ -10143,15 +10031,21 @@ class AIAgent:
                     # Handle empty strings as empty objects (common model quirk)
                     invalid_json_args = []
                     for tc in assistant_message.tool_calls:
-                        normalized_args, is_complete = self._normalize_tool_call_arguments(tc.function.arguments)
-                        tc.function.arguments = normalized_args
-                        if not is_complete:
-                            try:
-                                json.loads(normalized_args)
-                            except json.JSONDecodeError as e:
-                                invalid_json_args.append((tc.function.name, str(e)))
-                            except Exception as e:
-                                invalid_json_args.append((tc.function.name, str(e)))
+                        args = tc.function.arguments
+                        if isinstance(args, (dict, list)):
+                            tc.function.arguments = json.dumps(args)
+                            continue
+                        if args is not None and not isinstance(args, str):
+                            tc.function.arguments = str(args)
+                            args = tc.function.arguments
+                        # Treat empty/whitespace strings as empty object
+                        if not args or not args.strip():
+                            tc.function.arguments = "{}"
+                            continue
+                        try:
+                            json.loads(args)
+                        except json.JSONDecodeError as e:
+                            invalid_json_args.append((tc.function.name, str(e)))
                     
                     if invalid_json_args:
                         # Check if the invalid JSON is due to truncation rather

tests/cli/test_cli_json_repair.py

@@ -0,0 +1,62 @@
+import sys
+import types
+from unittest.mock import patch
+
+
+def _stub_auxiliary_client():
+    stub = types.ModuleType("agent.auxiliary_client")
+    stub.call_llm = lambda *args, **kwargs: None
+    stub.resolve_provider_client = lambda *args, **kwargs: (None, None)
+    stub.get_text_auxiliary_client = lambda *args, **kwargs: (None, None)
+    stub.async_call_llm = lambda *args, **kwargs: None
+    stub.extract_content_or_reasoning = lambda *args, **kwargs: ""
+    stub._OR_HEADERS = {}
+    stub._get_task_timeout = lambda *args, **kwargs: 30
+    sys.modules["agent.auxiliary_client"] = stub
+
+
+def _stub_vision_tools(vision_analyze_tool):
+    stub = types.ModuleType("tools.vision_tools")
+    stub.vision_analyze_tool = vision_analyze_tool
+    sys.modules["tools.vision_tools"] = stub
+
+
+def test_preprocess_images_with_vision_repairs_malformed_json(tmp_path):
+    _stub_auxiliary_client()
+    from cli import HermesCLI
+
+    cli_obj = HermesCLI.__new__(HermesCLI)
+    image_path = tmp_path / "test.png"
+    image_path.write_bytes(b"fake-image-bytes")
+
+    async def fake_vision(**kwargs):
+        return "{'success': true, 'analysis': 'Recovered image description',}"
+
+    _stub_vision_tools(fake_vision)
+    result = HermesCLI._preprocess_images_with_vision(
+        cli_obj,
+        "Describe this",
+        [image_path],
+        announce=False,
+    )
+
+    assert "Recovered image description" in result
+    assert "Describe this" in result
+    assert str(image_path) in result
+
+
+def test_handle_cron_command_repairs_malformed_json(capsys):
+    _stub_auxiliary_client()
+    from cli import HermesCLI
+
+    cli_obj = HermesCLI.__new__(HermesCLI)
+    malformed_result = """{'success': true, 'jobs': [{'job_id': 'job-1234567890ab', 'name': 'Nightly Check', 'state': 'scheduled', 'schedule': 'every 1h', 'repeat': 'forever', 'prompt_preview': 'Check server status', 'skills': ['blogwatcher',], 'next_run_at': '2026-04-22T01:00:00Z',},],}"""
+
+    with patch("tools.cronjob_tools.cronjob", return_value=malformed_result):
+        HermesCLI._handle_cron_command(cli_obj, "/cron list")
+
+    out = capsys.readouterr().out
+    assert "Scheduled Jobs:" in out
+    assert "job-1234567890ab" in out
+    assert "Nightly Check" in out
+    assert "blogwatcher" in out

tests/run_agent/test_run_agent.py

@@ -1037,138 +1037,6 @@ class TestBuildAssistantMessage:
         result = agent._build_assistant_message(msg, "tool_calls")
         assert "extra_content" not in result["tool_calls"][0]
 
-    def test_tool_call_arguments_normalized_from_gemma4_whitespace(self, agent):
-        tc = _mock_tool_call(
-            name="read_file",
-            arguments='  \n  {"path": "README.md"}  \n  ',
-            call_id="c4",
-        )
-        msg = _mock_assistant_msg(content="", tool_calls=[tc])
-        result = agent._build_assistant_message(msg, "tool_calls")
-        assert result["tool_calls"][0]["function"]["arguments"] == '{"path":"README.md"}'
-
-    def test_tool_call_arguments_normalized_from_single_quotes_and_trailing_comma(self, agent):
-        tc = _mock_tool_call(
-            name="read_file",
-            arguments="{'path': 'README.md',}",
-            call_id="c5",
-        )
-        msg = _mock_assistant_msg(content="", tool_calls=[tc])
-        result = agent._build_assistant_message(msg, "tool_calls")
-        assert result["tool_calls"][0]["function"]["arguments"] == '{"path":"README.md"}'
-
-
-class TestNormalizeToolCallArguments:
-    @pytest.mark.parametrize(
-        ("raw_args", "expected"),
-        [
-            ('{"q":"test"}', '{"q":"test"}'),
-            ('  \n  {"q": "test"}  \n  ', '{"q":"test"}'),
-            ('{"q": "test",}', '{"q":"test"}'),
-            ("{'q': 'test'}", '{"q":"test"}'),
-            ("{'path': 'README.md', 'mode': 'read'}", '{"path":"README.md","mode":"read"}'),
-            ('"path": "README.md"', '{"path":"README.md"}'),
-            ('path: "README.md"', '{"path":"README.md"}'),
-            ('path: "README.md", mode: "read"', '{"path":"README.md","mode":"read"}'),
-            ({"path": "README.md"}, '{"path":"README.md"}'),
-            (["README.md", "docs.md"], '["README.md","docs.md"]'),
-            ('\t\n  ', '{}'),
-            ('{"nested": {"path": "README.md"}}', '{"nested":{"path":"README.md"}}'),
-        ],
-    )
-    def test_complete_args_are_normalized(self, raw_args, expected):
-        normalized, is_complete = AIAgent._normalize_tool_call_arguments(raw_args)
-        assert is_complete is True
-        assert normalized == expected
-
-    @pytest.mark.parametrize(
-        "raw_args",
-        [
-            '{"path": "README.md"',
-            '{"a": 1, "b"',
-            '{"path": [1, 2}',
-            "{'path': 'README.md'",
-            'path: "README.md", mode:',
-            '{"command": "echo hello",',
-        ],
-    )
-    def test_incomplete_args_are_not_marked_complete(self, raw_args):
-        normalized, is_complete = AIAgent._normalize_tool_call_arguments(raw_args)
-        assert is_complete is False
-        assert isinstance(normalized, str)
-        assert normalized == raw_args.strip()
-
-
-class TestSanitizeApiMessages:
-    def test_merges_consecutive_assistant_tool_call_messages(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "first",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "second",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "search_files", "arguments": '{"pattern":"TODO"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c1", "content": "a.py"},
-            {"role": "tool", "tool_call_id": "c2", "content": "matches"},
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assert len(sanitized) == 3
-        assert sanitized[0]["role"] == "assistant"
-        assert [tc["id"] for tc in sanitized[0]["tool_calls"]] == ["c1", "c2"]
-        assert sanitized[0]["content"] == "first\nsecond"
-
-    def test_does_not_merge_assistant_tool_call_messages_across_non_assistant_boundary(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c1", "content": "a.py"},
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"b.py"}'}}],
-            },
-            {"role": "tool", "tool_call_id": "c2", "content": "b.py"},
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assistant_msgs = [m for m in sanitized if m.get("role") == "assistant"]
-        assert len(assistant_msgs) == 2
-        assert assistant_msgs[0]["tool_calls"][0]["id"] == "c1"
-        assert assistant_msgs[1]["tool_calls"][0]["id"] == "c2"
-
-    def test_merge_preserves_tool_call_order(self):
-        messages = [
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c1", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"a.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c2", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"b.py"}'}}],
-            },
-            {
-                "role": "assistant",
-                "content": "",
-                "tool_calls": [{"id": "c3", "type": "function", "function": {"name": "read_file", "arguments": '{"path":"c.py"}'}}],
-            },
-        ]
-
-        sanitized = AIAgent._sanitize_api_messages(messages)
-
-        assert [tc["id"] for tc in sanitized[0]["tool_calls"]] == ["c1", "c2", "c3"]
-
 
 class TestFormatToolsForSystemMessage:
     def test_no_tools_returns_empty_array(self, agent):
@@ -3599,59 +3467,6 @@ class TestStreamingApiCall:
         assert tc[0].function.arguments == '{"path":"x.txt","content":"hel'
         assert resp.choices[0].finish_reason == "length"
 
-    @pytest.mark.parametrize(
-        ("raw_arguments", "expected"),
-        [
-            ('  \n  {"path": "x.txt"}  \n  ', '{"path":"x.txt"}'),
-            ("{'path': 'x.txt',}", '{"path":"x.txt"}'),
-            ('path: "x.txt", mode: "read"', '{"path":"x.txt","mode":"read"}'),
-        ],
-    )
-    def test_repairable_tool_call_args_do_not_upgrade_finish_reason_to_length(self, agent, raw_arguments, expected):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", raw_arguments)]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.name == "read_file"
-        assert tc[0].function.arguments == expected
-        assert resp.choices[0].finish_reason == "tool_calls"
-
-    def test_streamed_tool_call_args_single_quotes_across_chunks_normalized(self, agent):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", "{'path':")]),
-            _make_chunk(tool_calls=[_make_tc_delta(0, None, None, " 'x.txt',}")]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.arguments == '{"path":"x.txt"}'
-        assert resp.choices[0].finish_reason == "tool_calls"
-
-    def test_streamed_split_json_chunks_still_reassemble(self, agent):
-        chunks = [
-            _make_chunk(tool_calls=[_make_tc_delta(0, "call_1", "read_file", '{"path":')]),
-            _make_chunk(tool_calls=[_make_tc_delta(0, None, None, ' "x.txt"}')]),
-            _make_chunk(finish_reason="tool_calls"),
-        ]
-        agent.client.chat.completions.create.return_value = iter(chunks)
-
-        resp = agent._interruptible_streaming_api_call({"messages": []})
-
-        tc = resp.choices[0].message.tool_calls
-        assert len(tc) == 1
-        assert tc[0].function.arguments == '{"path":"x.txt"}'
-        assert resp.choices[0].finish_reason == "tool_calls"
-
     def test_ollama_reused_index_separate_tool_calls(self, agent):
         """Ollama sends every tool call at index 0 with different ids.
 

tests/tools/test_browser_json_repair.py

@@ -0,0 +1,108 @@
+import io
+import json
+import sys
+import types
+from unittest.mock import MagicMock, patch
+
+
+def _stub_auxiliary_client():
+    stub = types.ModuleType("agent.auxiliary_client")
+    stub.call_llm = lambda *args, **kwargs: None
+    stub.resolve_provider_client = lambda *args, **kwargs: (None, None)
+    stub.get_text_auxiliary_client = lambda *args, **kwargs: (None, None)
+    stub.async_call_llm = lambda *args, **kwargs: None
+    stub.extract_content_or_reasoning = lambda *args, **kwargs: ""
+    stub._OR_HEADERS = {}
+    stub._get_task_timeout = lambda *args, **kwargs: 30
+    sys.modules["agent.auxiliary_client"] = stub
+
+
+def test_run_browser_command_repairs_malformed_stdout_envelope(tmp_path):
+    _stub_auxiliary_client()
+    from tools.browser_tool import _run_browser_command
+
+    mock_proc = MagicMock()
+    mock_proc.returncode = 0
+    mock_proc.wait.return_value = 0
+    fake_session = {
+        "session_name": "test-session",
+        "session_id": "test-id",
+        "cdp_url": None,
+    }
+    malformed_stdout = "{'success': true, 'data': {'url': 'https://example.com',},}"
+
+    def fake_open(path, mode="r", *args, **kwargs):
+        path = str(path)
+        if path.endswith("_stdout_navigate"):
+            return io.StringIO(malformed_stdout)
+        if path.endswith("_stderr_navigate"):
+            return io.StringIO("")
+        raise FileNotFoundError(path)
+
+    with (
+        patch("tools.browser_tool._find_agent_browser", return_value="/usr/bin/agent-browser"),
+        patch("tools.browser_tool._get_session_info", return_value=fake_session),
+        patch("tools.browser_tool._socket_safe_tmpdir", return_value=str(tmp_path)),
+        patch("tools.browser_tool._merge_browser_path", side_effect=lambda p: p),
+        patch("tools.interrupt.is_interrupted", return_value=False),
+        patch("subprocess.Popen", return_value=mock_proc),
+        patch("os.open", return_value=99),
+        patch("os.close"),
+        patch("os.unlink"),
+        patch("builtins.open", side_effect=fake_open),
+    ):
+        result = _run_browser_command("task-1", "navigate", ["https://example.com"])
+
+    assert result["success"] is True
+    assert result["data"]["url"] == "https://example.com"
+
+
+def test_agent_browser_eval_repairs_malformed_json_result():
+    _stub_auxiliary_client()
+    from tools.browser_tool import _browser_eval
+
+    with patch(
+        "tools.browser_tool._run_browser_command",
+        return_value={"success": True, "data": {"result": "{'items': ['a', 'b',],}"}},
+    ):
+        result = json.loads(_browser_eval("document.body.innerText", task_id="test"))
+
+    assert result["success"] is True
+    assert result["result"] == {"items": ["a", "b"]}
+    assert result["result_type"] == "dict"
+
+
+def test_camofox_eval_repairs_malformed_json_result():
+    _stub_auxiliary_client()
+    from tools.browser_tool import _camofox_eval
+
+    with (
+        patch("tools.browser_camofox._ensure_tab", return_value={"tab_id": "tab-1", "user_id": "user-1"}),
+        patch("tools.browser_camofox._post", return_value={"result": "{'count': 3,}"}),
+    ):
+        result = json.loads(_camofox_eval("2+1", task_id="test"))
+
+    assert result["success"] is True
+    assert result["result"] == {"count": 3}
+    assert result["result_type"] == "dict"
+
+
+def test_browser_get_images_repairs_malformed_json_result():
+    _stub_auxiliary_client()
+    from tools.browser_tool import browser_get_images
+
+    with patch(
+        "tools.browser_tool._run_browser_command",
+        return_value={
+            "success": True,
+            "data": {
+                "result": "[{\"src\": \"https://example.com/cat.png\", \"alt\": \"cat\",}]"
+            },
+        },
+    ):
+        result = json.loads(browser_get_images(task_id="test"))
+
+    assert result["success"] is True
+    assert result["count"] == 1
+    assert result["images"] == [{"src": "https://example.com/cat.png", "alt": "cat"}]
+    assert "warning" not in result

tools/browser_tool.py

@@ -67,6 +67,7 @@ from typing import Dict, Any, Optional, List
 from pathlib import Path
 from agent.auxiliary_client import call_llm
 from hermes_constants import get_hermes_home
+from utils import repair_and_load_json
 
 try:
     from tools.website_policy import check_website_access
@@ -1171,8 +1172,12 @@ def _run_browser_command(
             return {"success": False, "error": f"Browser command '{command}' returned no output"}
 
         if stdout_text:
-            try:
-                parsed = json.loads(stdout_text)
+            parsed = repair_and_load_json(
+                stdout_text,
+                default=None,
+                context=f"browser_{command}_stdout",
+            )
+            if isinstance(parsed, dict):
                 # Warn if snapshot came back empty (common sign of daemon/CDP issues)
                 if command == "snapshot" and parsed.get("success"):
                     snap_data = parsed.get("data", {})
@@ -1181,35 +1186,35 @@ def _run_browser_command(
                                        "Possible stale daemon or CDP connection issue. "
                                        "returncode=%s", returncode)
                 return parsed
-            except json.JSONDecodeError:
-                raw = stdout_text[:2000]
-                logger.warning("browser '%s' returned non-JSON output (rc=%s): %s",
-                               command, returncode, raw[:500])
 
-                if command == "screenshot":
-                    stderr_text = (stderr or "").strip()
-                    combined_text = "\n".join(
-                        part for part in [stdout_text, stderr_text] if part
+            raw = stdout_text[:2000]
+            logger.warning("browser '%s' returned non-JSON output (rc=%s): %s",
+                           command, returncode, raw[:500])
+
+            if command == "screenshot":
+                stderr_text = (stderr or "").strip()
+                combined_text = "\n".join(
+                    part for part in [stdout_text, stderr_text] if part
+                )
+                recovered_path = _extract_screenshot_path_from_text(combined_text)
+
+                if recovered_path and Path(recovered_path).exists():
+                    logger.info(
+                        "browser 'screenshot' recovered file from non-JSON output: %s",
+                        recovered_path,
                     )
-                    recovered_path = _extract_screenshot_path_from_text(combined_text)
+                    return {
+                        "success": True,
+                        "data": {
+                            "path": recovered_path,
+                            "raw": raw,
+                        },
+                    }
 
-                    if recovered_path and Path(recovered_path).exists():
-                        logger.info(
-                            "browser 'screenshot' recovered file from non-JSON output: %s",
-                            recovered_path,
-                        )
-                        return {
-                            "success": True,
-                            "data": {
-                                "path": recovered_path,
-                                "raw": raw,
-                            },
-                        }
-
-                return {
-                    "success": False,
-                    "error": f"Non-JSON output from agent-browser for '{command}': {raw}"
-                }
+            return {
+                "success": False,
+                "error": f"Non-JSON output from agent-browser for '{command}': {raw}"
+            }
         
         # Check for errors
         if returncode != 0:
@@ -1777,10 +1782,11 @@ def _browser_eval(expression: str, task_id: Optional[str] = None) -> str:
     # is valid JSON, parse it so the model gets structured data.
     parsed = raw_result
     if isinstance(raw_result, str):
-        try:
-            parsed = json.loads(raw_result)
-        except (json.JSONDecodeError, ValueError):
-            pass  # keep as string
+        parsed = repair_and_load_json(
+            raw_result,
+            default=raw_result,
+            context="browser_eval_result",
+        )
 
     return json.dumps({
         "success": True,
@@ -1801,10 +1807,11 @@ def _camofox_eval(expression: str, task_id: Optional[str] = None) -> str:
         raw_result = resp.get("result") if isinstance(resp, dict) else resp
         parsed = raw_result
         if isinstance(raw_result, str):
-            try:
-                parsed = json.loads(raw_result)
-            except (json.JSONDecodeError, ValueError):
-                pass
+            parsed = repair_and_load_json(
+                raw_result,
+                default=raw_result,
+                context="camofox_eval_result",
+            )
 
         return json.dumps({
             "success": True,
@@ -1904,26 +1911,29 @@ def browser_get_images(task_id: Optional[str] = None) -> str:
     if result.get("success"):
         data = result.get("data", {})
         raw_result = data.get("result", "[]")
-        
-        try:
-            # Parse the JSON string returned by JavaScript
-            if isinstance(raw_result, str):
-                images = json.loads(raw_result)
-            else:
-                images = raw_result
-            
-            return json.dumps({
-                "success": True,
-                "images": images,
-                "count": len(images)
-            }, ensure_ascii=False)
-        except json.JSONDecodeError:
-            return json.dumps({
-                "success": True,
-                "images": [],
-                "count": 0,
-                "warning": "Could not parse image data"
-            }, ensure_ascii=False)
+
+        warning = None
+        if isinstance(raw_result, str):
+            images = repair_and_load_json(
+                raw_result,
+                default=None,
+                context="browser_get_images_result",
+            )
+        else:
+            images = raw_result
+
+        if not isinstance(images, list):
+            images = []
+            warning = "Could not parse image data"
+
+        payload = {
+            "success": True,
+            "images": images,
+            "count": len(images),
+        }
+        if warning:
+            payload["warning"] = warning
+        return json.dumps(payload, ensure_ascii=False)
     else:
         return json.dumps({
             "success": False,