fix(agent): preserve bare email tool parity (#5075)

tests/test_live_strip_email_tool_fences.py

@@ -20,6 +20,7 @@ the behavioral tests exercise an equivalent Python regex built straight from the
 backend ``TOOL_TAGS`` — the same source the live regex now derives from — and
 source-level guards assert the frontend keeps no hard-coded list.
 """
+import json
 import re
 from pathlib import Path
 
@@ -46,18 +47,48 @@ def _exec_fence_regex() -> re.Pattern:
     derives from: the backend TOOL_TAGS (served via /api/tools) minus bash/python."""
     tags = _tool_tags() - _NON_STRIPPED
     assert tags, "TOOL_TAGS is empty"
-    return re.compile(r"```(?:" + "|".join(sorted(tags)) + r")\s*\n[\s\S]*?```", re.IGNORECASE)
+    return re.compile(
+        r"```(" + "|".join(re.escape(tag) for tag in sorted(tags)) + r")(?![\w-])"
+        r"[ \t]*([{\[][^\n]*?)?[ \t]*(?=\r?\n|```)\r?\n?([\s\S]*?)```",
+        re.IGNORECASE,
+    )
+
+
+def _strip_live_exec_fences(text: str) -> str:
+    rx = _exec_fence_regex()
+
+    def repl(match: re.Match) -> str:
+        inline = (match.group(2) or "").strip()
+        if not inline:
+            return ""
+        body = (match.group(3) or "").strip()
+        content = f"{inline}\n{body}" if body else inline
+        try:
+            json.loads(content)
+        except (TypeError, ValueError):
+            return match.group(0)
+        return ""
+
+    return rx.sub(repl, text)
 
 
 def test_strips_executed_email_tool_fences():
-    rx = _exec_fence_regex()
     # The exact shape the reporter observed lingering in the live bubble.
     text = 'Here are emails\n\n```list_emails\n{"max_results":10}\n```'
-    assert rx.sub("", text).strip() == "Here are emails"
+    assert _strip_live_exec_fences(text).strip() == "Here are emails"
+
+
+def test_strips_executed_inline_email_tool_fences():
+    text = 'Here are accounts\n\n```list_email_accounts {}\n```'
+    assert _strip_live_exec_fences(text).strip() == "Here are accounts"
+
+
+def test_strips_multiline_inline_json_email_fences():
+    text = 'Here are emails\n\n```list_emails {"folder": "INBOX",\n"max_results": 2}\n```'
+    assert _strip_live_exec_fences(text).strip() == "Here are emails"
 
 
 def test_strips_every_named_email_tool_fence():
-    rx = _exec_fence_regex()
     email_tools = [
         "list_email_accounts", "send_email", "list_emails", "read_email",
         "reply_to_email", "bulk_email", "archive_email", "delete_email",
@@ -65,22 +96,28 @@ def test_strips_every_named_email_tool_fence():
     ]
     for tool in email_tools:
         fence = f"```{tool}\n{{}}\n```"
-        assert rx.sub("", fence).strip() == "", f"{tool} fence not stripped"
+        assert _strip_live_exec_fences(fence).strip() == "", f"{tool} fence not stripped"
 
 
 def test_preserves_existing_web_search_stripping():
-    rx = _exec_fence_regex()
     fence = '```web_search\n{"q":"x"}\n```'
-    assert rx.sub("", fence).strip() == ""
+    assert _strip_live_exec_fences(fence).strip() == ""
 
 
 def test_does_not_strip_bash_or_python_code_examples():
     """bash/python fences are deliberately excluded — they are legitimate code
     examples a user may have asked the model to show, not tool invocations."""
-    rx = _exec_fence_regex()
     for lang in sorted(_NON_STRIPPED):
         example = f"```{lang}\nls -la\n```"
-        assert rx.sub("", example) == example, f"{lang} example wrongly stripped"
+        assert _strip_live_exec_fences(example) == example, f"{lang} example wrongly stripped"
+
+
+def test_does_not_strip_invalid_inline_json_metadata():
+    for example in (
+        '```list_email_accounts {title="setup"}\n```',
+        '```web_search {query="odysseus"}\n```',
+    ):
+        assert _strip_live_exec_fences(example) == example
 
 
 def test_frontend_keeps_no_hardcoded_tool_list():
@@ -99,6 +136,10 @@ def test_frontend_keeps_no_hardcoded_tool_list():
         "chatRenderer.js must fetch the tool set from /api/tools to build "
         "EXEC_FENCE_RE."
     )
+    assert "JSON.parse(content)" in source, (
+        "chatRenderer.js must validate inline JSON before stripping same-line "
+        "tool fences so Markdown metadata stays visible."
+    )
     # The bash/python carve-out must survive the move to the runtime list.
     m = re.search(r"EXEC_FENCE_NON_TOOL\s*=\s*new Set\(\[(?P<body>.*?)\]\)", source, re.DOTALL)
     assert m, "bash/python carve-out (EXEC_FENCE_NON_TOOL) not found in chatRenderer.js"