diff --git a/aidialog/_modidx.py b/aidialog/_modidx.py
index 8d93651..c86cd09 100644
--- a/aidialog/_modidx.py
+++ b/aidialog/_modidx.py
@@ -250,6 +250,7 @@
                                     'aidialog.msg_parts.InputFile': ('msg_parts.html#inputfile', 'aidialog/msg_parts.py'),
                                     'aidialog.msg_parts.InputImage': ('msg_parts.html#inputimage', 'aidialog/msg_parts.py'),
                                     'aidialog.msg_parts.InputVideo': ('msg_parts.html#inputvideo', 'aidialog/msg_parts.py'),
+                                    'aidialog.msg_parts.MdStr': ('msg_parts.html#mdstr', 'aidialog/msg_parts.py'),
                                     'aidialog.msg_parts.Media': ('msg_parts.html#media', 'aidialog/msg_parts.py'),
                                     'aidialog.msg_parts.Media.__init__': ('msg_parts.html#media.__init__', 'aidialog/msg_parts.py'),
                                     'aidialog.msg_parts.Media.ctext': ('msg_parts.html#media.ctext', 'aidialog/msg_parts.py'),
diff --git a/aidialog/dialog.py b/aidialog/dialog.py
index f809dba..7732e25 100644
--- a/aidialog/dialog.py
+++ b/aidialog/dialog.py
@@ -687,7 +687,7 @@ def _code_span(txt):
 _code_args = {'py':'code', 'python':'code', 'bash':'cmd'}
 
 def tool_md(d):
-    "Display markdown for one parsed `{.tool}` block: a folded details div labeled `func(params)→result`, code tools shown as code"
+    "Display markdown for one parsed `{.tool}` block: a folded details div labeled `func(params)→result`, code shown as code, the result fenced unless the block marks it `md`"
     params = ', '.join(f"{k}={_fmt_param(v)}" for k,v in (d.get('args') or {}).items())
     res = d.get('result')
     tail = f"→{_fmt_param(res)}" if res not in (None, '') else ''
@@ -696,7 +696,9 @@ def tool_md(d):
     if code is not None:
         body = f"Code:\n{fenced(str(code), d.get('name'))}"
         if str(res or '').strip(): body += f"\n\nOutput:\n\n{fenced(str(res))}"
-    else: body = fenced(dumps(d, indent=2, ensure_ascii=False), 'json')
+    else:
+        body = f"Args:\n{fenced(dumps(d.get('args') or {}, indent=2, ensure_ascii=False), 'json')}"
+        if str(res or '').strip(): body += f"\n\n{res}" if d.get('md') else f"\n\nOutput:\n\n{fenced(str(res))}"
     return fenced(f"## {label}\n\n{body}", ' {.details .tool-usage-details}', ch=':')
 
 def usage_md(d):
diff --git a/aidialog/msg_parts.py b/aidialog/msg_parts.py
index ffba55e..50448dc 100644
--- a/aidialog/msg_parts.py
+++ b/aidialog/msg_parts.py
@@ -10,7 +10,7 @@
            'msg2dict', 'dict2msg', 'ToolUse', 'ToolResult', 'display_list', 'Completion', 'mk_tool_res_msg', 'sys_text',
            'part_txt', 'data_url', 'url_mime', 'MediaUrl', 'mk_content', 'parse_tools', 'strip_tools', 'conv_tools',
            'extract_fence_call', 'mk_result_fence', 'split_fence_msgs', 'tool_text', 'fmt2hist', 'ToolResponse',
-           'StopResponse', 'FullResponse', 'trunc_str', 'mk_tr_details', 'hist2fmt', 'mk_msg', 'mk_msgs']
+           'StopResponse', 'FullResponse', 'MdStr', 'trunc_str', 'mk_tr_details', 'hist2fmt', 'mk_msg', 'mk_msgs']
 
 # %% ../nbs/00_msg_parts.ipynb #a616b4f5
 import base64, json, copy
@@ -46,7 +46,7 @@ def replace(self, **kw):
 
 # %% ../nbs/00_msg_parts.ipynb #48496c08
 class Text(Part, tag=PartType.text):
-    "Plain text content."
+    "Plain text content; `citations` lists the sources it rests on as `url_citation` dicts (`url`, `title`, optional `start_index`/`end_index`)"
     def __init__(self, text=None, citations=None, **kw):
         super().__init__(**kw)
         store_attr('text,citations')
@@ -268,7 +268,7 @@ def mk_content(o):
     return o
 
 # %% ../nbs/00_msg_parts.ipynb #ee9c825e
-tool_info = 'json {.tool}'     # fence info string of a tool block: {id, name, args, result} (+server; `error` reserved)
+tool_info = 'json {.tool}'     # fence info string of a tool block: {id, name, args, result} (+server, md; `error` reserved)
 usage_info = 'json {.usage}'   # fence info string of a usage block: UsageStats fields
 
 def parse_tools(s):
@@ -384,7 +384,8 @@ def _extract_tool_parts(d:dict):
     "Build (tool_use_part, tool_result_part) from a parsed `{.tool}` block"
     if not d or d.get('id') is None: return None
     tu = ToolUse   (id=d['id'], name=d['name'], arguments=d.get('args') or {}, server=d.get('server', False))
-    tr = ToolResult(id=d['id'], name=d['name'], text=tool_text(d.get('result')), server=d.get('server', False))
+    text = tool_text(d.get('result'))
+    tr = ToolResult(id=d['id'], name=d['name'], text=MdStr(text) if d.get('md') else text, server=d.get('server', False))
     return tu, tr
 
 # %% ../nbs/00_msg_parts.ipynb #916f8df0
@@ -428,6 +429,7 @@ def __hash__(self): return hash(str(self.content))
 # %% ../nbs/00_msg_parts.ipynb #86265805
 class StopResponse(str): pass
 class FullResponse(str): pass
+class MdStr(str): pass
 
 # %% ../nbs/00_msg_parts.ipynb #4f105e4d
 def trunc_str(s, mx=2000, skip=10, replace="TRUNCATED"):
@@ -470,6 +472,7 @@ def mk_tr_details(tr, mx=2000):
     args = {k:trunc_str(v, mx=None if mx is None else mx*5) if isinstance(v, str) else v for k,v in tr.arguments.items()}
     res = dict(id=tr.id, name=tr.name, args=args, result=trunc_str(tool_text(tr.text), mx=mx))
     if tr.server: res['server'] = True
+    if isinstance(tr.text, MdStr): res['md'] = True
     return "\n\n" + fenced(dumps(res, indent=2, ensure_ascii=False), tool_info) + "\n\n"
 
 # %% ../nbs/00_msg_parts.ipynb #70d1e8c3
diff --git a/nbs/00_msg_parts.ipynb b/nbs/00_msg_parts.ipynb
index d115940..a78538d 100644
--- a/nbs/00_msg_parts.ipynb
+++ b/nbs/00_msg_parts.ipynb
@@ -185,7 +185,7 @@
    "source": [
     "#| export\n",
     "class Text(Part, tag=PartType.text):\n",
-    "    \"Plain text content.\"\n",
+    "    \"Plain text content; `citations` lists the sources it rests on as `url_citation` dicts (`url`, `title`, optional `start_index`/`end_index`)\"\n",
     "    def __init__(self, text=None, citations=None, **kw):\n",
     "        super().__init__(**kw)\n",
     "        store_attr('text,citations')\n",
@@ -1066,7 +1066,7 @@
    "outputs": [],
    "source": [
     "#| export\n",
-    "tool_info = 'json {.tool}'     # fence info string of a tool block: {id, name, args, result} (+server; `error` reserved)\n",
+    "tool_info = 'json {.tool}'     # fence info string of a tool block: {id, name, args, result} (+server, md; `error` reserved)\n",
     "usage_info = 'json {.usage}'   # fence info string of a usage block: UsageStats fields\n",
     "\n",
     "def parse_tools(s):\n",
@@ -1635,7 +1635,8 @@
     "    \"Build (tool_use_part, tool_result_part) from a parsed `{.tool}` block\"\n",
     "    if not d or d.get('id') is None: return None\n",
     "    tu = ToolUse   (id=d['id'], name=d['name'], arguments=d.get('args') or {}, server=d.get('server', False))\n",
-    "    tr = ToolResult(id=d['id'], name=d['name'], text=tool_text(d.get('result')), server=d.get('server', False))\n",
+    "    text = tool_text(d.get('result'))\n",
+    "    tr = ToolResult(id=d['id'], name=d['name'], text=MdStr(text) if d.get('md') else text, server=d.get('server', False))\n",
     "    return tu, tr"
    ]
   },
@@ -1774,7 +1775,7 @@
    "id": "6a1e64c0",
    "metadata": {},
    "source": [
-    "`StopResponse` and `FullResponse` are `str` subclasses that mark a string's handling downstream: a `StopResponse` tool result ends a tool loop, and a `FullResponse` must never be truncated. `_trunc_str` honors the latter (along with fastcore's `Safe` and `PrettyString`, checked by class name so no import is needed), and the `𝍁...𝍁` marker is the same contract for strings that crossed a serialization boundary:"
+    "`StopResponse`, `FullResponse`, and `MdStr` are `str` subclasses that mark a string's handling downstream: a `StopResponse` tool result ends a tool loop, a `FullResponse` must never be truncated, and an `MdStr` is markdown a renderer may show as such rather than fenced (a tool that returns one has opted in, since untrusted text rendered as markdown can mislead). `_trunc_str` honors `FullResponse` (along with fastcore's `Safe` and `PrettyString`, checked by class name so no import is needed), and the `𝍁...𝍁` marker is the same contract for strings that crossed a serialization boundary:\n"
    ]
   },
   {
@@ -1786,7 +1787,8 @@
    "source": [
     "#| export\n",
     "class StopResponse(str): pass\n",
-    "class FullResponse(str): pass"
+    "class FullResponse(str): pass\n",
+    "class MdStr(str): pass"
    ]
   },
   {
@@ -1876,6 +1878,7 @@
     "    args = {k:trunc_str(v, mx=None if mx is None else mx*5) if isinstance(v, str) else v for k,v in tr.arguments.items()}\n",
     "    res = dict(id=tr.id, name=tr.name, args=args, result=trunc_str(tool_text(tr.text), mx=mx))\n",
     "    if tr.server: res['server'] = True\n",
+    "    if isinstance(tr.text, MdStr): res['md'] = True\n",
     "    return \"\\n\\n\" + fenced(dumps(res, indent=2, ensure_ascii=False), tool_info) + \"\\n\\n\""
    ]
   },
@@ -2291,7 +2294,7 @@
    "id": "bf492d17",
    "metadata": {},
    "source": [
-    "A server call, one the provider ran itself, has no result message of its own, so it renders as a completed block, with the call's `text` as the result when the provider gave one. Re-parsing gives a call and result pair with `server` kept, so a stored conversation replays the call as an ordinary tool call, and nothing tries to run it again:"
+    "A server call, one the provider ran itself, has no result message of its own, so it renders as a completed block, with the call's `text` as the result when the provider gave one. Re-parsing gives a call and result pair with `server` kept, so a stored conversation replays the call as an ordinary tool call, and nothing tries to run it again. A result that is markdown by construction, an `MdStr`, is marked `md` on the block and comes back as one:"
    ]
   },
   {
@@ -2301,12 +2304,12 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "srv = Msg('assistant', [Text('Let me check.'), ToolUse(id='s1', name='web_search', arguments={'query': 'otters'}, server=True, text='Otter: https://example.com/otter')])\n",
+    "srv = Msg('assistant', [Text('Let me check.'), ToolUse(id='s1', name='web_search', arguments={'query': 'otters'}, server=True, text=MdStr('Otter: https://example.com/otter'))])\n",
     "s = hist2fmt([srv])\n",
     "h3 = fmt2hist(s)\n",
     "test_eq([m.role for m in h3[:2]], ['assistant', 'tool'])\n",
     "test_eq((h3[0].content[1].server, h3[1].content[0].server), (True, True))\n",
-    "test_eq(h3[1].content[0].text, 'Otter: https://example.com/otter')\n",
+    "test_eq((h3[1].content[0].text, type(h3[1].content[0].text)), ('Otter: https://example.com/otter', MdStr))\n",
     "test_eq(hist2fmt(h3[:2]), s)\n",
     "Markdown(s)"
    ]
diff --git a/nbs/01_dialog.ipynb b/nbs/01_dialog.ipynb
index 566f4b6..9cb4c84 100644
--- a/nbs/01_dialog.ipynb
+++ b/nbs/01_dialog.ipynb
@@ -559,10 +559,10 @@
     {
      "data": {
       "text/markdown": [
-       "59c74047:n:in ⇒ out(3)"
+       "64ace59d:n:in ⇒ out(3)"
       ],
       "text/plain": [
-       "Message(id='59c74047', content='in', output='out', msg_type='note')"
+       "Message(id='64ace59d', content='in', output='out', msg_type='note')"
       ]
      },
      "execution_count": null,
@@ -584,9 +584,9 @@
     {
      "data": {
       "text/plain": [
-       "ac71fcba:p:q\n",
+       "4fdd721b:p:q\n",
        "> Edited reply\n",
-       "c398e29c:c:x=1 ⇒ out(601)"
+       "95575dfa:c:x=1 ⇒ out(601)"
       ]
      },
      "execution_count": null,
@@ -771,10 +771,10 @@
     {
      "data": {
       "text/markdown": [
-       "76fab414:c:8*8 ⇒ out(103)"
+       "d6b8bf64:c:8*8 ⇒ out(103)"
       ],
       "text/plain": [
-       "Message(id='76fab414', content='8*8', output=[{'output_type': 'execute_result', 'metadata': {}, 'data': {'text/plain': '64'}, 'execution_count': 1}], msg_type='code')"
+       "Message(id='d6b8bf64', content='8*8', output=[{'output_type': 'execute_result', 'metadata': {}, 'data': {'text/plain': '64'}, 'execution_count': 1}], msg_type='code')"
       ]
      },
      "execution_count": null,
@@ -2325,7 +2325,7 @@
    "id": "7e642860",
    "metadata": {},
    "source": [
-    "Solveit replies carry tool calls and usage stats as fenced JSON wire blocks (fastllm's format: a ```` ```json {.tool} ```` fence holding `{id, name, args, result}`, and ```` ```json {.usage} ```` holding usage fields). `fmt_tools` rewrites those blocks into folded `::: details` display markdown with a code-span label derived from the data, for renderers (solveit, viewmd) whose html exporter lowers `div.details` to a native `<details>` element. It is a pure formatter: parsing stays with the emitting library, and unparseable blocks are left untouched."
+    "Solveit replies carry tool calls and usage stats as fenced JSON wire blocks (fastllm's format: a ```` ```json {.tool} ```` fence holding `{id, name, args, result}`, and ```` ```json {.usage} ```` holding usage fields). `fmt_tools` rewrites those blocks into folded `::: details` display markdown with a code-span label derived from the data, for renderers (solveit, viewmd) whose html exporter lowers `div.details` to a native `<details>` element. A code tool shows its code and its output in fences. Any other tool shows its args as JSON and its result fenced, unless the block carries `\"md\": true`, the mark a tool result that is markdown by construction (a page fetch, a search) travels with as an `MdStr`; only then is it rendered as markdown. It is a pure formatter: parsing stays with the emitting library, and unparseable blocks are left untouched.\n"
    ]
   },
   {
@@ -2354,7 +2354,7 @@
     "_code_args = {'py':'code', 'python':'code', 'bash':'cmd'}\n",
     "\n",
     "def tool_md(d):\n",
-    "    \"Display markdown for one parsed `{.tool}` block: a folded details div labeled `func(params)→result`, code tools shown as code\"\n",
+    "    \"Display markdown for one parsed `{.tool}` block: a folded details div labeled `func(params)→result`, code shown as code, the result fenced unless the block marks it `md`\"\n",
     "    params = ', '.join(f\"{k}={_fmt_param(v)}\" for k,v in (d.get('args') or {}).items())\n",
     "    res = d.get('result')\n",
     "    tail = f\"→{_fmt_param(res)}\" if res not in (None, '') else ''\n",
@@ -2363,7 +2363,9 @@
     "    if code is not None:\n",
     "        body = f\"Code:\\n{fenced(str(code), d.get('name'))}\"\n",
     "        if str(res or '').strip(): body += f\"\\n\\nOutput:\\n\\n{fenced(str(res))}\"\n",
-    "    else: body = fenced(dumps(d, indent=2, ensure_ascii=False), 'json')\n",
+    "    else:\n",
+    "        body = f\"Args:\\n{fenced(dumps(d.get('args') or {}, indent=2, ensure_ascii=False), 'json')}\"\n",
+    "        if str(res or '').strip(): body += f\"\\n\\n{res}\" if d.get('md') else f\"\\n\\nOutput:\\n\\n{fenced(str(res))}\"\n",
     "    return fenced(f\"## {label}\\n\\n{body}\", ' {.details .tool-usage-details}', ch=':')\n",
     "\n",
     "def usage_md(d):\n",
@@ -2403,7 +2405,9 @@
     "assert 'Code:\\n```py\\n1+1\\n```' in disp             # code tools show code + output fences\n",
     "assert 'Output:' in disp\n",
     "other = fmt_tools('```json {.tool}\\n{\"id\": \"t3\", \"name\": \"web\", \"args\": {\"url\": \"x\"}, \"result\": \"ok\"}\\n```')\n",
-    "assert '```json' in other                            # non-code tools show the JSON\n",
+    "assert '```json\\n{\\n  \"url\": \"x\"\\n}\\n```' in other and 'Output:\\n\\n```\\nok\\n```' in other  # other tools show their args as JSON and the result fenced\n",
+    "md = fmt_tools('```json {.tool}\\n{\"id\": \"t5\", \"name\": \"read_url\", \"args\": {\"url\": \"x\"}, \"result\": \"# Title\\\\n\\\\nBody\", \"md\": true}\\n```')\n",
+    "assert '\\n\\n# Title\\n\\nBody\\n' in md and '```\\n# Title' not in md  # a result marked `md` renders as markdown\n",
     "test_eq(fmt_tools(disp), disp)                       # idempotent: display form has no wire blocks\n",
     "uwire = '```json {.usage}\\n{\"model\": \"m\", \"total_tokens\": 150, \"cost\": 0.0123}\\n```'\n",
     "udisp = fmt_tools(uwire)\n",
@@ -2798,7 +2802,7 @@
     {
      "data": {
       "text/plain": [
-       "c2599acf: ok\n",
+       "d9705aed: ok\n",
        "run: 1 msg ok"
       ]
      },
@@ -2833,7 +2837,7 @@
     {
      "data": {
       "text/plain": [
-       "c2599acf: NameError: name 'y' is not defined\n",
+       "d9705aed: NameError: name 'y' is not defined\n",
        "run: 0 msgs ok, 1 failed"
       ]
      },
@@ -2866,7 +2870,7 @@
      "data": {
       "text/plain": [
        "a9a03f99: ok\n",
-       "009b926b: ZeroDivisionError: division by zero\n",
+       "197f3ee9: ZeroDivisionError: division by zero\n",
        "run: 1 msg ok, 1 failed"
       ]
      },
@@ -2901,9 +2905,9 @@
      "data": {
       "text/plain": [
        "a9a03f99: ok\n",
-       "009b926b: ZeroDivisionError: division by zero\n",
+       "197f3ee9: ZeroDivisionError: division by zero\n",
        "9ed1a7ce: ok\n",
-       "c2599acf: ok\n",
+       "d9705aed: ok\n",
        "run: 3 msgs ok, 1 failed"
       ]
      },