feat(agents,backends,conversations,frontends,ui): progress thinking is told apart by its agent, raycast reasoning stays collapsed, read_conversation keeps progress

src/beaver_gateway/agents/base.py

@@ -36,3 +36,8 @@ class BaseAgent(BaseModel):
     system_prompt: str
     expose_mcps: tuple[ExposedMcp, ...] = ()
     accept_client_tools: bool = False
+
+    @property
+    def thinking_is_progress(self) -> bool:
+        """Whether non-empty thinking is text for the human rather than reasoning."""
+        return False

src/beaver_gateway/agents/claude.py

@@ -121,6 +121,11 @@ class ClaudeAgent(BaseAgent):
                 raise ValueError(msg)
         return self
 
+    @property
+    def thinking_is_progress(self) -> bool:
+        """Reasoning arrives empty unless ``display`` is ``summarized``."""
+        return (self.options.thinking or {}).get("display") != "summarized"
+
     def prompt_for(self, kind: Kind) -> tuple[PromptSource, ...] | None:
         return self.prompts.for_kind(kind)
 

src/beaver_gateway/backends/transcript.py

@@ -10,7 +10,7 @@ import hashlib
 import uuid as _uuid
 from collections.abc import Mapping
 from datetime import UTC, datetime
-from typing import TYPE_CHECKING, Any
+from typing import TYPE_CHECKING, Any, cast
 
 try:
     from claude_agent_sdk._cli_version import __cli_version__
@@ -31,6 +31,8 @@ __all__ = [
     "open_tool_uses",
     "prompt_count",
     "render_messages",
+    "reply_text",
+    "spoken_text",
     "strip_tool_entries",
     "text_of",
     "window_entries",
@@ -434,13 +436,45 @@ def prompt_count(entries: Iterable[Mapping[str, Any]]) -> int:
     return sum(1 for e in entries if _is_prompt_entry(e))
 
 
+def spoken_text(block: Any, *, thinking_is_progress: bool) -> str:
+    """What a content block (dict or model) says to the human.
+
+    Text always; thinking only when the agent's thinking is progress
+    written for the human, never when it is reasoning.
+    """
+    if isinstance(block, Mapping):
+        fields = cast("Mapping[str, Any]", block)
+        kind, text, thinking = (
+            fields.get("type"),
+            fields.get("text"),
+            fields.get("thinking"),
+        )
+    else:
+        kind = getattr(block, "type", None)
+        text, thinking = getattr(block, "text", None), getattr(block, "thinking", None)
+    if kind == "text":
+        return str(text or "").strip()
+    if kind == "thinking" and thinking_is_progress:
+        return str(thinking or "").strip()
+    return ""
+
+
+def reply_text(content: Iterable[Any], *, thinking_is_progress: bool) -> str:
+    """The reply as the human reads it: spoken blocks in order, a blank line apart."""
+    parts = (spoken_text(b, thinking_is_progress=thinking_is_progress) for b in content)
+    return "\n\n".join(p for p in parts if p)
+
+
 def render_messages(
-    messages: Iterable[Mapping[str, Any]], *, window: int | None = None
+    messages: Iterable[Mapping[str, Any]],
+    *,
+    thinking_is_progress: bool,
+    window: int | None = None,
 ) -> str:
     """Plain-text projection for ``read_conversation``: ``user:``/``assistant:`` turns.
 
     Tool calls collapse to a one-line summary per assistant turn; tool
-    results and thinking are dropped.
+    results and reasoning are dropped, progress thinking reads as text.
     """
     turns: list[str] = []
     tools: list[str] = []
@@ -466,8 +500,9 @@ def render_messages(
             turns.append("user:\n" + _text_of_content(content))
             continue
         for block in _assistant_blocks(content):
-            if block.get("type") == "text" and block.get("text"):
-                current.append(str(block["text"]))
+            text = spoken_text(block, thinking_is_progress=thinking_is_progress)
+            if text:
+                current.append(text)
             elif block.get("type") == "tool_use":
                 tools.append(str(block.get("name", "")))
     flush()

src/beaver_gateway/conversations/rows.py

@@ -352,6 +352,8 @@ class Rows(State):
             for b in await self.bindings(conv)
         ]
         live = self._pool.get(conv.external_id)
+        agent = self._agents.get(conv.agent_name)
+        out["thinking_is_progress"] = agent is not None and agent.thinking_is_progress
         out["live"] = live is not None
         out["busy"] = live.busy if live is not None else False
         runner = self._runners.get(cast("int", conv.id))
@@ -466,7 +468,13 @@ class Rows(State):
         return marked
 
     async def read(self, conv: Conversation, *, window: int | None = None) -> str:
-        return render_messages(await self.history(conv), window=window)
+        return render_messages(
+            await self.history(conv),
+            thinking_is_progress=self._claude_agent(
+                conv.agent_name
+            ).thinking_is_progress,
+            window=window,
+        )
 
     async def history(self, conv: Conversation) -> list[dict[str, Any]]:
         if conv.session_id is None:

src/beaver_gateway/conversations/turns.py

@@ -23,7 +23,7 @@ from claude_agent_sdk import (
 )
 
 from beaver_gateway.backends.capture import TurnCapture
-from beaver_gateway.backends.transcript import text_of
+from beaver_gateway.backends.transcript import reply_text, text_of
 from beaver_gateway.conversations import injects
 from beaver_gateway.conversations.seeds import Seeds
 from beaver_gateway.conversations.state import Runner, aware, iso
@@ -32,11 +32,11 @@ from beaver_gateway.frontends.accumulate import StreamAccumulator
 from beaver_gateway.storage.models import Conversation, InjectQueueItem, RateLimit
 
 if TYPE_CHECKING:
-    from collections.abc import AsyncIterator, Callable, Iterable, Sequence
+    from collections.abc import AsyncIterator, Callable, Sequence
 
     from beaver_gateway.events.stream import MessageStreamEvent
 
-__all__ = ["Turns", "content_of", "image_block", "reply_text"]
+__all__ = ["Turns", "content_of", "image_block"]
 
 _log = logging.getLogger(__name__)
 
@@ -164,7 +164,9 @@ class Turns(Seeds):
         ):
             acc.feed(event)
         message = acc.finalize(model=agent.model)
-        return reply_text(message.content), capture
+        return reply_text(
+            message.content, thinking_is_progress=agent.thinking_is_progress
+        ), capture
 
     async def before_turn(self, conv: Conversation) -> str | None:  # noqa: ARG002
         return None
@@ -484,22 +486,6 @@ class Turns(Seeds):
 IMAGE_MAX_BYTES = 5 * 1024 * 1024
 
 
-def reply_text(content: Iterable[Any]) -> str:
-    """Text blocks plus non-empty thinking blocks, in block order.
-
-    What the model writes between tool calls rides in thinking blocks; the
-    hidden reasoning in them is empty and drops out on its own.
-    """
-    parts: list[str] = []
-    for block in content:
-        kind = getattr(block, "type", "")
-        if kind == "text":
-            parts.append(getattr(block, "text", "") or "")
-        elif kind == "thinking":
-            parts.append(getattr(block, "thinking", "") or "")
-    return "\n\n".join(p for p in parts if p).strip()
-
-
 def content_of(
     text: str, attachments: Sequence[dict[str, Any]] | None
 ) -> str | list[dict[str, Any]]:

src/beaver_gateway/frontends/anthropic.py

@@ -17,7 +17,7 @@ from fastapi.responses import JSONResponse, StreamingResponse
 
 from beaver_gateway.agents.claude import ClaudeAgent
 from beaver_gateway.backends.capture import TurnCapture
-from beaver_gateway.backends.transcript import fingerprint, text_of
+from beaver_gateway.backends.transcript import fingerprint, reply_text, text_of
 from beaver_gateway.frontends.accumulate import StreamAccumulator
 from beaver_gateway.frontends.base import Frontend
 from beaver_gateway.frontends.bearer import require_token
@@ -29,6 +29,7 @@ if TYPE_CHECKING:
 
     from anthropic.types import Message, MessageParam
 
+    from beaver_gateway.agents.base import BaseAgent
     from beaver_gateway.conversations.service import Conversations
     from beaver_gateway.events.stream import MessageStreamEvent
     from beaver_gateway.frontends.base import GatewayRuntime
@@ -158,6 +159,7 @@ class AnthropicMessagesFrontend(Frontend):
                         capture=capture,
                         turn_id=turn_id,
                         history=history,
+                        thinking_is_progress=agent.thinking_is_progress,
                     )
             else:
                 options: dict[str, Any] = {}
@@ -170,7 +172,7 @@ class AnthropicMessagesFrontend(Frontend):
                 async def after(message: Message) -> None:
                     await _broadcast_turn(
                         runtime,
-                        agent_name=agent.name,
+                        agent=agent,
                         input_messages=messages,
                         output_message=message,
                         system=system_str,
@@ -235,12 +237,9 @@ async def _finish(
     capture: TurnCapture,
     turn_id: str,
     history: list[dict[str, str]] | None,
+    thinking_is_progress: bool,
 ) -> None:
-    reply = "\n\n".join(
-        getattr(b, "text", "")
-        for b in message.content
-        if getattr(b, "type", "") == "text"
-    ).strip()
+    reply = reply_text(message.content, thinking_is_progress=thinking_is_progress)
     tail = capture.synthesized_messages or [
         {"role": "assistant", "content": [{"type": "text", "text": reply}]}
     ]
@@ -291,7 +290,7 @@ async def _sse(
 async def _broadcast_turn(
     runtime: GatewayRuntime,
     *,
-    agent_name: str,
+    agent: BaseAgent,
     input_messages: list[MessageParam],
     output_message: Message,
     system: str | None,
@@ -299,9 +298,10 @@ async def _broadcast_turn(
     if not runtime.turn_log_handlers:
         return
     record = TurnRecord(
-        agent_name=agent_name,
+        agent_name=agent.name,
         input_messages=list(input_messages),
         output_message=output_message,
+        thinking_is_progress=agent.thinking_is_progress,
         system=system,
         source=FRONTEND,
     )

src/beaver_gateway/frontends/markdown/crossfront.py

@@ -142,14 +142,16 @@ class CrossFrontendLogger:
                 new_user = record.input_messages[-1]
                 new_block = renderer.render_user_param(new_user)
                 new_block = renderer.append_to_body(
-                    new_block, renderer.render_assistant_message(record.output_message)
+                    new_block,
+                    renderer.render_assistant_message(
+                        record.output_message,
+                        thinking_is_progress=record.thinking_is_progress,
+                    ),
                 )
                 new_body = renderer.append_to_body(body, new_block)
                 metadata = dict(parsed.metadata)
             else:
-                new_body = _render_full_history(
-                    record.input_messages, record.output_message
-                )
+                new_body = _render_full_history(record)
                 metadata = {}
 
             new_body = renderer.append_to_body(new_body, renderer.USER_SCAFFOLD)
@@ -204,10 +206,10 @@ def _flatten_text(message: Any) -> str:
     return "\n\n".join(c for c in chunks if c)
 
 
-def _render_full_history(messages: list[MessageParam], assistant: Any) -> str:
+def _render_full_history(record: TurnRecord) -> str:
     """Render an entire conversation (used when materializing a new logged file)."""
     blocks: list[str] = []
-    for msg in messages:
+    for msg in record.input_messages:
         role = msg.get("role")
         if role == "user":
             blocks.append(renderer.render_user_param(msg))
@@ -215,7 +217,11 @@ def _render_full_history(messages: list[MessageParam], assistant: Any) -> str:
             content = msg.get("content", "")
             text = content if isinstance(content, str) else _content_to_text(content)
             blocks.append(f"### Assistant:\n\n{text.strip()}\n" if text.strip() else "")
-    blocks.append(renderer.render_assistant_message(assistant))
+    blocks.append(
+        renderer.render_assistant_message(
+            record.output_message, thinking_is_progress=record.thinking_is_progress
+        )
+    )
     body = ""
     for block in blocks:
         if not block:

src/beaver_gateway/frontends/markdown/frontend.py

@@ -341,7 +341,11 @@ class MarkdownFrontend(Frontend):
             file_path=file_path,
             first_user_text=parsed.turns[0].text,
         )
-        outcome = diff_and_fork(stored=stored_msgs, incoming=parsed.turns)
+        outcome = diff_and_fork(
+            stored=stored_msgs,
+            incoming=parsed.turns,
+            thinking_is_progress=agent.thinking_is_progress,
+        )
         capture = TurnCapture()
         events = runtime.conversations.turn(
             conv,
@@ -356,6 +360,7 @@ class MarkdownFrontend(Frontend):
                 file_path=file_path,
                 parsed=parsed,
                 model=agent.model or agent.name,
+                thinking_is_progress=agent.thinking_is_progress,
                 filename=filename,
                 write_disk=write_disk,
             )
@@ -371,6 +376,7 @@ class MarkdownFrontend(Frontend):
             parsed=parsed,
             message=message,
             agent_name=agent.name,
+            thinking_is_progress=agent.thinking_is_progress,
             conv_external_id=conv_external_id,
             write_disk=write_disk,
         )
@@ -388,6 +394,7 @@ class MarkdownFrontend(Frontend):
             agent_name=agent.name,
             input_messages=list(parsed.messages),
             output_message=message,
+            thinking_is_progress=agent.thinking_is_progress,
             system=None,
             source="markdown",
         )
@@ -534,7 +541,11 @@ class MarkdownFrontend(Frontend):
             len(stored_msgs),
             len(parsed.turns),
         )
-        outcome = diff_and_fork(stored=stored_msgs, incoming=parsed.turns)
+        outcome = diff_and_fork(
+            stored=stored_msgs,
+            incoming=parsed.turns,
+            thinking_is_progress=agent.thinking_is_progress,
+        )
         _log.info(
             "chat/stream: file=%s diff_and_fork divergence=%s "
             "backend_msgs=%d persist_msgs=%d",
@@ -567,7 +578,9 @@ class MarkdownFrontend(Frontend):
             partial = acc.finalize(model=model)
             if not partial.content:
                 return None
-            rendered = renderer.render_assistant_message(partial)
+            rendered = renderer.render_assistant_message(
+                partial, thinking_is_progress=agent.thinking_is_progress
+            )
             new_body = renderer.append_to_body(parsed.body, rendered)
             return reattach_frontmatter(parsed.metadata, new_body)
 
@@ -593,7 +606,10 @@ class MarkdownFrontend(Frontend):
             new_body = parsed.body
             if partial.content:
                 new_body = renderer.append_to_body(
-                    new_body, renderer.render_assistant_message(partial)
+                    new_body,
+                    renderer.render_assistant_message(
+                        partial, thinking_is_progress=agent.thinking_is_progress
+                    ),
                 )
             new_body = renderer.append_to_body(new_body, _render_error_block(exc))
             if write_disk:
@@ -616,6 +632,7 @@ class MarkdownFrontend(Frontend):
             parsed=parsed,
             message=message,
             agent_name=agent.name,
+            thinking_is_progress=agent.thinking_is_progress,
             conv_external_id=conv_external_id,
             write_disk=write_disk,
         )
@@ -633,6 +650,7 @@ class MarkdownFrontend(Frontend):
             agent_name=agent.name,
             input_messages=list(parsed.messages),
             output_message=message,
+            thinking_is_progress=agent.thinking_is_progress,
             system=None,
             source="markdown",
         )
@@ -659,6 +677,7 @@ class MarkdownFrontend(Frontend):
         file_path: Path,
         parsed: parser.ParsedFile,
         model: str,
+        thinking_is_progress: bool,
         filename: str,
         write_disk: bool = True,
     ) -> Any:
@@ -675,7 +694,9 @@ class MarkdownFrontend(Frontend):
             partial = acc.finalize(model=model)
             if not partial.content:
                 return
-            rendered = renderer.render_assistant_message(partial)
+            rendered = renderer.render_assistant_message(
+                partial, thinking_is_progress=thinking_is_progress
+            )
             new_body = renderer.append_to_body(parsed.body, rendered)
             await write_atomic(
                 file_path, reattach_frontmatter(parsed.metadata, new_body)
@@ -698,7 +719,10 @@ class MarkdownFrontend(Frontend):
             new_body = parsed.body
             if partial.content:
                 new_body = renderer.append_to_body(
-                    new_body, renderer.render_assistant_message(partial)
+                    new_body,
+                    renderer.render_assistant_message(
+                        partial, thinking_is_progress=thinking_is_progress
+                    ),
                 )
             new_body = renderer.append_to_body(new_body, _render_error_block(exc))
             if write_disk:
@@ -715,11 +739,14 @@ class MarkdownFrontend(Frontend):
         parsed: parser.ParsedFile,
         message: Any,
         agent_name: str,
+        thinking_is_progress: bool,
         conv_external_id: str,
         write_disk: bool = True,
     ) -> str:
         """Render the assistant turn, refresh frontmatter, write if we own the file."""
-        rendered = renderer.render_assistant_message(message)
+        rendered = renderer.render_assistant_message(
+            message, thinking_is_progress=thinking_is_progress
+        )
         new_body = renderer.append_to_body(parsed.body, rendered)
         new_body = renderer.append_to_body(new_body, renderer.USER_SCAFFOLD)
         updated_metadata = dict(parsed.metadata)

src/beaver_gateway/frontends/markdown/history.py

@@ -14,6 +14,7 @@ from typing import TYPE_CHECKING, Any, cast
 from sqlalchemy import delete
 from sqlmodel import select
 
+from beaver_gateway.backends.transcript import spoken_text
 from beaver_gateway.storage.models import Conversation, ConversationMessage
 
 if TYPE_CHECKING:
@@ -180,7 +181,9 @@ class _StoredDisplayTurn:
     text_segment_count: int
 
 
-def _group_display_turns(stored: list[dict[str, Any]]) -> list[_StoredDisplayTurn]:
+def _group_display_turns(
+    stored: list[dict[str, Any]], *, thinking_is_progress: bool
+) -> list[_StoredDisplayTurn]:
     """Walk raw stored messages, group them into Obsidian-visible turns.
 
     A user-prompt message (no ``tool_result`` blocks) opens a new turn;
@@ -210,7 +213,9 @@ def _group_display_turns(stored: list[dict[str, Any]]) -> list[_StoredDisplayTur
                 break
             group.append(m)
             i += 1
-        spoken, skeleton, text_count = _summarize_assistant_group(group)
+        spoken, skeleton, text_count = _summarize_assistant_group(
+            group, thinking_is_progress=thinking_is_progress
+        )
         out.append(
             _StoredDisplayTurn(
                 role="assistant",
@@ -251,13 +256,13 @@ def _user_prompt_text(content: Any) -> str:
 
 
 def _summarize_assistant_group(
-    group: list[dict[str, Any]],
+    group: list[dict[str, Any]], *, thinking_is_progress: bool
 ) -> tuple[str, list[str], int]:
     """Compute (spoken_text, tool_skeleton, text_segment_count) for a display group.
 
     Must mirror what the renderer writes and the parser reads back: text and
-    non-empty thinking collapse into one segment, tool_use becomes a skeleton
-    entry, tool_result and empty thinking are invisible.
+    progress thinking collapse into one segment, tool_use becomes a skeleton
+    entry, tool_result, reasoning and empty thinking are invisible.
     """
     from beaver_gateway.frontends.markdown.parser import TextSegment, ToolSegment
 
@@ -283,7 +288,7 @@ def _summarize_assistant_group(
         for blk in content:
             if not isinstance(blk, dict):
                 continue
-            text = _spoken(blk)
+            text = spoken_text(blk, thinking_is_progress=thinking_is_progress)
             if text:
                 pending.append(text)
             elif blk.get("type") == "tool_use":
@@ -298,7 +303,10 @@ def _summarize_assistant_group(
 
 
 def diff_and_fork(
-    *, stored: list[dict[str, Any]], incoming: list[ParsedTurn]
+    *,
+    stored: list[dict[str, Any]],
+    incoming: list[ParsedTurn],
+    thinking_is_progress: bool,
 ) -> ForkOutcome:
     """Align the incoming parsed file against stored history.
 
@@ -313,11 +321,15 @@ def diff_and_fork(
         )
         raise ValueError(msg)
 
-    stored_groups = _group_display_turns(stored)
+    stored_groups = _group_display_turns(
+        stored, thinking_is_progress=thinking_is_progress
+    )
     new_user_turn = incoming[-1]
     prior_incoming = incoming[:-1]
 
-    spliced_groups, divergence, edited = _walk_prefix(prior_incoming, stored_groups)
+    spliced_groups, divergence, edited = _walk_prefix(
+        prior_incoming, stored_groups, thinking_is_progress=thinking_is_progress
+    )
 
     if divergence is None and len(prior_incoming) < len(stored_groups):
         if _file_lags_store(stored_groups, len(prior_incoming), new_user_turn):
@@ -357,7 +369,10 @@ def _file_lags_store(
 
 
 def _walk_prefix(
-    prior_incoming: list[ParsedTurn], stored_groups: list[_StoredDisplayTurn]
+    prior_incoming: list[ParsedTurn],
+    stored_groups: list[_StoredDisplayTurn],
+    *,
+    thinking_is_progress: bool,
 ) -> tuple[list[list[dict[str, Any]]], int | None, bool]:
     """Walk incoming vs stored side-by-side until first divergence.
 
@@ -390,7 +405,9 @@ def _walk_prefix(
             continue
         if inc_skeleton and inc_text_count != st.text_segment_count:
             return spliced_groups, i, edited
-        spliced = _splice_assistant_group(stored_group=st, incoming=inc)
+        spliced = _splice_assistant_group(
+            stored_group=st, incoming=inc, thinking_is_progress=thinking_is_progress
+        )
         if spliced is None:
             return spliced_groups, i, edited
         spliced_groups.append(spliced)
@@ -425,7 +442,10 @@ def _assemble_tail(
 
 
 def _splice_assistant_group(
-    *, stored_group: _StoredDisplayTurn, incoming: ParsedTurn
+    *,
+    stored_group: _StoredDisplayTurn,
+    incoming: ParsedTurn,
+    thinking_is_progress: bool,
 ) -> list[dict[str, Any]] | None:
     """Apply the file's prose to a stored assistant turn.
 
@@ -433,26 +453,21 @@ def _splice_assistant_group(
     markdown can't express. Falls back to rebuilding from the file's
     structure when the text blocks don't line up one-to-one.
     """
-    preserved = _splice_in_place(stored_group=stored_group, incoming=incoming)
+    preserved = _splice_in_place(
+        stored_group=stored_group,
+        incoming=incoming,
+        thinking_is_progress=thinking_is_progress,
+    )
     if preserved is not None:
         return preserved
     return _splice_by_rebuild(stored_group=stored_group, incoming=incoming)
 
 
-def _spoken(blk: object) -> str:
-    if not isinstance(blk, dict):
-        return ""
-    block = cast("dict[str, Any]", blk)
-    kind = block.get("type")
-    if kind == "text":
-        return str(block.get("text", "")).strip()
-    if kind == "thinking":
-        return str(block.get("thinking", "")).strip()
-    return ""
-
-
 def _splice_in_place(
-    *, stored_group: _StoredDisplayTurn, incoming: ParsedTurn
+    *,
+    stored_group: _StoredDisplayTurn,
+    incoming: ParsedTurn,
+    thinking_is_progress: bool,
 ) -> list[dict[str, Any]] | None:
     """Copy the stored messages, swapping only their text block contents.
 
@@ -472,7 +487,7 @@ def _splice_in_place(
         if msg["role"] != "assistant" or not isinstance(content, list):
             continue
         for bi, blk in enumerate(cast("list[dict[str, Any]]", content)):
-            if not _spoken(blk):
+            if not spoken_text(blk, thinking_is_progress=thinking_is_progress):
                 continue
             if blk["type"] == "thinking":
                 return None

src/beaver_gateway/frontends/markdown/renderer.py

@@ -1,7 +1,7 @@
 """Render Anthropic ``Message`` (and user turns) into markdown.
 
 One-way: produces the human-facing artifact only. The parser strips
-the tool/thinking callouts of older files when reshaping history for replay.
+tool/thinking callouts separately when reshaping history for backend replay.
 """
 
 from __future__ import annotations
@@ -43,17 +43,25 @@ def render_assistant_text(text: str) -> str:
     return f"### Assistant:\n\n{text.strip()}\n"
 
 
-def render_assistant_message(message: Message) -> str:
+def render_assistant_message(message: Message, *, thinking_is_progress: bool) -> str:
     """Render an assistant ``Message`` into a turn block.
 
-    Tool-use blocks and empty thinking render to nothing; a non-empty thinking
-    block is what the model wrote between tool calls and renders as text.
+    Tool-use blocks render to nothing. Progress thinking renders as text in
+    place and vanishes when empty; reasoning renders as a collapsed callout,
+    of which only the first between two texts reaches the file.
     """
     parts: list[str] = ["### Assistant:", ""]
+    after_reasoning = False
     for block in message.content:
+        if isinstance(block, ThinkingBlock) and not thinking_is_progress:
+            if not after_reasoning:
+                parts.extend((*_render_reasoning(block.thinking or ""), ""))
+            after_reasoning = True
+            continue
         text = _spoken(block)
         if text:
             parts.extend((text, ""))
+            after_reasoning = False
     return "\n".join(parts).rstrip() + "\n"
 
 
@@ -99,6 +107,11 @@ def summarize_tool_input(name: str, tool_input: object) -> str:
     return name
 
 
+def _render_reasoning(text: str) -> tuple[str, ...]:
+    lines = text.strip().splitlines() or [""]
+    return ("> [!thinking]-", *(f"> {line}" if line else ">" for line in lines))
+
+
 def _spoken(block: object) -> str:
     if isinstance(block, TextBlock):
         return (block.text or "").strip()

src/beaver_gateway/frontends/telegram/frontend.py

@@ -852,7 +852,10 @@ class TelegramFrontend(Frontend):
                 if event.get("origin") in ("user", "task"):
                     await self._open_draft(key, event, target)
             case "stream":
-                self._feed_draft(key, event)
+                agents = cast("GatewayRuntime", self._runtime).agents
+                agent = agents.get(conv.agent_name)
+                progress = agent is not None and agent.thinking_is_progress
+                self._feed_draft(key, event, thinking_is_progress=progress)
             case "tool":
                 draft = self._drafts.get(key)
                 if draft is not None and event.get("parent_tool_use_id") is None:
@@ -971,7 +974,9 @@ class TelegramFrontend(Frontend):
         self._drafts[key] = draft
         draft.start()
 
-    def _feed_draft(self, key: str, event: Event) -> None:
+    def _feed_draft(
+        self, key: str, event: Event, *, thinking_is_progress: bool
+    ) -> None:
         draft = self._drafts.get(key)
         if draft is None or event.get("parent_tool_use_id") is not None:
             return
@@ -984,7 +989,7 @@ class TelegramFrontend(Frontend):
                 draft.append(str(delta.get("text") or ""))
             elif delta.get("type") == "thinking_delta":
                 progress = str(delta.get("thinking") or "")
-                if progress:
+                if progress and thinking_is_progress:
                     draft.set_status(self.texts.writing)
                     draft.append(progress)
                 else:

src/beaver_gateway/frontends/turn_record.py

@@ -46,6 +46,7 @@ class TurnRecord:
     agent_name: str
     input_messages: list[MessageParam]
     output_message: Message
+    thinking_is_progress: bool
     system: str | None = None
     source: str = "anthropic"