fix(agents): only models that send progress updates read thinking as progress

Dockerfile

@@ -45,11 +45,30 @@ RUN apt-get update \
  && rm -rf /var/lib/apt/lists/*
 
 # Bun native binary (glibc): `bunx` runs the stdio MCP servers declared in
-# the user's config. The claude CLI itself ships inside the
+# the user's config.
-# claude-agent-sdk wheel (`_bundled/claude`), nothing to install here.
 COPY --from=oven/bun:1-slim /usr/local/bin/bun /usr/local/bin/bun
 RUN ln -s /usr/local/bin/bun /usr/local/bin/bunx
 
+# The claude-agent-sdk wheel bundles a CLI, but it trails the npm release:
+# 0.2.157 carries 2.1.277, which does not know claude-opus-5-5 and writes an
+# [claude-code:unrecognized_model] warning per turn. Install the npm build and
+# point BEAVER_CLAUDE_CLI at it; the adapter prefers it over the bundled one.
+#
+# Only the glibc platform package is pulled: the meta package also drags in the
+# musl build, and this image is glibc for the reason stated at the top. The
+# tree lands under /opt world-readable because the model runs as beaver-runner,
+# not as root.
+ARG CLAUDE_CODE_VERSION=2.1.280
+RUN mkdir -p /opt/claude-code \
+ && cd /opt/claude-code \
+ && echo '{"name":"claude-cli-holder","private":true}' > package.json \
+ && bun add @anthropic-ai/claude-code-linux-x64@${CLAUDE_CODE_VERSION} \
+ && ln -sf /opt/claude-code/node_modules/@anthropic-ai/claude-code-linux-x64/claude \
+       /usr/local/bin/claude-cli \
+ && chmod -R a+rX /opt/claude-code \
+ && /usr/local/bin/claude-cli --version
+ENV BEAVER_CLAUDE_CLI=/usr/local/bin/claude-cli
+
 # The model process runs as `beaver-runner`, not as the gateway: the
 # adapter spawns claude under this uid with a whitelisted environment.
 # `acl` lets the entrypoint grant it write access to the vault sub-mounts

pyproject.toml

@@ -14,7 +14,7 @@ dependencies = [
     "anthropic>=0.103.0",
     "anyio>=4.13.0",
     "argon2-cffi>=25.1.0",
-    "claude-agent-sdk>=0.2.150",
+    "claude-agent-sdk>=0.2.157",
     "croniter>=6.2.4",
     "fastapi>=0.136.1",
     "fastmcp>=3.3.1",

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

@@ -17,13 +17,19 @@ from beaver_gateway.agents.policy import PolicyRule  # noqa: TC001
 from beaver_gateway.agents.prompts import PromptSource  # noqa: TC001
 from beaver_gateway.conversations.kinds import KINDS, Kind
 
-__all__ = ["ClaudeAgent", "ClaudeOptions", "Prompts", "SkillSets"]
+__all__ = ["PROGRESS_MODELS", "ClaudeAgent", "ClaudeOptions", "Prompts", "SkillSets"]
+
+PROGRESS_MODELS = frozenset({"claude-opus-5-5", "claude-fable-5-1"})
+"""Models whose non-empty thinking is the text written between tool calls."""
 
 
 class ClaudeOptions(BaseModel):
     model_config = ConfigDict(frozen=True, arbitrary_types_allowed=True)
 
     effort: str | None = None
+    thinking: Mapping[str, str] | None = None
+    """Thinking config for the SDK; ``None`` keeps the CLI default."""
+
     include_partial_messages: bool = True
     """Token-level deltas on the wire; off means one delta per finished block."""
 
@@ -118,6 +124,12 @@ class ClaudeAgent(BaseAgent):
                 raise ValueError(msg)
         return self
 
+    @property
+    def thinking_is_progress(self) -> bool:
+        """A progress model, unless ``display: summarized`` brings reasoning."""
+        summarized = (self.options.thinking or {}).get("display") == "summarized"
+        return self.model in PROGRESS_MODELS and not summarized
+
     def prompt_for(self, kind: Kind) -> tuple[PromptSource, ...] | None:
         return self.prompts.for_kind(kind)
 

src/beaver_gateway/backends/claude_sdk.py

@@ -726,6 +726,7 @@ class ClaudeSdkBackend:
         return ClaudeAgentOptions(
             model=agent.model or None,
             effort=cast("Any", opt.effort),
+            thinking=cast("Any", dict(opt.thinking) if opt.thinking else None),
             system_prompt=system_prompt,
             setting_sources=[],
             strict_mcp_config=True,
@@ -933,7 +934,17 @@ async def _disconnect(live: Session) -> None:
         _log.exception("disconnect failed for session %s", live.session_id)
 
 
+CLI_ENV = "BEAVER_CLAUDE_CLI"
+"""Path to a ``claude`` CLI that wins over the one bundled in the wheel."""
+
+
 def _claude_binary() -> str:
+    override = os.environ.get(CLI_ENV)
+    if override:
+        if not Path(override).is_file():
+            msg = f"{CLI_ENV} points at {override!r}, which is not a file"
+            raise FileNotFoundError(msg)
+        return override
     bundled = Path(claude_agent_sdk.__file__).parent / "_bundled" / "claude"
     if bundled.is_file():
         return str(bundled)

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"):
+            text = spoken_text(block, thinking_is_progress=thinking_is_progress)
-                current.append(str(block["text"]))
+            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
@@ -164,12 +164,9 @@ class Turns(Seeds):
         ):
             acc.feed(event)
         message = acc.finalize(model=agent.model)
-        reply = "\n\n".join(
+        return reply_text(
-            getattr(b, "text", "")
+            message.content, thinking_is_progress=agent.thinking_is_progress
-            for b in message.content
+        ), capture
-            if getattr(b, "type", "") == "text"
-        ).strip()
-        return reply, capture
 
     async def before_turn(self, conv: Conversation) -> str | None:  # noqa: ARG002
         return None

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(
+    reply = reply_text(message.content, thinking_is_progress=thinking_is_progress)
-        getattr(b, "text", "")
-        for b in message.content
-        if getattr(b, "type", "") == "text"
-    ).strip()
     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(
+                new_body = _render_full_history(record)
-                    record.input_messages, record.output_message
-                )
                 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",
@@ -251,12 +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 ``parser.parse_assistant_structure``: text blocks collapse into
+    Must mirror what the renderer writes and the parser reads back: text and
-    one segment, tool_use becomes a skeleton entry, tool_result/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
 
@@ -282,12 +288,10 @@ def _summarize_assistant_group(
         for blk in content:
             if not isinstance(blk, dict):
                 continue
-            btype = blk.get("type")
+            text = spoken_text(blk, thinking_is_progress=thinking_is_progress)
-            if btype == "text":
+            if text:
-                text = str(blk.get("text", "")).strip()
+                pending.append(text)
-                if text:
+            elif blk.get("type") == "tool_use":
-                    pending.append(text)
-            elif btype == "tool_use":
                 _flush()
                 segments.append(ToolSegment(name=str(blk.get("name", ""))))
     _flush()
@@ -299,7 +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.
 
@@ -314,11 +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):
@@ -358,7 +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.
 
@@ -391,7 +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)
@@ -426,7 +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.
 
@@ -434,27 +453,28 @@ 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 _is_text_block(blk: object) -> bool:
-    if not isinstance(blk, dict):
-        return False
-    block = cast("dict[str, Any]", blk)
-    return block.get("type") == "text" and bool(str(block.get("text", "")).strip())
-
-
 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.
 
     Preserves message/thinking-block boundaries that a rebuild would lose
     (and that a cache-hit continuation depends on); returns ``None`` if
-    text blocks aren't 1:1 with incoming segments, so the caller rebuilds instead.
+    text blocks aren't 1:1 with incoming segments or the prose runs through
+    signed thinking blocks, so the caller rebuilds instead.
     """
     from beaver_gateway.frontends.markdown.parser import TextSegment
 
@@ -466,9 +486,12 @@ def _splice_in_place(
         content = msg.get("content")
         if msg["role"] != "assistant" or not isinstance(content, list):
             continue
-        positions.extend(
+        for bi, blk in enumerate(cast("list[dict[str, Any]]", content)):
-            (mi, bi) for bi, blk in enumerate(content) if _is_text_block(blk)
+            if not spoken_text(blk, thinking_is_progress=thinking_is_progress):
-        )
+                continue
+            if blk["type"] == "thinking":
+                return None
+            positions.append((mi, bi))
     if len(positions) != len(new_texts):
         return None
 

src/beaver_gateway/frontends/markdown/renderer.py

@@ -12,8 +12,6 @@ from typing import TYPE_CHECKING, Any, cast
 from anthropic.types import Message, TextBlock, ThinkingBlock
 
 if TYPE_CHECKING:
-    from collections.abc import Iterable
-
     from anthropic.types import MessageParam
 
 
@@ -45,24 +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 render to nothing; of the thinking blocks between two
+    Tool-use blocks render to nothing. Progress thinking renders as text in
-    texts only the first one reaches the file.
+    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_thinking = False
+    after_reasoning = False
     for block in message.content:
-        thinking = isinstance(block, ThinkingBlock)
+        if isinstance(block, ThinkingBlock) and not thinking_is_progress:
-        if thinking and after_thinking:
+            if not after_reasoning:
+                parts.extend((*_render_reasoning(block.thinking or ""), ""))
+            after_reasoning = True
             continue
-        lines = list(_render_block(block))
+        text = _spoken(block)
-        if not lines:
+        if text:
-            continue
+            parts.extend((text, ""))
-        after_thinking = thinking
+            after_reasoning = False
-        parts.extend(lines)
-        parts.append("")
     return "\n".join(parts).rstrip() + "\n"
 
 
@@ -108,21 +107,17 @@ def summarize_tool_input(name: str, tool_input: object) -> str:
     return name
 
 
-def _render_block(block: object) -> Iterable[str]:
+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):
-        text = (block.text or "").strip()
+        return (block.text or "").strip()
-        if text:
-            yield text
-        return
     if isinstance(block, ThinkingBlock):
-        yield from _render_thinking(block.thinking or "")
+        return (block.thinking or "").strip()
-        return
+    return ""
-
-
-def _render_thinking(text: str) -> Iterable[str]:
-    yield "> [!thinking]-"
-    for line in text.strip().splitlines() or [""]:
-        yield f"> {line}" if line else ">"
 
 
 def adaptive_fence(content: str) -> str:

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
@@ -983,7 +988,12 @@ class TelegramFrontend(Frontend):
                 draft.set_status(self.texts.writing)
                 draft.append(str(delta.get("text") or ""))
             elif delta.get("type") == "thinking_delta":
-                draft.set_status(self.texts.thinking)
+                progress = str(delta.get("thinking") or "")
+                if progress and thinking_is_progress:
+                    draft.set_status(self.texts.writing)
+                    draft.append(progress)
+                else:
+                    draft.set_status(self.texts.thinking)
         elif kind == "content_block_start":
             block = raw.get("content_block") or {}
             if block.get("type") == "tool_use":

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"
 

src/beaver_gateway/vault/watch.py

@@ -19,12 +19,14 @@ from pathlib import Path
 from typing import TYPE_CHECKING
 from zoneinfo import ZoneInfo
 
-from watchfiles import awatch
+from watchfiles import DefaultFilter, awatch
 
 if TYPE_CHECKING:
     from collections.abc import Iterable
 
-__all__ = ["Change", "VaultWatch", "WatchRules"]
+    from watchfiles import Change as FileChange
+
+__all__ = ["Change", "VaultFilter", "VaultWatch", "WatchRules"]
 
 _log = logging.getLogger("beaver_gateway.vault.watch")
 
@@ -74,6 +76,26 @@ class _Seen:
     count: int
 
 
+IGNORE_DIRS = (".obsidian", ".trash", ".git")
+"""Housekeeping directories whose churn never means an edit of the vault."""
+
+IGNORE_NAMES = (".sync.lock", "*.lock", "*.tmp", "*~")
+"""Lock and scratch file names, matched against the entity name."""
+
+
+class VaultFilter(DefaultFilter):
+    """Drops housekeeping paths inside ``awatch``, before the loop is woken."""
+
+    def __init__(self) -> None:
+        super().__init__(ignore_dirs=(*DefaultFilter.ignore_dirs, *IGNORE_DIRS))
+
+    def __call__(self, change: FileChange, path: str) -> bool:
+        name = Path(path).name
+        if any(fnmatch.fnmatchcase(name, pattern) for pattern in IGNORE_NAMES):
+            return False
+        return super().__call__(change, path)
+
+
 class VaultWatch:
     def __init__(
         self, root: Path, rules: WatchRules, *, tz: str = "UTC", debounce: float = 4.0
@@ -98,6 +120,7 @@ class VaultWatch:
         _log.info("watching %s (%d files in snapshot)", self.root, len(self._seen))
         async for changes in awatch(
             self.root,
+            watch_filter=VaultFilter(),
             debounce=int(self._debounce * 1000),
             stop_event=self._stop,
             ignore_permission_denied=True,

tests/test_conversations.py

@@ -20,7 +20,11 @@ from claude_agent_sdk import (
     project_key_for_directory,
 )
 
+from anthropic.types import TextBlock as AnthropicTextBlock
+from anthropic.types import ThinkingBlock
+
 from beaver_gateway.agents.claude import ClaudeAgent, ClaudeOptions
+from beaver_gateway.agents.raycast import RaycastAgent
 from beaver_gateway.backends.claude_sdk import ClaudeSdkBackend
 from beaver_gateway.events.bus import EventBus
 from beaver_gateway.conversations.service import (
@@ -36,6 +40,7 @@ from beaver_gateway.backends.transcript import (
     close_open_tool_uses,
     open_tool_uses,
     render_messages,
+    reply_text,
     strip_tool_entries,
     window_entries,
 )
@@ -714,12 +719,86 @@ def test_transcript_helpers() -> None:
     assert windowed[0]["message"]["content"] == "q2" and len(windowed) == 2
     assert (
         render_messages(
-            [{"role": "user", "content": "q"}, {"role": "assistant", "content": "a"}]
+            [{"role": "user", "content": "q"}, {"role": "assistant", "content": "a"}],
+            thinking_is_progress=True,
         )
         == "user:\nq\n\nassistant:\na"
     )
 
 
+def test_read_conversation_keeps_progress_and_drops_reasoning() -> None:
+    history = [
+        {"role": "user", "content": "status?"},
+        {
+            "role": "assistant",
+            "content": [
+                {"type": "text", "text": "Checking."},
+                {"type": "thinking", "thinking": "", "signature": "s"},
+                {"type": "tool_use", "id": "t1", "name": "komodo", "input": {}},
+            ],
+        },
+        {
+            "role": "user",
+            "content": [{"type": "tool_result", "tool_use_id": "t1", "content": "ok"}],
+        },
+        {
+            "role": "assistant",
+            "content": [
+                {
+                    "type": "thinking",
+                    "thinking": "All up, reading logs.",
+                    "signature": "s",
+                },
+                {"type": "tool_use", "id": "t2", "name": "komodo", "input": {}},
+            ],
+        },
+        {
+            "role": "user",
+            "content": [{"type": "tool_result", "tool_use_id": "t2", "content": "ok"}],
+        },
+        {"role": "assistant", "content": [{"type": "text", "text": "Healthy."}]},
+    ]
+    assert render_messages(history, thinking_is_progress=True) == (
+        "user:\nstatus?\n\nassistant:\nChecking.\n\nAll up, reading logs.\n\n"
+        "Healthy.\n(tools: komodo, komodo)"
+    )
+    assert render_messages(history, thinking_is_progress=False) == (
+        "user:\nstatus?\n\nassistant:\nChecking.\n\nHealthy.\n(tools: komodo, komodo)"
+    )
+
+
+@pytest.mark.parametrize(
+    ("model", "expected"),
+    [
+        ("claude-opus-5-5", True),
+        ("claude-fable-5-1", True),
+        ("claude-opus-5", False),
+        ("claude-sonnet-5", False),
+        ("claude-haiku-4-5-20251001", False),
+        ("claude-next-unknown", False),
+    ],
+)
+def test_only_progress_models_read_their_thinking_as_progress(
+    tmp_path: Path, model: str, expected: bool
+) -> None:
+    assert (
+        ClaudeAgent(name="c", model=model, cwd=tmp_path).thinking_is_progress
+        is expected
+    )
+
+
+def test_summarized_display_and_raycast_are_never_progress(tmp_path: Path) -> None:
+    raycast = RaycastAgent(name="r", model="google-gemini-3.1-pro", system_prompt="")
+    summarized = ClaudeAgent(
+        name="s",
+        model="claude-opus-5-5",
+        cwd=tmp_path,
+        options=ClaudeOptions(thinking={"type": "adaptive", "display": "summarized"}),
+    )
+    assert not raycast.thinking_is_progress
+    assert not summarized.thinking_is_progress
+
+
 async def test_agent_kinds_gate_create(world: World) -> None:
     with pytest.raises(ValueError, match="does not serve kind 'deep'"):
         await world.conversations.create(kind="deep", agent="a", origin="test")
@@ -958,3 +1037,18 @@ async def test_subagent_result_inside_a_turn_does_not_end_it(world: World) -> No
     assert client.prompts[-1].startswith("two")
     rows = await world.conversations.queue.recent(conv.id)
     assert [r.status for r in rows] == ["done", "done"]
+
+
+def test_reply_text_keeps_progress_updates_and_drops_hidden_reasoning() -> None:
+    blocks = [
+        ThinkingBlock(type="thinking", thinking="", signature="s"),
+        ThinkingBlock(type="thinking", thinking="looking it up", signature="s"),
+        AnthropicTextBlock(type="text", text="found it"),
+    ]
+    assert reply_text(blocks, thinking_is_progress=True) == "looking it up\n\nfound it"
+    assert reply_text(blocks, thinking_is_progress=False) == "found it"
+
+
+def test_reply_text_on_a_plain_answer_is_unchanged() -> None:
+    blocks = [AnthropicTextBlock(type="text", text="  just an answer  ")]
+    assert reply_text(blocks, thinking_is_progress=True) == "just an answer"

tests/test_markdown_renderer.py

@@ -1,6 +1,11 @@
+from pathlib import Path
+
 from anthropic.types import Message, TextBlock, ThinkingBlock, ToolUseBlock, Usage
 
-from beaver_gateway.frontends.markdown import renderer
+from beaver_gateway.frontends.markdown import history, parser, renderer
+from beaver_gateway.frontends.markdown.crossfront import CrossFrontendLogger
+from beaver_gateway.frontends.markdown.parser import TextSegment, ToolSegment
+from beaver_gateway.frontends.turn_record import TurnRecord
 
 
 def _message(*blocks: object) -> Message:
@@ -24,36 +29,159 @@ def test_tool_calls_leave_no_blank_lines() -> None:
         ToolUseBlock(type="tool_use", id="t3", name="Grep", input={"pattern": "y"}),
         TextBlock(type="text", text="second"),
     )
-    assert renderer.render_assistant_message(message) == (
+    for progress in (True, False):
-        "### Assistant:\n\nfirst\n\nsecond\n"
+        assert renderer.render_assistant_message(
+            message, thinking_is_progress=progress
+        ) == ("### Assistant:\n\nfirst\n\nsecond\n")
+
+
+def _progress_turn() -> Message:
+    return _message(
+        TextBlock(type="text", text="Checking the containers."),
+        ThinkingBlock(type="thinking", thinking="", signature="s"),
+        ToolUseBlock(type="tool_use", id="t1", name="komodo", input={}),
+        ThinkingBlock(type="thinking", thinking="", signature="s"),
+        ThinkingBlock(type="thinking", thinking="All up, reading logs.", signature="s"),
+        ToolUseBlock(type="tool_use", id="t2", name="komodo", input={}),
+        ThinkingBlock(type="thinking", thinking="Found it.\nOne more.", signature="s"),
+        ToolUseBlock(type="tool_use", id="t3", name="komodo", input={}),
+        TextBlock(type="text", text="The stack is healthy."),
     )
 
 
-def test_thinking_renders_as_collapsed_callout() -> None:
+def _reasoning_turn() -> Message:
-    message = _message(
+    return _message(
-        ThinkingBlock(type="thinking", thinking="a\nb", signature="s"),
+        ThinkingBlock(type="thinking", thinking="one", signature=""),
+        ToolUseBlock(type="tool_use", id="t1", name="web_search", input={}),
+        ThinkingBlock(type="thinking", thinking="two", signature=""),
+        ToolUseBlock(type="tool_use", id="t2", name="read_page", input={}),
+        ThinkingBlock(type="thinking", thinking="three", signature=""),
         TextBlock(type="text", text="answer"),
-    )
+        ThinkingBlock(type="thinking", thinking="four\nfive", signature=""),
-    assert renderer.render_assistant_message(message) == (
-        "### Assistant:\n\n> [!thinking]-\n> a\n> b\n\nanswer\n"
-    )
-
-
-def test_thinking_between_hidden_tool_calls_collapses_to_the_first() -> None:
-    message = _message(
-        ThinkingBlock(type="thinking", thinking="one", signature="s"),
-        ToolUseBlock(type="tool_use", id="t1", name="Read", input={}),
-        ThinkingBlock(type="thinking", thinking="two", signature="s"),
-        ToolUseBlock(type="tool_use", id="t2", name="Read", input={}),
-        ThinkingBlock(type="thinking", thinking="three", signature="s"),
-        TextBlock(type="text", text="answer"),
-        ThinkingBlock(type="thinking", thinking="four", signature="s"),
         TextBlock(type="text", text="more"),
     )
-    assert renderer.render_assistant_message(message) == (
+
-        "### Assistant:\n\n> [!thinking]-\n> one\n\nanswer\n\n"
+
-        "> [!thinking]-\n> four\n\nmore\n"
+REASONING_TURN_AS_BEFORE = (
+    "### Assistant:\n\n> [!thinking]-\n> one\n\nanswer\n\n"
+    "> [!thinking]-\n> four\n> five\n\nmore\n"
+)
+
+
+def _full_circle(
+    message: Message, *, thinking_is_progress: bool
+) -> history.ForkOutcome:
+    body = renderer.render_user_text("status?") + renderer.render_assistant_message(
+        message, thinking_is_progress=thinking_is_progress
     )
+    turns = parser.parse(body + renderer.render_user_text("thanks")).turns
+    stored = [
+        {"role": "user", "content": "status?"},
+        {"role": "assistant", "content": [b.model_dump() for b in message.content]},
+    ]
+    outcome = history.diff_and_fork(
+        stored=stored, incoming=turns, thinking_is_progress=thinking_is_progress
+    )
+    assert outcome.divergence_index is None
+    assert not outcome.edited
+    assert outcome.messages[1] == stored[1]
+    return outcome
+
+
+def test_progress_renders_as_text_in_place_and_empty_thinking_vanishes() -> None:
+    assert renderer.render_assistant_message(
+        _progress_turn(), thinking_is_progress=True
+    ) == (
+        "### Assistant:\n\nChecking the containers.\n\nAll up, reading logs.\n\n"
+        "Found it.\nOne more.\n\nThe stack is healthy.\n"
+    )
+
+
+def test_reasoning_renders_as_collapsed_callouts_as_before() -> None:
+    assert (
+        renderer.render_assistant_message(_reasoning_turn(), thinking_is_progress=False)
+        == REASONING_TURN_AS_BEFORE
+    )
+
+
+def test_progress_round_trips_through_the_parser_and_the_stored_history() -> None:
+    message = _progress_turn()
+    body = renderer.render_assistant_message(message, thinking_is_progress=True)
+    assert parser.parse_assistant_structure(body.removeprefix("### Assistant:")) == [
+        TextSegment(
+            text="Checking the containers.\n\nAll up, reading logs.\n\n"
+            "Found it.\nOne more.\n\nThe stack is healthy."
+        )
+    ]
+    _full_circle(message, thinking_is_progress=True)
+
+
+def test_reasoning_round_trips_through_the_parser_and_the_stored_history() -> None:
+    message = _reasoning_turn()
+    body = renderer.render_assistant_message(message, thinking_is_progress=False)
+    assert parser.parse_assistant_structure(body.removeprefix("### Assistant:")) == [
+        TextSegment(text="answer"),
+        TextSegment(text="more"),
+    ]
+    _full_circle(message, thinking_is_progress=False)
+
+
+def test_an_edited_progress_turn_rebuilds_instead_of_rewriting_signed_blocks() -> None:
+    message = _progress_turn()
+    body = renderer.render_assistant_message(
+        message, thinking_is_progress=True
+    ).replace("Found it", "Fixed it")
+    turns = parser.parse(
+        renderer.render_user_text("status?") + body + renderer.render_user_text("ok")
+    ).turns
+    stored = [
+        {"role": "user", "content": "status?"},
+        {"role": "assistant", "content": [b.model_dump() for b in message.content]},
+    ]
+    outcome = history.diff_and_fork(
+        stored=stored, incoming=turns, thinking_is_progress=True
+    )
+    assert outcome.edited
+    assert outcome.messages[1] == {
+        "role": "assistant",
+        "content": [{"type": "text", "text": turns[1].text}],
+    }
+
+
+async def test_a_raycast_archive_with_reasoning_renders_as_before(
+    tmp_path: Path,
+) -> None:
+    logger = CrossFrontendLogger(vault_path=tmp_path, logged_subdir="logged")
+    record = TurnRecord(
+        agent_name="beaver-gemini-pro-high",
+        input_messages=[{"role": "user", "content": "q"}],
+        output_message=_reasoning_turn(),
+        thinking_is_progress=False,
+    )
+    await logger.handle(record)
+    [written] = (tmp_path / "logged" / "beaver-gemini-pro-high").glob("*.md")
+    body = parser.parse(written.read_text()).body
+    assert f"---\n\n{REASONING_TURN_AS_BEFORE}\n---\n\n### User:" in body
+
+
+def test_a_file_with_collapsed_thinking_callouts_parses_as_before() -> None:
+    body = (
+        "### User:\n\nq\n\n---\n\n### Assistant:\n\n"
+        "> [!thinking]-\n> weighing it\n> twice\n\nanswer\n\n"
+        "> [!tool]- Read · notes.md\n> input\n\n"
+        "> [!thinking]-\n>\n\nmore\n\n---\n\n### User:\n\nnext\n"
+    )
+    parsed = parser.parse(body)
+    assert parsed.turns[1].structure == (
+        TextSegment(text="answer"),
+        ToolSegment(name="Read"),
+        TextSegment(text="more"),
+    )
+    assert parsed.messages == [
+        {"role": "user", "content": "q"},
+        {"role": "assistant", "content": "answer\n\nmore"},
+        {"role": "user", "content": "next"},
+    ]
 
 
 def test_user_scaffold_matches_rendered_user_turn_shape() -> None:

tests/test_telegram.py

@@ -20,6 +20,7 @@ from beaver_gateway.frontends.base import GatewayRuntime
 from beaver_gateway.frontends.files import SendFileError, SendFiles, method_for
 from beaver_gateway.frontends.telegram import Attachments, TelegramFrontend
 from beaver_gateway.frontends.telegram.drafts import Draft
+from beaver_gateway.frontends.telegram.texts import TelegramTexts
 from beaver_gateway.frontends.telegram.render import (
     LIMIT,
     chunks,
@@ -1194,3 +1195,74 @@ async def test_a_refused_upload_comes_back_to_the_agent(stack: Stack) -> None:
     outside = await _send(stack, master, path="/etc/passwd")
     assert outside["is_error"]
     assert "outside the allowed roots" in outside["content"][0]["text"]
+
+
+def _progress_draft() -> Draft:
+    return Draft(cast("Any", None), chat_id=USER, thread_id=None, turn_id="t")
+
+
+def _thinking_delta(text: str) -> Any:
+    return {
+        "event": {
+            "type": "content_block_delta",
+            "index": 0,
+            "delta": {"type": "thinking_delta", "thinking": text},
+        }
+    }
+
+
+def test_progress_updates_reach_the_draft_but_hidden_reasoning_does_not() -> None:
+    draft = _progress_draft()
+    texts = TelegramTexts()
+    stub = SimpleNamespace(_drafts={"k": draft}, texts=texts)
+    feed = functools.partial(
+        TelegramFrontend._feed_draft, cast("Any", stub), "k", thinking_is_progress=True
+    )
+
+    feed(_thinking_delta(""))
+    assert draft.text == ""
+    assert draft.status == texts.thinking
+
+    feed(_thinking_delta("checking the vault"))
+    assert draft.text == "checking the vault"
+    assert draft.status == texts.writing
+
+
+def test_reasoning_from_an_agent_whose_thinking_is_not_progress_stays_hidden() -> None:
+    draft = _progress_draft()
+    texts = TelegramTexts()
+    stub = SimpleNamespace(_drafts={"k": draft}, texts=texts)
+    TelegramFrontend._feed_draft(
+        cast("Any", stub),
+        "k",
+        _thinking_delta("weighing it"),
+        thinking_is_progress=False,
+    )
+    assert draft.text == ""
+    assert draft.status == texts.thinking
+
+
+def test_a_tool_call_between_progress_updates_starts_a_new_paragraph() -> None:
+    draft = _progress_draft()
+    texts = TelegramTexts()
+    stub = SimpleNamespace(
+        _drafts={"k": draft},
+        texts=texts,
+        _tool_status=lambda name, _input: texts.tool.format(label=name),
+    )
+    feed = functools.partial(
+        TelegramFrontend._feed_draft, cast("Any", stub), "k", thinking_is_progress=True
+    )
+
+    feed(_thinking_delta("reading the note"))
+    feed(
+        {
+            "event": {
+                "type": "content_block_start",
+                "index": 1,
+                "content_block": {"type": "tool_use", "name": "Read"},
+            }
+        }
+    )
+    feed(_thinking_delta("now writing it back"))
+    assert draft.text == "reading the note\n\nnow writing it back"

tests/test_watch.py

@@ -0,0 +1,63 @@
+import tempfile
+from collections.abc import AsyncIterator
+from pathlib import Path
+
+import pytest
+from watchfiles import Change as FileChange
+
+from beaver_gateway.vault import watch as watch_module
+from beaver_gateway.vault.watch import VaultFilter, VaultWatch, WatchRules
+
+VAULT = Path("/vault")
+FILTER = VaultFilter()
+
+
+def wakes(rel: str) -> bool:
+    return FILTER(FileChange.modified, str(VAULT / rel))
+
+
+def test_obsidian_never_wakes_the_watcher() -> None:
+    assert not wakes(".obsidian/.sync.lock")
+    assert not wakes(".obsidian/.sync.lock/data")
+    assert not wakes(".obsidian")
+    assert not wakes(".obsidian/workspace.json")
+    assert not wakes(".obsidian/plugins/obsidian-git/data.json")
+
+
+def test_other_service_paths_never_wake_the_watcher() -> None:
+    assert not wakes(".trash/📅 дни/2026-09-01.md")
+    assert not wakes(".git/index")
+
+
+def test_notes_still_wake_the_watcher() -> None:
+    assert wakes("📅 дни/2026-09-09.md")
+    assert wakes("мета/бобер/состояние.md")
+    assert wakes("заметка.md")
+    assert wakes("💻 проекты/бобер/план.md")
+
+
+def test_lock_beside_a_note_is_quiet_but_the_note_is_not() -> None:
+    assert not wakes("📅 дни/.2026-09-09.md.lock")
+    assert not wakes("📅 дни/2026-09-09.md.tmp")
+    assert not wakes("📅 дни/2026-09-09.md~")
+    assert wakes("📅 дни/2026-09-09.md")
+
+
+async def test_the_filter_is_applied_by_awatch_itself(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    seen: dict[str, object] = {}
+
+    async def fake_awatch(
+        *paths: object, **kwargs: object
+    ) -> AsyncIterator[set[object]]:
+        seen.update(kwargs)
+        seen["paths"] = paths
+        for change in ():
+            yield change
+
+    monkeypatch.setattr(watch_module, "awatch", fake_awatch)
+    root = Path(tempfile.mkdtemp(prefix="beaver-vault-"))
+    watch = VaultWatch(root, WatchRules(names=("**",)), tz="UTC")
+    await watch.run()
+    assert isinstance(seen["watch_filter"], VaultFilter)

ui/src/lib/api/types.ts

@@ -79,6 +79,7 @@ export interface ConversationInfo extends ConversationSummary {
   parent: string | null;
   question: PendingQuestion | null;
   queue: QueueItem[];
+  thinking_is_progress?: boolean;
   turn: TurnSnapshot | null;
 }
 

ui/src/lib/panel/activity.svelte.ts

@@ -233,8 +233,13 @@ export class ActivityModel {
     }
     if (sdk.type === "content_block_delta" && parent === null) {
       const delta = sdk.delta as Record<string, unknown> | undefined;
-      if (delta?.type === "text_delta" && typeof delta.text === "string") {
+      const progress = this.conversation?.thinking_is_progress === true;
-        turn.text = joinText(turn.text, delta.text, turn.paused);
+      const text =
+        delta?.type === "thinking_delta" && progress
+          ? delta.thinking
+          : delta?.text;
+      if (typeof text === "string" && text) {
+        turn.text = joinText(turn.text, text, turn.paused);
         turn.paused = false;
       }
     }

ui/src/lib/panel/chat-view.svelte

@@ -190,9 +190,20 @@
   }
 
   function blocks(message: HistoryMessage): ContentBlock[] {
-    return typeof message.content === "string"
+    if (typeof message.content === "string") {
-      ? [{ text: message.content, type: "text" }]
+      return [{ text: message.content, type: "text" }];
-      : message.content;
+    }
+    if (!model.conversation?.thinking_is_progress) {
+      return message.content;
+    }
+    return message.content.flatMap((block) => {
+      if (block.type !== "thinking") {
+        return [block];
+      }
+      return block.thinking?.trim()
+        ? [{ text: block.thinking, type: "text" }]
+        : [];
+    });
   }
 
   function resultText(block: ContentBlock): string {
@@ -287,7 +298,7 @@
               {/if}
             </button>
           </div>
-        {:else if !(isResultOnly && !showResults)}
+        {:else if parts.length > 0 && !(isResultOnly && !showResults)}
           <div
             class={cn(
             "group flex min-w-0 max-w-full flex-col gap-1.5",

uv.lock

@@ -338,7 +338,7 @@ requires-dist = [
     { name = "anthropic", specifier = ">=0.103.0" },
     { name = "anyio", specifier = ">=4.13.0" },
     { name = "argon2-cffi", specifier = ">=25.1.0" },
-    { name = "claude-agent-sdk", specifier = ">=0.2.150" },
+    { name = "claude-agent-sdk", specifier = ">=0.2.157" },
     { name = "croniter", specifier = ">=6.2.4" },
     { name = "fastapi", specifier = ">=0.136.1" },
     { name = "fastmcp", specifier = ">=3.3.1" },
@@ -450,7 +450,7 @@ wheels = [
 
 [[package]]
 name = "claude-agent-sdk"
-version = "0.2.150"
+version = "0.2.157"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "anyio" },
@@ -458,13 +458,12 @@ dependencies = [
     { name = "mcp" },
     { name = "sniffio" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/6d/1e/99bababecf24a4a55a0775665ff0be010aff26f653fc9be2f4f17a0b5ef0/claude_agent_sdk-0.2.150.tar.gz", hash = "sha256:3856e932f8d45514ddf5ddd540fb844678bb11638cabf9b939ce8b708e4eb352", size = 344692, upload-time = "2026-09-01T18:07:15.284Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/b4/e3/444df1691c864a8349067686719b43e714fc43d3853fb1b2ebc2097e2a6e/claude_agent_sdk-0.2.157.tar.gz", hash = "sha256:4ce6f41bc965fcb4a9d11bd7a3cac266975c75b632f3d74a26f55f0fffb3160a", size = 346579, upload-time = "2026-09-18T18:21:10.306Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/c3/13/b5b86d90c95e8a691143eeb45bbf2e1c250324817980464d6ea63ef0f789/claude_agent_sdk-0.2.150-py3-none-macosx_11_0_arm64.whl", hash = "sha256:8f58d746f3552faeac52883b5a502a8d9b5aef067ea0b68fae5c9bae0f39d73d", size = 85396845, upload-time = "2026-09-01T18:07:19.074Z" },
+    { url = "https://files.pythonhosted.org/packages/ec/e8/3e9cc23eb5a5a01a0e2ee931a17db2785f95407525d50366af06a4a5469c/claude_agent_sdk-0.2.157-py3-none-macosx_11_0_arm64.whl", hash = "sha256:dd6a2c7aa1901e43b827411d62f2327aed82f53ed8ca6b076eaa5be4a1fc681f", size = 92744001, upload-time = "2026-09-18T18:21:15.969Z" },
-    { url = "https://files.pythonhosted.org/packages/d9/8f/c17da4fa9cd8bddb8ac1538456dd0fde3a43abb1dfc28ca448f85e2a10d3/claude_agent_sdk-0.2.150-py3-none-macosx_11_0_x86_64.whl", hash = "sha256:fee3ae2dd457742d82070b5fa122070ee8e2e80b4f9b37a57ff0a8261ade656d", size = 89904483, upload-time = "2026-09-01T18:07:22.671Z" },
+    { url = "https://files.pythonhosted.org/packages/49/eb/e5095703a8efbe504ec468cab039543caabfc0ea3327c735a3d1b60e57ac/claude_agent_sdk-0.2.157-py3-none-macosx_11_0_x86_64.whl", hash = "sha256:ebf3f514c61db9be01bbf6d1acb0d993247340583b3559d2df75e3822f618bbb", size = 97352326, upload-time = "2026-09-18T18:21:21.007Z" },
-    { url = "https://files.pythonhosted.org/packages/8d/95/5b33717b807c6fbe29fca55a246eab9a09a80c9dd288ffeb31f05681b7fc/claude_agent_sdk-0.2.150-py3-none-manylinux_2_17_aarch64.whl", hash = "sha256:813535a0267271d91af29c456ec26137b34a8b7bc066647ea59efba6e3c26c93", size = 95270310, upload-time = "2026-09-01T18:07:26.61Z" },
+    { url = "https://files.pythonhosted.org/packages/d6/f4/64505beb3c757b863753d6556539554704ec9168cb96454dc6463b14df85/claude_agent_sdk-0.2.157-py3-none-manylinux_2_17_aarch64.whl", hash = "sha256:a58089e0c11364db2ec7a139e09b36a4fb9c95817a758bee7b28e0434f55c0b6", size = 102701030, upload-time = "2026-09-18T18:21:27.065Z" },
-    { url = "https://files.pythonhosted.org/packages/9f/de/5a8b59e6a713ea3f4be88816cfd487b1a5b08c8ad9f042abe56d2decc184/claude_agent_sdk-0.2.150-py3-none-manylinux_2_17_x86_64.whl", hash = "sha256:cb65ff33060baa098b420232c630fa287d0f95e9469e7ceb706d0b7d8841888e", size = 95460277, upload-time = "2026-09-01T18:07:31.092Z" },
+    { url = "https://files.pythonhosted.org/packages/03/8f/2a7c121469fd16d1d6465bec180726c54f6e151b12e50835283b4285d1ce/claude_agent_sdk-0.2.157-py3-none-manylinux_2_17_x86_64.whl", hash = "sha256:bd126b2edee3a247b2c7c4cee1f5a3ea68ef748288ec0650db88dd3918e5471a", size = 102830653, upload-time = "2026-09-18T18:21:32.948Z" },
-    { url = "https://files.pythonhosted.org/packages/dd/42/d147578a0df16c1b4b0deaebfb3d25e590ad5c16cb6b59a68f9dbb618be3/claude_agent_sdk-0.2.150-py3-none-win_amd64.whl", hash = "sha256:3b517a21c723854de778b38fa51660c5b68ee5fb6f28b2abf291a51471a7b378", size = 98078644, upload-time = "2026-09-01T18:07:35.734Z" },
 ]
 
 [[package]]