From bbce425eebb5f2ee82ba0d7c14b133baa61261f6 Mon Sep 17 00:00:00 2001 From: h Date: Tue, 22 Sep 2026 18:02:01 +0000 Subject: [PATCH] feat(backends,conversations,frontends): opus 5.5 progress updates reach telegram, current claude cli in the image --- Dockerfile | 23 +++++++- pyproject.toml | 2 +- src/beaver_gateway/agents/claude.py | 3 ++ src/beaver_gateway/backends/claude_sdk.py | 11 ++++ src/beaver_gateway/conversations/turns.py | 27 +++++++--- .../frontends/telegram/frontend.py | 7 ++- tests/test_conversations.py | 18 +++++++ tests/test_telegram.py | 54 +++++++++++++++++++ uv.lock | 15 +++--- 9 files changed, 140 insertions(+), 20 deletions(-) diff --git a/Dockerfile b/Dockerfile index 81fd41e..337b60b 100644 --- a/Dockerfile +++ b/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 -# claude-agent-sdk wheel (`_bundled/claude`), nothing to install here. +# the user's config. 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 diff --git a/pyproject.toml b/pyproject.toml index d83676a..a9bb22a 100644 --- a/pyproject.toml +++ b/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", diff --git a/src/beaver_gateway/agents/claude.py b/src/beaver_gateway/agents/claude.py index fa553f9..cd80383 100644 --- a/src/beaver_gateway/agents/claude.py +++ b/src/beaver_gateway/agents/claude.py @@ -24,6 +24,9 @@ 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.""" diff --git a/src/beaver_gateway/backends/claude_sdk.py b/src/beaver_gateway/backends/claude_sdk.py index a4d0efd..7424f2f 100644 --- a/src/beaver_gateway/backends/claude_sdk.py +++ b/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) diff --git a/src/beaver_gateway/conversations/turns.py b/src/beaver_gateway/conversations/turns.py index be2117d..fbd0586 100644 --- a/src/beaver_gateway/conversations/turns.py +++ b/src/beaver_gateway/conversations/turns.py @@ -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, Sequence + from collections.abc import AsyncIterator, Callable, Iterable, Sequence from beaver_gateway.events.stream import MessageStreamEvent -__all__ = ["Turns", "content_of", "image_block"] +__all__ = ["Turns", "content_of", "image_block", "reply_text"] _log = logging.getLogger(__name__) @@ -164,12 +164,7 @@ class Turns(Seeds): ): acc.feed(event) message = acc.finalize(model=agent.model) - reply = "\n\n".join( - getattr(b, "text", "") - for b in message.content - if getattr(b, "type", "") == "text" - ).strip() - return reply, capture + return reply_text(message.content), capture async def before_turn(self, conv: Conversation) -> str | None: # noqa: ARG002 return None @@ -489,6 +484,22 @@ 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]]: diff --git a/src/beaver_gateway/frontends/telegram/frontend.py b/src/beaver_gateway/frontends/telegram/frontend.py index 623b63d..c11d735 100644 --- a/src/beaver_gateway/frontends/telegram/frontend.py +++ b/src/beaver_gateway/frontends/telegram/frontend.py @@ -983,7 +983,12 @@ 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: + 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": diff --git a/tests/test_conversations.py b/tests/test_conversations.py index a7bad0e..19b14fb 100644 --- a/tests/test_conversations.py +++ b/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.conversations.turns import reply_text from beaver_gateway.backends.claude_sdk import ClaudeSdkBackend from beaver_gateway.events.bus import EventBus from beaver_gateway.conversations.service import ( @@ -958,3 +962,17 @@ 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) == "looking it up\n\nfound it" + + +def test_reply_text_on_a_plain_answer_is_unchanged() -> None: + blocks = [AnthropicTextBlock(type="text", text=" just an answer ")] + assert reply_text(blocks) == "just an answer" diff --git a/tests/test_telegram.py b/tests/test_telegram.py index 1dafdf6..6bc4304 100644 --- a/tests/test_telegram.py +++ b/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 +1195,56 @@ 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") + + 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_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") + + 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" diff --git a/uv.lock b/uv.lock index 52e7a29..d27a232 100644 --- a/uv.lock +++ b/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/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/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/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/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" }, + { 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/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/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/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" }, ] [[package]]