lionagi 0.13.2__py3-none-any.whl → 0.13.3__py3-none-any.whl

lionagi/service/connections/match_endpoint.py

@@ -2,6 +2,8 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 
+from lionagi.service.connections.endpoint_config import EndpointConfig
+
 from .endpoint import Endpoint
 
 
@@ -45,13 +47,29 @@ def match_endpoint(
         from .providers.perplexity_ import PerplexityChatEndpoint
 
         return PerplexityChatEndpoint(**kwargs)
-    if provider == "claude_code" and (
-        "query" in endpoint or "code" in endpoint
-    ):
-        from lionagi.service.connections.providers.claude_code_ import (
-            ClaudeCodeEndpoint,
-        )
+    if provider == "claude_code":
+        if "cli" in endpoint:
+            from .providers.claude_code_cli import ClaudeCodeCLIEndpoint
+
+            return ClaudeCodeCLIEndpoint(**kwargs)
+
+        if "query" in endpoint or "code" in endpoint:
+            from lionagi.service.connections.providers.claude_code_ import (
+                ClaudeCodeEndpoint,
+            )
+
+            return ClaudeCodeEndpoint(**kwargs)
+
+    from .providers.oai_ import OpenaiChatEndpoint
 
-        return ClaudeCodeEndpoint(**kwargs)
+    config = EndpointConfig(
+        provider=provider,
+        endpoint=endpoint or "chat/completions",
+        name="openai_compatible_chat",
+        auth_type="bearer",
+        content_type="application/json",
+        method="POST",
+        requires_tokens=True,
+    )
 
-    return None
+    return Endpoint(config, **kwargs)

lionagi/service/connections/providers/claude_code_.py

@@ -76,6 +76,7 @@ class ClaudeCodeRequest(BaseModel):
         description="Automatically finish the conversation after the first response",
     )
     verbose_output: bool = Field(default=False, exclude=True)
+    cli_display_theme: Literal["light", "dark"] = "light"
 
     # ------------------------ validators & helpers --------------------------
     @field_validator("permission_mode", mode="before")
@@ -141,8 +142,7 @@ class ClaudeCodeRequest(BaseModel):
     # ------------------------ CLI helpers -----------------------------------
     def as_cmd_args(self) -> list[str]:
         """Build argument list for the *Node* `claude` CLI."""
-        full_prompt = f"Human User: {self.prompt}\n\nAssistant:"
-        args: list[str] = ["-p", full_prompt, "--output-format", "stream-json"]
+        args: list[str] = ["-p", self.prompt, "--output-format", "stream-json"]
         if self.allowed_tools:
             args.append("--allowedTools")
             for tool in self.allowed_tools:
@@ -356,9 +356,8 @@ class ClaudeCodeEndpoint(Endpoint):
 
         # 1. stream the Claude Code response
         async for chunk in self._stream_claude_code(**payload):
-
             if request.verbose_output:
-                _display_message(chunk)
+                _display_message(chunk, theme=request.cli_display_theme)
 
             if isinstance(chunk, cc_types.SystemMessage):
                 system = chunk
@@ -391,7 +390,7 @@ class ClaudeCodeEndpoint(Endpoint):
                                 display_str=True,
                                 format_curly=True,
                                 max_panel_width=100,
-                                theme="light",
+                                theme=request.cli_display_theme,
                             )
 
                     responses.append(chunk)
@@ -400,7 +399,7 @@ class ClaudeCodeEndpoint(Endpoint):
         return self._parse_claude_code_response(responses)
 
 
-def _display_message(chunk):
+def _display_message(chunk, theme):
     if isinstance(
         chunk,
         cc_types.SystemMessage
@@ -415,7 +414,7 @@ def _display_message(chunk):
                     md=True,
                     display_str=True,
                     max_panel_width=100,
-                    theme="light",
+                    theme=theme,
                 )
             else:
                 as_readable(
@@ -423,7 +422,7 @@ def _display_message(chunk):
                     format_curly=True,
                     display_str=True,
                     max_panel_width=100,
-                    theme="light",
+                    theme=theme,
                 )
 
     if isinstance(chunk, cc_types.ResultMessage):
@@ -434,7 +433,7 @@ def _display_message(chunk):
             display_str=True,
             format_curly=True,
             max_panel_width=100,
-            theme="light",
+            theme=theme,
         )
 
 

lionagi/service/connections/providers/claude_code_cli.py

@@ -0,0 +1,414 @@
+from __future__ import annotations
+
+import asyncio
+import codecs
+import contextlib
+import dataclasses
+import json
+import logging
+import shutil
+from collections.abc import AsyncIterator, Callable
+from datetime import datetime
+from functools import partial
+from textwrap import shorten
+from typing import Any
+
+from json_repair import repair_json
+from pydantic import BaseModel
+
+from lionagi.libs.schema.as_readable import as_readable
+from lionagi.service.connections.endpoint import Endpoint, EndpointConfig
+from lionagi.utils import to_dict
+
+from .claude_code_ import ClaudeCodeRequest
+
+CLAUDE = shutil.which("claude") or "claude"
+if not shutil.which(CLAUDE):
+    raise RuntimeError(
+        "Claude CLI binary not found (npm i -g @anthropic-ai/claude-code)"
+    )
+logging.basicConfig(level=logging.INFO)
+log = logging.getLogger("claude-cli")
+
+
+@dataclasses.dataclass
+class ClaudeChunk:
+    """Low-level wrapper around every NDJSON object coming from the CLI."""
+
+    raw: dict[str, Any]
+    type: str
+    # convenience views
+    thinking: str | None = None
+    text: str | None = None
+    tool_use: dict[str, Any] | None = None
+    tool_result: dict[str, Any] | None = None
+
+
+@dataclasses.dataclass
+class ClaudeSession:
+    """Aggregated view of a whole CLI conversation."""
+
+    session_id: str | None = None
+    model: str | None = None
+
+    # chronological log
+    chunks: list[ClaudeChunk] = dataclasses.field(default_factory=list)
+
+    # materialised views
+    thinking_log: list[str] = dataclasses.field(default_factory=list)
+    messages: list[dict[str, Any]] = dataclasses.field(default_factory=list)
+    tool_uses: list[dict[str, Any]] = dataclasses.field(default_factory=list)
+    tool_results: list[dict[str, Any]] = dataclasses.field(
+        default_factory=list
+    )
+
+    # final summary
+    result: str = ""
+    usage: dict[str, Any] = dataclasses.field(default_factory=dict)
+    total_cost_usd: float | None = None
+    num_turns: int | None = None
+    duration_ms: int | None = None
+    duration_api_ms: int | None = None
+    is_error: bool = False
+
+
+# --------------------------------------------------------------------------- helpers
+
+
+async def ndjson_from_cli(request: ClaudeCodeRequest):
+    """
+    Yields each JSON object emitted by the *claude-code* CLI.
+
+    • Robust against UTF‑8 splits across chunks (incremental decoder).
+    • Robust against braces inside strings (uses json.JSONDecoder.raw_decode)
+    • Falls back to `json_repair.repair_json` when necessary.
+    """
+    workspace = request.cwd()
+    workspace.mkdir(parents=True, exist_ok=True)
+
+    proc = await asyncio.create_subprocess_exec(
+        CLAUDE,
+        *request.as_cmd_args(),
+        cwd=str(workspace),
+        stdout=asyncio.subprocess.PIPE,
+        stderr=asyncio.subprocess.PIPE,
+    )
+
+    decoder = codecs.getincrementaldecoder("utf-8")()
+    json_decoder = json.JSONDecoder()
+    buffer: str = ""  # text buffer that may hold >1 JSON objects
+
+    try:
+        while True:
+            chunk = await proc.stdout.read(4096)
+            if not chunk:
+                break
+
+            # 1) decode *incrementally* so we never split multibyte chars
+            buffer += decoder.decode(chunk)
+
+            # 2) try to peel off as many complete JSON objs as possible
+            while buffer:
+                buffer = buffer.lstrip()  # remove leading spaces/newlines
+                if not buffer:
+                    break
+                try:
+                    obj, idx = json_decoder.raw_decode(buffer)
+                    yield obj
+                    buffer = buffer[idx:]  # keep remainder for next round
+                except json.JSONDecodeError:
+                    # incomplete → need more bytes
+                    break
+
+        # 3) flush any tail bytes in the incremental decoder
+        buffer += decoder.decode(b"", final=True)
+        buffer = buffer.strip()
+        if buffer:
+            try:
+                obj, idx = json_decoder.raw_decode(buffer)
+                yield obj
+            except json.JSONDecodeError:
+                try:
+                    fixed = repair_json(buffer)
+                    yield json.loads(fixed)
+                    log.warning(
+                        "Repaired malformed JSON fragment at stream end"
+                    )
+                except Exception:
+                    log.error(
+                        "Skipped unrecoverable JSON tail: %.120s…", buffer
+                    )
+
+        # 4) propagate non‑zero exit code
+        if await proc.wait() != 0:
+            err = (await proc.stderr.read()).decode().strip()
+            raise RuntimeError(err or "CLI exited non‑zero")
+
+    finally:
+        with contextlib.suppress(ProcessLookupError):
+            proc.terminate()
+        await proc.wait()
+
+
+# --------------------------------------------------------------------------- SSE route
+async def stream_events(request: ClaudeCodeRequest):
+    async for obj in ndjson_from_cli(request):
+        yield obj
+    yield {"type": "done"}
+
+
+print_readable = partial(as_readable, md=True, display_str=True)
+
+
+def _pp_system(sys_obj: dict[str, Any], theme) -> None:
+    txt = (
+        f"◼️  **Claude Code Session**  \n"
+        f"- id: `{sys_obj.get('session_id', '?')}`  \n"
+        f"- model: `{sys_obj.get('model', '?')}`  \n"
+        f"- tools: {', '.join(sys_obj.get('tools', [])[:8])}"
+        + ("…" if len(sys_obj.get("tools", [])) > 8 else "")
+    )
+    print_readable(txt, border=False, theme=theme)
+
+
+def _pp_thinking(thought: str, theme) -> None:
+    text = f"""
+    🧠 Thinking:
+    {thought}
+    """
+    print_readable(text, border=True, theme=theme)
+
+
+def _pp_assistant_text(text: str, theme) -> None:
+    txt = f"""
+    > 🗣️ Claude:
+    {text}
+    """
+    print_readable(txt, theme=theme)
+
+
+def _pp_tool_use(tu: dict[str, Any], theme) -> None:
+    preview = shorten(str(tu["input"]).replace("\n", " "), 130)
+    body = f"- 🔧 Tool Use — {tu['name']}({tu['id']}) - input: {preview}"
+    print_readable(body, border=False, panel=False, theme=theme)
+
+
+def _pp_tool_result(tr: dict[str, Any], theme) -> None:
+    body_preview = shorten(str(tr["content"]).replace("\n", " "), 130)
+    status = "ERR" if tr.get("is_error") else "OK"
+    body = (
+        f"- 📄 Tool Result({tr['tool_use_id']}) - {status}\n\n"
+        f"\tcontent: {body_preview}"
+    )
+    print_readable(body, border=False, panel=False, theme=theme)
+
+
+def _pp_final(sess: ClaudeSession, theme) -> None:
+    usage = sess.usage or {}
+    txt = (
+        f"### ✅ Session complete - {datetime.utcnow().isoformat(timespec='seconds')} UTC\n"
+        f"**Result:**\n\n{sess.result or ''}\n\n"
+        f"- cost: **${sess.total_cost_usd:.4f}**  \n"
+        f"- turns: **{sess.num_turns}**  \n"
+        f"- duration: **{sess.duration_ms} ms** (API {sess.duration_api_ms} ms)  \n"
+        f"- tokens in/out: {usage.get('input_tokens', 0)}/{usage.get('output_tokens', 0)}"
+    )
+    print_readable(txt, theme=theme)
+
+
+# --------------------------------------------------------------------------- internal utils
+
+
+async def _maybe_await(func, *args, **kw):
+    """Call func which may be sync or async."""
+    res = func(*args, **kw) if func else None
+    if asyncio.iscoroutine(res):
+        await res
+
+
+# --------------------------------------------------------------------------- main parser
+
+
+async def stream_claude_code_cli(  # noqa: C901  (complexity from branching is fine here)
+    request: ClaudeCodeRequest,
+    session: ClaudeSession = ClaudeSession(),
+    *,
+    on_system: Callable[[dict[str, Any]], None] | None = None,
+    on_thinking: Callable[[str], None] | None = None,
+    on_text: Callable[[str], None] | None = None,
+    on_tool_use: Callable[[dict[str, Any]], None] | None = None,
+    on_tool_result: Callable[[dict[str, Any]], None] | None = None,
+    on_final: Callable[[ClaudeSession], None] | None = None,
+) -> AsyncIterator[ClaudeChunk | dict | ClaudeSession]:
+    """
+    Consume the ND‑JSON stream produced by ndjson_from_cli()
+    and return a fully‑populated ClaudeSession.
+
+    If callbacks are omitted a default pretty‑print is emitted.
+    """
+    stream = ndjson_from_cli(request)
+    theme = request.cli_display_theme or "light"
+
+    async for obj in stream:
+        typ = obj.get("type", "unknown")
+        chunk = ClaudeChunk(raw=obj, type=typ)
+        session.chunks.append(chunk)
+
+        # ------------------------ SYSTEM -----------------------------------
+        if typ == "system":
+            data = obj
+            session.session_id = data.get("session_id", session.session_id)
+            session.model = data.get("model", session.model)
+            await _maybe_await(on_system, data)
+            if request.verbose_output and on_system is None:
+                _pp_system(data, theme)
+            yield data
+
+        # ------------------------ ASSISTANT --------------------------------
+        elif typ == "assistant":
+            msg = obj["message"]
+            session.messages.append(msg)
+
+            for blk in msg.get("content", []):
+                btype = blk.get("type")
+                if btype == "thinking":
+                    thought = blk.get("thinking", "").strip()
+                    chunk.thinking = thought
+                    session.thinking_log.append(thought)
+                    await _maybe_await(on_thinking, thought)
+                    if request.verbose_output and on_thinking is None:
+                        _pp_thinking(thought, theme)
+
+                elif btype == "text":
+                    text = blk.get("text", "")
+                    chunk.text = text
+                    await _maybe_await(on_text, text)
+                    if request.verbose_output and on_text is None:
+                        _pp_assistant_text(text, theme)
+
+                elif btype == "tool_use":
+                    tu = {
+                        "id": blk["id"],
+                        "name": blk["name"],
+                        "input": blk["input"],
+                    }
+                    chunk.tool_use = tu
+                    session.tool_uses.append(tu)
+                    await _maybe_await(on_tool_use, tu)
+                    if request.verbose_output and on_tool_use is None:
+                        _pp_tool_use(tu, theme)
+
+                elif btype == "tool_result":
+                    tr = {
+                        "tool_use_id": blk["tool_use_id"],
+                        "content": blk["content"],
+                        "is_error": blk.get("is_error", False),
+                    }
+                    chunk.tool_result = tr
+                    session.tool_results.append(tr)
+                    await _maybe_await(on_tool_result, tr)
+                    if request.verbose_output and on_tool_result is None:
+                        _pp_tool_result(tr, theme)
+            yield chunk
+
+        # ------------------------ USER (tool_result containers) ------------
+        elif typ == "user":
+            msg = obj["message"]
+            session.messages.append(msg)
+            for blk in msg.get("content", []):
+                if blk.get("type") == "tool_result":
+                    tr = {
+                        "tool_use_id": blk["tool_use_id"],
+                        "content": blk["content"],
+                        "is_error": blk.get("is_error", False),
+                    }
+                    chunk.tool_result = tr
+                    session.tool_results.append(tr)
+                    await _maybe_await(on_tool_result, tr)
+                    if request.verbose_output and on_tool_result is None:
+                        _pp_tool_result(tr, theme)
+            yield chunk
+
+        # ------------------------ RESULT -----------------------------------
+        elif typ == "result":
+            session.result = obj.get("result", "").strip()
+            session.usage = obj.get("usage", {})
+            session.total_cost_usd = obj.get("total_cost_usd")
+            session.num_turns = obj.get("num_turns")
+            session.duration_ms = obj.get("duration_ms")
+            session.duration_api_ms = obj.get("duration_api_ms")
+            session.is_error = obj.get("is_error", False)
+
+        # ------------------------ DONE -------------------------------------
+        elif typ == "done":
+            break
+
+    # final pretty print
+    await _maybe_await(on_final, session)
+    if request.verbose_output and on_final is None:
+        _pp_final(session, theme)
+
+    yield session
+
+
+ENDPOINT_CONFIG = EndpointConfig(
+    name="claude_code_cli",
+    provider="claude_code",
+    base_url="internal",
+    endpoint="query_cli",
+    api_key="dummy",
+    request_options=ClaudeCodeRequest,
+    timeout=12000,  # 20 mins
+)
+
+
+class ClaudeCodeCLIEndpoint(Endpoint):
+    def __init__(self, config: EndpointConfig = ENDPOINT_CONFIG, **kwargs):
+        super().__init__(config=config, **kwargs)
+
+    def create_payload(self, request: dict | BaseModel, **kwargs):
+        req_dict = {**self.config.kwargs, **to_dict(request), **kwargs}
+        messages = req_dict.pop("messages")
+        req_obj = ClaudeCodeRequest.create(messages=messages, **req_dict)
+        return {"request": req_obj}, {}
+
+    async def stream(self, request: dict | BaseModel, **kwargs):
+        payload, _ = self.create_payload(request, **kwargs)["request"]
+        async for chunk in stream_claude_code_cli(payload):
+            yield chunk
+
+    async def _call(
+        self,
+        payload: dict,
+        headers: dict,  # type: ignore[unused-argument]
+        **kwargs,
+    ):
+        responses = []
+        request: ClaudeCodeRequest = payload["request"]
+        session: ClaudeSession = ClaudeSession()
+        system: dict = None
+
+        # 1. stream the Claude Code response
+        async for chunk in stream_claude_code_cli(request, session, **kwargs):
+            if isinstance(chunk, dict):
+                system = chunk
+            responses.append(chunk)
+
+        if request.auto_finish and not isinstance(
+            responses[-1], ClaudeSession
+        ):
+            req2 = request.model_copy(deep=True)
+            req2.max_turns = 1
+            req2.continue_conversation = True
+            if system:
+                req2.resume = system.get("session_id") if system else None
+
+            async for chunk in stream_claude_code_cli(req2, session, **kwargs):
+                responses.append(chunk)
+                if isinstance(chunk, ClaudeSession):
+                    break
+        print(
+            f"Session {session.session_id} finished with {len(responses)} chunks"
+        )
+
+        return to_dict(session, recursive=True)

lionagi/service/connections/providers/oai_.py

@@ -25,7 +25,7 @@ OPENAI_CHAT_ENDPOINT_CONFIG = EndpointConfig(
     provider="openai",
     base_url="https://api.openai.com/v1",
     endpoint="chat/completions",
-    kwargs={"model": "gpt-4o"},
+    kwargs={"model": "gpt-4.1-nano"},
     api_key=settings.OPENAI_API_KEY or "dummy-key-for-testing",
     auth_type="bearer",
     content_type="application/json",

lionagi/service/manager.py

@@ -9,7 +9,6 @@ from .imodel import iModel
 
 
 class iModelManager(Manager):
-
     def __init__(self, *args: iModel, **kwargs):
         super().__init__()
 

lionagi/service/rate_limited_processor.py

@@ -18,7 +18,6 @@ __all__ = (
 
 
 class RateLimitedAPIProcessor(Processor):
-
     event_type = APICalling
 
     def __init__(
@@ -127,7 +126,6 @@ class RateLimitedAPIProcessor(Processor):
 
 
 class RateLimitedAPIExecutor(Executor):
-
     processor_type = RateLimitedAPIProcessor
 
     def __init__(

lionagi/service/token_calculator.py

@@ -104,10 +104,8 @@ def get_image_pricing(model: str) -> dict:
 
 
 class TokenCalculator:
-
     @staticmethod
     def calculate_message_tokens(messages: list[dict], /, **kwargs) -> int:
-
         model = kwargs.get("model", "gpt-4o")
         tokenizer = tiktoken.get_encoding(get_encoding_name(model)).encode
 
@@ -149,7 +147,6 @@ class TokenCalculator:
         return_tokens: bool = False,
         return_decoded: bool = False,
     ) -> int | list[int]:
-
         if not s_:
             return 0
 

lionagi/session/branch.py

@@ -185,7 +185,6 @@ class Branch(Element, Communicatable, Relational):
                 use_lion_system_message,
             ]
         ):
-
             if use_lion_system_message:
                 system = f"Developer Prompt: {str(system)}" if system else ""
                 system = (LION_SYSTEM_MESSAGE + "\n\n" + system).strip()
@@ -647,7 +646,6 @@ class Branch(Element, Communicatable, Relational):
         """
         meta = {}
         if "clone_from" in self.metadata:
-
             # Provide some reference info about the source from which we cloned
             meta["clone_from"] = {
                 "id": str(self.metadata["clone_from"].id),

lionagi/session/session.py

@@ -86,7 +86,6 @@ class Session(Node, Communicatable, Relational):
         tools: Tool | Callable | list = None,
         **kwargs,  # additional branch parameters
     ) -> Branch:
-
         kwargs["system"] = system
         kwargs["system_sender"] = system_sender
         kwargs["system_datetime"] = system_datetime

lionagi/settings.py

@@ -56,7 +56,6 @@ LOG_CONFIG = {
 
 
 class Settings:
-
     class Config:
         TIMEZONE: timezone = timezone.utc
         LOG: dict = LOG_CONFIG