初始化：独立 oclaw 仓库首提交

- 在 oclaw/ 下重新初始化 Git 仓库
- 补齐子仓库 .gitignore，避免提交本地运行态数据（_local、node_modules、logs 等）
- 提交当前工程代码与配置

Made-with: Cursor

platform/llm/transports/__init__.py

@@ -0,0 +1 @@
+# oclaw.platform.llm.transports package

platform/llm/transports/anthropic_messages.py

@@ -0,0 +1,204 @@
+from __future__ import annotations
+
+import json
+import logging
+import os
+import uuid
+from typing import Any, Optional
+from collections.abc import Callable, Iterable
+
+from oclaw.platform.llm.transports.base import ChatModel, LLMResponse, LLMToolCall
+
+logger = logging.getLogger(__name__)
+
+
+def _as_dict(obj: Any) -> dict[str, Any] | None:
+    if obj is None:
+        return None
+    if isinstance(obj, dict):
+        return obj
+    if hasattr(obj, "model_dump"):
+        try:
+            d = obj.model_dump(mode="python")
+            return d if isinstance(d, dict) else None
+        except Exception:
+            return None
+    return None
+
+
+def parse_anthropic_stream_events(
+    events: Iterable[Any],
+    *,
+    on_token: Optional[Callable[[str], None]] = None,
+) -> tuple[str, list[LLMToolCall]]:
+    parts: list[str] = []
+    tool_by_id: dict[str, dict[str, Any]] = {}
+    current_block: dict[str, Any] | None = None
+
+    for ev in events:
+        d = _as_dict(ev) or {}
+        typ = str(d.get("type") or "")
+        if typ in ("message_start", "message_delta", "message_stop"):
+            continue
+        if typ == "content_block_start":
+            cb = d.get("content_block") if isinstance(d.get("content_block"), dict) else {}
+            current_block = cb
+            # tool_use blocks carry name/id/input in start
+            if str(cb.get("type") or "") == "tool_use":
+                tid = str(cb.get("id") or "") or f"call_{uuid.uuid4().hex}"
+                tool_by_id[tid] = {
+                    "id": tid,
+                    "name": str(cb.get("name") or ""),
+                    "input_json": cb.get("input") if isinstance(cb.get("input"), dict) else {},
+                    "_input_buf": "",
+                }
+            continue
+        if typ == "content_block_delta":
+            delta = d.get("delta") if isinstance(d.get("delta"), dict) else {}
+            dt = str(delta.get("type") or "")
+            if dt == "text_delta":
+                t = str(delta.get("text") or "")
+                if t:
+                    parts.append(t)
+                    if on_token:
+                        on_token(t)
+                continue
+            if dt in ("input_json_delta", "tool_use_delta"):
+                partial = delta.get("partial_json") or delta.get("partial") or ""
+                s = str(partial or "")
+                # attach to last tool_use block if any
+                # Anthropics uses content block index; we just buffer by most recent tool_use id.
+                if current_block and str(current_block.get("type") or "") == "tool_use":
+                    tid = str(current_block.get("id") or "")
+                    if tid and tid in tool_by_id:
+                        tool_by_id[tid]["_input_buf"] = str(tool_by_id[tid].get("_input_buf") or "") + s
+                continue
+            continue
+        if typ == "content_block_stop":
+            current_block = None
+            continue
+
+    tool_calls: list[LLMToolCall] = []
+    for tid, slot in tool_by_id.items():
+        args = slot.get("input_json") if isinstance(slot.get("input_json"), dict) else {}
+        buf = str(slot.get("_input_buf") or "").strip()
+        if buf:
+            try:
+                parsed = json.loads(buf)
+                if isinstance(parsed, dict):
+                    args = parsed
+                else:
+                    args = {"_raw": parsed}
+            except Exception:
+                # keep best-effort
+                pass
+        tool_calls.append(LLMToolCall(id=tid, name=str(slot.get("name") or ""), arguments=args, thought_signature=None))
+
+    return "".join(parts), tool_calls
+
+
+class AnthropicMessagesModel(ChatModel):
+    def __init__(self, *, model: str | None = None, api_key: str | None = None, base_url: str | None = None):
+        self.model = (model or os.getenv("ANTHROPIC_MODEL") or "claude-3-5-sonnet-latest").strip()
+        self.api_key = (api_key or os.getenv("ANTHROPIC_API_KEY") or "").strip()
+        self.base_url = (base_url or os.getenv("ANTHROPIC_BASE_URL") or "").strip() or None
+        if not self.api_key:
+            raise RuntimeError("未设置 ANTHROPIC_API_KEY，无法使用 Anthropic 模型")
+        try:
+            import anthropic
+        except Exception as exc:
+            raise RuntimeError("未安装 anthropic 依赖，请先 pip install -r requirements.txt") from exc
+        kw: dict[str, Any] = {"api_key": self.api_key}
+        if self.base_url:
+            kw["base_url"] = self.base_url
+        self._client = anthropic.Anthropic(**kw)
+
+    @staticmethod
+    def _to_anthropic_messages(messages: list[dict[str, Any]]) -> tuple[str, list[dict[str, Any]]]:
+        system_chunks: list[str] = []
+        out: list[dict[str, Any]] = []
+        for m in messages or []:
+            if not isinstance(m, dict):
+                continue
+            role = str(m.get("role") or "")
+            content = m.get("content")
+            if role == "system":
+                if isinstance(content, str) and content.strip():
+                    system_chunks.append(content)
+                continue
+            if role in ("user", "assistant"):
+                if isinstance(content, str):
+                    out.append({"role": role, "content": content})
+                else:
+                    out.append({"role": role, "content": str(content)})
+                continue
+            if role == "tool":
+                # Anthropic expects tool results as user content blocks; we stringify.
+                name = str(m.get("name") or m.get("tool_name") or m.get("toolName") or "tool").strip()
+                out.append({"role": "user", "content": f"[tool_result:{name}] {str(content)}"})
+                continue
+        return "\n\n".join(system_chunks).strip(), out
+
+    @staticmethod
+    def _to_anthropic_tools(tools: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        out: list[dict[str, Any]] = []
+        for t in tools or []:
+            if not isinstance(t, dict) or str(t.get("type") or "") != "function":
+                continue
+            fn = t.get("function")
+            if not isinstance(fn, dict):
+                continue
+            name = str(fn.get("name") or "").strip()
+            if not name:
+                continue
+            out.append(
+                {
+                    "name": name,
+                    "description": str(fn.get("description") or ""),
+                    "input_schema": fn.get("parameters") if isinstance(fn.get("parameters"), dict) else {"type": "object"},
+                }
+            )
+        return out
+
+    def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        *,
+        on_token: Optional[Callable[[str], None]] = None,
+    ) -> LLMResponse:
+        system, msgs = self._to_anthropic_messages(messages)
+        atools = self._to_anthropic_tools(tools)
+        try:
+            stream = self._client.messages.stream(
+                model=self.model,
+                system=system or None,
+                messages=msgs,
+                tools=atools or None,
+                max_tokens=2048,
+            )
+            text, tool_calls = parse_anthropic_stream_events(stream, on_token=on_token)
+            return LLMResponse(content=text, tool_calls=tool_calls)
+        except Exception as exc:
+            logger.info("anthropic stream failed; fallback to non-stream (%s)", exc)
+            resp = self._client.messages.create(
+                model=self.model,
+                system=system or None,
+                messages=msgs,
+                tools=atools or None,
+                max_tokens=2048,
+            )
+            d = _as_dict(resp) or {}
+            content = d.get("content")
+            text = ""
+            if isinstance(content, list):
+                for blk in content:
+                    if isinstance(blk, dict) and blk.get("type") == "text":
+                        text += str(blk.get("text") or "")
+            if on_token and text:
+                on_token(text)
+            return LLMResponse(content=text, tool_calls=[])
+
+
+__all__ = ["AnthropicMessagesModel", "parse_anthropic_stream_events"]
+

platform/llm/transports/base.py

@@ -0,0 +1,63 @@
+from __future__ import annotations
+
+import base64
+import json
+from dataclasses import dataclass
+from typing import Any, Optional
+from collections.abc import Callable
+
+
+@dataclass(frozen=True)
+class LLMToolCall:
+    id: str
+    name: str
+    arguments: dict[str, Any]
+    # Gemini OpenAI-compat and native transports may require preserving a signature across tool loops.
+    thought_signature: str | None = None
+
+
+@dataclass(frozen=True)
+class LLMResponse:
+    content: str
+    tool_calls: list[LLMToolCall]
+    reasoning_content: str = ""
+
+
+class ChatModel:
+    def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        *,
+        on_token: Optional[Callable[[str], None]] = None,
+    ) -> LLMResponse:
+        raise NotImplementedError
+
+
+def coerce_thought_signature_for_storage(v: Any) -> str | None:
+    if v is None:
+        return None
+    if isinstance(v, str):
+        return v
+    if isinstance(v, (bytes, bytearray)):
+        try:
+            return bytes(v).decode("utf-8")
+        except UnicodeDecodeError:
+            return base64.b64encode(bytes(v)).decode("ascii")
+    if isinstance(v, (dict, list)):
+        return json.dumps(v, separators=(",", ":"), ensure_ascii=False)
+    return str(v)
+
+
+def normalize_image_b64_payload(raw: Any) -> str:
+    if raw is None:
+        return ""
+    if isinstance(raw, (bytes, bytearray)):
+        return base64.b64encode(bytes(raw)).decode("ascii")
+    s = str(raw).strip()
+    if not s:
+        return ""
+    if s.startswith("data:") and ";base64," in s:
+        s = s.split(";base64,", 1)[-1].strip()
+    return "".join(s.split())
+

platform/llm/transports/google_gemini_sse.py

@@ -0,0 +1,265 @@
+from __future__ import annotations
+
+import json
+import os
+import uuid
+from typing import Any, Optional
+from collections.abc import Callable
+
+from oclaw.platform.llm.transports.base import ChatModel, LLMResponse, LLMToolCall
+
+
+class GoogleGeminiChatModel(ChatModel):
+    """Native Google Gemini streaming transport (SSE), modeled after OpenClaw's google transport."""
+
+    def __init__(self, *, model: str, api_key: str, base_url: str | None = None):
+        self.model = (model or "").strip() or "gemini-2.5-pro"
+        self.api_key = (api_key or "").strip()
+        self.base_url = (base_url or "").strip() or "https://generativelanguage.googleapis.com/v1beta"
+        if not self.api_key:
+            raise RuntimeError("未设置 GOOGLE_API_KEY，无法使用 Gemini 原生接口")
+
+    def _tool_decls_from_openai_tools(self, tools: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        decls: list[dict[str, Any]] = []
+        for t in tools or []:
+            if not isinstance(t, dict) or str(t.get("type") or "") != "function":
+                continue
+            fn = t.get("function")
+            if not isinstance(fn, dict):
+                continue
+            name = str(fn.get("name") or "").strip()
+            if not name:
+                continue
+            decls.append(
+                {
+                    "name": name,
+                    "description": str(fn.get("description") or ""),
+                    # OpenClaw uses `parametersJsonSchema`
+                    "parametersJsonSchema": fn.get("parameters") if isinstance(fn.get("parameters"), dict) else {"type": "object"},
+                }
+            )
+        return decls
+
+    @staticmethod
+    def _thinking_config_from_env() -> dict[str, Any] | None:
+        enabled = str(os.getenv("AIA_GEMINI_THINKING") or "").strip().lower()
+        if enabled in ("0", "false", "no", "off"):
+            return {"includeThoughts": False}
+        if enabled in ("1", "true", "yes", "on"):
+            cfg: dict[str, Any] = {"includeThoughts": True}
+            lvl = str(os.getenv("AIA_GEMINI_THINKING_LEVEL") or "").strip()
+            if lvl:
+                cfg["thinkingLevel"] = lvl
+            bud = str(os.getenv("AIA_GEMINI_THINKING_BUDGET") or "").strip()
+            if bud.isdigit():
+                cfg["thinkingBudget"] = int(bud)
+            return cfg
+        return None
+
+    @staticmethod
+    def _parse_tool_result_payload(raw: Any) -> Any:
+        if raw is None:
+            return {}
+        if isinstance(raw, (dict, list)):
+            return raw
+        s = str(raw or "").strip()
+        if not s:
+            return {}
+        try:
+            return json.loads(s)
+        except Exception:
+            return {"raw": s}
+
+    @staticmethod
+    def _openai_messages_to_gemini_contents(messages: list[dict[str, Any]]) -> tuple[dict[str, Any] | None, list[dict[str, Any]]]:
+        system_parts: list[dict[str, Any]] = []
+        contents: list[dict[str, Any]] = []
+
+        def _append(role: str, parts: list[dict[str, Any]]) -> None:
+            if not parts:
+                return
+            contents.append({"role": role, "parts": parts})
+
+        for m in messages or []:
+            if not isinstance(m, dict):
+                continue
+            role = str(m.get("role") or "")
+            content = m.get("content")
+            if role == "system":
+                if isinstance(content, str) and content.strip():
+                    system_parts.append({"text": content})
+                continue
+            if role == "user":
+                parts: list[dict[str, Any]] = []
+                if isinstance(content, str):
+                    if content.strip():
+                        parts.append({"text": content})
+                elif isinstance(content, list):
+                    for item in content:
+                        if not isinstance(item, dict):
+                            continue
+                        if item.get("type") == "text" and str(item.get("text") or "").strip():
+                            parts.append({"text": str(item.get("text") or "")})
+                _append("user", parts)
+                continue
+            if role == "assistant":
+                parts: list[dict[str, Any]] = []
+                if isinstance(content, str) and content.strip():
+                    parts.append({"text": content})
+                _append("model", parts)
+                continue
+            if role == "tool":
+                name = str(m.get("name") or "").strip() or str(m.get("tool_name") or m.get("toolName") or "").strip()
+                payload = GoogleGeminiChatModel._parse_tool_result_payload(content)
+                _append("user", [{"functionResponse": {"name": name or "tool", "response": payload}}])
+                continue
+
+        system_instruction = {"parts": system_parts} if system_parts else None
+        return system_instruction, contents
+
+    def _build_url(self) -> str:
+        base = self._normalize_base_url_for_native_sse(self.base_url).rstrip("/")
+        model_path = self.model
+        if not model_path.startswith("models/") and not model_path.startswith("tunedModels/"):
+            model_path = f"models/{model_path}"
+        # Prefer header auth (OpenClaw-style); avoid forcing key into query string.
+        return f"{base}/{model_path}:streamGenerateContent?alt=sse"
+
+    @staticmethod
+    def _normalize_base_url_for_native_sse(base_url: str) -> str:
+        """Normalize unified gateway base_url for Gemini native SSE.
+
+        Many unified gateways expose OpenAI-compat under `/v1/*`, but Gemini native endpoints
+        are usually implemented under `/v1beta/*` semantics.
+        """
+        raw = str(base_url or "").strip()
+        if not raw:
+            return raw
+        # Strip trailing slashes first.
+        while raw.endswith("/"):
+            raw = raw[:-1]
+        # Common unified gateway layout: https://host/v1  -> https://host/v1beta
+        if raw.lower().endswith("/v1"):
+            return raw[:-3] + "/v1beta"
+        # Some gateways keep an `/openai` suffix for compat endpoints.
+        if raw.lower().endswith("/openai"):
+            raw = raw[:-7]
+            if raw.lower().endswith("/v1"):
+                return raw[:-3] + "/v1beta"
+            return raw
+        return raw
+
+    def _auth_headers(self) -> dict[str, str]:
+        k = str(self.api_key or "").strip()
+        if not k:
+            return {}
+        # If user stored a bearer token in the profile secret.
+        if k.lower().startswith("bearer "):
+            return {"authorization": k}
+        # Gemini API supports x-goog-api-key; unified gateways often accept this too.
+        return {"x-goog-api-key": k}
+
+    def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        *,
+        on_token: Optional[Callable[[str], None]] = None,
+    ) -> LLMResponse:
+        system_instruction, contents = self._openai_messages_to_gemini_contents(messages)
+        decls = self._tool_decls_from_openai_tools(tools)
+        req: dict[str, Any] = {"contents": contents}
+        if system_instruction is not None:
+            req["systemInstruction"] = system_instruction
+        if decls:
+            req["tools"] = [{"functionDeclarations": decls}]
+        thinking_cfg = self._thinking_config_from_env()
+        if thinking_cfg is not None:
+            req["generationConfig"] = {"thinkingConfig": thinking_cfg}
+
+        url = self._build_url()
+        try:
+            import httpx
+        except Exception as exc:
+            raise RuntimeError("missing dependency: httpx") from exc
+
+        thinking_parts: list[str] = []
+        thinking_signature: str | None = None
+        out_text: list[str] = []
+        tool_calls: list[LLMToolCall] = []
+
+        def _stream_once(stream_url: str) -> None:
+            with httpx.Client(timeout=httpx.Timeout(60.0, connect=10.0)) as client:
+                with client.stream(
+                    "POST",
+                    stream_url,
+                    json=req,
+                    headers={
+                        "Content-Type": "application/json",
+                        "accept": "text/event-stream",
+                        **self._auth_headers(),
+                    },
+                ) as r:
+                    r.raise_for_status()
+                    for line in r.iter_lines():
+                        if not line or not str(line).startswith("data:"):
+                            continue
+                        data = str(line)[len("data:") :].strip()
+                        if not data or data == "[DONE]":
+                            continue
+                        try:
+                            chunk = json.loads(data)
+                        except Exception:
+                            continue
+                        cands = chunk.get("candidates") if isinstance(chunk, dict) else None
+                        if not isinstance(cands, list) or not cands:
+                            continue
+                        c0 = cands[0] if isinstance(cands[0], dict) else {}
+                        content0 = c0.get("content") if isinstance(c0.get("content"), dict) else {}
+                        parts = content0.get("parts") if isinstance(content0.get("parts"), list) else []
+                        for p in parts:
+                            if not isinstance(p, dict):
+                                continue
+                            if p.get("thought") is True:
+                                txt = str(p.get("text") or "")
+                                if txt:
+                                    thinking_parts.append(txt)
+                                sig = p.get("thoughtSignature")
+                                if isinstance(sig, str) and sig:
+                                    thinking_signature = sig
+                                continue
+                            fc = p.get("functionCall") if isinstance(p.get("functionCall"), dict) else None
+                            if fc is not None:
+                                name = str(fc.get("name") or "").strip()
+                                args = fc.get("args") if isinstance(fc.get("args"), dict) else {}
+                                tid = str(fc.get("id") or "") or f"call_{uuid.uuid4().hex}"
+                                tool_calls.append(
+                                    LLMToolCall(id=tid, name=name, arguments=args, thought_signature=thinking_signature)
+                                )
+                                continue
+                            txt = str(p.get("text") or "")
+                            if txt:
+                                out_text.append(txt)
+                                if on_token:
+                                    on_token(txt)
+
+        try:
+            _stream_once(url)
+        except Exception as exc:
+            # If gateway returns 500 on /v1* SSE path, try /v1beta* fallback once.
+            msg = str(exc or "").lower()
+            if "500" in msg and "/v1/" in str(url).lower():
+                alt = str(url).replace("/v1/", "/v1beta/")
+                _stream_once(alt)
+            else:
+                raise
+
+        return LLMResponse(
+            content="".join(out_text),
+            tool_calls=tool_calls,
+            reasoning_content="".join(thinking_parts).strip(),
+        )
+
+
+__all__ = ["GoogleGeminiChatModel"]
+

platform/llm/transports/openai_chat_completions.py

@@ -0,0 +1,309 @@
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import uuid
+from typing import Any, Optional
+from collections.abc import Callable
+
+from oclaw.platform.llm.transports.base import ChatModel, LLMResponse, LLMToolCall, normalize_image_b64_payload, coerce_thought_signature_for_storage
+
+logger = logging.getLogger(__name__)
+
+
+def _likely_gemini_openai_compat_base_url(url: str | None) -> bool:
+    if not url:
+        return False
+    u = url.lower()
+    return "generativelanguage.googleapis.com" in u or "aiplatform.googleapis.com" in u
+
+
+def _model_id_suggests_gemini(model: str | None) -> bool:
+    return "gemini" in (model or "").lower()
+
+
+def _find_thought_signature_in_obj(o: Any) -> str | None:
+    if isinstance(o, dict):
+        for k in ("thought_signature", "thoughtSignature"):
+            v = o.get(k)
+            if v is not None:
+                coerced = coerce_thought_signature_for_storage(v)
+                if coerced is not None:
+                    return coerced
+        for v in o.values():
+            found = _find_thought_signature_in_obj(v)
+            if found is not None:
+                return found
+    elif isinstance(o, list):
+        for v in o:
+            found = _find_thought_signature_in_obj(v)
+            if found is not None:
+                return found
+    return None
+
+
+def _extract_thought_signature_from_message_tool_call(tc: Any) -> str | None:
+    if hasattr(tc, "model_dump"):
+        try:
+            d = tc.model_dump(mode="python")
+            ec = d.get("extra_content")
+            if isinstance(ec, dict):
+                g = ec.get("google")
+                if isinstance(g, dict) and "thought_signature" in g:
+                    return coerce_thought_signature_for_storage(g.get("thought_signature"))
+        except Exception:
+            pass
+    ec = getattr(tc, "extra_content", None)
+    if isinstance(ec, dict):
+        g = ec.get("google")
+        if isinstance(g, dict) and "thought_signature" in g:
+            return coerce_thought_signature_for_storage(g.get("thought_signature"))
+    return _extract_thought_signature_from_tool_delta(tc)
+
+
+def _extract_thought_signature_from_tool_delta(tc: Any) -> str | None:
+    for attr in ("thought_signature", "thoughtSignature"):
+        v = getattr(tc, attr, None)
+        if v is not None:
+            return coerce_thought_signature_for_storage(v)
+    extra = getattr(tc, "model_extra", None)
+    if isinstance(extra, dict):
+        for k in ("thought_signature", "thoughtSignature"):
+            ev = extra.get(k)
+            if ev is not None:
+                return coerce_thought_signature_for_storage(ev)
+    pyd_ex = getattr(tc, "__pydantic_extra__", None)
+    if isinstance(pyd_ex, dict):
+        for k in ("thought_signature", "thoughtSignature"):
+            ev = pyd_ex.get(k)
+            if ev is not None:
+                return coerce_thought_signature_for_storage(ev)
+    if hasattr(tc, "model_dump"):
+        try:
+            found = _find_thought_signature_in_obj(tc.model_dump(mode="python"))
+            if found is not None:
+                return found
+        except Exception:
+            pass
+    return None
+
+
+class OpenAIChatModel(ChatModel):
+    def __init__(
+        self,
+        model: str | None = None,
+        api_key: str | None = None,
+        base_url: str | None = None,
+    ):
+        self.model = model or os.getenv("OPENAI_MODEL") or "gpt-4o-mini"
+        self.api_key = api_key or os.getenv("OPENAI_API_KEY")
+        self.base_url = base_url or os.getenv("OPENAI_BASE_URL")
+
+        if not self.api_key:
+            raise RuntimeError("未设置 OPENAI_API_KEY，无法使用 OpenAI 模型")
+
+        try:
+            from openai import OpenAI
+        except Exception as e:
+            raise RuntimeError("未安装 openai 依赖，请先 pip install -r requirements.txt") from e
+
+        client_kwargs: dict[str, Any] = {"api_key": self.api_key}
+        if self.base_url:
+            client_kwargs["base_url"] = self.base_url
+        self._client = OpenAI(**client_kwargs)
+        self._skip_tools: bool = False
+
+    def _create_chat_completion(self, norm_msgs: list[dict[str, Any]], tools: list[dict[str, Any]], *, stream: bool):
+        # replay_policy/tool_wire_policy live in sibling modules; import lazily to avoid cycles.
+        try:
+            from oclaw.platform.llm.replay_policy import apply_replay_policy_to_messages, resolve_replay_policy
+
+            policy = resolve_replay_policy(self.base_url, self.model)
+            if policy.enabled:
+                norm_msgs = apply_replay_policy_to_messages(norm_msgs, policy)
+        except Exception as exc:
+            logger.warning("replay_policy apply failed (%s); continuing without rewrite", exc)
+
+        use_tools = bool(tools)
+        cleaned_msgs: list[dict[str, Any]] = []
+        for m in norm_msgs or []:
+            if not isinstance(m, dict):
+                continue
+            if str(m.get("role") or "") == "tool":
+                mm = dict(m)
+                for k in ("tool_call_id", "call_id"):
+                    if k in mm:
+                        v = mm.get(k)
+                        try:
+                            sv = str(v).strip() if v is not None else ""
+                        except Exception:
+                            sv = ""
+                        if not sv:
+                            mm.pop(k, None)
+                        else:
+                            mm[k] = sv
+                cleaned_msgs.append(mm)
+            else:
+                cleaned_msgs.append(m)
+
+        kwargs: dict[str, Any] = {"model": self.model, "messages": cleaned_msgs, "stream": stream}
+        if use_tools:
+            try:
+                from oclaw.platform.config.paths import db_path
+                from oclaw.platform.persistence.sqlite_store import SqliteStore
+                from oclaw.tools.exposure_plan import build_llm_tools_plan
+
+                plan = build_llm_tools_plan(
+                    store=SqliteStore(db_path()),
+                    role="",
+                    base_url=self.base_url,
+                    max_json_bytes=_default_max_openai_tools_json_bytes(self.base_url),
+                    include_mcp=False,
+                    preview_internal=False,
+                    raw_openai_tools_override=tools,
+                )
+                kwargs["tools"] = plan.tools_wired
+            except Exception:
+                kwargs["tools"] = tools
+        return self._client.chat.completions.create(**kwargs)
+
+    def _llm_response_from_completion(self, completion: Any, *, on_token: Optional[Callable[[str], None]]) -> LLMResponse:
+        msg = completion.choices[0].message
+        reasoning_parts = getattr(msg, "reasoning_content", None) or ""
+        reasoning_text = str(reasoning_parts).strip() if reasoning_parts else ""
+        content = msg.content or ""
+        if on_token and content:
+            on_token(content)
+
+        tool_calls: list[LLMToolCall] = []
+        if msg.tool_calls:
+            for tc in msg.tool_calls:
+                tid = str(getattr(tc, "id", "") or "")
+                fn = getattr(tc, "function", None)
+                name = str(fn.name) if fn and getattr(fn, "name", None) else ""
+                args_raw = str(fn.arguments) if fn and getattr(fn, "arguments", None) is not None else "{}"
+                if not name and not tid:
+                    continue
+                if not tid:
+                    tid = f"call_{uuid.uuid4().hex}"
+                try:
+                    args = json.loads(args_raw)
+                except json.JSONDecodeError:
+                    args = {"_raw": args_raw}
+                sig = _extract_thought_signature_from_message_tool_call(tc)
+                if sig is None and fn is not None:
+                    sig = _extract_thought_signature_from_tool_delta(fn)
+                tool_calls.append(LLMToolCall(id=tid, name=name, arguments=args, thought_signature=sig))
+        return LLMResponse(content=content, tool_calls=tool_calls, reasoning_content=reasoning_text)
+
+    def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        *,
+        on_token: Optional[Callable[[str], None]] = None,
+    ) -> LLMResponse:
+        # Normalize input_image to image_url.
+        norm_msgs = []
+        for m in messages:
+            if m.get("role") == "user" and isinstance(m.get("content"), list):
+                norm_content = []
+                for item in m["content"]:
+                    if isinstance(item, dict) and item.get("type") == "input_image":
+                        mime = str(item.get("mime") or "image/jpeg").strip() or "image/jpeg"
+                        b64 = normalize_image_b64_payload(item.get("image_base64") or item.get("data"))
+                        if not b64:
+                            continue
+                        norm_content.append({"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}})
+                    else:
+                        norm_content.append(item)
+                norm_msgs.append({"role": m["role"], "content": norm_content})
+            else:
+                norm_msgs.append(m)
+
+        try:
+            stream = self._create_chat_completion(norm_msgs, tools, stream=True)
+        except Exception as exc:
+            logger.info("Streaming chat completion failed; fallback to non-stream (%s)", exc)
+            completion = self._create_chat_completion(norm_msgs, tools, stream=False)
+            return self._llm_response_from_completion(completion, on_token=on_token)
+
+        content_parts: list[str] = []
+        tool_acc: dict[int, dict[str, Any]] = {}
+        reasoning_parts: list[str] = []
+
+        def _consume_stream(stream_obj: Any) -> None:
+            nonlocal content_parts, tool_acc, reasoning_parts
+            for chunk in stream_obj:
+                if not chunk.choices:
+                    continue
+                delta = chunk.choices[0].delta
+                rc = getattr(delta, "reasoning_content", None) or ""
+                if rc:
+                    reasoning_parts.append(rc)
+                if delta.tool_calls:
+                    for tc in delta.tool_calls:
+                        idx = int(tc.index)
+                        slot = tool_acc.setdefault(idx, {"id": None, "name": None, "arguments": "", "thought_signature": None})
+                        if tc.id:
+                            slot["id"] = tc.id
+                        if tc.function:
+                            if tc.function.name:
+                                slot["name"] = tc.function.name
+                            if tc.function.arguments:
+                                slot["arguments"] = f"{slot['arguments']}{tc.function.arguments}"
+                        sig = _extract_thought_signature_from_message_tool_call(tc)
+                        if sig is None:
+                            sig = _extract_thought_signature_from_tool_delta(tc)
+                        if sig is None and tc.function is not None:
+                            sig = _extract_thought_signature_from_tool_delta(tc.function)
+                        if sig is not None:
+                            slot["thought_signature"] = sig
+                if delta.content:
+                    content_parts.append(delta.content)
+                    if on_token:
+                        on_token(delta.content)
+
+        try:
+            _consume_stream(stream)
+        except Exception:
+            raise
+
+        content = "".join(content_parts)
+        reasoning_text = "".join(reasoning_parts).strip()
+
+        tool_calls: list[LLMToolCall] = []
+        for idx in sorted(tool_acc.keys()):
+            slot = tool_acc[idx]
+            tid = slot.get("id") or ""
+            name = slot.get("name") or ""
+            args_raw = slot.get("arguments") or "{}"
+            if not name and not tid:
+                continue
+            if not tid:
+                tid = f"call_{uuid.uuid4().hex}"
+            try:
+                args = json.loads(args_raw)
+            except json.JSONDecodeError:
+                args = {"_raw": args_raw}
+            ts_raw = slot.get("thought_signature")
+            tsig = coerce_thought_signature_for_storage(ts_raw) if ts_raw is not None else None
+            tool_calls.append(LLMToolCall(id=str(tid), name=str(name), arguments=args, thought_signature=tsig))
+
+        return LLMResponse(content=content, tool_calls=tool_calls, reasoning_content=reasoning_text)
+
+
+__all__ = ["OpenAIChatModel", "_likely_gemini_openai_compat_base_url", "_model_id_suggests_gemini"]
+
+
+def _default_max_openai_tools_json_bytes(base_url: str | None) -> int | None:
+    u = str(base_url or "").strip().lower()
+    if not u:
+        return None
+    # Conservative caps for some OpenAI-compatible gateways.
+    if "dashscope" in u or "aliyuncs" in u:
+        return 28_000
+    return None
+

platform/llm/transports/openai_responses.py

@@ -0,0 +1,236 @@
+from __future__ import annotations
+
+import json
+import logging
+import os
+import uuid
+from typing import Any, Optional
+from collections.abc import Callable, Iterable
+
+from oclaw.platform.llm.transports.base import ChatModel, LLMResponse, LLMToolCall, normalize_image_b64_payload
+
+logger = logging.getLogger(__name__)
+
+
+def _as_dict(obj: Any) -> dict[str, Any] | None:
+    if obj is None:
+        return None
+    if isinstance(obj, dict):
+        return obj
+    if hasattr(obj, "model_dump"):
+        try:
+            d = obj.model_dump(mode="python")
+            return d if isinstance(d, dict) else None
+        except Exception:
+            return None
+    return None
+
+
+def _collect_tool_calls_from_response_dict(resp: dict[str, Any]) -> list[LLMToolCall]:
+    out: list[LLMToolCall] = []
+    items = resp.get("output")
+    if not isinstance(items, list):
+        return out
+    for it in items:
+        d = _as_dict(it)
+        if not d:
+            continue
+        if str(d.get("type") or "") in ("function_call", "tool_call"):
+            call_id = str(d.get("call_id") or d.get("id") or "") or f"call_{uuid.uuid4().hex}"
+            name = str(d.get("name") or "") or str((d.get("function") or {}).get("name") or "")
+            args = d.get("arguments") or d.get("input") or {}
+            if isinstance(args, str):
+                try:
+                    args = json.loads(args)
+                except Exception:
+                    args = {"_raw": args}
+            if not isinstance(args, dict):
+                args = {"_raw": args}
+            out.append(LLMToolCall(id=call_id, name=name, arguments=args, thought_signature=None))
+    return out
+
+
+def parse_openai_responses_stream_events(
+    events: Iterable[Any],
+    *,
+    on_token: Optional[Callable[[str], None]] = None,
+) -> tuple[str, list[LLMToolCall], dict[str, Any] | None]:
+    """Pure stream parser (offline-testable).
+
+    Returns: (assembled_text, tool_calls, final_response_dict?)
+    """
+    parts: list[str] = []
+    final_resp: dict[str, Any] | None = None
+    for ev in events:
+        d = _as_dict(ev) or {}
+        typ = str(d.get("type") or "")
+        # Text deltas
+        if typ in ("response.output_text.delta", "response.output_text"):
+            delta = d.get("delta")
+            if delta is None:
+                delta = d.get("text")
+            s = str(delta or "")
+            if s:
+                parts.append(s)
+                if on_token:
+                    on_token(s)
+            continue
+        # Sometimes the SDK emits generic `response.delta` with nested segments.
+        if typ == "response.delta":
+            delta = d.get("delta")
+            if isinstance(delta, dict):
+                txt = delta.get("output_text")
+                if isinstance(txt, str) and txt:
+                    parts.append(txt)
+                    if on_token:
+                        on_token(txt)
+            continue
+        # Terminal response object
+        if typ in ("response.completed", "response.complete", "response.done"):
+            resp = d.get("response") or d.get("data") or d.get("result")
+            final_resp = _as_dict(resp) or final_resp
+            continue
+        if typ == "response.created":
+            continue
+        # Some SDK versions yield the response object directly (no explicit event type).
+        if d.get("output") is not None and d.get("id") is not None:
+            final_resp = d
+
+    text = "".join(parts)
+    tool_calls = _collect_tool_calls_from_response_dict(final_resp) if final_resp else []
+    return text, tool_calls, final_resp
+
+
+class OpenAIResponsesModel(ChatModel):
+    """OpenAI Responses API transport (OpenAI-compatible gateways may implement this surface)."""
+
+    def __init__(self, *, model: str | None = None, api_key: str | None = None, base_url: str | None = None):
+        self.model = (model or os.getenv("OPENAI_MODEL") or "gpt-4o-mini").strip()
+        self.api_key = (api_key or os.getenv("OPENAI_API_KEY") or "").strip()
+        self.base_url = (base_url or os.getenv("OPENAI_BASE_URL") or "").strip() or None
+        if not self.api_key:
+            raise RuntimeError("未设置 OPENAI_API_KEY，无法使用 OpenAI Responses")
+        try:
+            from openai import OpenAI
+        except Exception as e:
+            raise RuntimeError("未安装 openai 依赖，请先 pip install -r requirements.txt") from e
+        kw: dict[str, Any] = {"api_key": self.api_key}
+        if self.base_url:
+            kw["base_url"] = self.base_url
+        self._client = OpenAI(**kw)
+
+    @staticmethod
+    def _normalize_messages(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        """
+        Normalize to the strictest OpenAI-compatible `responses` shape:
+        - input.messages[*].role MUST be "user"
+        - input.messages[*].content MUST be a list (content blocks)
+        - text blocks use {"type":"text","text":...}
+        - image blocks use {"type":"image_url","image_url":{"url":"..."}} or {"type":"input_image","image_url":"..."} depending on gateway;
+          we prefer the common "image_url" block here.
+        """
+        out: list[dict[str, Any]] = []
+        for m in messages or []:
+            if not isinstance(m, dict):
+                continue
+            role = str(m.get("role") or "user").strip().lower() or "user"
+            content = m.get("content")
+            norm_content: list[dict[str, Any]] = []
+            if isinstance(content, list):
+                for item in content:
+                    if isinstance(item, dict) and item.get("type") == "input_image":
+                        mime = str(item.get("mime") or "image/jpeg").strip() or "image/jpeg"
+                        b64 = normalize_image_b64_payload(item.get("image_base64") or item.get("data"))
+                        if not b64:
+                            continue
+                        norm_content.append({"type": "image_url", "image_url": {"url": f"data:{mime};base64,{b64}"}})
+                        continue
+                    if isinstance(item, dict) and item.get("type") in ("input_text", "text"):
+                        t = str(item.get("text") or "").strip()
+                        if t:
+                            norm_content.append({"type": "text", "text": t})
+                        continue
+                    if isinstance(item, dict):
+                        # Best-effort coerce unknown blocks into valid text blocks.
+                        s = str(item.get("text") or "").strip()
+                        if s:
+                            norm_content.append({"type": "text", "text": s})
+                        continue
+                    s = str(item or "").strip()
+                    if s:
+                        norm_content.append({"type": "text", "text": s})
+            elif isinstance(content, str):
+                txt = content.strip()
+                if txt:
+                    norm_content.append({"type": "text", "text": txt})
+            elif content is not None:
+                s = str(content).strip()
+                if s:
+                    norm_content.append({"type": "text", "text": s})
+            if not norm_content:
+                # Keep shape valid even for empty messages.
+                continue
+
+            # Enforce role=user and preserve other roles via prefix tag.
+            if role != "user":
+                prefix = "assistant" if role == "assistant" else ("system" if role == "system" else role)
+                norm_content.insert(0, {"type": "text", "text": f"[{prefix}]"})
+            out.append({"role": "user", "content": norm_content})
+        return out
+
+    def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        *,
+        on_token: Optional[Callable[[str], None]] = None,
+    ) -> LLMResponse:
+        norm = self._normalize_messages(messages)
+        # OpenAI-compatible gateways differ: some require input={"messages":[...]} with role=user only.
+        stream_errors: list[str] = []
+        stream_variants: list[dict[str, Any]] = [
+            {"model": self.model, "input": {"messages": norm}, "tools": tools or None, "stream": True},
+            {"model": self.model, "input": norm, "tools": tools or None, "stream": True},
+        ]
+        try:
+            for payload in stream_variants:
+                try:
+                    stream = self._client.responses.create(**payload)
+                    text, tool_calls, final_resp = parse_openai_responses_stream_events(stream, on_token=on_token)
+                    if (not text.strip()) and final_resp:
+                        ot = final_resp.get("output_text")
+                        if isinstance(ot, str) and ot.strip():
+                            text = ot
+                            if on_token:
+                                on_token(text)
+                    return LLMResponse(content=text, tool_calls=tool_calls)
+                except Exception as exc:
+                    stream_errors.append(str(exc))
+                    continue
+            raise RuntimeError("; ".join(stream_errors) or "responses_stream_all_variants_failed")
+        except Exception as exc:
+            logger.info("responses stream failed; fallback to non-stream (%s)", exc)
+            nonstream_errors: list[str] = []
+            for payload in (
+                {"model": self.model, "input": {"messages": norm}, "tools": tools or None},
+                {"model": self.model, "input": norm, "tools": tools or None},
+            ):
+                try:
+                    resp = self._client.responses.create(**payload)
+                    d = _as_dict(resp) or {}
+                    text = str(d.get("output_text") or "")
+                    tool_calls = _collect_tool_calls_from_response_dict(d)
+                    if on_token and text:
+                        on_token(text)
+                    return LLMResponse(content=text, tool_calls=tool_calls)
+                except Exception as e2:
+                    nonstream_errors.append(str(e2))
+                    continue
+            raise RuntimeError(
+                "openai_responses_request_failed: "
+                + " | ".join([str(exc)] + nonstream_errors[-2:])
+            ) from exc
+
+
+__all__ = ["OpenAIResponsesModel", "parse_openai_responses_stream_events"]
+

platform/llm/transports/simple.py

@@ -0,0 +1,144 @@
+from __future__ import annotations
+
+import json
+import re
+from typing import Any, Optional
+from collections.abc import Callable
+
+from oclaw.platform.llm.transports.base import ChatModel, LLMResponse, LLMToolCall
+
+
+class StaticTextChatModel(ChatModel):
+    def __init__(self, text: str):
+        self._text = text or ""
+
+    def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        *,
+        on_token: Optional[Callable[[str], None]] = None,
+    ) -> LLMResponse:
+        if on_token and self._text:
+            on_token(self._text)
+        return LLMResponse(content=self._text, tool_calls=[])
+
+
+class RuleBasedChatModel(ChatModel):
+    _IP_RE = re.compile(
+        r"(?P<ip>(?:\d{1,3}\.){3}\d{1,3})|(?P<ipv6>(?:[0-9a-fA-F]{0,4}:){2,}[0-9a-fA-F]{0,4})"
+    )
+
+    def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        *,
+        on_token: Optional[Callable[[str], None]] = None,
+    ) -> LLMResponse:
+        def make_call(name: str, arguments: dict[str, Any]) -> LLMResponse:
+            return LLMResponse(
+                content="",
+                tool_calls=[LLMToolCall(id=f"rb_{name}_1", name=name, arguments=arguments)],
+            )
+
+        def emit(resp: LLMResponse) -> LLMResponse:
+            if on_token and (resp.content or "").strip():
+                on_token(resp.content)
+            return resp
+
+        if messages:
+            last = messages[-1]
+            if last.get("role") == "tool":
+                raw = str(last.get("content") or "")
+                try:
+                    data = json.loads(raw)
+                except json.JSONDecodeError:
+                    data = {"ok": True, "raw": raw}
+
+                if isinstance(data, dict) and data.get("ok") is False:
+                    return emit(LLMResponse(content=f"工具执行失败：{data.get('error')}", tool_calls=[]))
+
+                if isinstance(data, dict) and "route" in data:
+                    r = data.get("route") or {}
+                    return emit(
+                        LLMResponse(
+                            content=f"路由查询结果：目的 {data.get('destination')} 命中 {r.get('prefix')}，下一跳 {r.get('next_hop')}，出接口 {r.get('out_if')}。",
+                            tool_calls=[],
+                        )
+                    )
+
+                if isinstance(data, dict) and "hops" in data:
+                    hops = data.get("hops") or []
+                    return emit(
+                        LLMResponse(
+                            content=f"路径计算结果：{data.get('src')} → {data.get('dst')} 共 {data.get('hop_count')} 跳，路径为：{' -> '.join(map(str, hops))}",
+                            tool_calls=[],
+                        )
+                    )
+
+                if isinstance(data, dict) and "diff" in data:
+                    diff = data.get("diff") or ""
+                    if not diff:
+                        return emit(LLMResponse(content="配置对比结果：两段配置一致。", tool_calls=[]))
+                    return emit(LLMResponse(content=f"配置对比结果：\n\n```diff\n{diff}\n```", tool_calls=[]))
+
+                if isinstance(data, dict) and "level_count" in data:
+                    lc = data.get("level_count") or {}
+                    return emit(LLMResponse(content=f"日志分析结果：统计 {lc}。建议优先查看 ERROR 样例与日志尾部定位根因。", tool_calls=[]))
+
+                if isinstance(data, dict) and "reachable" in data:
+                    reach = bool(data.get("reachable"))
+                    avg = data.get("avg_ms")
+                    if reach:
+                        tail = f"平均时延 {avg}ms" if isinstance(avg, int) else "时延解析失败"
+                        return emit(LLMResponse(content=f"设备状态：可达，{tail}。", tool_calls=[]))
+                    return emit(LLMResponse(content="设备状态：不可达。建议检查 ACL/路由/DNS/防火墙策略，并确认目标在线。", tool_calls=[]))
+
+                return emit(LLMResponse(content=f"工具执行结果：{json.dumps(data, ensure_ascii=False)}", tool_calls=[]))
+
+        last_user = ""
+        for m in reversed(messages):
+            if m.get("role") == "user":
+                content = m.get("content")
+                if isinstance(content, list):
+                    texts = [item["text"] for item in content if isinstance(item, dict) and item.get("type") == "text"]
+                    last_user = "\n".join(texts)
+                else:
+                    last_user = str(content or "")
+                break
+
+        m = self._IP_RE.search(last_user)
+        if "路由" in last_user and m:
+            ip = m.group("ip") or m.group("ipv6")
+            return make_call("query_route", {"destination": ip})
+
+        if ("路径" in last_user or "链路" in last_user) and ("到" in last_user or "->" in last_user):
+            m2 = re.search(r"([A-Za-z0-9_-]+)\s*(?:到|->)\s*([A-Za-z0-9_-]+)", last_user)
+            if m2:
+                src, dst = m2.group(1), m2.group(2)
+                return make_call("get_path", {"src": src, "dst": dst})
+
+        if "对比" in last_user or "diff" in last_user.lower():
+            if "left:" in last_user.lower() and "right:" in last_user.lower():
+                left = last_user.split("left:", 1)[1].split("right:", 1)[0].strip()
+                right = last_user.split("right:", 1)[1].strip()
+                return make_call("config_diff", {"left": left, "right": right})
+
+        if "日志" in last_user and ("分析" in last_user or "统计" in last_user):
+            return make_call("log_analysis", {"log": last_user})
+
+        if ("设备" in last_user or "主机" in last_user) and ("状态" in last_user or "可达" in last_user or "ping" in last_user.lower()) and m:
+            ip = m.group("ip") or m.group("ipv6")
+            return make_call("device_status", {"host": ip})
+
+        return emit(
+            LLMResponse(
+                content="我可以进行路由查询、路径计算、配置对比、设备状态检查、日志分析。请直接描述需求并提供必要输入。",
+                tool_calls=[],
+            )
+        )
+
+
+__all__ = ["StaticTextChatModel", "RuleBasedChatModel"]
+