refactor: root-package imports (svc/runtime/interfaces) and fix PYTHONPATH

- Rename platform/ to svc/ to avoid shadowing stdlib platform.
- Replace from oclaw.* with from svc/runtime/interfaces; update -m CLI paths.
- tests/conftest: prepend repo root to sys.path (no parent-folder package name).
- CI: paths and offline_eval script under repo root.
- Ops scripts: PYTHONPATH must be repo root for python -m runtime.* (fixes gateway/WhatsApp sidecar startup).
- Fix default oclaw.json path in tabular/file attachment limits; stabilize attachment test config.

Co-authored-by: Cursor <cursoragent@cursor.com>

svc/llm/MIT_LICENSE.txt

@@ -0,0 +1,27 @@
+MIT License
+
+Copyright (c) 2025 Peter Steinberger (Oclaw)
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+---
+
+The Python modules `tool_call_id.py` and `replay_policy.py` in this directory
+implement replay / tool-call-id handling inspired by Oclaw's
+`src/agents/tool-call-id.ts` and `src/agents/transcript-policy.ts`.

svc/llm/__init__.py

@@ -0,0 +1 @@
+# svc.llm package

svc/llm/chat_models.py

@@ -0,0 +1,69 @@
+from __future__ import annotations
+
+"""Compatibility re-export layer.
+
+The codebase imports `ChatModel/OpenAIChatModel/...` from this module.
+Implementations live under `src.platform.llm.transports.*`.
+"""
+
+import os
+
+from svc.llm.tool_schema import shrink_openai_tools_payload_for_api
+from svc.llm.tool_schema import openai_tools_json_byte_length as _openai_tools_json_byte_length
+from svc.llm.transports.base import ChatModel, LLMResponse, LLMToolCall, normalize_image_b64_payload
+from svc.llm.transports.google_gemini_sse import GoogleGeminiChatModel
+from svc.llm.transports.openai_chat_completions import (
+    OpenAIChatModel,
+    _likely_gemini_openai_compat_base_url,
+    _model_id_suggests_gemini,
+)
+from svc.llm.transports.simple import RuleBasedChatModel, StaticTextChatModel
+
+# Backward-compat alias used by `src/chat/agent.py`.
+_normalize_image_b64_payload = normalize_image_b64_payload
+
+
+def gemini_openai_compat_client(model: ChatModel) -> bool:
+    if not isinstance(model, OpenAIChatModel):
+        return False
+    return _likely_gemini_openai_compat_base_url(getattr(model, "base_url", None)) or _model_id_suggests_gemini(
+        getattr(model, "model", None)
+    )
+
+
+def build_default_model() -> ChatModel:
+    mode = (os.getenv("AIA_ASSISTANT_MODE") or "").strip().lower()
+    if mode == "rule":
+        return RuleBasedChatModel()
+    if mode == "ollama":
+        try:
+            from runtime.agents.factory import DEFAULT_OLLAMA_BASE_URL, DEFAULT_OLLAMA_MODEL
+
+            return OpenAIChatModel(model=DEFAULT_OLLAMA_MODEL, api_key="ollama", base_url=DEFAULT_OLLAMA_BASE_URL)
+        except Exception:
+            return RuleBasedChatModel()
+    if mode == "openai" or mode == "":
+        try:
+            return OpenAIChatModel()
+        except Exception:
+            return RuleBasedChatModel()
+    # Provider selection is normally handled via LLM profiles in `src/agents/factory.py`.
+    return RuleBasedChatModel()
+
+
+__all__ = [
+    "ChatModel",
+    "LLMResponse",
+    "LLMToolCall",
+    "OpenAIChatModel",
+    "GoogleGeminiChatModel",
+    "RuleBasedChatModel",
+    "StaticTextChatModel",
+    "build_default_model",
+    "gemini_openai_compat_client",
+    "normalize_image_b64_payload",
+    "_normalize_image_b64_payload",
+    "shrink_openai_tools_payload_for_api",
+    "_openai_tools_json_byte_length",
+]
+

svc/llm/image_http_common.py

@@ -0,0 +1,550 @@
+"""Shared HTTP helpers for OCR and legacy image lanes (internal)."""
+
+from __future__ import annotations
+
+import base64
+import io
+import json
+import os
+import re
+import time
+from typing import Any
+
+import httpx
+from PIL import Image
+
+
+def join_url(base: str, path: str) -> str:
+    b = (base or "").rstrip("/")
+    p = (path or "").lstrip("/")
+    return f"{b}/{p}"
+
+
+def dashscope_native_multimodal_url_from_compatible_base(base_url: str) -> str | None:
+    """
+    ``qwen-image`` and similar models often return an **empty** OpenAI ``chat/completions`` message on
+    ``/compatible-mode/v1``. Native HTTP uses ``/api/v1/services/aigc/multimodal-generation/generation``
+    and returns ``output.choices[].message.content[].image`` URLs.
+
+    Override: ``AIA_IMAGE_EXPERT_DASHSCOPE_NATIVE_URL`` (full URL).
+    Path suffix: ``AIA_IMAGE_EXPERT_DASHSCOPE_NATIVE_PATH`` (default multimodal-generation path).
+    """
+    explicit = (os.getenv("AIA_IMAGE_EXPERT_DASHSCOPE_NATIVE_URL") or "").strip().rstrip("/")
+    if explicit:
+        return explicit
+    b = (base_url or "").strip().rstrip("/")
+    if not b or "compatible-mode" not in b.lower():
+        return None
+    root = re.sub(r"/compatible-mode/v\d+(?:/.*)?$", "", b, flags=re.IGNORECASE).rstrip("/")
+    if not root or root == b:
+        root = re.sub(r"/compatible-mode/?$", "", b, flags=re.IGNORECASE).rstrip("/")
+    if not root:
+        return None
+    path = (os.getenv("AIA_IMAGE_EXPERT_DASHSCOPE_NATIVE_PATH") or "").strip().lstrip("/")
+    if not path:
+        path = "api/v1/services/aigc/multimodal-generation/generation"
+    return f"{root}/{path}"
+
+
+def env_ocr_lane_api_key() -> str:
+    return (os.getenv("AIA_OCR_API_KEY") or "").strip()
+
+
+def env_ocr_lane_base_url() -> str:
+    return (os.getenv("AIA_OCR_BASE_URL") or "").strip()
+
+
+def env_ocr_lane_model() -> str:
+    return (os.getenv("AIA_OCR_MODEL") or "").strip()
+
+
+def env_ocr_lane_chat_endpoint() -> str:
+    raw = (os.getenv("AIA_OCR_CHAT_ENDPOINT") or "").strip()
+    return raw or "/chat/completions"
+
+
+def env_image_expert_api_key() -> str:
+    """Bearer key for **image specialist** multimodal/gen HTTP only (never shared with OCR ``AIA_OCR_*``)."""
+
+    return (os.getenv("AIA_IMAGE_EXPERT_API_KEY") or "").strip()
+
+
+def env_image_expert_base_url() -> str:
+    return (os.getenv("AIA_IMAGE_EXPERT_BASE_URL") or "").strip()
+
+
+def env_image_expert_model() -> str:
+    return (os.getenv("AIA_IMAGE_EXPERT_MODEL") or "").strip()
+
+
+def env_image_expert_chat_endpoint() -> str:
+    raw = (os.getenv("AIA_IMAGE_EXPERT_CHAT_ENDPOINT") or "").strip()
+    return raw or "/chat/completions"
+
+
+def is_data_url(s: str) -> bool:
+    return s.startswith("data:") and ";base64," in s
+
+
+def compress_data_url_image(
+    data_url: str,
+    *,
+    max_side: int = 1600,
+    max_bytes: int = 2 * 1024 * 1024,
+    jpeg_quality: int = 82,
+) -> str:
+    if not is_data_url(data_url):
+        return data_url
+    try:
+        header, b64 = data_url.split(";base64,", 1)
+        mime = header.replace("data:", "", 1).strip() or "image/jpeg"
+        raw = base64.b64decode(b64.encode("ascii"))
+        if len(raw) <= max_bytes:
+            return data_url
+
+        with Image.open(io.BytesIO(raw)) as im:
+            im = im.convert("RGB")
+            w, h = im.size
+            longest = max(w, h)
+            if longest > max_side:
+                scale = max_side / float(longest)
+                nw, nh = max(1, int(w * scale)), max(1, int(h * scale))
+                im = im.resize((nw, nh), Image.Resampling.LANCZOS)
+
+            out = io.BytesIO()
+            im.save(out, format="JPEG", quality=max(30, min(95, int(jpeg_quality))), optimize=True)
+            blob = out.getvalue()
+            if not blob:
+                return data_url
+            b64_new = base64.b64encode(blob).decode("ascii")
+            return f"data:{'image/jpeg' if mime.startswith('image/') else mime};base64,{b64_new}"
+    except Exception:
+        return data_url
+
+
+def post_with_retry(
+    client: httpx.Client,
+    *,
+    url: str,
+    headers: dict[str, str],
+    payload: dict[str, Any],
+    retries: int = 3,
+    backoff_sec: float = 0.8,
+) -> httpx.Response:
+    last_exc: Exception | None = None
+    for attempt in range(1, max(1, retries) + 1):
+        try:
+            return client.post(url, headers=headers, json=payload)
+        except (httpx.ReadError, httpx.ConnectError, httpx.TimeoutException) as e:
+            last_exc = e
+            if attempt >= retries:
+                break
+            time.sleep(backoff_sec * (2 ** (attempt - 1)))
+    assert last_exc is not None
+    raise last_exc
+
+
+def env_image_expert_download_timeout_sec() -> float:
+    """Read timeout for DashScope/OSS signed result URLs (official samples use ~300s)."""
+    raw = (os.getenv("AIA_IMAGE_EXPERT_DOWNLOAD_TIMEOUT_SEC") or "").strip()
+    if not raw:
+        return 300.0
+    try:
+        v = float(raw)
+    except ValueError:
+        return 300.0
+    return max(15.0, min(v, 900.0))
+
+
+def download_http_url_bytes(
+    url: str,
+    *,
+    timeout_sec: float | None = None,
+    user_agent: str | None = None,
+) -> tuple[bytes, str]:
+    """Streamed GET with ``raise_for_status`` — matches DashScope OSS download guidance (long reads)."""
+    t = float(timeout_sec) if timeout_sec is not None else env_image_expert_download_timeout_sec()
+    ua = (user_agent or "").strip() or (
+        "Mozilla/5.0 (compatible; oclaw-image-expert/1.0; +https://github.com/)"
+    )
+    req_headers = {"User-Agent": ua, "Accept": "*/*"}
+    connect_cap = min(45.0, max(10.0, t / 10.0))
+    timeout = httpx.Timeout(t, connect=connect_cap)
+    with httpx.Client(timeout=timeout, follow_redirects=True) as client:
+        with client.stream("GET", url, headers=req_headers) as r:
+            r.raise_for_status()
+            ctype = str(r.headers.get("content-type") or "").split(";", 1)[0].strip()
+            parts: list[bytes] = []
+            for chunk in r.iter_bytes(chunk_size=65_536):
+                parts.append(chunk)
+    return b"".join(parts), ctype
+
+
+def dashscope_multimodal_http_ok(body: dict[str, Any]) -> tuple[bool, str]:
+    """DashScope HTTP often uses HTTP 200 with a business-level ``code`` for failures."""
+    sc = body.get("status_code")
+    if isinstance(sc, str) and sc.strip().isdigit():
+        sc = int(sc.strip())
+    if isinstance(sc, int) and sc >= 400:
+        return False, str(body.get("message") or body.get("msg") or f"status_code={sc}")
+    code = body.get("code")
+    if code is None:
+        return True, ""
+    if isinstance(code, int):
+        if code == 200:
+            return True, ""
+        msg = str(body.get("message") or body.get("msg") or "").strip()
+        return False, msg or str(code)
+    cs = str(code).strip()
+    if cs == "":
+        return True, ""
+    if cs.lower() in ("success", "ok", "200"):
+        return True, ""
+    msg = str(body.get("message") or body.get("msg") or "").strip()
+    return False, msg or cs
+
+
+def _extract_response_roots(resp_json: dict[str, Any]) -> list[dict[str, Any]]:
+    """Collect dict slices that may carry ``choices`` / ``messages`` (DashScope nests under ``output``)."""
+    roots: list[dict[str, Any]] = []
+    seen: set[int] = set()
+
+    def push(d: dict[str, Any]) -> None:
+        i = id(d)
+        if i in seen:
+            return
+        seen.add(i)
+        roots.append(d)
+
+    push(resp_json)
+    nested = resp_json.get("output")
+    if isinstance(nested, dict):
+        push(nested)
+        deeper = nested.get("output")
+        if isinstance(deeper, dict):
+            push(deeper)
+    result = resp_json.get("result")
+    if isinstance(result, dict):
+        push(result)
+    data_obj = resp_json.get("data")
+    if isinstance(data_obj, dict):
+        push(data_obj)
+    return roots
+
+
+def _append_image_from_any(images: list[str], raw: Any) -> None:
+    if isinstance(raw, str) and raw.strip():
+        s = raw.strip()
+        if s.startswith(("http://", "https://", "data:")):
+            images.append(s)
+        return
+    if isinstance(raw, dict):
+        u = raw.get("url")
+        if isinstance(u, str) and u.strip() and u.strip().startswith(("http://", "https://")):
+            images.append(u.strip())
+
+
+def _parse_content_part(it: Any, *, text_parts: list[str], images: list[str]) -> None:
+    """Parse one multimodal element (DashScope shorthand, OpenAI-style typed blocks, etc.)."""
+    if isinstance(it, str) and it.strip():
+        text_parts.append(it.strip())
+        return
+    if not isinstance(it, dict):
+        return
+    typ = str(it.get("type") or "").strip().lower()
+    if typ == "text":
+        tx = it.get("text")
+        if isinstance(tx, str) and tx.strip():
+            text_parts.append(tx.strip())
+        return
+    if typ == "input_text":
+        tx = it.get("text")
+        if isinstance(tx, str) and tx.strip():
+            text_parts.append(tx.strip())
+        return
+    if typ == "image_url":
+        iu = it.get("image_url")
+        if isinstance(iu, dict):
+            _append_image_from_any(images, iu)
+        elif isinstance(iu, str):
+            _append_image_from_any(images, iu)
+        return
+    if typ == "input_image":
+        iu = it.get("image_url")
+        if isinstance(iu, dict):
+            _append_image_from_any(images, iu)
+        elif isinstance(iu, str):
+            _append_image_from_any(images, iu)
+        return
+    if typ == "image" and it.get("image") is not None:
+        _append_image_from_any(images, it.get("image"))
+        return
+    # Untyped / DashScope shorthand blocks
+    tx = it.get("text")
+    if isinstance(tx, str) and tx.strip():
+        text_parts.append(tx.strip())
+    if it.get("image") is not None:
+        _append_image_from_any(images, it.get("image"))
+    ius = it.get("image_url")
+    if isinstance(ius, str):
+        _append_image_from_any(images, ius)
+    elif isinstance(ius, dict):
+        _append_image_from_any(images, ius)
+    uu = it.get("url")
+    if isinstance(uu, str) and uu.strip().startswith(("http://", "https://")):
+        images.append(uu.strip())
+    if isinstance(it.get("b64_json"), str) and str(it.get("b64_json")).strip():
+        images.append(f"data:image/png;base64,{it['b64_json']}")
+
+
+def _try_parse_content_json_string(raw: str, *, text_parts: list[str], images: list[str]) -> bool:
+    """Some proxies stringify multimodal ``content`` as JSON."""
+    s = raw.strip()
+    if not s or s[0] not in "[{":
+        return False
+    try:
+        parsed = json.loads(s)
+    except Exception:
+        return False
+    if isinstance(parsed, list):
+        for it in parsed:
+            _parse_content_part(it, text_parts=text_parts, images=images)
+        return True
+    if isinstance(parsed, dict):
+        _parse_content_part(parsed, text_parts=text_parts, images=images)
+        return True
+    return False
+
+
+def _harvest_http_urls_from_value(val: Any, images: list[str], *, depth: int, max_depth: int = 6) -> None:
+    """Last resort: collect image URLs nested under ``message`` (unknown multimodal block shapes)."""
+    if depth > max_depth or val is None:
+        return
+    if isinstance(val, str):
+        s = val.strip()
+        if s.startswith(("http://", "https://")) and s not in images:
+            images.append(s)
+        return
+    if isinstance(val, dict):
+        for _k, v in val.items():
+            _harvest_http_urls_from_value(v, images, depth=depth + 1, max_depth=max_depth)
+        return
+    if isinstance(val, list):
+        for it in val:
+            _harvest_http_urls_from_value(it, images, depth=depth + 1, max_depth=max_depth)
+
+
+def _parse_choice_message_content(msg: dict[str, Any], *, text_parts: list[str], images: list[str]) -> None:
+    c = msg.get("content")
+    if c is None:
+        _harvest_http_urls_from_value(msg, images, depth=0)
+        return
+    if isinstance(c, str):
+        if c.strip() and _try_parse_content_json_string(c, text_parts=text_parts, images=images):
+            _harvest_http_urls_from_value(msg, images, depth=0)
+            return
+        if c.strip():
+            text_parts.append(c.strip())
+        _harvest_http_urls_from_value(msg, images, depth=0)
+        return
+    if isinstance(c, dict):
+        _parse_content_part(c, text_parts=text_parts, images=images)
+        _harvest_http_urls_from_value(msg, images, depth=0)
+        return
+    if isinstance(c, list):
+        for it in c:
+            _parse_content_part(it, text_parts=text_parts, images=images)
+        _harvest_http_urls_from_value(msg, images, depth=0)
+        return
+
+
+def _parse_messages_array_for_assistant(
+    messages: list[Any], *, text_parts: list[str], images: list[str]
+) -> None:
+    for m in reversed(messages):
+        if not isinstance(m, dict):
+            continue
+        if str(m.get("role") or "").strip().lower() != "assistant":
+            continue
+        _parse_choice_message_content(m, text_parts=text_parts, images=images)
+        break
+
+
+def extract_text_and_images(resp_json: dict[str, Any]) -> tuple[str, list[str]]:
+    text_parts: list[str] = []
+    images: list[str] = []
+    roots = _extract_response_roots(resp_json)
+
+    for root in roots:
+        ot = root.get("text")
+        if isinstance(ot, str) and ot.strip():
+            text_parts.append(ot.strip())
+
+    choices: list[Any] | None = None
+    for root in roots:
+        ch = root.get("choices")
+        if isinstance(ch, list) and ch:
+            choices = ch
+            break
+
+    if isinstance(choices, list) and choices:
+        for choice in choices:
+            if not isinstance(choice, dict):
+                continue
+            legacy_txt = choice.get("text")
+            if isinstance(legacy_txt, str) and legacy_txt.strip():
+                text_parts.append(legacy_txt.strip())
+            msg = choice.get("message")
+            if isinstance(msg, dict):
+                _parse_choice_message_content(msg, text_parts=text_parts, images=images)
+            elif not isinstance(msg, dict):
+                _harvest_http_urls_from_value(choice, images, depth=0)
+
+    if not text_parts and not images:
+        for root in roots:
+            msgs = root.get("messages")
+            if isinstance(msgs, list) and msgs:
+                _parse_messages_array_for_assistant(msgs, text_parts=text_parts, images=images)
+                if text_parts or images:
+                    break
+
+    if not images:
+        for root in roots:
+            data = root.get("data")
+            if not isinstance(data, list):
+                continue
+            for it in data:
+                if not isinstance(it, dict):
+                    continue
+                if isinstance(it.get("url"), str) and str(it.get("url")).strip():
+                    images.append(str(it.get("url")).strip())
+                if isinstance(it.get("b64_json"), str) and str(it.get("b64_json")).strip():
+                    images.append(f"data:image/png;base64,{it['b64_json']}")
+
+    if not images:
+        for root in roots:
+            for key in ("results", "artifacts", "task_outputs"):
+                arr = root.get(key)
+                if not isinstance(arr, list):
+                    continue
+                for it in arr:
+                    if not isinstance(it, dict):
+                        continue
+                    for ik in ("image", "url", "output_image_url", "image_url"):
+                        v = it.get(ik)
+                        if isinstance(v, str) and v.strip().startswith(("http://", "https://")):
+                            images.append(v.strip())
+                        elif isinstance(v, dict) and isinstance(v.get("url"), str):
+                            u = str(v.get("url")).strip()
+                            if u.startswith(("http://", "https://")):
+                                images.append(u)
+
+    # De-dupe URLs while preserving order
+    seen_u: set[str] = set()
+    uniq_images: list[str] = []
+    for u in images:
+        if u not in seen_u:
+            seen_u.add(u)
+            uniq_images.append(u)
+    images = uniq_images
+
+    return "\n".join([x for x in text_parts if x]).strip(), images
+
+
+def _diag_fill_choices(out: dict[str, Any], ch: Any, *, prefix: str) -> None:
+    out[f"{prefix}choices_typename"] = type(ch).__name__
+    out[f"{prefix}choices_len"] = len(ch) if isinstance(ch, list) else None
+    if not isinstance(ch, list) or not ch or not isinstance(ch[0], dict):
+        return
+    c0 = ch[0]
+    out[f"{prefix}choice0_keys"] = sorted(c0.keys())[:24]
+    msg = c0.get("message")
+    if isinstance(msg, dict):
+        out[f"{prefix}msg_keys"] = sorted(msg.keys())[:24]
+        c = msg.get("content")
+        out[f"{prefix}content_typename"] = type(c).__name__
+        if isinstance(c, list):
+            out[f"{prefix}content_len"] = len(c)
+            if c:
+                out[f"{prefix}content0_typename"] = type(c[0]).__name__
+                if isinstance(c[0], dict):
+                    out[f"{prefix}content0_keys"] = sorted(c[0].keys())[:24]
+        elif isinstance(c, dict):
+            out[f"{prefix}content_keys"] = sorted(c.keys())[:24]
+        elif c is None:
+            out[f"{prefix}content_is_null"] = True
+
+
+def build_extract_diag_empty(obj: dict[str, Any]) -> dict[str, Any]:
+    """Compact shape hints when extraction yielded nothing (stderr / chat debug)."""
+    out: dict[str, Any] = {"top_level_keys": sorted(obj.keys())[:40]}
+    ch_top = obj.get("choices")
+    if ch_top is not None:
+        _diag_fill_choices(out, ch_top, prefix="")
+    outp = obj.get("output")
+    if isinstance(outp, dict):
+        out["output_keys"] = sorted(outp.keys())[:40]
+        ch = outp.get("choices")
+        _diag_fill_choices(out, ch, prefix="output_")
+    return out
+
+
+def redact_response_for_debug(obj: Any, *, max_chars: int = 2800) -> str:
+    """JSON preview for error surfaces (truncate; redact long base64 / data URLs)."""
+
+    def _walk(x: Any) -> Any:
+        if isinstance(x, dict):
+            return {str(k): _walk(v) for k, v in x.items()}
+        if isinstance(x, list):
+            return [_walk(v) for v in x[:80]]
+        if isinstance(x, str):
+            s = x
+            if ";base64," in s and s.strip().startswith("data:") and len(s) > 120:
+                h, _, _ = s.partition(";base64,")
+                return f"{h};base64,<redacted {len(s)} chars>"
+            if len(s) > 500:
+                return s[:400] + f"...<truncated {len(s) - 400}>"
+            return s
+        return x
+
+    try:
+        txt = json.dumps(_walk(obj), ensure_ascii=False, default=str)
+    except Exception:
+        txt = str(obj)
+    if len(txt) > max_chars:
+        return txt[: max_chars - 24] + "\n…<truncated>"
+    return txt
+
+
+def format_extract_diag(diag: dict[str, Any], *, max_chars: int = 900) -> str:
+    try:
+        s = json.dumps(diag, ensure_ascii=False, default=str)
+    except Exception:
+        s = str(diag)
+    if len(s) > max_chars:
+        return s[: max_chars - 20] + "…<truncated>"
+    return s
+
+
+__all__ = [
+    "build_extract_diag_empty",
+    "compress_data_url_image",
+    "dashscope_multimodal_http_ok",
+    "dashscope_native_multimodal_url_from_compatible_base",
+    "download_http_url_bytes",
+    "env_image_expert_download_timeout_sec",
+    "redact_response_for_debug",
+    "env_image_expert_api_key",
+    "env_image_expert_base_url",
+    "env_image_expert_chat_endpoint",
+    "env_image_expert_model",
+    "env_ocr_lane_api_key",
+    "env_ocr_lane_base_url",
+    "env_ocr_lane_chat_endpoint",
+    "env_ocr_lane_model",
+    "extract_text_and_images",
+    "format_extract_diag",
+    "is_data_url",
+    "join_url",
+    "post_with_retry",
+]

svc/llm/image_legacy_client.py

@@ -0,0 +1,733 @@
+"""Legacy image+text payloads for DashScope-style gateways.
+
+Uses ``{"image":...}/{"text":...}`` or typed compatible-mode blocks on ``/chat/completions`` only.
+The **image specialist** uses this module from:
+
+- :mod:`~runtime.agents.specialist_agent` (orchestration temp sessions)
+- :mod:`~runtime.direct_loop` when ``skill_binding_role=="image"`` (**gateway /chat UI**), so vision
+  turns never hit :class:`~svc.llm.transports.openai_responses.OpenAIResponsesModel` unless explicitly disabled via env.
+
+Alignment with Alibaba ``dashscope.MultiModalConversation`` examples (**message shape**):
+    ``messages = [{"role": "user", "content": [{"image": "<url or data:...>"}, {"text": "..."}]}]``
+matches our non--``typed`` branch (same ``image`` / ``text`` keys as the SDK doc).
+
+Lane separation (**not** OCR):
+- Resolved as ``kwargs …`` from the user's **chosen chat model/profile** first, then ``AIA_IMAGE_EXPERT_*`` when a field is empty.
+  The **`AIA_OCR_*`** variables remain for **`query_image_attachment`** / OCR downgrade only.
+- Bearer HTTP targets ``BASE_URL`` + ``AIA_IMAGE_EXPERT_CHAT_ENDPOINT`` (default ``/chat/completions``), not ``MultiModalConversation`` SDK.
+
+SDK-style extras:
+- Optional top-level fields (``stream``, ``n``, ``watermark``, ``negative_prompt``, ``prompt_extend``, ``size``, …):
+  use ``DASHSCOPE_IMAGE_*`` env vars or JSON in ``AIA_IMAGE_EXPERT_REQUEST_EXTRA`` (alias: ``AIA_LEGACY_IMAGE_REQUEST_EXTRA``).
+
+Compatibility roots:
+- ``compatible-mode/v1`` **默认**使用 OpenAI Chat Completions 视觉块（每条含 ``type``：``image_url`` / ``text``）；百炼兼容网关否则会报 ``missing_required_parameter … content[n].type``。若网关只吃 DashScope 文档形态（无 ``type`` 的 ``{"image"}``/``{"text"}``），设置 ``AIA_IMAGE_EXPERT_COMPAT_USE_DASHSCOPE_NATIVE_BLOCKS=1``。
+
+For OpenAI-style ``image_url`` chat payloads (tool OCR / multimodal downgrade), use :mod:`svc.llm.image_ocr_client`.
+
+Chat UI **图片专家**端到端说明（与其它链路隔离的变更边界）见仓库内 ``docs/IMAGE_SPECIALIST_LANE.md``.
+"""
+
+from __future__ import annotations
+
+import base64
+import json
+import os
+import re
+import sys
+from typing import Any
+
+import httpx
+
+from svc.files.attachment_assets import AttachmentAssetStore
+from svc.llm.image_http_common import (
+    build_extract_diag_empty,
+    compress_data_url_image,
+    dashscope_multimodal_http_ok,
+    dashscope_native_multimodal_url_from_compatible_base,
+    download_http_url_bytes,
+    env_image_expert_api_key,
+    env_image_expert_base_url,
+    env_image_expert_chat_endpoint,
+    env_image_expert_model,
+    extract_text_and_images,
+    format_extract_diag,
+    is_data_url,
+    join_url,
+    post_with_retry,
+    redact_response_for_debug,
+)
+from runtime.prompt_templates import render_prompt
+
+IMAGE_SPECIALIST_DEFAULT_PROMPT_ZH = (
+    "请根据用户上传的图片作答：描述可见场景、物体与文字；不确定处请标明。"
+)
+
+
+def _max_legacy_input_images() -> int:
+    try:
+        v = int(os.getenv("AIA_IMAGE_EXPERT_MAX_INPUT_IMAGES") or "8")
+    except Exception:
+        v = 8
+    return max(1, min(int(v), 12))
+
+
+def parse_message_attachments_json(raw: Any) -> list[dict[str, Any]]:
+    """Decode ``chat_message.attachments`` (JSON string, list, or dict) into attachment dict rows."""
+    if raw is None:
+        return []
+    if isinstance(raw, list):
+        return [x for x in raw if isinstance(x, dict)]
+    if isinstance(raw, dict):
+        return [raw]
+    if isinstance(raw, str):
+        s = raw.strip()
+        if not s or s.lower() == "null":
+            return []
+        try:
+            v = json.loads(s)
+            if isinstance(v, list):
+                return [x for x in v if isinstance(x, dict)]
+            if isinstance(v, dict):
+                return [v]
+        except Exception:
+            pass
+    return []
+
+
+def collect_legacy_lane_images_from_attachments(
+    attachments: list[dict[str, Any]] | None,
+    *,
+    max_images: int | None = None,
+) -> list[str]:
+    """Normalize incoming UI/store attachments to URLs/data URLs for :func:`send_legacy_image_messages`."""
+    from svc.files.attachment_assets import attachment_id_to_data_url
+
+    cap = max(1, min(int(max_images if max_images is not None else _max_legacy_input_images()), 12))
+    out: list[str] = []
+    for att in attachments or []:
+        if not isinstance(att, dict):
+            continue
+        t = str(att.get("type") or "").strip().lower()
+        if t == "image_ref":
+            aid = str(att.get("attachment_id") or "").strip()
+            if not aid:
+                continue
+            data_url = attachment_id_to_data_url(aid, mime=str(att.get("mime") or ""))
+            if data_url:
+                out.append(data_url)
+        elif t in ("input_image", "image"):
+            raw = str(att.get("image_base64") or att.get("data") or "").strip()
+            if raw:
+                mime = str(att.get("mime") or "image/jpeg")
+                if raw.startswith("data:"):
+                    out.append(raw)
+                else:
+                    out.append(f"data:{mime};base64,{raw}")
+        elif t == "image_url":
+            u = str(att.get("url") or "").strip()
+            if u:
+                out.append(u)
+        elif t == "relay_pointer":
+            mime = str(att.get("mime") or "").strip().lower()
+            if mime and not mime.startswith("image/"):
+                continue
+            uri = str(att.get("pointer_uri") or "").strip()
+            aid = str(att.get("attachment_id") or "").strip()
+            if not aid and uri:
+                m = re.match(r"^relay://attachments/[^/]+/([a-f0-9]{8,128})$", uri, re.I)
+                if m:
+                    aid = str(m.group(1)).lower()
+            if aid:
+                data_url = attachment_id_to_data_url(aid, mime=str(att.get("mime") or ""))
+                if data_url:
+                    out.append(data_url)
+        if len(out) >= cap:
+            break
+    return out
+
+
+def collect_legacy_lane_images_with_session_fallback(
+    *,
+    store: Any,
+    session_id: str,
+    attachments: list[dict[str, Any]] | None,
+    max_images: int | None = None,
+) -> tuple[list[str], str]:
+    """Resolve legacy multimodal image URLs for the current turn, then session history.
+
+    When the current request has no embedded/URL images, scan persisted messages (newest→oldest):
+    1) assistant rows with usable image attachments (e.g. last model output);
+    2) user rows with images.
+
+    The trailing user bubble for this turn (text-only) is skipped when scanning.
+
+    Returns ``(images, source)`` where ``source`` is ``current``, ``assistant_history``,
+    ``user_history``, or empty when nothing was found.
+
+    Disable with ``AIA_IMAGE_SPECIALIST_SESSION_IMAGE_FALLBACK=0`` (default: on).
+    """
+    cap = max_images if max_images is not None else _max_legacy_input_images()
+    primary = collect_legacy_lane_images_from_attachments(attachments, max_images=cap)
+    if primary:
+        return primary, "current"
+    if str(os.getenv("AIA_IMAGE_SPECIALIST_SESSION_IMAGE_FALLBACK") or "").strip().lower() in (
+        "0",
+        "false",
+        "no",
+        "off",
+    ):
+        return [], ""
+
+    try:
+        scan_n = int(os.getenv("AIA_IMAGE_SPECIALIST_FALLBACK_SCAN_MESSAGES") or "400")
+    except Exception:
+        scan_n = 400
+    scan_n = max(1, min(scan_n, 2000))
+
+    try:
+        msgs = store.get_messages(str(session_id or "").strip(), limit=scan_n)
+    except Exception:
+        msgs = []
+
+    if not msgs:
+        return [], ""
+
+    hi = len(msgs) - 1
+    if hi >= 0:
+        last = msgs[hi]
+        if str(getattr(last, "role", "") or "").strip().lower() == "user":
+            la = parse_message_attachments_json(getattr(last, "attachments", None))
+            if not collect_legacy_lane_images_from_attachments(la, max_images=1):
+                hi -= 1
+
+    for phase in ("assistant", "user"):
+        for i in range(hi, -1, -1):
+            m = msgs[i]
+            if str(getattr(m, "role", "") or "").strip().lower() != phase:
+                continue
+            la = parse_message_attachments_json(getattr(m, "attachments", None))
+            got = collect_legacy_lane_images_from_attachments(la, max_images=cap)
+            if got:
+                return got, f"{phase}_history"
+    return [], ""
+
+
+def normalize_legacy_output_image_urls(resp_images: Any, *, max_items: int = 12) -> list[str]:
+    """Flatten provider ``images`` / content parts to HTTP/data URLs (strings only)."""
+    cap = max(1, min(int(max_items), 24))
+    out: list[str] = []
+    if resp_images is None:
+        return out
+    if isinstance(resp_images, str):
+        s = resp_images.strip()
+        if s:
+            out.append(s)
+        return out[:cap]
+    if not isinstance(resp_images, list):
+        return out
+    for it in resp_images[:cap]:
+        if isinstance(it, str):
+            s = it.strip()
+            if s:
+                out.append(s)
+            continue
+        if not isinstance(it, dict):
+            continue
+        u = it.get("image")
+        if isinstance(u, str) and u.strip():
+            out.append(u.strip())
+            continue
+        u = it.get("url")
+        if isinstance(u, str) and u.strip():
+            out.append(u.strip())
+            continue
+        iu = it.get("image_url")
+        if isinstance(iu, str) and iu.strip():
+            out.append(iu.strip())
+        elif isinstance(iu, dict):
+            u2 = iu.get("url")
+            if isinstance(u2, str) and u2.strip():
+                out.append(u2.strip())
+    return out[:cap]
+
+
+def materialize_legacy_response_output_attachments(
+    resp_images: Any,
+    *,
+    max_images: int = 3,
+) -> list[dict[str, Any]]:
+    """Persist remote/base64 model outputs as ``image_ref`` / ``image_url`` rows for chat UI."""
+    cap = max(1, min(int(max_images), 12))
+    produced: list[dict[str, Any]] = []
+    urls = normalize_legacy_output_image_urls(resp_images, max_items=cap)
+    if not urls:
+        return produced
+    store = AttachmentAssetStore()
+    for idx, item in enumerate(urls[:cap], start=1):
+        s = str(item or "").strip()
+        if not s:
+            continue
+        if s.startswith("data:") and ";base64," in s:
+            head, b64 = s.split(";base64,", 1)
+            mime = head.replace("data:", "", 1) or "image/png"
+            try:
+                blob = base64.b64decode(b64.encode("ascii"))
+            except Exception:
+                continue
+            meta = store.save_bytes(blob, filename=f"image-output-{idx}.png", mime=mime)
+            produced.append(
+                {
+                    "type": "image_ref",
+                    "attachment_id": meta.attachment_id,
+                    "name": meta.name,
+                    "mime": meta.mime,
+                    "bytes": meta.bytes,
+                    "width": meta.width,
+                    "height": meta.height,
+                }
+            )
+        elif s.startswith("http://") or s.startswith("https://"):
+            try:
+                blob, ctype = download_http_url_bytes(s)
+                if blob:
+                    mime = (ctype.split(";", 1)[0].strip() if ctype else "") or "image/png"
+                    ext = ".png"
+                    if mime == "image/jpeg":
+                        ext = ".jpg"
+                    elif mime == "image/webp":
+                        ext = ".webp"
+                    elif mime == "image/gif":
+                        ext = ".gif"
+                    meta = store.save_bytes(blob, filename=f"image-output-{idx}{ext}", mime=mime)
+                    produced.append(
+                        {
+                            "type": "image_ref",
+                            "attachment_id": meta.attachment_id,
+                            "name": meta.name,
+                            "mime": meta.mime,
+                            "bytes": meta.bytes,
+                            "width": meta.width,
+                            "height": meta.height,
+                        }
+                    )
+                    continue
+            except Exception:
+                pass
+            produced.append({"type": "image_url", "url": s, "name": f"image-output-{idx}.png"})
+    return produced
+
+
+def legacy_image_turn_bundle(resp: dict[str, Any]) -> tuple[bool, str, list[dict[str, Any]]]:
+    """Interpret ``send_legacy_image_messages`` result for persistence (text-only vision answers allowed)."""
+    ok = bool(resp.get("ok"))
+    text = str(resp.get("text") or "").strip()
+    if not ok:
+        err = str(resp.get("error") or "").strip()
+        return False, f"Image generation failed: {err or 'unknown error'}", []
+    raw_urls = normalize_legacy_output_image_urls(resp.get("images"), max_items=6)
+    imgs = materialize_legacy_response_output_attachments(raw_urls, max_images=3)
+    if not imgs and raw_urls:
+        imgs = [
+            {"type": "image_url", "url": u, "name": f"image-output-{i}.png"}
+            for i, u in enumerate(raw_urls[:3], start=1)
+            if u.startswith(("http://", "https://", "data:"))
+        ]
+    if imgs:
+        return True, text, imgs
+    if text:
+        return True, text, []
+    diag = resp.get("extract_diag")
+    hint = ""
+    if isinstance(diag, dict) and diag:
+        hint = format_extract_diag(diag, max_chars=1400)
+    red = resp.get("provider_response_redacted")
+    if isinstance(red, str) and red.strip():
+        hint = f"{hint}\nprovider_json={red.strip()}" if hint else f"provider_json={red.strip()}"
+    base = "Image specialist failed: empty response from provider."
+    out_msg = f"{base} {hint}".strip()
+    if len(out_msg) > 12_000:
+        out_msg = out_msg[:11_980] + "\n…<truncated>"
+    return False, out_msg, []
+
+
+def legacy_image_assistant_body_with_placeholder(
+    *,
+    lang: str | None,
+    body_text: str,
+    produced: list[dict[str, Any]] | None,
+) -> str:
+    """If the model returned images but no visible text, use the standard chat placeholder (ZH/EN).
+
+    Shared by ``direct_loop`` (gateway /chat) and ``specialist_agent`` (temp sessions).
+    """
+    if str(body_text or "").strip():
+        return str(body_text or "")
+    if produced:
+        return (
+            "Generated image (see attachment below)."
+            if str(lang or "").startswith("en")
+            else "已生成图片（见下方附件）。"
+        )
+    return str(body_text or "")
+
+
+def _truthy_env(name: str) -> bool:
+    return str(os.getenv(name) or "").strip().lower() in ("1", "true", "yes", "on")
+
+
+def _redact_payload_for_stderr(obj: Any) -> Any:
+    if isinstance(obj, dict):
+        return {str(k): _redact_payload_for_stderr(v) for k, v in obj.items()}
+    if isinstance(obj, list):
+        return [_redact_payload_for_stderr(x) for x in obj]
+    if isinstance(obj, str):
+        s = obj
+        if ";base64," in s and s.strip().startswith("data:") and len(s) > 120:
+            h, _, _ = s.partition(";base64,")
+            return f"{h};base64,<redacted ~{len(s) - len(h) - 8} chars>"
+        return s
+    return obj
+
+
+def _stderr_debug_image_legacy(url: str, payload: dict[str, Any]) -> None:
+    if not _truthy_env("AIA_IMAGE_EXPERT_DEBUG_PRINT_PAYLOAD"):
+        return
+    try:
+        txt = json.dumps(_redact_payload_for_stderr(dict(payload)), ensure_ascii=False, indent=2, default=str)
+        sys.stderr.write(f"\n[oclaw image_legacy] POST {url}\n{txt}\n\n")
+        sys.stderr.flush()
+    except Exception:
+        pass
+
+
+def _extra_request_fields_from_env() -> dict[str, Any]:
+    """Merge JSON from ``AIA_IMAGE_EXPERT_REQUEST_EXTRA`` (or legacy alias ``AIA_LEGACY_IMAGE_REQUEST_EXTRA``)."""
+    raw = (os.getenv("AIA_IMAGE_EXPERT_REQUEST_EXTRA") or os.getenv("AIA_LEGACY_IMAGE_REQUEST_EXTRA") or "").strip()
+    if not raw:
+        return {}
+    try:
+        parsed = json.loads(raw)
+        return parsed if isinstance(parsed, dict) else {}
+    except Exception:
+        return {}
+
+
+def _dashscope_image_env_kw() -> dict[str, Any]:
+    """Map `_local/system.env.example` ``DASHSCOPE_IMAGE_*`` vars to multimodal/top-level kwargs (parity with SDK samples)."""
+    out: dict[str, Any] = {}
+    raw_n = (os.getenv("DASHSCOPE_IMAGE_N") or "").strip()
+    if raw_n.isdigit():
+        out["n"] = max(1, min(int(raw_n), 6))
+
+    wm = (os.getenv("DASHSCOPE_IMAGE_WATERMARK") or "").strip().lower()
+    if wm in ("1", "true", "yes", "on"):
+        out["watermark"] = True
+    elif wm in ("0", "false", "no", "off"):
+        out["watermark"] = False
+
+    raw_stream = (os.getenv("DASHSCOPE_IMAGE_STREAM") or "").strip().lower()
+    if raw_stream in ("1", "true", "yes", "on"):
+        out["stream"] = True
+    elif raw_stream in ("0", "false", "no", "off"):
+        out["stream"] = False
+
+    neg = os.getenv("DASHSCOPE_IMAGE_NEGATIVE_PROMPT")
+    if neg is not None:
+        ns = str(neg)
+        if ns.strip():
+            out["negative_prompt"] = ns
+
+    pe = (os.getenv("DASHSCOPE_IMAGE_PROMPT_EXTEND") or "").strip().lower()
+    if pe in ("1", "true", "yes", "on"):
+        out["prompt_extend"] = True
+    elif pe in ("0", "false", "no", "off"):
+        out["prompt_extend"] = False
+
+    size = (os.getenv("DASHSCOPE_IMAGE_SIZE") or "").strip()
+    if size:
+        out["size"] = size
+
+    return out
+
+
+def _openai_compatible_vision_content(images: list[str], prompt_text: str) -> list[dict[str, Any]]:
+    """OpenAI Chat Completions vision shape (``type`` + ``image_url`` / ``text``).
+
+    Required for typical ``compatible-mode/v1`` ``/chat/completions`` (each part must have ``type``).
+    Image URLs may be ``https://…`` or ``data:image/…;base64,…``.
+    """
+    blocks: list[dict[str, Any]] = []
+    for img in images:
+        blocks.append({"type": "image_url", "image_url": {"url": img}})
+    blocks.append({"type": "text", "text": str(prompt_text or "").strip()})
+    return blocks
+
+
+def _compat_use_dashscope_native_blocks() -> bool:
+    """When ``1``, ``compatible-mode`` URLs use DashScope doc blocks without ``type`` (see :func:`_http_content_blocks`)."""
+    return str(os.getenv("AIA_IMAGE_EXPERT_COMPAT_USE_DASHSCOPE_NATIVE_BLOCKS") or "").strip().lower() in (
+        "1",
+        "true",
+        "yes",
+        "on",
+    )
+
+
+def _model_triggers_dashscope_native_fallback(model_name: str) -> bool:
+    """``qwen-image`` on OpenAI-compat ``/chat/completions`` often returns ``message.content=null``."""
+    if str(os.getenv("AIA_IMAGE_EXPERT_DISABLE_DASHSCOPE_NATIVE_FALLBACK") or "").strip().lower() in (
+        "1",
+        "true",
+        "yes",
+        "on",
+    ):
+        return False
+    if str(os.getenv("AIA_IMAGE_EXPERT_FORCE_DASHSCOPE_NATIVE") or "").strip().lower() in (
+        "1",
+        "true",
+        "yes",
+        "on",
+    ):
+        return True
+    m = (model_name or "").strip().lower()
+    needle = (os.getenv("AIA_IMAGE_EXPERT_NATIVE_FALLBACK_MODEL_SUBSTR") or "qwen-image").strip().lower()
+    return bool(needle) and needle in m
+
+
+def _http_content_blocks(images: list[str], prompt: str, *, typed: bool) -> list[dict[str, Any]]:
+    prompt_text = str(prompt or "").strip() or render_prompt("image/default_edit_prompt.zh.md", strict=True)
+    if not typed:
+        blocks: list[dict[str, Any]] = [{"image": img} for img in images]
+        blocks.append({"text": prompt_text})
+        return blocks
+    blocks_typed: list[dict[str, Any]] = [{"type": "image", "image": img} for img in images]
+    blocks_typed.append({"type": "text", "text": prompt_text})
+    return blocks_typed
+
+
+def send_legacy_image_messages(
+    *,
+    images: list[str],
+    prompt: str,
+    model: str | None = None,
+    timeout_sec: int = 60,
+    api_key: str | None = None,
+    base_url: str | None = None,
+) -> dict[str, Any]:
+    """Legacy multimodal HTTP (non--OpenAI-``image_url`` schema); image specialist uses this entry point."""
+    resolved_base_url = (base_url or env_image_expert_base_url()).strip()
+    resolved_api_key = (api_key or env_image_expert_api_key()).strip()
+    model_name = ((model or "").strip() or env_image_expert_model())
+    endpoint = env_image_expert_chat_endpoint()
+    url = join_url(resolved_base_url, endpoint)
+    if not resolved_api_key or not resolved_base_url:
+        return {
+            "ok": False,
+            "error": "missing AIA_IMAGE_EXPERT_API_KEY or AIA_IMAGE_EXPERT_BASE_URL (or pass api_key and base_url)",
+        }
+    if not model_name:
+        return {
+            "ok": False,
+            "error": "missing model (chosen profile/model=… or set AIA_IMAGE_EXPERT_MODEL)",
+        }
+    if not images:
+        return {"ok": False, "error": "at least one image input is required"}
+
+    raw_selected = [str(x).strip() for x in images if str(x).strip()][: _max_legacy_input_images()]
+    selected: list[str] = []
+    input_kind: list[str] = []
+    for img in raw_selected:
+        if is_data_url(img):
+            selected.append(compress_data_url_image(img))
+            input_kind.append("data_url")
+            continue
+        if img.startswith("http://") or img.startswith("https://"):
+            selected.append(img)
+            input_kind.append("url")
+            continue
+    if not selected:
+        return {"ok": False, "error": "no usable image input (expected URL or data URL)"}
+
+    # compatible-mode /chat/completions is OpenAI-shaped: every content part needs ``type`` (else 400 missing content[n].type).
+    # Native DashScope HTTP (non-compatible base) uses ``{"image"}`` / ``{"text"}`` without ``type``.
+    use_compatible_base = "compatible-mode" in resolved_base_url.lower()
+    prompt_plain = str(prompt or "").strip() or render_prompt("image/default_edit_prompt.zh.md", strict=True)
+    if use_compatible_base and not _compat_use_dashscope_native_blocks():
+        content_multi = _openai_compatible_vision_content(selected, prompt_plain)
+    else:
+        content_multi = _http_content_blocks(selected, prompt, typed=False)
+
+    headers = {
+        "Authorization": f"Bearer {resolved_api_key}",
+        "Content-Type": "application/json",
+    }
+    extra_ds = _dashscope_image_env_kw()
+    extra_json = _extra_request_fields_from_env()
+    extra = {**extra_ds, **extra_json}
+    core_multi = {
+        "model": model_name,
+        "messages": [{"role": "user", "content": content_multi}],
+    }
+    payload_multi = {**extra, **core_multi}
+
+    _stderr_debug_image_legacy(url, payload_multi)
+
+    with httpx.Client(timeout=float(timeout_sec)) as client:
+        try:
+            r = post_with_retry(client, url=url, headers=headers, payload=payload_multi)
+        except Exception as e:
+            return {"ok": False, "error": f"http request failed: {type(e).__name__}: {e}", "backend_shape": "multi"}
+
+        if r.status_code >= 400:
+            return {"ok": False, "error": f"http {r.status_code}: {r.text[:800]}", "backend_shape": "multi"}
+
+        try:
+            obj = r.json()
+        except Exception:
+            return {"ok": False, "error": f"non-json response: {r.text[:500]}", "backend_shape": "multi"}
+        if not isinstance(obj, dict):
+            return {
+                "ok": False,
+                "error": f"expected JSON object from provider, got {type(obj).__name__}",
+                "backend_shape": "multi",
+                "input_kind": input_kind if input_kind else ["data_url"],
+            }
+        body = obj
+        ds_ok, ds_err = dashscope_multimodal_http_ok(body)
+        if not ds_ok:
+            return {
+                "ok": False,
+                "error": ds_err or "provider rejected request (DashScope code/message)",
+                "backend_shape": "multi",
+                "input_kind": input_kind if input_kind else ["data_url"],
+            }
+        text, out_images = extract_text_and_images(body)
+        compat_extract_diag: dict[str, Any] | None = None
+        native_fallback_meta: dict[str, Any] = {}
+
+        if (
+            not str(text or "").strip()
+            and not out_images
+            and use_compatible_base
+            and _model_triggers_dashscope_native_fallback(model_name)
+        ):
+            compat_extract_diag = build_extract_diag_empty(body)
+            native_url = dashscope_native_multimodal_url_from_compatible_base(resolved_base_url)
+            if not native_url:
+                native_fallback_meta = {
+                    "attempted": False,
+                    "hint": "set AIA_IMAGE_EXPERT_DASHSCOPE_NATIVE_URL or use a compatible-mode base_url",
+                }
+            else:
+                content_native = _http_content_blocks(selected, prompt, typed=False)
+                param_merge = {**_dashscope_image_env_kw(), **_extra_request_fields_from_env()}
+                param_merge.pop("stream", None)
+                payload_native = {
+                    "model": model_name,
+                    "input": {"messages": [{"role": "user", "content": content_native}]},
+                    "parameters": {"result_format": "message", **param_merge},
+                }
+                _stderr_debug_image_legacy(native_url, payload_native)
+                try:
+                    r2 = post_with_retry(client, url=native_url, headers=headers, payload=payload_native)
+                except Exception as e:
+                    native_fallback_meta = {
+                        "attempted": True,
+                        "url": native_url,
+                        "error": f"{type(e).__name__}: {e}",
+                    }
+                else:
+                    if r2.status_code >= 400:
+                        native_fallback_meta = {
+                            "attempted": True,
+                            "url": native_url,
+                            "http_status": int(r2.status_code),
+                            "body_head": r2.text[:600],
+                        }
+                    else:
+                        try:
+                            b2 = r2.json()
+                        except Exception as e:
+                            native_fallback_meta = {
+                                "attempted": True,
+                                "url": native_url,
+                                "error": f"json: {type(e).__name__}: {e}",
+                            }
+                        else:
+                            if isinstance(b2, dict):
+                                ok2, err2 = dashscope_multimodal_http_ok(b2)
+                                if not ok2:
+                                    native_fallback_meta = {
+                                        "attempted": True,
+                                        "url": native_url,
+                                        "dashscope_error": err2 or "business code",
+                                    }
+                                else:
+                                    t2, im2 = extract_text_and_images(b2)
+                                    if str(t2 or "").strip() or im2:
+                                        text, out_images = t2, im2
+                                        body = b2
+                                        native_fallback_meta = {
+                                            "attempted": True,
+                                            "url": native_url,
+                                            "succeeded": True,
+                                        }
+                                    else:
+                                        native_fallback_meta = {
+                                            "attempted": True,
+                                            "url": native_url,
+                                            "succeeded": False,
+                                            "native_diag": build_extract_diag_empty(b2),
+                                        }
+                            else:
+                                native_fallback_meta = {
+                                    "attempted": True,
+                                    "url": native_url,
+                                    "error": "native response not a JSON object",
+                                }
+
+        extract_diag: dict[str, Any] | None = None
+        provider_response_redacted: str | None = None
+        if not str(text or "").strip() and not out_images:
+            extract_diag = build_extract_diag_empty(body)
+            if compat_extract_diag is not None:
+                extract_diag["openai_compat_empty"] = compat_extract_diag
+            if native_fallback_meta:
+                extract_diag["dashscope_native_fallback"] = native_fallback_meta
+            provider_response_redacted = redact_response_for_debug(body, max_chars=3200)
+            if _truthy_env("AIA_IMAGE_EXPERT_DEBUG_PRINT_PAYLOAD"):
+                try:
+                    sys.stderr.write(
+                        "\n[oclaw image_legacy] extract_empty extract_diag="
+                        + format_extract_diag(extract_diag, max_chars=4000)
+                        + "\n"
+                        + (provider_response_redacted or "")
+                        + "\n\n"
+                    )
+                    sys.stderr.flush()
+                except Exception:
+                    pass
+        out_images = normalize_legacy_output_image_urls(out_images, max_items=12)
+        return {
+            "ok": True,
+            "text": text,
+            "images": out_images,
+            "backend_shape": "multi",
+            "input_kind": input_kind if input_kind else ["data_url"],
+            "extract_diag": extract_diag,
+            "provider_response_redacted": provider_response_redacted,
+        }
+
+
+__all__ = [
+    "IMAGE_SPECIALIST_DEFAULT_PROMPT_ZH",
+    "collect_legacy_lane_images_from_attachments",
+    "collect_legacy_lane_images_with_session_fallback",
+    "parse_message_attachments_json",
+    "legacy_image_assistant_body_with_placeholder",
+    "legacy_image_turn_bundle",
+    "materialize_legacy_response_output_attachments",
+    "normalize_legacy_output_image_urls",
+    "send_legacy_image_messages",
+]

svc/llm/image_ocr_client.py

@@ -0,0 +1,200 @@
+"""OCR / vision lane: OpenAI-compatible multimodal ``image_url`` + ``text`` only."""
+
+from __future__ import annotations
+
+from typing import Any
+
+import httpx
+
+from svc.llm.image_http_common import (
+    compress_data_url_image,
+    env_ocr_lane_api_key,
+    env_ocr_lane_base_url,
+    env_ocr_lane_model,
+    env_ocr_lane_chat_endpoint,
+    extract_text_and_images,
+    is_data_url,
+    join_url,
+    post_with_retry,
+)
+from runtime.prompt_templates import render_prompt
+
+# Shared with query_image_attachment and OpenAI-compat multimodal→text downgrade.
+VISION_OCR_EXTRACT_PROMPT_ZH = (
+    "请只提取图片中可见文字并按阅读顺序输出。"
+    "如果有表格，保持行列结构；不确定的内容标注为[unclear]。"
+)
+VISION_DESCRIBE_PROMPT_ZH = (
+    "请详细描述这张图片的主要内容、对象、场景和可见文字。"
+    "回答请使用要点列表，避免臆测。"
+)
+
+
+def _http_content_blocks_openai_vision(images: list[str], prompt: str) -> list[dict[str, Any]]:
+    prompt_text = str(prompt or "").strip() or render_prompt("image/default_edit_prompt.zh.md", strict=True)
+    blocks: list[dict[str, Any]] = []
+    for img in images:
+        s = str(img or "").strip()
+        if not s:
+            continue
+        blocks.append({"type": "image_url", "image_url": {"url": s}})
+    if prompt_text:
+        blocks.append({"type": "text", "text": prompt_text})
+    return blocks
+
+
+def vision_llm_backend_status(
+    *,
+    api_key: str | None = None,
+    base_url: str | None = None,
+    model: str | None = None,
+) -> dict[str, Any]:
+    ak = (api_key or env_ocr_lane_api_key()).strip()
+    bu = (base_url or env_ocr_lane_base_url()).strip()
+    md = ((model or "").strip() or env_ocr_lane_model())
+    if ak and bu and md:
+        return {"ok": True, "backend": "aia_ocr_http"}
+    return {
+        "ok": False,
+        "backend": None,
+        "hint_zh": (
+            "未配置 OCR/看图通道：请同时设置 AIA_OCR_BASE_URL、AIA_OCR_API_KEY、AIA_OCR_MODEL"
+            "（OpenAI-compatible 多模态 /chat/completions）。不设默认模型 id。"
+        ),
+        "hint_en": (
+            "OCR/vision lane not configured: set AIA_OCR_BASE_URL, AIA_OCR_API_KEY, and AIA_OCR_MODEL. "
+            "No default model id."
+        ),
+    }
+
+
+def send_ocr_image_messages(
+    *,
+    images: list[str],
+    prompt: str,
+    model: str | None = None,
+    timeout_sec: int = 60,
+    api_key: str | None = None,
+    base_url: str | None = None,
+) -> dict[str, Any]:
+    """OpenAI Chat Completions multimodal only (``/chat/completions``): multi-image then single-image."""
+    resolved_base_url = (base_url or env_ocr_lane_base_url()).strip()
+    resolved_api_key = (api_key or env_ocr_lane_api_key()).strip()
+    model_name = ((model or "").strip() or env_ocr_lane_model())
+    endpoint = env_ocr_lane_chat_endpoint()
+    url = join_url(resolved_base_url, endpoint)
+    if not resolved_api_key or not resolved_base_url:
+        return {
+            "ok": False,
+            "error": "missing AIA_OCR_API_KEY or AIA_OCR_BASE_URL (or pass api_key and base_url)",
+        }
+    if not model_name:
+        return {
+            "ok": False,
+            "error": "missing AIA_OCR_MODEL (or pass model=...) — no default model id",
+        }
+    if not images:
+        return {"ok": False, "error": "at least one image input is required"}
+
+    raw_selected = [str(x).strip() for x in images if str(x).strip()][:3]
+    selected: list[str] = []
+    input_kind: list[str] = []
+    for img in raw_selected:
+        if is_data_url(img):
+            selected.append(compress_data_url_image(img))
+            input_kind.append("data_url")
+            continue
+        if img.startswith("http://") or img.startswith("https://"):
+            selected.append(img)
+            input_kind.append("url")
+            continue
+    if not selected:
+        return {"ok": False, "error": "no usable image input (expected URL or data URL)"}
+
+    content_openai = _http_content_blocks_openai_vision(selected, prompt)
+    content_openai_single = _http_content_blocks_openai_vision(selected[-1:], prompt) if selected else []
+
+    headers = {
+        "Authorization": f"Bearer {resolved_api_key}",
+        "Content-Type": "application/json",
+    }
+    payload_openai_multi = {
+        "model": model_name,
+        "messages": [{"role": "user", "content": content_openai}],
+    }
+    payload_openai_single = {
+        "model": model_name,
+        "messages": [{"role": "user", "content": content_openai_single}],
+    }
+
+    def _ok_response(resp: httpx.Response) -> tuple[dict[str, Any] | None, bool]:
+        if resp.status_code >= 400:
+            return None, False
+        try:
+            obj = resp.json()
+        except Exception:
+            return None, False
+        if not isinstance(obj, dict):
+            return None, False
+        text, out_imgs = extract_text_and_images(obj)
+        if not str(text or "").strip():
+            return obj, False
+        return {"text": text, "images": out_imgs}, True
+
+    err_last = ""
+    with httpx.Client(timeout=float(timeout_sec)) as client:
+        try:
+            r0 = post_with_retry(client, url=url, headers=headers, payload=payload_openai_multi)
+        except Exception as e:
+            return {"ok": False, "error": f"http request failed: {type(e).__name__}: {e}", "backend_shape": "openai-multi"}
+        if r0 is not None:
+            if r0.status_code >= 400:
+                err_last = f"http {r0.status_code}: {r0.text[:500]}"
+            else:
+                parsed, ok = _ok_response(r0)
+                if ok and isinstance(parsed, dict):
+                    return {
+                        "ok": True,
+                        "text": str(parsed.get("text") or ""),
+                        "images": list(parsed.get("images") or []),
+                        "backend_shape": "openai-multi",
+                        "input_kind": input_kind if input_kind else ["data_url"],
+                    }
+
+        try:
+            r1 = post_with_retry(client, url=url, headers=headers, payload=payload_openai_single)
+        except Exception as e:
+            return {
+                "ok": False,
+                "error": err_last or f"openai-single request failed: {type(e).__name__}: {e}",
+                "backend_shape": "openai-single-failed",
+            }
+        if r1.status_code >= 400:
+            return {
+                "ok": False,
+                "error": err_last or f"http {r1.status_code}: {r1.text[:500]}",
+                "backend_shape": "openai-single-failed",
+            }
+        try:
+            obj1 = r1.json()
+        except Exception:
+            return {"ok": False, "error": f"non-json response: {r1.text[:500]}", "backend_shape": "openai-single"}
+        text, out_images = extract_text_and_images(obj1 if isinstance(obj1, dict) else {})
+        out: dict[str, Any] = {
+            "ok": bool(str(text or "").strip()),
+            "text": text,
+            "images": out_images,
+            "backend_shape": "openai-single",
+            "input_kind": input_kind[-1:] if input_kind else ["data_url"],
+        }
+        if not str(text or "").strip():
+            out["error"] = "empty assistant text after openai-single"
+        return out
+
+
+__all__ = [
+    "VISION_DESCRIBE_PROMPT_ZH",
+    "VISION_OCR_EXTRACT_PROMPT_ZH",
+    "send_ocr_image_messages",
+    "vision_llm_backend_status",
+]

svc/llm/replay_policy.py

@@ -0,0 +1,91 @@
+"""Transcript replay policy for OpenAI-compatible Chat Completions.
+
+Inspired by Oclaw (MIT) `src/agents/transcript-policy.ts` defaults for
+`openai-completions`: enable strict tool-call-id sanitization so proxies and
+multi-provider histories do not break on id format / length.
+
+See OCLAW_MIT_LICENSE.txt in this package.
+"""
+
+from __future__ import annotations
+
+import os
+from dataclasses import dataclass
+from typing import Any
+
+from svc.llm.tool_call_id import DEFAULT_MAX_OPENAI_TOOL_CALL_ID_LEN
+
+
+@dataclass(frozen=True)
+class ReplayPolicy:
+    """Controls pre-request normalization of chat messages for tool-use flow."""
+
+    enabled: bool
+    sanitize_tool_call_ids: bool
+    repair_tool_pairing: bool
+    max_tool_call_id_len: int
+
+
+def _env_bool(name: str, default: bool) -> bool:
+    raw = str(os.getenv(name) or "").strip().lower()
+    if raw in ("0", "false", "no", "off"):
+        return False
+    if raw in ("1", "true", "yes", "on"):
+        return True
+    return default
+
+
+def resolve_replay_policy(
+    base_url: str | None,
+    model_id: str | None,
+    *,
+    llm_profile_mode: str | None = None,
+) -> ReplayPolicy:
+    """Resolve replay policy for OpenAI-compatible chat completions.
+
+    When enabled (default): sanitize assistant/tool ids before every request.
+    This matches oclaw's strict OpenAI-compatible replay defaults.
+    """
+    del base_url, model_id, llm_profile_mode  # reserved for provider-specific overrides
+
+    if not _env_bool("AIA_REPLAY_POLICY_ENABLED", True):
+        return ReplayPolicy(
+            enabled=False,
+            sanitize_tool_call_ids=False,
+            repair_tool_pairing=False,
+            max_tool_call_id_len=DEFAULT_MAX_OPENAI_TOOL_CALL_ID_LEN,
+        )
+
+    max_len = DEFAULT_MAX_OPENAI_TOOL_CALL_ID_LEN
+    raw_ml = str(os.getenv("AIA_TOOL_CALL_ID_MAX_LEN") or "").strip()
+    if raw_ml.isdigit():
+        max_len = max(8, min(int(raw_ml), 128))
+
+    repair = _env_bool("AIA_REPLAY_REPAIR_TOOL_PAIRING", True)
+
+    return ReplayPolicy(
+        enabled=True,
+        sanitize_tool_call_ids=True,
+        repair_tool_pairing=repair,
+        max_tool_call_id_len=max_len,
+    )
+
+
+def apply_replay_policy_to_messages(
+    messages: list[dict[str, Any]],
+    policy: ReplayPolicy,
+) -> list[dict[str, Any]]:
+    """Apply repair + id sanitization in order."""
+    if not policy.enabled or not messages:
+        return messages
+    from svc.llm.tool_call_id import (
+        repair_orphan_tool_messages,
+        rewrite_openai_chat_messages_tool_ids,
+    )
+
+    out = messages
+    if policy.repair_tool_pairing:
+        out = repair_orphan_tool_messages(out)
+    if policy.sanitize_tool_call_ids:
+        out = rewrite_openai_chat_messages_tool_ids(out, max_len=policy.max_tool_call_id_len)
+    return out

svc/llm/tool_call_id.py

@@ -0,0 +1,160 @@
+"""OpenAI-compatible tool_call_id / tool result id rewriting.
+
+Inspired by Oclaw (MIT) `src/agents/tool-call-id.ts`: sanitize ids to strict
+alphanumeric form and keep assistant.tool_calls[].id paired with role=tool
+tool_call_id in encounter order.
+
+See OCLAW_MIT_LICENSE.txt in this package.
+"""
+
+from __future__ import annotations
+
+import hashlib
+import re
+from collections import deque
+from typing import Any, Literal
+
+ToolCallIdMode = Literal["strict", "strict9"]
+
+# OpenAI Chat Completions: tool_calls[].id commonly capped ~40 chars on some gateways.
+DEFAULT_MAX_OPENAI_TOOL_CALL_ID_LEN = 40
+
+
+def _short_hash(text: str, length: int = 8) -> str:
+    return hashlib.sha256(text.encode("utf-8", errors="replace")).hexdigest()[:length]
+
+
+def sanitize_tool_call_id(raw: str | None, *, mode: ToolCallIdMode = "strict") -> str:
+    """Strip to [a-zA-Z0-9] only; empty input gets a deterministic placeholder."""
+    if not raw or not isinstance(raw, str):
+        return "sanitizedtoolid" if mode == "strict" else "defaultid"
+    s = raw.strip()
+    if not s:
+        return "sanitizedtoolid" if mode == "strict" else "defaultid"
+    alnum = re.sub(r"[^a-zA-Z0-9]", "", s)
+    if mode == "strict9":
+        if len(alnum) >= 9:
+            return alnum[:9]
+        if alnum:
+            return _short_hash(alnum, 9)
+        return _short_hash("sanitized", 9)
+    return alnum if alnum else "sanitizedtoolid"
+
+
+def _make_unique(seed: str, used: set[str], *, max_len: int) -> str:
+    base = sanitize_tool_call_id(seed, mode="strict")[:max_len]
+    if base and base not in used:
+        used.add(base)
+        return base
+    h = _short_hash(seed, 8)
+    for i in range(1000):
+        cand = f"{base[: max(0, max_len - len(h) - 1)]}{h}"[:max_len]
+        if i > 0:
+            suf = f"x{i}"
+            cand = f"{base[: max(0, max_len - len(suf) - len(h))]}{suf}{h}"[:max_len]
+        if cand not in used:
+            used.add(cand)
+            return cand
+    cand = _short_hash(f"{seed}:{id(used)}", max_len)[:max_len]
+    used.add(cand)
+    return cand
+
+
+def rewrite_openai_chat_messages_tool_ids(
+    messages: list[dict[str, Any]],
+    *,
+    max_len: int = DEFAULT_MAX_OPENAI_TOOL_CALL_ID_LEN,
+) -> list[dict[str, Any]]:
+    """Rewrite assistant tool_calls ids and matching tool tool_call_id/call_id.
+
+    Uses a per-raw-id queue (oclaw-style) so repeated raw ids map to distinct
+    new ids in order, and tool results consume the queue FIFO.
+    """
+    if not messages:
+        return messages
+
+    pending: dict[str, deque[str]] = {}
+    used: set[str] = set()
+
+    def _alloc_for_raw(raw: str) -> str:
+        raw = str(raw or "").strip()
+        seed = raw if raw else "empty"
+        return _make_unique(seed, used, max_len=max_len)
+
+    out: list[dict[str, Any]] = []
+    for m in messages:
+        if not isinstance(m, dict):
+            continue
+        role = str(m.get("role") or "")
+        if role == "assistant" and m.get("tool_calls"):
+            mm = dict(m)
+            tcs = mm.get("tool_calls")
+            if isinstance(tcs, list):
+                new_tcs: list[dict[str, Any]] = []
+                for tc in tcs:
+                    if not isinstance(tc, dict):
+                        new_tcs.append(tc)
+                        continue
+                    raw_id = str(tc.get("id") or "").strip()
+                    new_id = _alloc_for_raw(raw_id)
+                    q = pending.setdefault(raw_id, deque())
+                    q.append(new_id)
+                    new_tcs.append({**tc, "id": new_id})
+                mm["tool_calls"] = new_tcs
+            out.append(mm)
+            continue
+        if role == "tool":
+            mm = dict(m)
+            raw_ref = (
+                str(mm.get("tool_call_id") or mm.get("call_id") or "").strip()
+            )
+            new_ref: str | None = None
+            if raw_ref and raw_ref in pending and pending[raw_ref]:
+                new_ref = pending[raw_ref].popleft()
+            elif raw_ref:
+                # Orphan or ordering mismatch: allocate fresh id (best-effort pairing)
+                new_ref = _alloc_for_raw(raw_ref)
+            if new_ref:
+                mm["tool_call_id"] = new_ref
+                # Keep call_id in sync when both present (some gateways expect call_id)
+                if "call_id" in mm:
+                    mm["call_id"] = new_ref
+            else:
+                mm.pop("tool_call_id", None)
+                mm.pop("call_id", None)
+            out.append(mm)
+            continue
+        out.append(dict(m) if isinstance(m, dict) else m)
+    return out
+
+
+def repair_orphan_tool_messages(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """Drop tool_call_id/call_id from tool rows that reference no prior assistant tool_calls id."""
+    allowed: set[str] = set()
+    for m in messages:
+        if not isinstance(m, dict):
+            continue
+        if str(m.get("role") or "") != "assistant":
+            continue
+        tcs = m.get("tool_calls")
+        if not isinstance(tcs, list):
+            continue
+        for tc in tcs:
+            if isinstance(tc, dict) and tc.get("id"):
+                allowed.add(str(tc.get("id")))
+
+    out: list[dict[str, Any]] = []
+    for m in messages:
+        if not isinstance(m, dict):
+            out.append(m)
+            continue
+        if str(m.get("role") or "") != "tool":
+            out.append(m)
+            continue
+        mm = dict(m)
+        tid = str(mm.get("tool_call_id") or mm.get("call_id") or "").strip()
+        if tid and tid not in allowed:
+            mm.pop("tool_call_id", None)
+            mm.pop("call_id", None)
+        out.append(mm)
+    return out

svc/llm/tool_schema.py

@@ -0,0 +1,205 @@
+from __future__ import annotations
+
+import json
+import logging
+import os
+from typing import Any
+
+logger = logging.getLogger(__name__)
+
+# Full minimal object schema for ``function.parameters`` (wire / shrink / tier_minimal).
+# A bare ``{type, additionalProperties}`` alone has led strict gateways to emit or validate
+# ``required: null`` / missing array fields — always send explicit ``properties`` + ``required``.
+_MIN_TOOL_PARAMETERS: dict[str, Any] = {
+    "type": "object",
+    "properties": {},
+    "required": [],
+    "additionalProperties": True,
+}
+
+_COMBO_KEYS = frozenset({"allOf", "anyOf", "oneOf", "prefixItems"})
+
+
+def _schema_type_tags(t: Any) -> set[str]:
+    if t is None:
+        return set()
+    if isinstance(t, list):
+        return {str(x).strip().lower() for x in t if x is not None and str(x).strip()}
+    s = str(t).strip().lower()
+    return {s} if s else set()
+
+
+def complete_json_schema_for_openai_tools(obj: Any) -> Any:
+    """Recursively ensure JSON-schema fragments match what strict OpenAI-compat gateways expect.
+
+    This is **structural completion** for our wire path (tier/shrink/SDK), not a general MCP fix-all:
+    every ``type: object`` node gets explicit ``properties`` + ``required`` (list); every
+    ``type: array`` gets ``items`` (object). Combo keywords that must be arrays and are ``null``
+    are dropped; list elements that are ``null`` are removed.
+    """
+
+    if isinstance(obj, dict):
+        out: dict[str, Any] = {}
+        for k, v in obj.items():
+            key = str(k)
+            if key in _COMBO_KEYS and v is None:
+                if key in ("allOf", "prefixItems"):
+                    out[key] = []
+                continue
+            out[key] = complete_json_schema_for_openai_tools(v)
+
+        for combo in _COMBO_KEYS:
+            if combo not in out:
+                continue
+            val = out[combo]
+            if not isinstance(val, list):
+                continue
+            cleaned = [complete_json_schema_for_openai_tools(x) for x in val if x is not None]
+            if combo in ("anyOf", "oneOf") and not cleaned:
+                out.pop(combo, None)
+            else:
+                out[combo] = cleaned
+
+        if isinstance(out.get("properties"), dict):
+            t0 = out.get("type")
+            if t0 is None or (isinstance(t0, str) and not str(t0).strip()):
+                out["type"] = "object"
+
+        tags = _schema_type_tags(out.get("type"))
+        if "object" in tags:
+            if not isinstance(out.get("properties"), dict):
+                out["properties"] = {}
+            r = out.get("required")
+            if r is None or not isinstance(r, list):
+                out["required"] = []
+        if "array" in tags:
+            if out.get("items") is None:
+                out["items"] = {}
+            elif isinstance(out.get("items"), dict):
+                out["items"] = complete_json_schema_for_openai_tools(out["items"])
+        return out
+    if isinstance(obj, list):
+        return [complete_json_schema_for_openai_tools(x) for x in obj]
+    return obj
+
+
+def complete_openai_tools_wire_parameters(tools: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """Apply :func:`complete_json_schema_for_openai_tools` to each function's ``parameters``."""
+    out: list[dict[str, Any]] = []
+    for ent in tools or []:
+        if not isinstance(ent, dict):
+            continue
+        if str(ent.get("type") or "") != "function":
+            out.append(ent)
+            continue
+        fn = ent.get("function")
+        if not isinstance(fn, dict):
+            out.append(ent)
+            continue
+        params = fn.get("parameters")
+        if not isinstance(params, dict):
+            row = dict(ent)
+            row["function"] = {**dict(fn), "parameters": dict(_MIN_TOOL_PARAMETERS)}
+            out.append(row)
+            continue
+        row = dict(ent)
+        row["function"] = {**dict(fn), "parameters": complete_json_schema_for_openai_tools(dict(params))}
+        out.append(row)
+    return out
+
+
+def openai_tools_json_byte_length(tools: list[dict[str, Any]]) -> int:
+    try:
+        return len(json.dumps(tools, ensure_ascii=False, default=str))
+    except Exception:
+        return 0
+
+
+def shrink_openai_tools_payload_for_api(tools: list[dict[str, Any]], *, max_json_bytes: int) -> list[dict[str, Any]]:
+    """Shrink OpenAI ``tools`` JSON for strict OpenAI-compatible gateways."""
+    if not tools or max_json_bytes <= 0:
+        return tools
+    if openai_tools_json_byte_length(tools) <= max_json_bytes:
+        return tools
+    flat: list[dict[str, Any]] = []
+    for t in tools:
+        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
+        desc = str(fn.get("description") or "")
+        flat.append({"type": "function", "function": {"name": name, "description": desc, "parameters": dict(_MIN_TOOL_PARAMETERS)}})
+    if not flat:
+        return tools
+
+    def _serialized_with_desc_cap(cap: int) -> str:
+        trimmed: list[dict[str, Any]] = []
+        for x in flat:
+            fn0 = x["function"]
+            trimmed.append(
+                {
+                    "type": "function",
+                    "function": {
+                        "name": fn0["name"],
+                        "description": (fn0.get("description") or "")[: max(0, cap)],
+                        "parameters": fn0["parameters"],
+                    },
+                }
+            )
+        return json.dumps(trimmed, ensure_ascii=False, default=str)
+
+    lo, hi = 0, min(8000, max_json_bytes)
+    best_cap = 0
+    while lo <= hi:
+        mid = (lo + hi) // 2
+        s = _serialized_with_desc_cap(mid)
+        if len(s) <= max_json_bytes:
+            best_cap = mid
+            lo = mid + 1
+        else:
+            hi = mid - 1
+    out_obj: list[dict[str, Any]] = json.loads(_serialized_with_desc_cap(best_cap))
+    blob = json.dumps(out_obj, ensure_ascii=False, default=str)
+    dropped = 0
+    while len(out_obj) > 8 and len(blob) > max_json_bytes:
+        out_obj.pop()
+        dropped += 1
+        blob = json.dumps(out_obj, ensure_ascii=False, default=str)
+    logger.warning(
+        "Shrunk OpenAI tools[] for API size limits: count=%d max_json_bytes=%d desc_cap=%d dropped=%d final_bytes=%d",
+        len(out_obj),
+        max_json_bytes,
+        best_cap,
+        dropped,
+        len(blob),
+    )
+    return out_obj
+
+
+def default_max_openai_tools_json_bytes(base_url: str | None) -> int | None:
+    raw = str(os.getenv("AIA_OPENAI_TOOLS_MAX_JSON_CHARS") or "").strip()
+    if raw.isdigit():
+        return max(2000, int(raw))
+    u = (base_url or "").lower()
+    if "dashscope.aliyuncs.com" in u:
+        return 28000
+    if str(os.getenv("AIA_SHRINK_OPENAI_TOOLS") or "").strip().lower() in ("1", "true", "yes", "on"):
+        mx = str(os.getenv("AIA_SHRINK_OPENAI_TOOLS_MAX_JSON") or "28000").strip()
+        return max(2000, int(mx)) if mx.isdigit() else 28000
+    return None
+
+
+__all__ = [
+    "MIN_OPENAI_FUNCTION_PARAMETERS",
+    "complete_json_schema_for_openai_tools",
+    "complete_openai_tools_wire_parameters",
+    "default_max_openai_tools_json_bytes",
+    "openai_tools_json_byte_length",
+    "shrink_openai_tools_payload_for_api",
+]
+
+MIN_OPENAI_FUNCTION_PARAMETERS = _MIN_TOOL_PARAMETERS

svc/llm/tool_wire_policy.py

@@ -0,0 +1,787 @@
+"""Graduated OpenAI ``tools[]`` shaping + usage-aware omission / per-tool admin policies.
+
+Uses ``tool_log`` aggregates, ``app_setting`` for admin config / per-tool levels / penalty state.
+Wire payload differs from local :class:`~src.tools.base.ToolSpec`; validation still uses full schemas.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+from datetime import datetime, timedelta, timezone
+from typing import Any
+
+from svc.llm.tool_schema import MIN_OPENAI_FUNCTION_PARAMETERS, complete_openai_tools_wire_parameters
+
+logger = logging.getLogger(__name__)
+
+SETTINGS_KEY_PENALTY_STATE = "mcp_tool_wire_penalty_state"
+SETTINGS_KEY_TOOL_POLICIES = "mcp_tool_wire_tool_policies"
+SETTINGS_KEY_ADMIN_CONFIG = "mcp_tool_wire_admin_config"
+
+# Role-scoped (per expert) overrides. Shape:
+# - policies: { "<role>": { "mcp__server__tool": <int level> } }
+# - penalty:  { "<role>": { "mcp__server__tool": <penalty state dict> } }
+SETTINGS_KEY_PENALTY_STATE_BY_ROLE = "mcp_tool_wire_penalty_state_by_role"
+SETTINGS_KEY_TOOL_POLICIES_BY_ROLE = "mcp_tool_wire_tool_policies_by_role"
+SETTINGS_KEY_ROLE_MODE_BY_ROLE = "mcp_tool_wire_role_mode_by_role"
+
+_MIN_PARAMETERS: dict[str, Any] = dict(MIN_OPENAI_FUNCTION_PARAMETERS)
+
+
+def _env_prefixed(name_suffix: str, default: str = "") -> str:
+    raw = str(os.getenv(f"AIA_{name_suffix}") or "").strip()
+    if raw:
+        return raw
+    raw = str(os.getenv(f"OPS_{name_suffix}") or "").strip()
+    if raw:
+        return raw
+    return str(default).strip()
+
+
+def wire_policy_enabled(base_url: str | None) -> bool:
+    raw = _env_prefixed("MCP_WIRE_USAGE_POLICY", "").lower()
+    if raw in ("1", "true", "yes", "on"):
+        return True
+    if raw in ("0", "false", "no", "off"):
+        return False
+    u = (base_url or "").lower()
+    return "dashscope.aliyuncs.com" in u
+
+
+def penalty_disabled() -> bool:
+    return _env_prefixed("MCP_WIRE_PENALTY_DISABLE", "").lower() in ("1", "true", "yes", "on")
+
+
+def _admin_defaults_from_env() -> dict[str, Any]:
+    return {
+        "wire_policy": "inherit",
+        "top_n_full": int(_env_prefixed("MCP_WIRE_TOP_N_FULL", "20") or "20"),
+        "stale_hours": float(_env_prefixed("MCP_WIRE_STALE_HOURS", "3") or "3"),
+        "penalty_minutes": float(_env_prefixed("MCP_WIRE_PENALTY_MINUTES", "30") or "30"),
+        "medium_rank_start": int(_env_prefixed("MCP_WIRE_MEDIUM_RANK_START", "21") or "21"),
+        "medium_rank_end": int(_env_prefixed("MCP_WIRE_MEDIUM_RANK_END", "50") or "50"),
+        "medium_desc_chars": int(_env_prefixed("MCP_WIRE_MEDIUM_DESC_CHARS", "520") or "520"),
+        "minimal_desc_cap": int(_env_prefixed("MCP_WIRE_MINIMAL_DESC_CAP", "80") or "80"),
+        "penalty_disable": False,
+    }
+
+
+def load_merged_admin_config(store: Any) -> dict[str, Any]:
+    out = _admin_defaults_from_env()
+    try:
+        raw = store.get_setting(SETTINGS_KEY_ADMIN_CONFIG)
+        if not raw:
+            return out
+        d = json.loads(raw) if isinstance(raw, str) else raw
+        if isinstance(d, dict):
+            for k, v in d.items():
+                if k in out or k == "wire_policy":
+                    out[k] = v
+    except Exception:
+        pass
+    pd = out.get("penalty_disable")
+    if isinstance(pd, str):
+        out["penalty_disable"] = pd.strip().lower() in ("1", "true", "yes", "on")
+    return out
+
+
+def load_tool_policies_dict(store: Any) -> dict[str, int]:
+    try:
+        raw = store.get_setting(SETTINGS_KEY_TOOL_POLICIES) or "{}"
+        d = json.loads(raw) if isinstance(raw, str) else raw
+    except Exception:
+        return {}
+    if not isinstance(d, dict):
+        return {}
+    out: dict[str, int] = {}
+    for k, v in d.items():
+        nm = str(k or "").strip()
+        if not nm.startswith("mcp__"):
+            continue
+        try:
+            n = int(v)
+        except (TypeError, ValueError):
+            continue
+        if n == 9999:
+            out[nm] = 9999
+        elif n <= 0:
+            out[nm] = 0
+        else:
+            out[nm] = min(n, 9998)
+    return out
+
+
+def load_tool_policies_dict_for_role(store: Any, *, role: str | None) -> dict[str, int]:
+    """Role-scoped policies; falls back to global if absent/unparseable."""
+    r = str(role or "").strip().lower()
+    if not r:
+        return load_tool_policies_dict(store)
+    try:
+        raw = store.get_setting(SETTINGS_KEY_TOOL_POLICIES_BY_ROLE) or "{}"
+        outer = json.loads(raw) if isinstance(raw, str) else raw
+    except Exception:
+        outer = {}
+    if not isinstance(outer, dict):
+        return load_tool_policies_dict(store)
+    inner = outer.get(r)
+    if not isinstance(inner, dict):
+        return load_tool_policies_dict(store)
+    # Reuse same coercion rules as global loader.
+    out: dict[str, int] = {}
+    for k, v in inner.items():
+        nm = str(k or "").strip()
+        if not nm.startswith("mcp__"):
+            continue
+        try:
+            n = int(v)
+        except (TypeError, ValueError):
+            continue
+        if n == 9999:
+            out[nm] = 9999
+        elif n <= 0:
+            out[nm] = 0
+        else:
+            out[nm] = min(n, 9998)
+    return out
+
+
+def load_penalty_state_for_role(store: Any, *, role: str | None) -> dict[str, Any]:
+    """Role-scoped penalty state; falls back to global if absent/unparseable."""
+    r = str(role or "").strip().lower()
+    if not r:
+        raw_pen = store.get_setting(SETTINGS_KEY_PENALTY_STATE) or "{}"
+        try:
+            pen = json.loads(raw_pen) if isinstance(raw_pen, str) else {}
+        except Exception:
+            pen = {}
+        if not isinstance(pen, dict):
+            pen = {}
+        return migrate_legacy_penalty_store(pen)
+    try:
+        raw = store.get_setting(SETTINGS_KEY_PENALTY_STATE_BY_ROLE) or "{}"
+        outer = json.loads(raw) if isinstance(raw, str) else raw
+    except Exception:
+        outer = {}
+    if not isinstance(outer, dict):
+        return load_penalty_state_for_role(store, role=None)
+    inner = outer.get(r)
+    if not isinstance(inner, dict):
+        return load_penalty_state_for_role(store, role=None)
+    return migrate_legacy_penalty_store(inner)
+
+
+def _persist_penalty_state_for_role(store: Any, *, role: str | None, penalty_state: dict[str, Any]) -> None:
+    r = str(role or "").strip().lower()
+    clean: dict[str, Any] = {}
+    for k, v in (penalty_state or {}).items():
+        if v is None:
+            continue
+        if isinstance(v, dict):
+            clean[str(k)] = v
+    if not r:
+        store.set_setting(SETTINGS_KEY_PENALTY_STATE, json.dumps(clean, ensure_ascii=False))
+        return
+    try:
+        raw = store.get_setting(SETTINGS_KEY_PENALTY_STATE_BY_ROLE) or "{}"
+        outer = json.loads(raw) if isinstance(raw, str) else raw
+    except Exception:
+        outer = {}
+    if not isinstance(outer, dict):
+        outer = {}
+    outer[r] = clean
+    store.set_setting(SETTINGS_KEY_PENALTY_STATE_BY_ROLE, json.dumps(outer, ensure_ascii=False))
+
+
+def wire_graduation_effective(base_url: str | None, admin: dict[str, Any]) -> bool:
+    mode = str(admin.get("wire_policy") or "inherit").strip().lower()
+    if mode in ("always", "on", "true", "1", "yes"):
+        return True
+    if mode in ("never", "off", "false", "0", "no"):
+        return False
+    return wire_policy_enabled(base_url)
+
+
+def penalty_effective_disabled(admin: dict[str, Any]) -> bool:
+    if bool(admin.get("penalty_disable")):
+        return True
+    return penalty_disabled()
+
+
+def migrate_legacy_penalty_store(raw: dict[str, Any]) -> dict[str, Any]:
+    """Older rows used ``omit_until`` without ``phase``; add ``kind`` for global tier."""
+    out: dict[str, Any] = {}
+    for k, v in (raw or {}).items():
+        if not isinstance(v, dict):
+            continue
+        v2 = dict(v)
+        if str(v2.get("phase") or "") in ("active", "done"):
+            if "kind" not in v2:
+                v2["kind"] = "d"
+            out[str(k)] = v2
+            continue
+        ou_s = str(v2.get("omit_until") or "")
+        ws = str(v2.get("wave_ts") or "")
+        ou = _parse_iso(ou_s)
+        if ou_s and ou:
+            if _utc_now() < ou:
+                v2 = {"phase": "active", "omit_until": ou_s, "wave_ts": ws, "kind": v2.get("kind") or "d"}
+            else:
+                v2 = {"phase": "done", "wave_ts": ws, "kind": v2.get("kind") or "d"}
+        elif ws:
+            v2 = {"phase": "done", "wave_ts": ws, "kind": v2.get("kind") or "d"}
+        else:
+            continue
+        out[str(k)] = v2
+    return out
+
+
+def _parse_iso(ts: str) -> datetime | None:
+    s = str(ts or "").strip()
+    if not s:
+        return None
+    try:
+        if s.endswith("Z"):
+            s = s[:-1] + "+00:00"
+        return datetime.fromisoformat(s)
+    except Exception:
+        return None
+
+
+def _utc_now() -> datetime:
+    return datetime.now(timezone.utc)
+
+
+def _wire_json_size(tools: list[dict[str, Any]]) -> int:
+    try:
+        return len(json.dumps(tools, ensure_ascii=False, default=str))
+    except Exception:
+        return 0
+
+
+def _strip_nested_descriptions(obj: Any) -> Any:
+    if isinstance(obj, dict):
+        out: dict[str, Any] = {}
+        for k, v in obj.items():
+            if str(k) == "description":
+                continue
+            out[str(k)] = _strip_nested_descriptions(v)
+        return out
+    if isinstance(obj, list):
+        return [_strip_nested_descriptions(x) for x in obj]
+    return obj
+
+
+def _tier_medium(fn: dict[str, Any], *, cfg: dict[str, Any]) -> dict[str, Any]:
+    mid_desc_len = int(cfg.get("medium_desc_chars") or 520)
+    mid_desc_len = max(80, min(mid_desc_len, 4000))
+    desc = str(fn.get("description") or "")[:mid_desc_len]
+    params = fn.get("parameters")
+    if isinstance(params, dict):
+        params_out = dict(_strip_nested_descriptions(params))
+    else:
+        params_out = dict(_MIN_PARAMETERS)
+    return {"name": fn.get("name"), "description": desc, "parameters": params_out}
+
+
+def _tier_minimal(fn: dict[str, Any], *, desc_cap: int) -> dict[str, Any]:
+    desc = str(fn.get("description") or "")[: max(0, desc_cap)]
+    return {"name": fn.get("name"), "description": desc, "parameters": dict(_MIN_PARAMETERS)}
+
+
+def _clone_tool_entry(t: dict[str, Any]) -> dict[str, Any]:
+    try:
+        return json.loads(json.dumps(t, ensure_ascii=False, default=str))
+    except Exception:
+        return dict(t)
+
+
+def _is_builtin_tool_name(name: str) -> bool:
+    return not str(name).startswith("mcp__")
+
+
+def _normalize_role_mode(v: Any) -> str:
+    s = str(v or "").strip().lower()
+    if s in {"unrestricted", "forbidden", "restricted"}:
+        return s
+    return "restricted"
+
+
+def load_role_mode_for_role(store: Any, *, role: str | None) -> str:
+    r = str(role or "").strip().lower()
+    if not r:
+        return "restricted"
+    try:
+        raw = store.get_setting(SETTINGS_KEY_ROLE_MODE_BY_ROLE) or "{}"
+        obj = json.loads(raw) if isinstance(raw, str) else raw
+    except Exception:
+        obj = {}
+    if not isinstance(obj, dict):
+        return "restricted"
+    return _normalize_role_mode(obj.get(r))
+
+
+def _is_stale_hours(last_ts_iso: str | None, stale_hours: float) -> bool:
+    if not last_ts_iso:
+        return False
+    dt = _parse_iso(last_ts_iso)
+    if dt is None:
+        return False
+    return (_utc_now() - dt).total_seconds() > stale_hours * 3600.0
+
+
+def _idle_minutes_exceeded(last_ts_iso: str | None, idle_minutes: float) -> bool:
+    if not last_ts_iso:
+        return False
+    dt = _parse_iso(last_ts_iso)
+    if dt is None:
+        return False
+    return (_utc_now() - dt).total_seconds() > idle_minutes * 60.0
+
+
+def _wave_penalty_omit(
+    name: str,
+    last_ts: str | None,
+    idle_hit: bool,
+    penalty_minutes: float,
+    st_nm: Any,
+    out_penalty_updates: dict[str, Any],
+    *,
+    kind: str,
+) -> bool:
+    pw = last_ts or ""
+    mins = max(1.0, min(float(penalty_minutes), 24 * 60))
+    if not idle_hit:
+        return False
+    prev = out_penalty_updates.get(name)
+    if isinstance(prev, dict) and str(prev.get("kind") or "") != str(kind):
+        out_penalty_updates.pop(name, None)
+        st_nm = None
+    else:
+        st_nm = prev
+
+    omit = False
+    if isinstance(st_nm, dict) and str(st_nm.get("phase") or "") == "active":
+        ou = _parse_iso(str(st_nm.get("omit_until") or ""))
+        if ou and _utc_now() < ou:
+            omit = True
+        else:
+            out_penalty_updates[name] = {"phase": "done", "wave_ts": pw, "kind": str(kind)}
+    elif isinstance(st_nm, dict) and str(st_nm.get("phase") or "") == "done" and str(st_nm.get("wave_ts") or "") == pw:
+        pass
+    else:
+        until = (_utc_now() + timedelta(minutes=mins)).isoformat()
+        out_penalty_updates[name] = {"phase": "active", "omit_until": until, "wave_ts": pw, "kind": str(kind)}
+        omit = True
+    return omit
+
+
+def filter_permanent_ban_mcp_tools(tools: list[dict[str, Any]], policies: dict[str, int]) -> list[dict[str, Any]]:
+    out: list[dict[str, Any]] = []
+    for entry in tools:
+        if not isinstance(entry, dict) or str(entry.get("type") or "") != "function":
+            out.append(entry)
+            continue
+        fn = entry.get("function")
+        if not isinstance(fn, dict):
+            out.append(entry)
+            continue
+        name = str(fn.get("name") or "").strip()
+        if not name or _is_builtin_tool_name(name):
+            out.append(entry)
+            continue
+        if policies.get(name) == 9999:
+            continue
+        out.append(entry)
+    return out
+
+
+def prepare_openai_tools_for_llm_api(
+    tools: list[dict[str, Any]],
+    *,
+    base_url: str | None,
+    max_json_bytes: int | None,
+    store: Any | None = None,
+    role: str | None = None,
+) -> list[dict[str, Any]]:
+    """Build wire ``tools[]``: optional graduated policy + byte cap."""
+
+    def _fallback_shrink(raw: list[dict[str, Any]], cap: int) -> list[dict[str, Any]]:
+        from svc.llm.tool_schema import shrink_openai_tools_payload_for_api
+
+        return shrink_openai_tools_payload_for_api(raw, max_json_bytes=cap)
+
+    if not tools:
+        return tools
+
+    tools = complete_openai_tools_wire_parameters(tools)
+
+    def _finalize(result: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        """Re-apply completion after tier/shrink/squeeze (those paths emit fresh ``parameters`` dicts)."""
+        return complete_openai_tools_wire_parameters(result)
+
+    policies: dict[str, int] = {}
+    admin: dict[str, Any] = _admin_defaults_from_env()
+    role_mode = "restricted"
+    try:
+        if store is None:
+            from svc.config.paths import db_path
+            from svc.persistence.sqlite_store import SqliteStore
+
+            store = SqliteStore(db_path())
+        # Role-scoped policies if configured; otherwise fall back to global.
+        policies = load_tool_policies_dict_for_role(store, role=str(role or "").strip().lower() or None)
+        admin = load_merged_admin_config(store)
+        role_mode = load_role_mode_for_role(store, role=str(role or "").strip().lower() or None)
+    except Exception as exc:
+        logger.warning("tool_wire_policy: load admin/policies (%s)", exc)
+
+    def _filter_mcp_out(raw: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        out: list[dict[str, Any]] = []
+        for entry in raw:
+            if not isinstance(entry, dict) or str(entry.get("type") or "") != "function":
+                out.append(entry)
+                continue
+            fn = entry.get("function")
+            if not isinstance(fn, dict):
+                out.append(entry)
+                continue
+            nm = str(fn.get("name") or "").strip()
+            if nm.startswith("mcp__"):
+                continue
+            out.append(entry)
+        return out
+
+    tools_pass1 = filter_permanent_ban_mcp_tools(tools, policies)
+
+    if role_mode == "forbidden":
+        out = _filter_mcp_out(tools)
+        if max_json_bytes is not None and max_json_bytes > 0:
+            return _finalize(_fallback_shrink(out, max_json_bytes))
+        return _finalize(out)
+    if role_mode == "unrestricted":
+        if max_json_bytes is not None and max_json_bytes > 0:
+            return _finalize(_fallback_shrink(tools_pass1, max_json_bytes))
+        return _finalize(tools_pass1)
+
+    if not wire_graduation_effective(base_url, admin):
+        if max_json_bytes is not None and max_json_bytes > 0:
+            return _finalize(_fallback_shrink(tools_pass1, max_json_bytes))
+        return _finalize(tools_pass1)
+
+    top_n = int(admin.get("top_n_full") or 20)
+    top_n = max(3, min(top_n, 80))
+    stale_h = float(admin.get("stale_hours") or 3)
+    stale_h = max(0.25, min(stale_h, 720.0))
+    penalty_min_global = float(admin.get("penalty_minutes") or 30)
+    penalty_min_global = max(1.0, min(penalty_min_global, 24 * 60))
+    medium_floor = int(admin.get("medium_rank_start") or 21)
+    medium_ceiling = int(admin.get("medium_rank_end") or 50)
+    if medium_floor > medium_ceiling:
+        medium_floor, medium_ceiling = medium_ceiling, medium_floor
+
+    usage_map: dict[str, dict[str, Any]] = {}
+    penalty_state: dict[str, Any] = {}
+    try:
+        usage_map = store.list_mcp_tool_aggregate_usage()
+        penalty_state = load_penalty_state_for_role(store, role=str(role or "").strip().lower() or None)
+    except Exception as exc:
+        logger.warning("tool_wire_policy: usage store unavailable (%s); fallback shrink only", exc)
+        if max_json_bytes is not None and max_json_bytes > 0:
+            return _finalize(_fallback_shrink(tools_pass1, max_json_bytes))
+        return _finalize(tools_pass1)
+
+    ranked = sorted(
+        usage_map.items(),
+        key=lambda kv: (-int(kv[1].get("count") or 0), str(kv[1].get("last_ts") or "")),
+    )
+    rank_by_name = {nm: idx + 1 for idx, (nm, _) in enumerate(ranked)}
+    top_full = {nm for nm, _ in ranked[:top_n]}
+
+    out_penalty_updates: dict[str, Any] = dict(penalty_state)
+    omit_set: set[str] = set()
+
+    for nm in list(out_penalty_updates.keys()):
+        meta = usage_map.get(nm)
+        cur_ts = str(meta.get("last_ts") or "") if meta else ""
+        st = out_penalty_updates.get(nm)
+        if not isinstance(st, dict):
+            continue
+        w_ts = str(st.get("wave_ts") or "")
+        if w_ts and cur_ts and w_ts != cur_ts:
+            out_penalty_updates.pop(nm, None)
+
+    pen_glob = penalty_effective_disabled(admin)
+    built: list[dict[str, Any]] = []
+    for entry in tools_pass1:
+        if not isinstance(entry, dict) or str(entry.get("type") or "") != "function":
+            built.append(entry)
+            continue
+        fn = entry.get("function")
+        if not isinstance(fn, dict):
+            built.append(entry)
+            continue
+        name = str(fn.get("name") or "").strip()
+        if not name:
+            built.append(entry)
+            continue
+
+        if _is_builtin_tool_name(name):
+            built.append(_clone_tool_entry(entry))
+            continue
+
+        usage = usage_map.get(name) or {}
+        last_ts = str(usage.get("last_ts") or "") or None
+        rk = rank_by_name.get(name)
+
+        policy_lv = policies.get(name)
+        st_nm = out_penalty_updates.get(name)
+
+        # policy_lv semantics (punishment only):
+        # - None: inherit global graduation + idle penalty.
+        # - 0: exempt this tool from idle penalty (but still participates in graduation tiers).
+        # - 1..9998: per-tool idle/penalty tuning (N * 10 minutes) + still participates in graduation tiers.
+        # - 9999: permanent ban (filtered earlier in filter_permanent_ban_mcp_tools).
+        omit = False
+        if policy_lv == 0:
+            omit = False
+        elif policy_lv is not None and policy_lv >= 1:
+            idle_m = float(policy_lv) * 10.0
+            pen_m = float(policy_lv) * 10.0
+            idle_hit = _idle_minutes_exceeded(last_ts, idle_m)
+            kind = f"L{int(policy_lv)}"
+            if not pen_glob:
+                omit = _wave_penalty_omit(
+                    name, last_ts, idle_hit, pen_m, st_nm, out_penalty_updates, kind=kind
+                )
+        else:
+            idle_hit = _is_stale_hours(last_ts, stale_h)
+            exempt = name in top_full
+            if not pen_glob and idle_hit and not exempt:
+                omit = _wave_penalty_omit(
+                    name,
+                    last_ts,
+                    idle_hit,
+                    penalty_min_global,
+                    st_nm,
+                    out_penalty_updates,
+                    kind="d",
+                )
+
+        if omit:
+            omit_set.add(name)
+            continue
+
+        tier = "minimal"
+        if name in top_full:
+            tier = "full"
+        elif rk is not None and medium_floor <= rk <= medium_ceiling:
+            tier = "medium"
+        else:
+            tier = "minimal"
+
+        if tier == "full":
+            built.append(_clone_tool_entry(entry))
+        elif tier == "medium":
+            mf = _tier_medium(fn, cfg=admin)
+            built.append({"type": "function", "function": mf})
+        else:
+            cap0 = int(admin.get("minimal_desc_cap") or 80)
+            cap0 = max(0, min(cap0, 2000))
+            mf = _tier_minimal(fn, desc_cap=cap0)
+            built.append({"type": "function", "function": mf})
+
+    try:
+        _persist_penalty_state_for_role(
+            store,
+            role=str(role or "").strip().lower() or None,
+            penalty_state=out_penalty_updates,
+        )
+    except Exception as exc:
+        logger.warning("tool_wire_policy: persist penalty state failed: %s", exc)
+
+    if omit_set:
+        logger.info(
+            "tool_wire_policy: omitted %d MCP tools during penalty window: %s",
+            len(omit_set),
+            sorted(omit_set)[:12],
+        )
+
+    if max_json_bytes is not None and max_json_bytes > 0 and _wire_json_size(built) > max_json_bytes:
+        return _finalize(_squeeze_to_budget(built, max_json_bytes, admin=admin))
+
+    return _finalize(built)
+
+
+def _squeeze_to_budget(
+    built: list[dict[str, Any]], max_json_bytes: int, *, admin: dict[str, Any]
+) -> list[dict[str, Any]]:
+    work = json.loads(json.dumps(built, ensure_ascii=False, default=str))
+
+    def sz() -> int:
+        return _wire_json_size(work)
+
+    if sz() <= max_json_bytes:
+        return work
+    for i, ent in enumerate(work):
+        fn = ent.get("function") if isinstance(ent.get("function"), dict) else {}
+        ps = fn.get("parameters")
+        if isinstance(ps, dict) and len(json.dumps(ps, ensure_ascii=False)) > 400:
+            nm = str(fn.get("name") or "")
+            if nm.startswith("mcp__"):
+                cap = int(admin.get("minimal_desc_cap") or 80)
+                work[i] = {
+                    "type": "function",
+                    "function": _tier_minimal(fn, desc_cap=max(0, cap)),
+                }
+        if sz() <= max_json_bytes:
+            return work
+
+    for cap in (60, 40, 24, 12, 8, 4, 0):
+        for i, ent in enumerate(work):
+            fn = ent.get("function") if isinstance(ent.get("function"), dict) else {}
+            nm = str(fn.get("name") or "")
+            if nm.startswith("mcp__"):
+                work[i] = {"type": "function", "function": _tier_minimal(fn, desc_cap=cap)}
+        if sz() <= max_json_bytes:
+            return work
+
+    idxs = [i for i, e in enumerate(work) if str((e.get("function") or {}).get("name") or "").startswith("mcp__")]
+    for i in reversed(idxs):
+        if sz() <= max_json_bytes:
+            break
+        work.pop(i)
+
+    if sz() > max_json_bytes:
+        from svc.llm.tool_schema import shrink_openai_tools_payload_for_api
+
+        work = shrink_openai_tools_payload_for_api(work, max_json_bytes=max_json_bytes)
+    return work
+
+
+def penalty_row_status(
+    *,
+    wire_name: str,
+    policy_level: int | None,
+    penalty: dict[str, Any] | None,
+    last_ts: str | None,
+) -> dict[str, Any]:
+    """Human-readable row for Admin UI."""
+    if policy_level is None:
+        return {
+            "phase": "inherit",
+            "omit_until": None,
+            "unblock_hint": "未配置：走全局线侧与闲置惩罚（与显式写入 0 不同）",
+            "kind": None,
+        }
+    if policy_level == 9999:
+        return {
+            "phase": "permanent_ban",
+            "omit_until": None,
+            "unblock_hint": "永久不上送（需将策略改为非 9999）",
+            "kind": None,
+        }
+    if policy_level == 0:
+        return {
+            "phase": "exempt",
+            "omit_until": None,
+            "unblock_hint": "策略 0：不参与闲置惩罚（仍可能受线侧分层/压缩影响）",
+            "kind": None,
+        }
+    st = penalty.get(wire_name) if isinstance(penalty, dict) else None
+    if not isinstance(st, dict):
+        return {"phase": "none", "omit_until": None, "unblock_hint": "-", "kind": None}
+    ph = str(st.get("phase") or "")
+    ou = str(st.get("omit_until") or "")
+    kd = str(st.get("kind") or "")
+    if ph == "active" and ou:
+        dt = _parse_iso(ou)
+        hint = f"惩罚中，预计解封（omit_until）: {ou}"
+        if dt and _utc_now() >= dt:
+            hint = "惩罚窗口应已结束；下次请求会写入 done"
+        return {"phase": "active", "omit_until": ou, "unblock_hint": hint, "kind": kd}
+    if ph == "done":
+        return {
+            "phase": "done",
+            "omit_until": None,
+            "unblock_hint": "本轮已服完刑；再次闲置达到阈值后会重新惩罚",
+            "kind": kd,
+        }
+    return {"phase": ph or "unknown", "omit_until": ou or None, "unblock_hint": "-", "kind": kd or None}
+
+
+def build_tool_wire_snapshot(store: Any, *, role: str | None = None) -> dict[str, Any]:
+    """Aggregate MCP install list + usage + policies for Admin GET."""
+    from runtime.tools.mcp.registry import McpRegistry
+
+    admin = load_merged_admin_config(store)
+    role_norm = str(role or "").strip().lower() or None
+    role_mode = load_role_mode_for_role(store, role=role_norm)
+    policies = load_tool_policies_dict_for_role(store, role=role_norm)
+    pen = load_penalty_state_for_role(store, role=role_norm)
+    usage = store.list_mcp_tool_aggregate_usage()
+    servers = McpRegistry(store).list_servers(enabled_only=False)
+    tools_out: list[dict[str, Any]] = []
+    for s in servers:
+        sid = str(s.get("server_id") or "").strip()
+        if not sid:
+            continue
+        for t in store.list_mcp_server_tools(server_id=sid):
+            short = str(t.get("tool_name") or "").strip()
+            if not short:
+                continue
+            wire = f"mcp__{sid}__{short}"
+            agg = usage.get(wire) or {}
+            in_db = wire in policies
+            lv = int(policies[wire]) if in_db else None
+            st = penalty_row_status(
+                wire_name=wire,
+                policy_level=lv,
+                penalty=pen,
+                last_ts=str(agg.get("last_ts") or "") or None,
+            )
+            tools_out.append(
+                {
+                    "server_id": sid,
+                    "mcp_tool_name": short,
+                    "wire_name": wire,
+                    "policy_level": lv,
+                    "policy_in_db": in_db,
+                    "count": int(agg.get("count") or 0),
+                    "last_ts": str(agg.get("last_ts") or ""),
+                    "penalty": st,
+                    "raw_penalty": pen.get(wire) if isinstance(pen.get(wire), dict) else None,
+                }
+            )
+    tools_out.sort(key=lambda x: (x["server_id"], x["wire_name"]))
+    return {
+        "ok": True,
+        "role": role_norm or "",
+        "role_mode": role_mode,
+        "config": admin,
+        "policies": policies,
+        "penalty_state": pen,
+        "tools": tools_out,
+    }
+
+
+__all__ = [
+    "SETTINGS_KEY_PENALTY_STATE",
+    "SETTINGS_KEY_TOOL_POLICIES",
+    "SETTINGS_KEY_ADMIN_CONFIG",
+    "SETTINGS_KEY_PENALTY_STATE_BY_ROLE",
+    "SETTINGS_KEY_TOOL_POLICIES_BY_ROLE",
+    "SETTINGS_KEY_ROLE_MODE_BY_ROLE",
+    "build_tool_wire_snapshot",
+    "filter_permanent_ban_mcp_tools",
+    "load_merged_admin_config",
+    "load_tool_policies_dict",
+    "load_tool_policies_dict_for_role",
+    "load_penalty_state_for_role",
+    "load_role_mode_for_role",
+    "migrate_legacy_penalty_store",
+    "penalty_row_status",
+    "prepare_openai_tools_for_llm_api",
+    "wire_policy_enabled",
+]

svc/llm/transports/__init__.py

@@ -0,0 +1 @@
+# svc.llm.transports package

svc/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 svc.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"]
+

svc/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())
+

svc/llm/transports/google_gemini_sse.py

@@ -0,0 +1,267 @@
+from __future__ import annotations
+
+import json
+import os
+import uuid
+from typing import Any, Optional
+from collections.abc import Callable
+
+from svc.llm.transports.base import ChatModel, LLMResponse, LLMToolCall
+
+
+class GoogleGeminiChatModel(ChatModel):
+    """Native Google Gemini streaming transport (SSE), modeled after Oclaw'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 ""),
+                    # Oclaw 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 (Oclaw-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)
+                                    if on_token:
+                                        on_token(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"]
+

svc/llm/transports/openai_chat_completions.py

@@ -0,0 +1,728 @@
+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 svc.llm.tool_schema import complete_openai_tools_wire_parameters
+from svc.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 _is_minimax_compat(model: str | None, base_url: str | None) -> bool:
+    m = str(model or "").strip().lower()
+    b = str(base_url or "").strip().lower()
+    return ("minimax" in m) or ("minimax" in b)
+
+
+def _truthy_env(name: str, default: str = "") -> bool:
+    v = os.getenv(name)
+    if v is None:
+        v = default
+    return str(v or "").strip().lower() in ("1", "true", "yes", "on")
+
+
+def _should_disable_thinking(base_url: str | None) -> bool:
+    # Some OpenAI-compatible gateways enable "thinking" mode and require replaying
+    # reasoning_content in subsequent turns. When the client does not preserve it,
+    # the gateway returns HTTP 400. Allow disabling thinking at request level.
+    #
+    # Safety: do NOT send unknown fields to official OpenAI endpoints by default.
+    if _truthy_env("AIA_LLM_THINKING_FORCE_DISABLED", "0"):
+        return True
+    if _truthy_env("AIA_LLM_THINKING_FORCE_ENABLED", "0"):
+        return False
+    b = str(base_url or "").strip().lower()
+    if not b:
+        return False
+    if "api.openai.com" in b:
+        return False
+    # Default-on for non-official OpenAI-compatible gateways.
+    return _truthy_env("AIA_LLM_THINKING_DISABLED", "1")
+
+
+def _should_enable_thinking(base_url: str | None, *, thinking_mode_enabled: bool = False) -> bool:
+    if _truthy_env("AIA_LLM_THINKING_FORCE_ENABLED", "0"):
+        return True
+    if _truthy_env("AIA_LLM_THINKING_FORCE_DISABLED", "0"):
+        return False
+    if not bool(thinking_mode_enabled):
+        return False
+    b = str(base_url or "").strip().lower()
+    if not b:
+        return False
+    if "api.openai.com" in b:
+        return False
+    return True
+
+
+def _should_force_text_only_messages(base_url: str | None) -> bool:
+    del base_url
+    # Opt-in only. Avoid hard-coding vendor/domain specific heuristics.
+    return _truthy_env("AIA_OPENAI_FORCE_TEXT_CONTENT", "0")
+
+
+def _multimodal_proactive_downgrade_enabled() -> bool:
+    raw = str(os.getenv("AIA_OPENAI_MULTIMODAL_PROACTIVE_DOWNGRADE") or "1").strip().lower()
+    return raw not in {"0", "false", "no", "off"}
+
+
+def _text_only_multimodal_substrings() -> tuple[str, ...]:
+    raw = os.getenv("AIA_OPENAI_TEXT_ONLY_MULTIMODAL_SUBSTRINGS")
+    if raw is None:
+        return ("deepseek",)
+    s = str(raw).strip().lower()
+    if s in {"", "-", "none", "off"}:
+        return ()
+    return tuple(part.strip().lower() for part in str(raw).split(",") if part.strip())
+
+
+def _messages_contain_list_with_image(messages: list[dict[str, Any]]) -> bool:
+    for m in messages or []:
+        if not isinstance(m, dict):
+            continue
+        c = m.get("content")
+        if not isinstance(c, list):
+            continue
+        for item in c:
+            if not isinstance(item, dict):
+                continue
+            typ = str(item.get("type") or "").strip().lower()
+            if typ in ("image_url", "input_image"):
+                return True
+    return False
+
+
+def _deepseek_strict_tool_mode_enabled(*, base_url: str | None, model: str | None) -> bool:
+    """DeepSeek OpenAI-compat: optional ``strict_tool_mode`` in request body (via ``extra_body``).
+
+    Upstream documents this as reducing DSML-in-text tool output in favor of native ``tool_calls``.
+    Opt out with ``AIA_DEEPSEEK_STRICT_TOOL_MODE=0`` when a proxy rejects unknown fields.
+    """
+    raw = str(os.getenv("AIA_DEEPSEEK_STRICT_TOOL_MODE") or "").strip().lower()
+    if raw in ("0", "false", "no", "off"):
+        return False
+    hay = f"{base_url or ''} {model or ''}".lower()
+    return "deepseek" in hay
+
+
+def _merge_deepseek_strict_tool_mode_extra_body(
+    kwargs: dict[str, Any],
+    *,
+    base_url: str | None,
+    model: str | None,
+    use_tools: bool,
+) -> None:
+    if not use_tools or not _deepseek_strict_tool_mode_enabled(base_url=base_url, model=model):
+        return
+    extra_body = kwargs.get("extra_body") if isinstance(kwargs.get("extra_body"), dict) else {}
+    extra_body = dict(extra_body)
+    extra_body["strict_tool_mode"] = True
+    kwargs["extra_body"] = extra_body
+
+
+def _should_proactively_downgrade_multimodal_messages(
+    messages: list[dict[str, Any]],
+    *,
+    model: str | None,
+    base_url: str | None,
+) -> bool:
+    if _should_force_text_only_messages(base_url):
+        return _messages_contain_list_with_image(messages)
+    if not _multimodal_proactive_downgrade_enabled():
+        return False
+    if not _messages_contain_list_with_image(messages):
+        return False
+    parts = _text_only_multimodal_substrings()
+    if not parts:
+        return False
+    hay = f"{model or ''} {base_url or ''}".lower()
+    return any(p and p in hay for p in parts)
+
+
+def _wire_error_message(exc: BaseException) -> str:
+    chunks: list[str] = [str(exc)]
+    em = getattr(exc, "message", None)
+    if isinstance(em, str) and em.strip():
+        chunks.append(em)
+    body = getattr(exc, "body", None)
+    if isinstance(body, dict):
+        try:
+            chunks.append(json.dumps(body, ensure_ascii=False, default=str))
+        except Exception:
+            chunks.append(repr(body))
+    elif isinstance(body, str) and body.strip():
+        chunks.append(body)
+    resp = getattr(exc, "response", None)
+    if resp is not None:
+        t = ""
+        try:
+            rt = getattr(resp, "text", "")
+            if isinstance(rt, str):
+                t = rt
+        except Exception:
+            pass
+        if not t.strip():
+            try:
+                ct = getattr(resp, "content", None)
+                if isinstance(ct, (bytes, bytearray)):
+                    t = bytes(ct).decode("utf-8", errors="replace")
+            except Exception:
+                pass
+        if t.strip():
+            chunks.append(t)
+    inner = getattr(exc, "__cause__", None)
+    if inner is not None:
+        chunks.append(str(inner))
+    return "\n".join(chunks)
+
+
+def _is_text_only_gateway_error(msg: str) -> bool:
+    m = str(msg or "").strip().lower()
+    if not m:
+        return False
+    if "unknown variant image_url" in m:
+        return True
+    if "image_url" in m and "expected text" in m:
+        return True
+    if "failed to deserialize" in m and ("expected text" in m or "unknown variant image_url" in m):
+        return True
+    if "deserialize" in m and "image_url" in m and ("expected text" in m or "messages[" in m):
+        return True
+    if "invalid_request_error" in m and ("image_url" in m or "expected text" in m):
+        return True
+    return False
+
+
+def _flatten_message_content_to_text(content: Any) -> str:
+    if isinstance(content, str):
+        return content
+    if not isinstance(content, list):
+        return str(content or "")
+    parts: list[str] = []
+    for item in content:
+        if isinstance(item, str):
+            if item.strip():
+                parts.append(item)
+            continue
+        if not isinstance(item, dict):
+            continue
+        typ = str(item.get("type") or "").strip().lower()
+        if typ in ("text", "input_text"):
+            txt = str(item.get("text") or item.get("content") or "").strip()
+            if txt:
+                parts.append(txt)
+            continue
+        if typ in ("image_url", "input_image"):
+            # Proactive / forced text-only path without OCR injection (or OCR disabled).
+            parts.append(
+                "【图片】本消息含图片，但当前以纯文本发往上游模型，像素未传入；"
+                "若需要图中细节，请改用支持多模态的模型或开启看图 OCR 降级通道。"
+            )
+            continue
+    return "\n".join([p for p in parts if p]).strip()
+
+
+def _multimodal_downgrade_ocr_enabled() -> bool:
+    """When True, text-only downgrade replaces image blocks with OCR text via AIA_OCR_* lane."""
+    return str(os.getenv("AIA_MULTIMODAL_DOWNGRADE_OCR") or "1").strip().lower() not in {
+        "0",
+        "false",
+        "no",
+        "off",
+    }
+
+
+def _image_block_to_url(item: dict[str, Any]) -> str | None:
+    typ = str(item.get("type") or "").strip().lower()
+    if typ == "image_url":
+        u = item.get("image_url")
+        if isinstance(u, dict):
+            s = str(u.get("url") or "").strip()
+            return s or None
+        if isinstance(u, str) and u.strip():
+            return u.strip()
+    if typ == "input_image":
+        b64 = normalize_image_b64_payload(item.get("image_base64") or item.get("data"))
+        mime = str(item.get("mime") or "image/jpeg").strip() or "image/jpeg"
+        if b64:
+            return f"data:{mime};base64,{b64}"
+    return None
+
+
+def _message_list_contains_image_block(content: Any) -> bool:
+    if not isinstance(content, list):
+        return False
+    for item in content:
+        if not isinstance(item, dict):
+            continue
+        typ = str(item.get("type") or "").strip().lower()
+        if typ in ("image_url", "input_image"):
+            return True
+    return False
+
+
+def _downgrade_ocr_wrap_for_llm(ocr_text: str) -> str:
+    """User-facing copy injected into the main model when multimodal was OCR-downgraded to text."""
+    head = (
+        "【图片内容·OCR】当前主对话模型仅接收纯文本，无法直接读图；"
+        "下面是由独立看图通道从用户图片中转写提取的正文。请把它和用户的文字问题一起当作依据来回答；"
+        "转写可能有漏字、错行或与截图不完全一致，重要结论可请用户核对原图或补充说明。"
+    )
+    return f"{head}\n\n{ocr_text.strip()}"
+
+
+def _flatten_message_content_for_text_gateway(content: Any) -> str:
+    """Flatten list content to a string; image blocks may become OCR text when configured."""
+    if isinstance(content, str):
+        return content
+    if not isinstance(content, list):
+        return str(content or "")
+    if _multimodal_downgrade_ocr_enabled() and _message_list_contains_image_block(content):
+        try:
+            from svc.llm.image_ocr_client import (
+                VISION_OCR_EXTRACT_PROMPT_ZH,
+                send_ocr_image_messages,
+                vision_llm_backend_status,
+            )
+        except Exception:
+            return _flatten_message_content_to_text(content)
+        if vision_llm_backend_status().get("ok"):
+            parts: list[str] = []
+            for item in content:
+                if isinstance(item, str):
+                    if item.strip():
+                        parts.append(item)
+                    continue
+                if not isinstance(item, dict):
+                    continue
+                typ = str(item.get("type") or "").strip().lower()
+                if typ in ("image_url", "input_image"):
+                    url = _image_block_to_url(item)
+                    if not url:
+                        parts.append(
+                            "【图片】无法从本消息中还原图片数据，图中细节未传入本轮模型；"
+                            "若作答需要像素级信息，请提示用户重发或改用支持多模态的模型配置。"
+                        )
+                        continue
+                    try:
+                        out = send_ocr_image_messages(images=[url], prompt=VISION_OCR_EXTRACT_PROMPT_ZH)
+                    except Exception as exc:
+                        parts.append(
+                            f"【图片】独立看图通道 OCR 调用异常（{type(exc).__name__}），"
+                            "本张图的正文未能写入对话；可请用户改述图中要点或稍后重试。"
+                        )
+                        continue
+                    if not out.get("ok"):
+                        err = str(out.get("error") or "unknown")
+                        parts.append(
+                            f"【图片】看图通道返回失败（{err}），本张图未转成文字；"
+                            "请据用户文字说明作答，或提示检查 OCR 配置/重发图片。"
+                        )
+                        continue
+                    t = str(out.get("text") or "").strip()
+                    if not t:
+                        parts.append(
+                            "【图片】看图通道未返回有效文字（可能模型拒识或图不清晰）。"
+                            "请提示用户补充文字说明或重发更清晰的截图。"
+                        )
+                    else:
+                        parts.append(_downgrade_ocr_wrap_for_llm(t))
+                    continue
+                if typ in ("text", "input_text"):
+                    txt = str(item.get("text") or item.get("content") or "").strip()
+                    if txt:
+                        parts.append(txt)
+                    continue
+            return "\n\n".join([p for p in parts if p]).strip()
+    return _flatten_message_content_to_text(content)
+
+
+def _normalize_messages_for_text_only_gateway(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    out: list[dict[str, Any]] = []
+    for m in messages or []:
+        if not isinstance(m, dict):
+            continue
+        mm = dict(m)
+        if isinstance(mm.get("content"), list):
+            mm["content"] = _flatten_message_content_for_text_gateway(mm.get("content"))
+        out.append(mm)
+    return out
+
+
+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,
+        thinking_mode_enabled: bool = False,
+        reasoning_effort: 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")
+        self.thinking_mode_enabled = bool(thinking_mode_enabled)
+        eff = str(reasoning_effort or "").strip().lower()
+        self.reasoning_effort = eff if eff in ("low", "medium", "high") else ""
+
+        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 svc.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)
+
+        minimax_compat = _is_minimax_compat(self.model, self.base_url)
+        use_tools = bool(tools)
+        cleaned_msgs: list[dict[str, Any]] = []
+        for m in norm_msgs or []:
+            if not isinstance(m, dict):
+                continue
+            role = str(m.get("role") or "")
+            if minimax_compat and role == "assistant" and isinstance(m.get("tool_calls"), list):
+                # MiniMax OpenAI-compat can reject long-history tool_use/tool_result replays.
+                # Keep text semantics, but remove historical wire-level tool_calls.
+                mm = dict(m)
+                mm.pop("tool_calls", None)
+                cleaned_msgs.append(mm)
+                continue
+            if minimax_compat and role == "tool":
+                # Downgrade history tool_result rows to plain assistant text context,
+                # avoiding strict tool_result sequence validation on replay.
+                tcid = str(m.get("tool_call_id") or m.get("call_id") or "").strip()
+                tname = str(m.get("name") or "").strip()
+                raw = str(m.get("content") or "")
+                prefix = "[tool_result replay]"
+                if tname:
+                    prefix += f" name={tname}"
+                if tcid:
+                    prefix += f" id={tcid}"
+                cleaned_msgs.append({"role": "assistant", "content": f"{prefix}\n{raw}".strip()})
+                continue
+            if role == "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)
+
+        msgs_wire = cleaned_msgs
+        if _should_force_text_only_messages(self.base_url) or _should_proactively_downgrade_multimodal_messages(
+            cleaned_msgs,
+            model=self.model,
+            base_url=self.base_url,
+        ):
+            msgs_wire = _normalize_messages_for_text_only_gateway(cleaned_msgs)
+        kwargs: dict[str, Any] = {"model": self.model, "messages": msgs_wire, "stream": stream}
+        if _should_enable_thinking(self.base_url, thinking_mode_enabled=bool(getattr(self, "thinking_mode_enabled", False))):
+            extra_body = kwargs.get("extra_body") if isinstance(kwargs.get("extra_body"), dict) else {}
+            extra_body = dict(extra_body)
+            extra_body["thinking"] = {"type": "enabled"}
+            kwargs["extra_body"] = extra_body
+            eff = str(getattr(self, "reasoning_effort", "") or "").strip().lower()
+            if eff in ("low", "medium", "high"):
+                kwargs["reasoning_effort"] = eff
+        elif _should_disable_thinking(self.base_url):
+            extra_body = kwargs.get("extra_body") if isinstance(kwargs.get("extra_body"), dict) else {}
+            extra_body = dict(extra_body)
+            extra_body["thinking"] = {"type": "disabled"}
+            kwargs["extra_body"] = extra_body
+        if use_tools:
+            try:
+                from svc.config.paths import db_path
+                from svc.persistence.sqlite_store import SqliteStore
+                from runtime.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"] = complete_openai_tools_wire_parameters(tools)
+        _merge_deepseek_strict_tool_mode_extra_body(
+            kwargs, base_url=self.base_url, model=self.model, use_tools=use_tools
+        )
+        try:
+            return self._client.chat.completions.create(**kwargs)
+        except Exception as exc:
+            msg = _wire_error_message(exc)
+            if _is_text_only_gateway_error(msg) and isinstance(kwargs.get("messages"), list):
+                kwargs["messages"] = _normalize_messages_for_text_only_gateway(kwargs["messages"])
+                return self._client.chat.completions.create(**kwargs)
+            if "reasoning_content" in msg and "thinking mode" in msg and "must be passed back" in msg:
+                # Provider requires replaying assistant.reasoning_content in thinking mode.
+                # As a safety fallback, force-disable thinking and retry once.
+                extra_body = kwargs.get("extra_body") if isinstance(kwargs.get("extra_body"), dict) else {}
+                extra_body = dict(extra_body)
+                extra_body["thinking"] = {"type": "disabled"}
+                kwargs["extra_body"] = extra_body
+                kwargs.pop("reasoning_effort", None)
+                return self._client.chat.completions.create(**kwargs)
+            raise
+
+    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:
+            if reasoning_text:
+                on_token(reasoning_text)
+            if 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 on_token:
+                        on_token(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",
+    "_deepseek_strict_tool_mode_enabled",
+    "_merge_deepseek_strict_tool_mode_extra_body",
+]
+
+
+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
+

svc/llm/transports/openai_responses.py

@@ -0,0 +1,811 @@
+from __future__ import annotations
+
+import json
+import logging
+import os
+import sys
+import uuid
+from typing import Any, Optional
+from collections.abc import Callable, Iterable
+
+from svc.llm.transports.base import ChatModel, LLMResponse, LLMToolCall, normalize_image_b64_payload
+
+logger = logging.getLogger(__name__)
+
+
+def _redact_nested_json_preview(obj: Any, *, max_chars: int = 4000) -> str:
+    """Best-effort JSON preview for logs (truncate; redact mega data URLs)."""
+
+    try:
+        s = json.dumps(obj, ensure_ascii=False, default=str)
+    except Exception:
+        s = str(obj)
+
+    def _clip_data_urls(t: str) -> str:
+        out_parts: list[str] = []
+        i = 0
+        needle = "data:"
+        while i < len(t):
+            j = t.find(needle, i)
+            if j < 0:
+                out_parts.append(t[i:])
+                break
+            out_parts.append(t[i:j])
+            k = j + len(needle)
+            while k < len(t) and t[k] not in "?;, \n\r\t\"]":
+                k += 1
+            if k < len(t) and t[k:k + 8] == ";base64,":
+                end = k + 8
+                while end < len(t) and t[end] not in "\"}] \n\r\t":
+                    end += 1
+                seg = end - (k + 8)
+                out_parts.append(f"data:<redacted base64 ~{seg} chars>")
+                i = end
+                continue
+            out_parts.append(t[j:k])
+            i = k
+        return "".join(out_parts)
+
+    s = _clip_data_urls(s)
+    if len(s) > max_chars:
+        return s[: max_chars - 40] + "\n...<<truncated>>\n..."
+    return s
+
+
+def _deep_redact_for_debug(obj: Any) -> Any:
+    """Recursive copy for logs: shorten giant ``data:...;base64,...`` and very long strings."""
+
+    if isinstance(obj, dict):
+        return {str(k): _deep_redact_for_debug(v) for k, v in obj.items()}
+    if isinstance(obj, list):
+        return [_deep_redact_for_debug(x) for x in obj]
+    if isinstance(obj, str):
+        s = obj
+        if ";base64," in s and s.strip().startswith("data:") and len(s) > 160:
+            head, _, _tail = s.partition(";base64,")
+            return f"{head};base64,<redacted ~{len(s) - len(head) - 8} b64 chars>"
+        if len(s) > 16_000:
+            return s[:12_000] + f"...<truncated {len(s) - 12_000} chars>"
+        return s
+    return obj
+
+
+def _safe_int(raw: str | None, default: int, *, max_value: int = 2_000_000) -> int:
+    try:
+        value = int(str(raw or "").strip())
+    except Exception:
+        return default
+    if value < 1:
+        return default
+    return min(value, max_value)
+
+
+def _log_openai_responses_wire_kwargs(payload: dict[str, Any], *, variant: str, phase: str) -> None:
+    """Temporary diagnostics: log kwargs passed into ``OpenAI.responses.create`` (after our assembly).
+
+    Enable: ``AIA_RESPONSES_DEBUG_PRINT_PAYLOAD=1`` — emits ``logging.warning`` **and** the same JSON on **stderr**
+    (gateway控制台默认可见，不依赖 logging level)。
+
+    Notes:
+
+    - OpenAI SDK may still apply minor JSON transforms on send; this matches **our** arguments.
+    - Base64 / long ``data:`` URLs are redacted to keep logs readable.
+    """
+    raw = str(os.getenv("AIA_RESPONSES_DEBUG_PRINT_PAYLOAD") or "").strip().lower()
+    if raw not in ("1", "true", "yes", "on"):
+        return
+    try:
+        dbg = _deep_redact_for_debug(dict(payload))
+        txt = json.dumps(dbg, ensure_ascii=False, indent=2, default=str)
+        cap = _safe_int(os.getenv("AIA_RESPONSES_DEBUG_PRINT_MAX_CHARS"), 120_000, max_value=500_000)
+        if len(txt) > cap:
+            txt = txt[: max(cap - 80, 0)] + "\n...<<truncated>>\n"
+        logger.warning(
+            "openai_responses DEBUG wire kwargs [%s] variant=%s (%d chars):\n%s",
+            phase,
+            variant,
+            len(txt),
+            txt,
+        )
+        try:
+            sys.stderr.write(
+                f"\n[oclaw openai_responses DEBUG] phase={phase} variant={variant} chars={len(txt)}\n{txt}\n\n"
+            )
+            sys.stderr.flush()
+        except Exception:
+            pass
+    except Exception as exc:
+        logger.warning("openai_responses DEBUG payload serialization failed: %s", exc)
+
+
+def _openai_sdk_diagnostic_text(exc: BaseException) -> str:
+    """``str(APIStatusError)`` is often only ``Error code: 400``; validation detail lives in ``body['message']``."""
+    chunks: list[str] = [str(exc)]
+    msg = getattr(exc, "message", None)
+    if isinstance(msg, str) and msg.strip() and msg not in chunks:
+        chunks.append(msg)
+    bod = getattr(exc, "body", None)
+    if isinstance(bod, dict):
+        bm = bod.get("message")
+        if isinstance(bm, str) and bm.strip():
+            chunks.append(bm)
+        err = bod.get("error")
+        if isinstance(err, dict):
+            em = err.get("message")
+            if isinstance(em, str) and em.strip():
+                chunks.append(em)
+    return "\n".join(chunks)
+
+
+def _is_input_messages_validation_error(exc: BaseException) -> bool:
+    """Third-party gateways often return 400/422 with Pydantic paths like ``input.messages.0.role``."""
+    m = _openai_sdk_diagnostic_text(exc).lower()
+    if "input.messages" in m:
+        return True
+    if "input should be 'user'" in m and "content" in m:
+        return True
+    return False
+
+
+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 _stream_reasoning_delta_text(ev: dict[str, Any]) -> str:
+    """Extract a textual chunk from a Responses streaming reasoning-related event."""
+    delta = ev.get("delta")
+    if isinstance(delta, dict):
+        for k in ("text", "delta", "summary", "content"):
+            v = delta.get(k)
+            if isinstance(v, str) and v:
+                return v
+        return ""
+    if delta is not None:
+        s = str(delta)
+        return s
+    for k in ("text", "summary", "content"):
+        v = ev.get(k)
+        if isinstance(v, str) and v:
+            return v
+    return ""
+
+
+def _reasoning_from_completed_response(resp: dict[str, Any]) -> str:
+    """Best-effort reasoning string from a terminal Responses ``response`` object."""
+    for key in ("reasoning", "reasoning_summary", "reasoning_text", "reasoning_content"):
+        v = resp.get(key)
+        if isinstance(v, str) and v.strip():
+            return v.strip()
+    items = resp.get("output")
+    if not isinstance(items, list):
+        return ""
+    chunks: list[str] = []
+    for it in items:
+        d = _as_dict(it) or {}
+        typ = str(d.get("type") or "").lower()
+        if "reasoning" not in typ:
+            continue
+        for k in ("summary", "text", "content"):
+            x = d.get(k)
+            if isinstance(x, str) and x.strip():
+                chunks.append(x.strip())
+                break
+    return "\n".join(chunks).strip()
+
+
+def parse_openai_responses_stream_events(
+    events: Iterable[Any],
+    *,
+    on_token: Optional[Callable[[str], None]] = None,
+) -> tuple[str, list[LLMToolCall], dict[str, Any] | None, str]:
+    """Pure stream parser (offline-testable).
+
+    Returns: (assembled_output_text, tool_calls, final_response_dict?, reasoning_text)
+
+    Reasoning models emit ``response.reasoning_summary_text.delta`` (and similar) events; those are
+    forwarded to ``on_token`` for live UI, accumulated into ``reasoning_text`` for persistence, and
+    are **not** merged into the returned ``assembled_output_text`` (assistant body).
+    """
+    parts: list[str] = []
+    reasoning_parts: list[str] = []
+    final_resp: dict[str, Any] | None = None
+
+    def _emit_reasoning_chunk(s: str) -> None:
+        s = str(s or "")
+        if not s:
+            return
+        reasoning_parts.append(s)
+        if on_token:
+            on_token(s)
+
+    for ev in events:
+        d = _as_dict(ev) or {}
+        typ = str(d.get("type") or "")
+        # Reasoning / thinking summary stream (do not append to output_text body).
+        if typ in (
+            "response.reasoning_summary_text.delta",
+            "response.reasoning_summary_text",
+            "response.reasoning_text.delta",
+            "response.reasoning_text",
+            "response.reasoning_summary.delta",
+        ) or (typ.startswith("response.reasoning") and typ.endswith(".delta")):
+            rs = _stream_reasoning_delta_text(d)
+            if rs:
+                _emit_reasoning_chunk(rs)
+            continue
+        # 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)
+                for rkey in ("reasoning_summary_text", "reasoning_text", "reasoning"):
+                    rv = delta.get(rkey)
+                    if isinstance(rv, str) and rv:
+                        _emit_reasoning_chunk(rv)
+            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 []
+    reasoning_text = "".join(reasoning_parts).strip()
+    if not reasoning_text and final_resp:
+        reasoning_text = _reasoning_from_completed_response(final_resp)
+    return text, tool_calls, final_resp, reasoning_text
+
+
+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,
+        thinking_mode_enabled: bool = False,
+        reasoning_effort: 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
+        self.thinking_mode_enabled = bool(thinking_mode_enabled)
+        eff = str(reasoning_effort or "").strip().lower()
+        self.reasoning_effort = eff if eff in ("low", "medium", "high") else ""
+        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 _strip_leading_system_messages(messages: list[dict[str, Any]]) -> tuple[str, list[dict[str, Any]]]:
+        """Separate leading ``role=system`` rows for ``instructions=` (Responses API expects user-shaped ``input.messages``)."""
+        msgs = messages or []
+        sys_chunks: list[str] = []
+        i = 0
+        while i < len(msgs):
+            m = msgs[i]
+            if not isinstance(m, dict):
+                break
+            if str(m.get("role") or "").strip().lower() != "system":
+                break
+            c = m.get("content")
+            if isinstance(c, str) and c.strip():
+                sys_chunks.append(c.strip())
+            elif isinstance(c, list):
+                texts: list[str] = []
+                for it in c:
+                    if not isinstance(it, dict):
+                        continue
+                    tt = str(it.get("type") or "").strip().lower()
+                    if tt in ("text", "input_text"):
+                        xs = str(it.get("text") or "").strip()
+                        if xs:
+                            texts.append(xs)
+                    elif isinstance(it.get("text"), str) and str(it.get("text")).strip():
+                        texts.append(str(it.get("text")).strip())
+                if texts:
+                    sys_chunks.append("\n".join(texts))
+            i += 1
+        joined = "\n\n".join(sys_chunks).strip()
+        return joined, list(msgs[i:])
+
+    @staticmethod
+    def _normalize_messages(
+        messages: list[dict[str, Any]],
+        *,
+        envelope_openai_message: bool = True,
+        content_chat_completions_parts: bool = False,
+        image_detail_auto: bool = True,
+    ) -> list[dict[str, Any]]:
+        """
+        Build Response input rows used under ``input`` (flat array) or ``input.messages``.
+
+        - **Responses content** (default when ``content_chat_completions_parts=False``): each part is
+          ``{"type":"input_text",...}`` / ``{"type":"input_image","image_url":"<str>","detail":"auto"}``
+          (`OpenAI response_input_*` TypedDict surface).
+        - **Chat completions content** (``content_chat_completions_parts=True``): ``text`` /
+          ``image_url`` multimodal blocks.
+        - **Envelope** ``envelope_openai_message=True``: ``{"type":"message","role","content":[...]}``;
+          ``False``: ``{"role","content":[...]}`` (per some Bailian curls).
+        """
+        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]] = []
+
+            def _append_image_part(url_value: str) -> None:
+                if content_chat_completions_parts:
+                    norm_content.append({"type": "image_url", "image_url": {"url": url_value}})
+                else:
+                    p: dict[str, Any] = {"type": "input_image", "image_url": url_value}
+                    if image_detail_auto:
+                        p["detail"] = "auto"
+                    norm_content.append(p)
+
+            def _append_text_part(txt: str) -> None:
+                if content_chat_completions_parts:
+                    norm_content.append({"type": "text", "text": txt})
+                else:
+                    norm_content.append({"type": "input_text", "text": txt})
+
+            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
+                        _append_image_part(f"data:{mime};base64,{b64}")
+                        continue
+                    if isinstance(item, dict) and str(item.get("type") or "").strip().lower() == "image_url":
+                        iu = item.get("image_url")
+                        url = ""
+                        if isinstance(iu, dict):
+                            url = str(iu.get("url") or "").strip()
+                        elif isinstance(iu, str):
+                            url = iu.strip()
+                        if url.startswith(("http://", "https://", "data:")):
+                            _append_image_part(url)
+                        continue
+                    if isinstance(item, dict) and item.get("type") in ("input_text", "text"):
+                        t = str(item.get("text") or "").strip()
+                        if t:
+                            _append_text_part(t)
+                        continue
+                    if isinstance(item, dict):
+                        itype = str(item.get("type") or "").strip().lower()
+                        img_key = item.get("image")
+                        if isinstance(img_key, str) and img_key.strip() and itype != "input_image":
+                            uu = img_key.strip()
+                            if uu.startswith(("http://", "https://", "data:")):
+                                _append_image_part(uu)
+                                continue
+                        if isinstance(item.get("text"), str) and str(item.get("text") or "").strip():
+                            _append_text_part(str(item.get("text") or "").strip())
+                            continue
+                        s = str(item.get("text") or "").strip()
+                        if s:
+                            _append_text_part(s)
+                        continue
+                    s = str(item or "").strip()
+                    if s:
+                        _append_text_part(s)
+            elif isinstance(content, str):
+                txt = content.strip()
+                if txt:
+                    _append_text_part(txt)
+            elif content is not None:
+                s = str(content).strip()
+                if s:
+                    _append_text_part(s)
+
+            if not norm_content:
+                continue
+
+            coerced_role = "user"
+            if role != "user":
+                prefix = "assistant" if role == "assistant" else ("system" if role == "system" else role)
+                norm_content.insert(
+                    0,
+                    OpenAIResponsesModel._role_prefix_part(prefix, content_chat_completions_parts),
+                )
+
+            if envelope_openai_message:
+                out.append({"type": "message", "role": coerced_role, "content": norm_content})
+            else:
+                out.append({"role": coerced_role, "content": norm_content})
+        return out
+
+    @staticmethod
+    def _role_prefix_part(prefix: str, chat_parts: bool) -> dict[str, Any]:
+        t = f"[{prefix}]"
+        if chat_parts:
+            return {"type": "text", "text": t}
+        return {"type": "input_text", "text": t}
+
+    @staticmethod
+    def _responses_input_candidates(
+        msgs: list[dict[str, Any]],
+        *,
+        flat_responses: bool,
+        prefer_envelope: bool,
+        prefer_chat_parts: bool,
+    ) -> list[tuple[str, Any]]:
+        """Several gateways validate ``input.messages`` differently; try a small deterministic set."""
+
+        def N(env: bool, chat: bool) -> list[dict[str, Any]]:
+            return OpenAIResponsesModel._normalize_messages(
+                msgs,
+                envelope_openai_message=env,
+                content_chat_completions_parts=chat,
+            )
+
+        keys_seen: set[str] = set()
+        out: list[tuple[str, Any]] = []
+
+        def push(tag: str, inp: Any) -> None:
+            try:
+                sk = json.dumps(inp, ensure_ascii=False, sort_keys=True, default=str)
+            except Exception:
+                sk = repr(inp)
+            if sk in keys_seen:
+                return
+            keys_seen.add(sk)
+            out.append((tag, inp))
+
+        if flat_responses:
+            for env in (prefer_envelope, not prefer_envelope):
+                push(f"flat_envelope_{env}", N(env, False))
+            return out
+
+        # Prefer profile/env defaults first (``AIA_RESPONSES_NESTED_CHAT_PARTS`` / envelope toggles), then a fixed
+        # fallback ring so picky gateways still get Chat-shaped ``input.messages`` without dropping multimodal pixels.
+        primary_combo = (prefer_envelope, prefer_chat_parts)
+        fallback_ring = [(False, True), (False, False), (True, False), (True, True)]
+        combos = [primary_combo] + [p for p in fallback_ring if p != primary_combo]
+        for env, chat in combos:
+            norm = N(env, chat)
+            push(f"e{int(env)}c{int(chat)}_messages", {"messages": norm})
+            push(f"e{int(env)}c{int(chat)}_flat_input", norm)
+        return out
+
+    def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]],
+        *,
+        on_token: Optional[Callable[[str], None]] = None,
+    ) -> LLMResponse:
+        sys_instructions, rest = OpenAIResponsesModel._strip_leading_system_messages(messages)
+        def _env_truthy(name: str) -> bool:
+            return str(os.getenv(name) or "").strip().lower() in ("1", "true", "yes", "on")
+
+        if _env_truthy("AIA_RESPONSES_DEBUG_PRINT_PAYLOAD"):
+            try:
+                sys.stderr.write(
+                    "[oclaw openai_responses] chat() ENTERED "
+                    f"model={self.model!r} base_url={self.base_url!r} "
+                    f"messages={len(messages)} tools={len(tools or [])}\n"
+                )
+                sys.stderr.flush()
+            except Exception:
+                pass
+
+        # Default: nested ``input.messages`` rows use **Responses** parts (``input_text`` / ``input_image``).
+        # Flat Bailian ``input`` array: ``AIA_RESPONSES_INPUT_IS_FLAT_LIST=1`` (optional ``AIA_RESPONSES_FLAT_OPENAI_ENVELOPE``).
+        flat_raw = os.getenv("AIA_RESPONSES_INPUT_IS_FLAT_LIST")
+        if flat_raw is not None and str(flat_raw).strip():
+            flat_responses = _env_truthy("AIA_RESPONSES_INPUT_IS_FLAT_LIST")
+        else:
+            legacy = os.getenv("AIA_RESPONSES_INPUT_USE_MESSAGES_ARRAY")
+            if legacy is None or not str(legacy).strip():
+                flat_responses = False
+            else:
+                flat_responses = not _env_truthy("AIA_RESPONSES_INPUT_USE_MESSAGES_ARRAY")
+
+        # Nested ``input.messages`` defaults to **Responses** content parts (``input_*``) + ``type:message``,
+        # matching OpenAI ``EasyInputMessageParam`` / Model Studio expanded examples. Some proxies wrongly
+        # expect Chat ``text``/``image_url`` parts → ``AIA_RESPONSES_NESTED_CHAT_PARTS=1``.
+        if flat_responses:
+            envelope_openai_message = _env_truthy("AIA_RESPONSES_FLAT_OPENAI_ENVELOPE")
+            nested_chat_parts = False
+        else:
+            envelope_openai_message = True
+            nested_chat_parts = _env_truthy("AIA_RESPONSES_NESTED_CHAT_PARTS")
+
+        def _normalize_batch(msgs: list[dict[str, Any]]) -> list[dict[str, Any]]:
+            return OpenAIResponsesModel._normalize_messages(
+                msgs,
+                envelope_openai_message=envelope_openai_message,
+                content_chat_completions_parts=nested_chat_parts,
+            )
+
+        msgs_for_norm: list[dict[str, Any]]
+        if rest:
+            msgs_for_norm = rest
+            norm = _normalize_batch(rest)
+            inst_kw: dict[str, Any] = {}
+            if sys_instructions.strip():
+                si = sys_instructions.strip()
+                if len(si) > 80_000:
+                    si = si[:80_000] + "\n...[truncated]"
+                inst_kw = {"instructions": si}
+            if not norm:
+                msgs_for_norm = messages
+                norm = _normalize_batch(messages)
+                inst_kw = {}
+        else:
+            msgs_for_norm = messages
+            norm = _normalize_batch(messages)
+            inst_kw = {}
+
+        explicit_variants = os.getenv("AIA_RESPONSES_INPUT_VARIANTS")
+        bu_norm = (self.base_url or "").strip().lower()
+        # CRITICAL: do **not** treat empty ``base_url`` as "official" here. Clients often omit storing the
+        # default URL in profiles while still hitting third-party gateways via env/SDK defaults—but then we
+        # would wrongly disable alternate ``input`` shapes and only ever send ``primary``.
+        host_is_explicit_official_api = bool(bu_norm) and ("api.openai.com" in bu_norm)
+        if explicit_variants is not None and str(explicit_variants).strip():
+            multi_shape = _env_truthy("AIA_RESPONSES_INPUT_VARIANTS")
+        else:
+            multi_shape = not host_is_explicit_official_api
+
+        if flat_responses:
+            primary_input: Any = norm
+        else:
+            primary_input = {"messages": norm}
+
+        input_candidates: list[tuple[str, Any]]
+        if multi_shape and not _env_truthy("AIA_RESPONSES_DISABLE_INPUT_VARIANTS"):
+            input_candidates = OpenAIResponsesModel._responses_input_candidates(
+                msgs_for_norm,
+                flat_responses=flat_responses,
+                prefer_envelope=envelope_openai_message,
+                prefer_chat_parts=nested_chat_parts,
+            )
+        else:
+            input_candidates = [("primary", primary_input)]
+
+        stream_errors: list[str] = []
+        b = str(self.base_url or "").strip().lower()
+        force_disable = str(os.getenv("AIA_LLM_THINKING_FORCE_DISABLED") or "").strip().lower() in ("1", "true", "yes", "on")
+        force_enable = str(os.getenv("AIA_LLM_THINKING_FORCE_ENABLED") or "").strip().lower() in ("1", "true", "yes", "on")
+        mode_enabled = bool(getattr(self, "thinking_mode_enabled", False))
+        extra_body: dict[str, Any] = {}
+        if b and ("api.openai.com" not in b):
+            if force_disable:
+                extra_body["thinking"] = {"type": "disabled"}
+            elif force_enable or mode_enabled:
+                extra_body["thinking"] = {"type": "enabled"}
+            else:
+                # Default: disable thinking for non-official gateways unless explicitly enabled.
+                if str(os.getenv("AIA_LLM_THINKING_DISABLED") or "1").strip().lower() in ("1", "true", "yes", "on"):
+                    extra_body["thinking"] = {"type": "disabled"}
+        thinking = {"extra_body": extra_body} if extra_body else {}
+        reasoning_effort = str(getattr(self, "reasoning_effort", "") or "").strip().lower()
+        if reasoning_effort not in ("low", "medium", "high"):
+            reasoning_effort = ""
+        base_stream_template: dict[str, Any] = {
+            **thinking,
+            **inst_kw,
+            **({"reasoning_effort": reasoning_effort} if reasoning_effort else {}),
+            "model": self.model,
+            "tools": tools or None,
+            "stream": True,
+        }
+        try:
+            for cand_label, responses_input in input_candidates:
+                stream_payload = {**base_stream_template, "input": responses_input}
+                if _env_truthy("AIA_RESPONSES_LOG_PAYLOAD_SUMMARY"):
+                    logger.warning(
+                        "openai_responses [%s] payload summary (redacted): %s",
+                        cand_label,
+                        _redact_nested_json_preview(
+                            {
+                                "model": self.model,
+                                "input": responses_input,
+                                "has_instructions": bool(inst_kw.get("instructions")),
+                                "tools_n": len(tools or []),
+                                "stream": True,
+                            }
+                        ),
+                    )
+                payload = stream_payload
+                _log_openai_responses_wire_kwargs(payload, variant=str(cand_label), phase="stream")
+                try:
+                    stream = self._client.responses.create(**payload)
+                    text, tool_calls, final_resp, reasoning_text = 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)
+                    if cand_label != input_candidates[0][0]:
+                        logger.info("openai_responses: succeeded with input variant %s", cand_label)
+                    return LLMResponse(content=text, tool_calls=tool_calls, reasoning_content=reasoning_text)
+                except Exception as exc:
+                    emsg = str(exc)
+                    # Fallback: provider thinking-mode replay contract.
+                    if "reasoning_content" in emsg and "thinking mode" in emsg and "must be passed back" in emsg:
+                        try:
+                            forced = dict(payload)
+                            eb = forced.get("extra_body") if isinstance(forced.get("extra_body"), dict) else {}
+                            eb = dict(eb)
+                            eb["thinking"] = {"type": "disabled"}
+                            forced["extra_body"] = eb
+                            forced.pop("reasoning_effort", None)
+                            _log_openai_responses_wire_kwargs(
+                                forced,
+                                variant=f"{cand_label}_thinking_disabled",
+                                phase="stream_retry_thinking",
+                            )
+                            stream = self._client.responses.create(**forced)
+                            text, tool_calls, final_resp, reasoning_text = 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)
+                            if cand_label != input_candidates[0][0]:
+                                logger.info("openai_responses: succeeded with input variant %s", cand_label)
+                            return LLMResponse(content=text, tool_calls=tool_calls, reasoning_content=reasoning_text)
+                        except Exception:
+                            pass
+                    if _env_truthy("AIA_RESPONSES_LOG_API_ERROR_DETAIL"):
+                        eb = getattr(exc, "body", None)
+                        if eb is not None:
+                            logger.warning(
+                                "responses.create stream [%s] body: %s", cand_label, eb
+                            )
+                    if len(input_candidates) > 1 and _is_input_messages_validation_error(exc):
+                        stream_errors.append(f"{cand_label}:{emsg}")
+                        continue
+                    stream_errors.append(emsg)
+                    break
+            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] = []
+            base_nonstream = {
+                **thinking,
+                **inst_kw,
+                **({"reasoning_effort": reasoning_effort} if reasoning_effort else {}),
+                "model": self.model,
+                "tools": tools or None,
+            }
+            for cand_label, responses_input in input_candidates:
+                payload = {**base_nonstream, "input": responses_input}
+                _log_openai_responses_wire_kwargs(payload, variant=str(cand_label), phase="non_stream")
+                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)
+                    reasoning_text = _reasoning_from_completed_response(d)
+                    if on_token and text:
+                        on_token(text)
+                    if cand_label != input_candidates[0][0]:
+                        logger.info("openai_responses non-stream: succeeded with input variant %s", cand_label)
+                    return LLMResponse(content=text, tool_calls=tool_calls, reasoning_content=reasoning_text)
+                except Exception as e2:
+                    emsg2 = str(e2)
+                    if "reasoning_content" in emsg2 and "thinking mode" in emsg2 and "must be passed back" in emsg2:
+                        try:
+                            forced = dict(payload)
+                            eb = forced.get("extra_body") if isinstance(forced.get("extra_body"), dict) else {}
+                            eb = dict(eb)
+                            eb["thinking"] = {"type": "disabled"}
+                            forced["extra_body"] = eb
+                            forced.pop("reasoning_effort", None)
+                            _log_openai_responses_wire_kwargs(
+                                forced,
+                                variant=f"{cand_label}_thinking_disabled",
+                                phase="non_stream_retry_thinking",
+                            )
+                            resp = self._client.responses.create(**forced)
+                            d = _as_dict(resp) or {}
+                            text = str(d.get("output_text") or "")
+                            tool_calls = _collect_tool_calls_from_response_dict(d)
+                            reasoning_text = _reasoning_from_completed_response(d)
+                            if on_token and text:
+                                on_token(text)
+                            if cand_label != input_candidates[0][0]:
+                                logger.info(
+                                    "openai_responses non-stream: succeeded with input variant %s",
+                                    cand_label,
+                                )
+                            return LLMResponse(content=text, tool_calls=tool_calls, reasoning_content=reasoning_text)
+                        except Exception:
+                            pass
+                    if _env_truthy("AIA_RESPONSES_LOG_API_ERROR_DETAIL"):
+                        eb = getattr(e2, "body", None)
+                        if eb is not None:
+                            logger.warning(
+                                "responses.create non-stream [%s] body: %s", cand_label, eb
+                            )
+                    if len(input_candidates) > 1 and _is_input_messages_validation_error(e2):
+                        nonstream_errors.append(f"{cand_label}:{emsg2}")
+                        continue
+                    nonstream_errors.append(emsg2)
+                    break
+            raise RuntimeError(
+                "openai_responses_request_failed: "
+                + " | ".join([str(exc)] + nonstream_errors[-2:])
+            ) from exc
+
+
+__all__ = ["OpenAIResponsesModel", "parse_openai_responses_stream_events"]
+

svc/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 svc.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"]
+

svc/llm/video_generation_client.py

@@ -0,0 +1,466 @@
+"""DashScope / Model Studio **text-to-video** and **image-to-video** (async: create task → poll).
+
+Used by the **video** specialist (Admin Chat + orchestration) to bypass the generic LLM tool loop,
+analogous to :mod:`svc.llm.image_legacy_client` for the image specialist.
+
+**Text-to-video:** ``input`` only ``prompt`` (plus optional ``audio_url`` / ``negative_prompt`` for some models).
+Use a **t2v** model id.
+
+**Image-to-video (Wan 2.6 and earlier):** ``input.img_url`` — HTTPS or ``data:image/...;base64,...``.
+
+**Image-to-video (Wan 2.7 general API):** ``input.media``, e.g.
+``[{"type": "first_frame", "url": "<https or data url>"}]``.
+Do **not** send ``img_url`` for that path (see Alibaba *image-to-video-general* API reference).
+
+Override with ``AIA_VIDEO_I2V_INPUT_STYLE=media|img_url``; default is **auto** (``wan2.7`` in model id → ``media``, else ``img_url``).
+
+HTTP contract (region of ``base_url`` must match the API key):
+
+- ``POST {root}/api/v1/services/aigc/video-generation/video-synthesis`` with header
+  ``X-DashScope-Async: enable``
+- ``GET {root}/api/v1/tasks/{task_id}`` until ``output.task_status`` is terminal
+
+References:
+- https://www.alibabacloud.com/help/en/model-studio/text-to-video-api-reference/
+- https://www.alibabacloud.com/help/en/model-studio/image-to-video-api-reference/
+- https://www.alibabacloud.com/help/en/model-studio/image-to-video-general-api-reference/
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import sys
+import time
+from typing import Any, Callable, Literal
+
+import httpx
+
+from svc.files.attachment_assets import AttachmentAssetStore
+from svc.llm.image_http_common import (
+    compress_data_url_image,
+    dashscope_multimodal_http_ok,
+    download_http_url_bytes,
+    is_data_url,
+    join_url,
+    post_with_retry,
+)
+
+VIDEO_SPECIALIST_DEFAULT_PROMPT_ZH = "请根据描述生成一段短视频，画面稳定、主体清晰。"
+
+_DEFAULT_SYNTHESIS_PATH = "api/v1/services/aigc/video-generation/video-synthesis"
+
+
+def _truthy_env(name: str) -> bool:
+    return str(os.getenv(name) or "").strip().lower() in ("1", "true", "yes", "on")
+
+
+def env_video_expert_api_key() -> str:
+    return (os.getenv("AIA_VIDEO_EXPERT_API_KEY") or "").strip()
+
+
+def env_video_expert_base_url() -> str:
+    return (os.getenv("AIA_VIDEO_EXPERT_BASE_URL") or "").strip()
+
+
+def env_video_expert_model() -> str:
+    return (os.getenv("AIA_VIDEO_EXPERT_MODEL") or "").strip()
+
+
+def _effective_video_model_for_request(model_name: str, *, img_url: str | None) -> str:
+    """DashScope Wan: **text-to-video** ids use ``*-t2v-*``; **image-to-video** needs ``*-i2v-*``.
+
+    Chat profiles often configure only a **t2v** model; if we still submit ``input.img_url``, the gateway
+    may largely follow the prompt and ignore the photograph (looks like unrelated output). When a first
+    frame is present we therefore coerce ``-t2v-`` → ``-i2v-`` unless disabled or overridden.
+    """
+    m = (model_name or "").strip()
+    if not m:
+        return m
+    if not str(img_url or "").strip():
+        return m
+    if _truthy_env("AIA_VIDEO_EXPERT_DISABLE_I2V_MODEL_COERCION"):
+        return m
+    override = (os.getenv("AIA_VIDEO_EXPERT_I2V_MODEL") or "").strip()
+    if override:
+        return override
+    if "-t2v-" in m:
+        coerced = m.replace("-t2v-", "-i2v-", 1)
+        if coerced != m:
+            return coerced
+    low = m.lower().rstrip()
+    if low.endswith("-t2v"):
+        coerced = m[: -len("-t2v")] + "-i2v"
+        if coerced != m:
+            return coerced
+    return m
+
+
+def _i2v_first_frame_input_style(*, model_name: str) -> Literal["img_url", "media"]:
+    """How to pass the first-frame image: legacy ``img_url`` vs Wan 2.7 ``media`` array."""
+    raw = (os.getenv("AIA_VIDEO_I2V_INPUT_STYLE") or "").strip().lower()
+    if raw in ("media", "wan27", "general"):
+        return "media"
+    if raw in ("img_url", "legacy", "wan26"):
+        return "img_url"
+    ml = (model_name or "").strip().lower()
+    return "media" if "wan2.7" in ml else "img_url"
+
+
+def env_video_expert_synthesis_path() -> str:
+    raw = (os.getenv("AIA_VIDEO_EXPERT_SYNTHESIS_PATH") or "").strip().lstrip("/")
+    return raw or _DEFAULT_SYNTHESIS_PATH
+
+
+def _poll_interval_sec() -> float:
+    raw = (os.getenv("AIA_VIDEO_EXPERT_POLL_INTERVAL_SEC") or "").strip()
+    if not raw:
+        return 15.0
+    try:
+        return max(3.0, min(float(raw), 120.0))
+    except ValueError:
+        return 15.0
+
+
+def _max_wait_sec() -> float:
+    raw = (os.getenv("AIA_VIDEO_EXPERT_MAX_WAIT_SEC") or "").strip()
+    if not raw:
+        return 900.0
+    try:
+        return max(30.0, min(float(raw), 3600.0))
+    except ValueError:
+        return 900.0
+
+
+def dashscope_api_root_from_base_url(base_url: str) -> str:
+    """Strip OpenAI-compatible suffixes so native DashScope paths can be joined."""
+    b = (base_url or "").strip().rstrip("/")
+    if not b:
+        return ""
+    root = re.sub(r"/compatible-mode/v\d+(?:/.*)?$", "", b, flags=re.IGNORECASE).rstrip("/")
+    if not root or root == b:
+        root = re.sub(r"/compatible-mode/?$", "", b, flags=re.IGNORECASE).rstrip("/")
+    return root or b
+
+
+def _parameters_extra_from_env() -> dict[str, Any]:
+    raw = (os.getenv("AIA_VIDEO_EXPERT_PARAMETERS_EXTRA") or "").strip()
+    if not raw:
+        return {}
+    try:
+        v = json.loads(raw)
+        return v if isinstance(v, dict) else {}
+    except Exception:
+        return {}
+
+
+def _input_extra_from_env() -> dict[str, Any]:
+    raw = (os.getenv("AIA_VIDEO_EXPERT_INPUT_EXTRA") or "").strip()
+    if not raw:
+        return {}
+    try:
+        v = json.loads(raw)
+        return v if isinstance(v, dict) else {}
+    except Exception:
+        return {}
+
+
+def _dashscope_video_env_parameters() -> dict[str, Any]:
+    out: dict[str, Any] = {}
+    size = (os.getenv("DASHSCOPE_VIDEO_SIZE") or "").strip()
+    if size:
+        out["size"] = size
+    raw_dur = (os.getenv("DASHSCOPE_VIDEO_DURATION") or "").strip()
+    if raw_dur.isdigit():
+        out["duration"] = max(1, min(int(raw_dur), 600))
+    pe = (os.getenv("DASHSCOPE_VIDEO_PROMPT_EXTEND") or "").strip().lower()
+    if pe in ("1", "true", "yes", "on"):
+        out["prompt_extend"] = True
+    elif pe in ("0", "false", "no", "off"):
+        out["prompt_extend"] = False
+    st = (os.getenv("DASHSCOPE_VIDEO_SHOT_TYPE") or "").strip()
+    if st:
+        out["shot_type"] = st
+    wm = (os.getenv("DASHSCOPE_VIDEO_WATERMARK") or "").strip().lower()
+    if wm in ("1", "true", "yes", "on"):
+        out["watermark"] = True
+    elif wm in ("0", "false", "no", "off"):
+        out["watermark"] = False
+    seed = (os.getenv("DASHSCOPE_VIDEO_SEED") or "").strip()
+    if seed.isdigit():
+        out["seed"] = max(0, min(int(seed), 2_147_483_647))
+    return out
+
+
+def _dashscope_video_env_input() -> dict[str, Any]:
+    out: dict[str, Any] = {}
+    neg = os.getenv("DASHSCOPE_VIDEO_NEGATIVE_PROMPT")
+    if neg is not None and str(neg).strip():
+        out["negative_prompt"] = str(neg)
+    audio = (os.getenv("DASHSCOPE_VIDEO_AUDIO_URL") or "").strip()
+    if audio:
+        out["audio_url"] = audio
+    return out
+
+
+def _extract_output(body: dict[str, Any]) -> dict[str, Any]:
+    out = body.get("output")
+    return out if isinstance(out, dict) else {}
+
+
+def _stderr_debug_video(url: str, payload: dict[str, Any]) -> None:
+    if not _truthy_env("AIA_VIDEO_EXPERT_DEBUG_PRINT_PAYLOAD"):
+        return
+    try:
+        txt = json.dumps(payload, ensure_ascii=False, indent=2, default=str)
+        sys.stderr.write(f"\n[oclaw video_generation] POST {url}\n{txt}\n\n")
+        sys.stderr.flush()
+    except Exception:
+        pass
+
+
+def send_video_generation_request(
+    *,
+    prompt: str,
+    model: str | None = None,
+    api_key: str | None = None,
+    base_url: str | None = None,
+    img_url: str | None = None,
+    timeout_submit_sec: float = 120.0,
+    on_progress: Callable[[str], None] | None = None,
+    should_stop: Callable[[], bool] | None = None,
+) -> dict[str, Any]:
+    """Create an async video task and poll until terminal state or timeout.
+
+    When ``img_url`` is set (HTTPS URL or data URL), it becomes the first-frame reference: for Wan **2.7**
+    i2v models this is sent as ``input.media`` with ``type: first_frame`` and ``url``; for older i2v models
+    as ``input.img_url``.     Caller image overrides any overlapping keys from ``AIA_VIDEO_EXPERT_INPUT_EXTRA`` after merge
+    (style selected by model id and ``AIA_VIDEO_I2V_INPUT_STYLE``).
+    """
+    resolved_key = (api_key or env_video_expert_api_key()).strip()
+    raw_base = (base_url or env_video_expert_base_url()).strip()
+    root = dashscope_api_root_from_base_url(raw_base)
+    model_name = ((model or "").strip() or env_video_expert_model()).strip()
+    if not resolved_key or not root:
+        return {
+            "ok": False,
+            "error": "missing AIA_VIDEO_EXPERT_API_KEY or AIA_VIDEO_EXPERT_BASE_URL (or pass api_key and base_url)",
+        }
+    if not model_name:
+        return {
+            "ok": False,
+            "error": "missing model (chosen profile/model=… or set AIA_VIDEO_EXPERT_MODEL)",
+        }
+    prompt_plain = str(prompt or "").strip()
+    if not prompt_plain:
+        prompt_plain = VIDEO_SPECIALIST_DEFAULT_PROMPT_ZH
+
+    model_name = _effective_video_model_for_request(model_name, img_url=img_url)
+
+    path = env_video_expert_synthesis_path()
+    url_submit = join_url(root, path)
+    input_body: dict[str, Any] = {
+        **_input_extra_from_env(),
+        **_dashscope_video_env_input(),
+        "prompt": prompt_plain,
+    }
+    iu = str(img_url or "").strip()
+    if iu:
+        if is_data_url(iu):
+            iu = compress_data_url_image(iu)
+        _style = _i2v_first_frame_input_style(model_name=model_name)
+        if _style == "media":
+            input_body.pop("img_url", None)
+            first = {"type": "first_frame", "url": iu}
+            prev = input_body.get("media")
+            rest: list[dict[str, Any]] = []
+            if isinstance(prev, list):
+                for item in prev:
+                    if isinstance(item, dict) and str(item.get("type") or "").strip().lower() != "first_frame":
+                        rest.append(dict(item))
+            input_body["media"] = [first, *rest]
+        else:
+            input_body.pop("media", None)
+            input_body["img_url"] = iu
+    parameters: dict[str, Any] = {
+        **_dashscope_video_env_parameters(),
+        **_parameters_extra_from_env(),
+    }
+    payload = {"model": model_name, "input": input_body, "parameters": parameters}
+    headers = {
+        "Authorization": f"Bearer {resolved_key}",
+        "Content-Type": "application/json",
+        "X-DashScope-Async": "enable",
+    }
+    _stderr_debug_video(url_submit, payload)
+
+    submit_timeout = httpx.Timeout(max(30.0, float(timeout_submit_sec)), connect=45.0)
+    with httpx.Client(timeout=submit_timeout, follow_redirects=True) as client:
+        try:
+            r = post_with_retry(client, url=url_submit, headers=headers, payload=payload)
+        except Exception as e:
+            return {"ok": False, "error": f"submit failed: {type(e).__name__}: {e}"}
+        if r.status_code >= 400:
+            return {"ok": False, "error": f"submit http {r.status_code}: {r.text[:800]}"}
+        try:
+            body = r.json()
+        except Exception:
+            return {"ok": False, "error": f"submit non-json: {r.text[:500]}"}
+        if not isinstance(body, dict):
+            return {"ok": False, "error": "submit response is not a JSON object"}
+        ok0, err0 = dashscope_multimodal_http_ok(body)
+        if not ok0:
+            return {"ok": False, "error": err0 or "submit rejected by provider"}
+        out0 = _extract_output(body)
+        task_id = str(out0.get("task_id") or "").strip()
+        st0 = str(out0.get("task_status") or "").strip().upper()
+        v0 = str(out0.get("video_url") or "").strip()
+        if st0 == "SUCCEEDED" and v0.startswith(("http://", "https://")):
+            return {"ok": True, "video_urls": [v0], "task_id": task_id or None}
+        if st0 in {"FAILED", "UNKNOWN"}:
+            msg = str(out0.get("message") or out0.get("msg") or "").strip()
+            code = str(out0.get("code") or "").strip()
+            bit = f"{code}: {msg}".strip(": ").strip() or st0
+            return {"ok": False, "error": bit, "task_id": task_id or None}
+        if not task_id:
+            return {"ok": False, "error": "submit response missing output.task_id", "raw": body}
+
+    poll_interval = _poll_interval_sec()
+    max_wait = _max_wait_sec()
+    deadline = time.monotonic() + max_wait
+    task_url = join_url(root, f"api/v1/tasks/{task_id}")
+    headers_get = {"Authorization": f"Bearer {resolved_key}"}
+    poll_timeout = httpx.Timeout(120.0, connect=45.0)
+    with httpx.Client(timeout=poll_timeout, follow_redirects=True) as client:
+        while time.monotonic() < deadline:
+            if should_stop and should_stop():
+                return {"ok": False, "error": "stopped", "task_id": task_id}
+            try:
+                gr = client.get(task_url, headers=headers_get)
+            except Exception as e:
+                return {"ok": False, "error": f"poll failed: {type(e).__name__}: {e}", "task_id": task_id}
+            if gr.status_code >= 400:
+                return {
+                    "ok": False,
+                    "error": f"poll http {gr.status_code}: {gr.text[:800]}",
+                    "task_id": task_id,
+                }
+            try:
+                tb = gr.json()
+            except Exception:
+                return {"ok": False, "error": f"poll non-json: {gr.text[:500]}", "task_id": task_id}
+            if not isinstance(tb, dict):
+                return {"ok": False, "error": "poll response not a JSON object", "task_id": task_id}
+            okp, errp = dashscope_multimodal_http_ok(tb)
+            if not okp:
+                return {"ok": False, "error": errp or "poll rejected", "task_id": task_id}
+            tout = _extract_output(tb)
+            st = str(tout.get("task_status") or "").strip().upper()
+            vurl = str(tout.get("video_url") or "").strip()
+            if st == "SUCCEEDED" and vurl.startswith(("http://", "https://")):
+                return {"ok": True, "video_urls": [vurl], "task_id": task_id}
+            if st in {"FAILED", "UNKNOWN", "CANCELED"}:
+                msg = str(tout.get("message") or tout.get("msg") or "").strip()
+                code = str(tout.get("code") or "").strip()
+                bit = f"{code}: {msg}".strip(": ").strip() or st
+                return {"ok": False, "error": bit, "task_id": task_id}
+            if on_progress and st in {"PENDING", "RUNNING"}:
+                on_progress(f"oclaw: video task {st.lower()}…")
+            time.sleep(poll_interval)
+
+    return {"ok": False, "error": f"poll timeout after {int(max_wait)}s", "task_id": task_id}
+
+
+def materialize_video_output_attachments(
+    video_urls: Any,
+    *,
+    max_videos: int = 1,
+) -> list[dict[str, Any]]:
+    """Download remote MP4 (etc.) into ``video_ref`` rows; on failure fall back to best-effort metadata."""
+    cap = max(1, min(int(max_videos), 4))
+    produced: list[dict[str, Any]] = []
+    if not isinstance(video_urls, list):
+        return produced
+    urls = [str(u).strip() for u in video_urls if str(u).strip().startswith(("http://", "https://"))]
+
+    store = AttachmentAssetStore()
+    ua = "Mozilla/5.0 (compatible; oclaw-video-expert/1.0; +https://github.com/)"
+    for idx, u in enumerate(urls[:cap], start=1):
+        try:
+            blob, ctype = download_http_url_bytes(u, user_agent=ua)
+            if not blob:
+                raise ValueError("empty body")
+            mime = (ctype.split(";", 1)[0].strip() if ctype else "") or "video/mp4"
+            ext = ".mp4"
+            if mime == "video/webm":
+                ext = ".webm"
+            elif mime in ("video/quicktime", "video/mov"):
+                ext = ".mov"
+            meta = store.save_bytes(blob, filename=f"video-output-{idx}{ext}", mime=mime)
+            produced.append(
+                {
+                    "type": "video_ref",
+                    "attachment_id": meta.attachment_id,
+                    "name": meta.name,
+                    "mime": meta.mime,
+                    "bytes": meta.bytes,
+                }
+            )
+        except Exception:
+            produced.append(
+                {
+                    "type": "video_ref",
+                    "name": f"video-output-{idx}.mp4",
+                    "mime": "video/mp4",
+                    "url": u,
+                }
+            )
+    return produced
+
+
+def legacy_video_turn_bundle(resp: dict[str, Any]) -> tuple[bool, str, list[dict[str, Any]]]:
+    ok = bool(resp.get("ok"))
+    text = str(resp.get("text") or "").strip()
+    if not ok:
+        err = str(resp.get("error") or "").strip()
+        return False, f"Video generation failed: {err or 'unknown error'}", []
+    urls = resp.get("video_urls")
+    if not isinstance(urls, list):
+        urls = []
+    atts = materialize_video_output_attachments(urls, max_videos=1)
+    if atts:
+        return True, text, atts
+    if text:
+        return True, text, []
+    return False, "Video specialist failed: empty response from provider.", []
+
+
+def legacy_video_assistant_body_with_placeholder(
+    *,
+    lang: str | None,
+    body_text: str,
+    produced: list[dict[str, Any]] | None,
+) -> str:
+    if str(body_text or "").strip():
+        return str(body_text or "")
+    if produced:
+        return (
+            "Generated video (see attachment below)."
+            if str(lang or "").startswith("en")
+            else "已生成视频（见下方附件）。"
+        )
+    return str(body_text or "")
+
+
+__all__ = [
+    "VIDEO_SPECIALIST_DEFAULT_PROMPT_ZH",
+    "dashscope_api_root_from_base_url",
+    "env_video_expert_api_key",
+    "env_video_expert_base_url",
+    "env_video_expert_model",
+    "legacy_video_assistant_body_with_placeholder",
+    "legacy_video_turn_bundle",
+    "materialize_video_output_attachments",
+    "send_video_generation_request",
+]