# SPDX-License-Identifier: AGPL-3.0-only # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0 """ Async HTTP client for proxying chat completions to external LLM providers. Most registry providers expose OpenAI-compatible /v1/chat/completions endpoints; Anthropic uses native Messages API with translation in this client. """ import json as _json import re from typing import Any, AsyncGenerator, Literal, NamedTuple, Optional import httpx import structlog # Use structlog so INFO-level diagnostics actually surface in the # studio backend's JSON log stream. The stdlib root logger defaults to # WARNING and is not configured with handlers, so plain # `logging.getLogger(__name__).info(...)` was being silently dropped — # only WARNING/ERROR made it through (because they bypassed the root # level threshold via uvicorn's stderr capture). All existing call # sites use printf-style positional args, which structlog accepts. logger = structlog.get_logger(__name__) # Claude 4.7 (Opus/Sonnet/Haiku) removed temperature, top_p, and top_k — # the API returns 400 " is deprecated for this model" if any of # them is set to a non-default value. The "Sampling parameters removed" # section of the 4.7 release notes is the authoritative reference: # https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-7 # 3.x and 4.5/4.6 still accept all three; match the 4-7 line strictly so # the knobs keep working on earlier families. The trailing -4-7[-.]/EOL # anchor keeps future versions (e.g. claude-opus-5) unaffected. _ANTHROPIC_4_7_SAMPLING_REMOVED = re.compile( r"^claude-(?:opus|sonnet|haiku)-4-7(?:[-.]|$)" ) _OPENAI_REASONING_SUMMARY_UNSUPPORTED = re.compile(r"^o3(?:[-.]|$)") class _AnthropicThinkingSpec(NamedTuple): prefixes: tuple[str, ...] kind: Literal["adaptive", "manual"] efforts: tuple[str, ...] _ANTHROPIC_THINKING_SPECS = ( _AnthropicThinkingSpec( prefixes = ("claude-opus-4-7",), kind = "adaptive", efforts = ("none", "low", "medium", "high", "xhigh", "max"), ), _AnthropicThinkingSpec( prefixes = ("claude-opus-4-6", "claude-sonnet-4-6"), kind = "adaptive", efforts = ("none", "low", "medium", "high", "xhigh", "max"), ), _AnthropicThinkingSpec( prefixes = ("claude-opus-4-5", "claude-sonnet-4-5", "claude-haiku-4-5"), kind = "manual", efforts = ("none", "low", "medium", "high"), ), ) def _anthropic_thinking_spec(model: str) -> Optional[_AnthropicThinkingSpec]: for spec in _ANTHROPIC_THINKING_SPECS: if model.startswith(spec.prefixes): return spec return None class _MistralThinkingSpec(NamedTuple): models: tuple[str, ...] style: Literal["prompt_mode", "reasoning_effort", "disabled"] efforts: tuple[str, ...] = () _MISTRAL_THINKING_SPECS = ( _MistralThinkingSpec( models = ("magistral-medium-latest",), style = "prompt_mode", ), _MistralThinkingSpec( models = ("mistral-small-latest", "mistral-vibe-cli-latest"), style = "reasoning_effort", efforts = ("none", "high"), ), ) _OPENROUTER_MANDATORY_REASONING_MODELS = frozenset( { "~google/gemini-pro-latest", "baidu/cobuddy:free", "inclusionai/ring-2.6-1t:free", "deepseek/deepseek-r1", } ) def _mistral_thinking_spec(model: str) -> _MistralThinkingSpec: for spec in _MISTRAL_THINKING_SPECS: if model in spec.models: return spec return _MistralThinkingSpec(models = (), style = "disabled") def _apply_mistral_reasoning_controls( body: dict[str, Any], model: str, enable_thinking: Optional[bool], reasoning_effort: Optional[str], ) -> None: """ Translate generic reasoning controls into Mistral's model-specific shape. Current contract: - magistral-medium-latest: baseline (no extra field) or `prompt_mode="reasoning"` for the explicit reasoning mode. - mistral-small-latest / mistral-vibe-cli-latest: `reasoning_effort` in {"none", "high"}. - all other tested Mistral models: no reasoning/thinking params. """ model_for_matching = model.rsplit("/", 1)[-1].strip().lower() spec = _mistral_thinking_spec(model_for_matching) body.pop("prompt_mode", None) body.pop("reasoning_effort", None) if spec.style == "prompt_mode": # Magistral baseline is already reasoning-capable. The explicit # prompt_mode path is only used for the "high" UI selection. if enable_thinking is True or reasoning_effort == "high": body["prompt_mode"] = "reasoning" return if spec.style == "reasoning_effort": if reasoning_effort in spec.efforts: body["reasoning_effort"] = reasoning_effort elif enable_thinking is False: body["reasoning_effort"] = "none" elif enable_thinking is True: body["reasoning_effort"] = "high" # Shared client reused across all requests for HTTP connection pooling. # Auth headers and timeouts are passed per-request, so a single client # handles every provider without storing credentials. _http_client = httpx.AsyncClient() def _build_kimi_tool_end( synthetic_chunk_fn: Any, tool_call_id: str, citations: list[dict[str, str]], ) -> str: """Format Kimi web_search citations into the tool_end payload. Same shape parseSourcesFromResult on the frontend expects for the other built-in web_search providers: `Title: ...\\nURL: ...\\n Snippet: ...\\n---\\n...`. If no citations were emitted, fall back to a generic "(search complete)" string so the UI still shows the tool card transitioning to a completed state. """ blocks: list[str] = [] for cit in citations: line = f"Title: {cit['title']}\nURL: {cit['url']}" if cit.get("snippet"): line += f"\nSnippet: {cit['snippet']}" blocks.append(line) return synthetic_chunk_fn( { "type": "tool_end", "tool_call_id": tool_call_id, "result": "\n---\n".join(blocks) if blocks else "(search complete)", } ) class ExternalProviderClient: """Async proxy for OpenAI-compatible external LLM APIs.""" def __init__( self, provider_type: str, base_url: str, api_key: str, timeout: float = 120.0, ): self.provider_type = provider_type self.base_url = base_url.rstrip("/") self.api_key = api_key self._timeout = httpx.Timeout(timeout, connect = 10.0) # Separate timeout for SSE streams: reasoning-heavy providers # (Anthropic Opus 4.7 with adaptive thinking, OpenAI gpt-5.x via # /v1/responses) can pause for tens of seconds between bytes # while the model is internally thinking. httpx's read timeout is # the *gap* between successive reads, not a wall clock — so # disabling it lets long thinks complete without cutting the # stream prematurely. connect/write/pool keep the 10s / 120s # bounds so genuine network failures still surface. self._stream_timeout = httpx.Timeout(timeout, connect = 10.0, read = None) def _auth_headers(self) -> dict[str, str]: """Build authentication headers using the provider's registry config.""" from core.inference.providers import get_provider_info provider_info = get_provider_info(self.provider_type) or {} auth_header = provider_info.get("auth_header", "Authorization") auth_prefix = provider_info.get("auth_prefix", "Bearer ") headers = {"Content-Type": "application/json"} # Skip auth header when api_key is empty (optional for local providers); # httpx rejects an empty `Bearer ` value as "Illegal header value". if self.api_key: headers[auth_header] = f"{auth_prefix}{self.api_key}" # Merge any provider-specific extra headers (e.g. anthropic-version, OpenRouter attribution) headers.update(provider_info.get("extra_headers", {})) return headers def _is_openai_compatible(self) -> bool: """Return False for providers that need request/response translation (e.g. Anthropic).""" from core.inference.providers import get_provider_info info = get_provider_info(self.provider_type) or {} return info.get("openai_compatible", True) async def stream_chat_completion( self, messages: list[dict[str, Any]], model: str, temperature: float = 0.7, top_p: float = 0.95, max_tokens: Optional[int] = None, presence_penalty: float = 0.0, top_k: Optional[int] = None, enable_thinking: Optional[bool] = None, reasoning_effort: Optional[str] = None, enabled_tools: Optional[list[str]] = None, enable_prompt_caching: Optional[bool] = None, openai_code_exec_container_id: Optional[str] = None, stream: bool = True, ) -> AsyncGenerator[str, None]: """ Yield OpenAI-format SSE lines from the external provider. For OpenAI-compatible providers, lines are forwarded verbatim. For Anthropic, the native Messages API SSE is translated to OpenAI format. ``top_k`` and ``presence_penalty`` are forwarded only when the caller supplies a value the provider accepts — the frontend's provider-capability map already filters these per provider, so we treat them as opt-in here. """ if not self._is_openai_compatible(): async for line in self._stream_anthropic( messages, model, temperature, top_p, max_tokens, top_k, enable_thinking, reasoning_effort, enabled_tools, enable_prompt_caching, ): yield line return # OpenAI moved their flagship models (gpt-5.x) off /v1/chat/completions # — those endpoints return 404 with "This is not a chat model" for the # new families. Route all OpenAI traffic through /v1/responses instead; # we translate the Responses SSE back into Chat Completions chunks so # the frontend stays endpoint-agnostic. if self.provider_type == "openai": async for line in self._stream_openai_responses( messages, model, temperature, top_p, max_tokens, enable_thinking, reasoning_effort, enabled_tools, enable_prompt_caching, openai_code_exec_container_id, ): yield line return # Kimi's $web_search is a builtin_function that requires a client # round-trip: the first call returns a tool_calls envelope with # function.arguments populated; the caller echoes those arguments # back as a role=tool message; the second call streams the final # answer with the search incorporated. The doc also mandates # disabling thinking while $web_search is active. Route to a # dedicated helper so the default OAI-compat path stays single-pass. # https://platform.kimi.ai/docs/guide/use-web-search if ( self.provider_type == "kimi" and enabled_tools and "web_search" in enabled_tools ): async for line in self._stream_kimi_web_search( messages, model, max_tokens, ): yield line return body: dict[str, Any] = { "model": model, "messages": messages, "stream": stream, "temperature": temperature, "top_p": top_p, "presence_penalty": presence_penalty, } if max_tokens is not None: # OpenAI newer models (gpt-4o, gpt-5.x) reject max_tokens if self.provider_type == "openai": body["max_completion_tokens"] = max_tokens else: body["max_tokens"] = max_tokens # Strip body fields a provider's registry entry declares unusable — # reasoning-class models that lock these to fixed defaults (e.g. # Kimi k2.5/k2.6 only accept temperature=1, top_p=1) 400 otherwise. # The frontend capability map already hides the matching sliders; # this is the matching guard for the pydantic default that the # route layer would otherwise still fill in. from core.inference.providers import get_provider_info provider_info = get_provider_info(self.provider_type) or {} for field in provider_info.get("body_omit", ()): body.pop(field, None) # Kimi (kimi-k2.6, kimi-k2-thinking) accepts a boolean thinking toggle # via a top-level `thinking` field (the docs show it nested under # extra_body, but that is an OpenAI Python SDK convention; on the # wire it merges into the request body). # - kimi-k2.6 defaults to thinking enabled; clients can pass # {"type": "disabled"} to suppress it. # - kimi-k2-thinking is always on; we never send disabled there. # `keep: all` retains every thinking chunk through the stream, which # is what we need so our frontend can wrap reasoning_content into # the chat reasoning panel. if self.provider_type == "kimi" and enable_thinking is not None: if model == "kimi-k2-thinking": # Always on; ignore client toggle to avoid an API-level reject. pass elif enable_thinking: body["thinking"] = {"type": "enabled", "keep": "all"} else: body["thinking"] = {"type": "disabled"} elif self.provider_type == "mistral": _apply_mistral_reasoning_controls( body, model, enable_thinking, reasoning_effort ) elif self.provider_type == "vllm" and enable_thinking is not None: # vLLM gates thinking via chat_template_kwargs.enable_thinking. tpl_kw = body.get("chat_template_kwargs") if not isinstance(tpl_kw, dict): tpl_kw = {} tpl_kw["enable_thinking"] = bool(enable_thinking) body["chat_template_kwargs"] = tpl_kw # OpenRouter exposes a unified `reasoning` parameter on every # chat-completion request — the gateway routes it to whichever # underlying model actually supports reasoning, and silently # no-ops for ones that don't. Documented at # https://openrouter.ai/docs/guides/best-practices/reasoning-tokens # Shape: `reasoning: {enabled?: bool, effort?: low|medium|high, # max_tokens?: N, exclude?: bool}` with effort and max_tokens # mutually exclusive. We forward either an effort level (when # the user picked one) or a bare {enabled: true}. A small set of # known routes rejects explicit disable with 400 ("Reasoning is # mandatory for this endpoint ..."), so only those omit "off". if self.provider_type == "openrouter": normalized_or_model = model.strip().lower() if reasoning_effort in ("low", "medium", "high"): body["reasoning"] = {"effort": reasoning_effort} elif enable_thinking is True: body["reasoning"] = {"enabled": True} elif enable_thinking is False: if normalized_or_model in _OPENROUTER_MANDATORY_REASONING_MODELS: body.pop("reasoning", None) else: body["reasoning"] = {"enabled": False} # OpenRouter web-search plugin — universal shape that works # for every model id, including the `openrouter/free` and # `openrouter/auto` meta-routers. Documented at # https://openrouter.ai/docs/guides/features/plugins/web-search # The `:online` model-suffix shortcut is "exactly equivalent # to" this plugin per the same doc, but only works on # concrete model ids — meta-routers reject the suffix. # `plugins: [{id: "web"}]` works everywhere, no model id # rewrite needed, and idempotent if some future call site # adds the entry first. if enabled_tools and "web_search" in enabled_tools: plugins = list(body.get("plugins") or []) if not any( isinstance(p, dict) and p.get("id") == "web" for p in plugins ): plugins.append({"id": "web"}) body["plugins"] = plugins logger.info( "OpenRouter web_search: attached plugins=[{id: 'web'}] " "(model=%s)", body.get("model"), ) url = f"{self.base_url}/chat/completions" logger.info( "Proxying chat completion to %s (provider=%s, model=%s)", url, self.provider_type, model, ) try: async with _http_client.stream( "POST", url, json = body, headers = self._auth_headers(), timeout = self._stream_timeout, ) as response: if response.status_code != 200: error_body = await response.aread() error_text = error_body.decode("utf-8", errors = "replace") logger.error( "External provider returned %d: %s", response.status_code, error_text[:500], ) yield _error_sse_line( response.status_code, error_text, self.provider_type ) return # NOTE: manual __anext__ loop instead of `async for` is intentional. # On Python 3.13 + httpcore 1.0.x, `async for` auto-calls aclose() on # early exit (break/return/GeneratorExit) BEFORE our finally block runs. # That propagates GeneratorExit into PoolByteStream.__aiter__() while it # calls `await self.aclose()` inside `with AsyncShieldCancellation()`, # triggering "RuntimeError: async generator ignored GeneratorExit". # Fix: call response.aclose() FIRST (sets PoolByteStream._closed=True), # then lines_gen.aclose() is a no-op and GeneratorExit re-raises cleanly. lines_gen = response.aiter_lines().__aiter__() # Best-effort diagnostics for the default OAI-compat path. Without # this, OpenRouter mid-stream errors (200 OK + error event in the # SSE body) and OpenRouter-router model selection were invisible # in the backend logs — the user only saw "Provider returned # error" in the UI with no trail on the server side. event_counts: dict[str, int] = {} chosen_model: Optional[str] = None # Web-search tool-card synthesis for OpenRouter. The gateway # doesn't emit structured web_search_call events — citations # come back as `annotations` of type=url_citation on delta / # message objects. Mirror the OpenAI/Anthropic UX by yielding # a synthetic tool_start at stream open and tool_end at # stream close with the collected citation list. web_search_active = ( self.provider_type == "openrouter" and bool(enabled_tools) and "web_search" in (enabled_tools or []) ) web_search_tool_id = "openrouter_web_search" web_search_citations: list[dict[str, str]] = [] web_search_tool_started = False web_search_tool_ended = False def _emit_synthetic_tool_event(payload: dict[str, Any]) -> str: chunk = { "id": f"chatcmpl-{self.provider_type}-synthetic", "object": "chat.completion.chunk", "choices": [ { "index": 0, "delta": {}, "finish_reason": None, } ], "_toolEvent": payload, } return f"data: {_json.dumps(chunk)}" def _record_or_url_citation(payload: Any) -> None: if not isinstance(payload, dict): return if payload.get("type") != "url_citation": return # OpenRouter (and OpenAI Chat Completions web_search) # nest the citation under url_citation; some variants # ship the fields flat on the annotation itself. Accept # both. cit = payload.get("url_citation") if not isinstance(cit, dict): cit = payload url = cit.get("url", "") if isinstance(cit, dict) else "" if not url or not isinstance(url, str): return if any(c["url"] == url for c in web_search_citations): return title = cit.get("title") or url snippet = cit.get("content") or cit.get("snippet") or "" web_search_citations.append( { "url": url, "title": title, "snippet": snippet if isinstance(snippet, str) else "", } ) def _build_web_search_tool_end() -> str: blocks: list[str] = [] for cit in web_search_citations: line = f"Title: {cit['title']}\nURL: {cit['url']}" if cit.get("snippet"): line += f"\nSnippet: {cit['snippet']}" blocks.append(line) return _emit_synthetic_tool_event( { "type": "tool_end", "tool_call_id": web_search_tool_id, "result": ( "\n---\n".join(blocks) if blocks else "(search complete)" ), } ) if web_search_active: yield _emit_synthetic_tool_event( { "type": "tool_start", "tool_name": "web_search", "tool_call_id": web_search_tool_id, "arguments": {}, } ) web_search_tool_started = True try: while True: try: line = await lines_gen.__anext__() except StopAsyncIteration: break if not line.strip(): continue if line.startswith("data:"): data_str = line[len("data:") :].strip() if data_str == "[DONE]": event_counts["done"] = event_counts.get("done", 0) + 1 # Emit synthetic tool_end with collected # citations BEFORE forwarding [DONE], so the # tool-card transitions to "complete" in the # UI before the stream closes. if ( web_search_active and web_search_tool_started and not web_search_tool_ended ): yield _build_web_search_tool_end() web_search_tool_ended = True elif data_str: try: parsed = _json.loads(data_str) except Exception: parsed = None if isinstance(parsed, dict): # Mid-stream provider error event. OpenRouter # in particular returns 200 then surfaces the # actual failure as an SSE error event. if "error" in parsed: event_counts["error"] = ( event_counts.get("error", 0) + 1 ) logger.warning( "%s SSE error event: %s", self.provider_type, parsed.get("error"), ) else: event_counts["delta"] = ( event_counts.get("delta", 0) + 1 ) # OpenRouter (and most OAI-compat providers) # report the underlying model that handled # the request in every chunk's `model` field. # Latch the first non-empty value so the # router-picked model surfaces in logs and # is available to the proxy caller. if chosen_model is None and isinstance( parsed.get("model"), str ): chosen_model = parsed["model"] # When the user has web_search on, scan # every chunk's delta and message # objects for url_citation annotations. # Different OpenRouter upstreams place # them in different spots. if web_search_active: choices = parsed.get("choices") or [] if isinstance(choices, list): for choice in choices: if not isinstance(choice, dict): continue for envelope in ( choice.get("delta"), choice.get("message"), ): if not isinstance(envelope, dict): continue for ann in ( envelope.get("annotations") or [] ): _record_or_url_citation(ann) yield line # Stream ended without [DONE] (some upstreams just close # the connection). Emit tool_end so the card doesn't # stay in "running" forever. if ( web_search_active and web_search_tool_started and not web_search_tool_ended ): yield _build_web_search_tool_end() web_search_tool_ended = True except GeneratorExit: await response.aclose() # set PoolByteStream._closed=True FIRST await lines_gen.aclose() # now safe — aclose() is a no-op raise finally: logger.info( "%s stream complete (model=%s, chosen=%s, " "web_search_requested=%s, citations=%s, events=%s)", self.provider_type, model, chosen_model, web_search_active, len(web_search_citations), event_counts, ) await response.aclose() await lines_gen.aclose() except httpx.ConnectError as exc: logger.error("Connection error to %s: %s", self.provider_type, exc) yield _error_sse_line( 502, f"Failed to connect to {self.provider_type}: {exc}", self.provider_type, ) except httpx.ReadTimeout as exc: logger.error("Read timeout from %s: %s", self.provider_type, exc) yield _error_sse_line( 504, f"Timeout waiting for {self.provider_type} response", self.provider_type, ) except httpx.HTTPError as exc: logger.error("HTTP error from %s: %s", self.provider_type, exc) yield _error_sse_line( 502, f"Error communicating with {self.provider_type}: {exc}", self.provider_type, ) async def _stream_kimi_web_search( self, messages: list[dict[str, Any]], model: str, max_tokens: Optional[int], ) -> AsyncGenerator[str, None]: """ Kimi $web_search round-trip. Wire flow (per https://platform.kimi.ai/docs/guide/use-web-search): 1. POST messages with tools=[{type: "builtin_function", function: {name: "$web_search"}}] and thinking=disabled. 2. Stream the first response — accumulate function.arguments across tool_call deltas until finish_reason="tool_calls". Do NOT forward those tool_call chunks to the client (they are an internal protocol step, not user-visible output). 3. Build a second request: original messages + the assistant message carrying the tool_calls + a role=tool message that echoes the same arguments back verbatim (per Kimi docs, the caller "just needs to submit tool_call.function.arguments to Kimi as they are" — the server actually runs the search). 4. Stream the second response — that is the final answer the user sees, with search results already incorporated. We synthesize tool_start (with the parsed query) when step (2) completes, and tool_end (with any url_citation annotations the second stream emits) before [DONE], so the chat UI shows the same web-search tool card as the other providers. """ url = f"{self.base_url}/chat/completions" body: dict[str, Any] = { "model": model, "messages": messages, "stream": True, # $web_search forbids thinking; sending the toggle silently # would have the server reject the request with 400. "thinking": {"type": "disabled"}, "tools": [ {"type": "builtin_function", "function": {"name": "$web_search"}} ], } if max_tokens is not None: body["max_tokens"] = max_tokens # Strip body fields the Kimi registry declares unusable # (temperature/top_p — see body_omit in providers.py). from core.inference.providers import get_provider_info provider_info = get_provider_info(self.provider_type) or {} for field in provider_info.get("body_omit", ()): body.pop(field, None) tool_call_id = "kimi_web_search" synthetic_id = f"chatcmpl-{self.provider_type}-synthetic" def _synthetic_chunk(payload: dict[str, Any]) -> str: chunk = { "id": synthetic_id, "object": "chat.completion.chunk", "choices": [{"index": 0, "delta": {}, "finish_reason": None}], "_toolEvent": payload, } return f"data: {_json.dumps(chunk)}" logger.info( "Kimi $web_search round-trip starting (model=%s, url=%s)", model, url, ) # ---- First call: collect the model's $web_search tool_call ---- tool_calls_acc: dict[int, dict[str, Any]] = {} try: async with _http_client.stream( "POST", url, json = body, headers = self._auth_headers(), timeout = self._stream_timeout, ) as response: if response.status_code != 200: error_body = await response.aread() error_text = error_body.decode("utf-8", errors = "replace") logger.error( "Kimi first-call returned %d: %s", response.status_code, error_text[:500], ) yield _error_sse_line( response.status_code, error_text, self.provider_type ) return lines_gen = response.aiter_lines().__aiter__() try: while True: try: line = await lines_gen.__anext__() except StopAsyncIteration: break if not line.strip() or not line.startswith("data:"): continue data_str = line[len("data:") :].strip() if data_str == "[DONE]": break try: parsed = _json.loads(data_str) except Exception: continue for choice in parsed.get("choices") or []: if not isinstance(choice, dict): continue delta = choice.get("delta") or {} for tc in delta.get("tool_calls") or []: if not isinstance(tc, dict): continue idx = tc.get("index", 0) slot = tool_calls_acc.setdefault( idx, { "id": tc.get("id") or f"call_{idx}", "type": "function", "function": {"name": "", "arguments": ""}, }, ) if tc.get("id"): slot["id"] = tc["id"] fn = tc.get("function") or {} if fn.get("name"): slot["function"]["name"] = fn["name"] if fn.get("arguments"): slot["function"]["arguments"] += fn["arguments"] if choice.get("finish_reason") == "tool_calls": break except GeneratorExit: await response.aclose() await lines_gen.aclose() raise finally: await response.aclose() await lines_gen.aclose() except httpx.HTTPError as exc: logger.error("Kimi first-call HTTP error: %s", exc) yield _error_sse_line( 502, f"Error communicating with kimi: {exc}", self.provider_type, ) return # If the model decided not to search, fall back to a plain # streaming call without the builtin tool. That mirrors the UX # of every other provider when web_search is on but the model # didn't actually need it. search_calls = [ tc for tc in tool_calls_acc.values() if tc["function"]["name"] == "$web_search" ] if not search_calls: logger.info( "Kimi $web_search: model did not invoke search; " "falling back to plain stream" ) fallback_body = dict(body) fallback_body.pop("tools", None) try: async with _http_client.stream( "POST", url, json = fallback_body, headers = self._auth_headers(), timeout = self._stream_timeout, ) as response: if response.status_code != 200: error_body = await response.aread() error_text = error_body.decode("utf-8", errors = "replace") logger.error( "Kimi fallback returned %d: %s", response.status_code, error_text[:500], ) yield _error_sse_line( response.status_code, error_text, self.provider_type ) return # Manual __anext__ loop instead of `async for` — see the # comment in stream_chat_completion for the Python 3.13 + # httpcore 1.0.x GeneratorExit interaction this avoids. lines_gen = response.aiter_lines().__aiter__() try: while True: try: line = await lines_gen.__anext__() except StopAsyncIteration: break if line.strip(): yield line except GeneratorExit: await response.aclose() await lines_gen.aclose() raise finally: await response.aclose() await lines_gen.aclose() except httpx.HTTPError as exc: logger.error("Kimi fallback HTTP error: %s", exc) yield _error_sse_line( 502, f"Error communicating with kimi: {exc}", self.provider_type, ) return # Synthesize tool_start with the parsed search query so the # chat UI's web-search card shows "Searching for: ...". first_args_raw = search_calls[0]["function"]["arguments"] or "{}" try: first_args = _json.loads(first_args_raw) except Exception: first_args = {} # Log the raw arguments so we can confirm the server actually # ran the search. The shape is documented loosely but in practice # the model emits `{"search_result":{"search_id":...}, # "usage":{"total_tokens":N}}` — an opaque receipt where N is the # token cost of the injected search context. The query string is # NOT present; Kimi runs the search server-side during the first # call and bakes the results straight into the model's context. logger.info( "Kimi $web_search: %d tool_call(s), args[0]=%s", len(search_calls), first_args_raw[:500], ) first_args_search_tokens: Optional[int] = None if isinstance(first_args, dict): usage_block = first_args.get("usage") if isinstance(usage_block, dict): tok = usage_block.get("total_tokens") if isinstance(tok, int): first_args_search_tokens = tok yield _synthetic_chunk( { "type": "tool_start", "tool_name": "web_search", "tool_call_id": tool_call_id, "arguments": first_args if isinstance(first_args, dict) else {}, } ) # Kimi's search has already executed server-side by the time the # first call returns (the tool_call envelope encodes the search # result reference, not a query for us to dispatch). Emit # tool_end NOW so the UI's web-search card transitions to # "complete" before the second call starts streaming the # answer, instead of after — otherwise the card sits in # "running" all the way through the answer streaming and the # user perceives the model answering before search finishes. yield _build_kimi_tool_end(_synthetic_chunk, tool_call_id, []) # ---- Second call: echo the tool_calls back and stream answer ---- assistant_msg = { "role": "assistant", "content": "", "tool_calls": list(tool_calls_acc.values()), } tool_msgs = [ { "role": "tool", "tool_call_id": tc["id"], "name": tc["function"]["name"], "content": tc["function"]["arguments"], } for tc in tool_calls_acc.values() ] followup_body = dict(body) followup_body["messages"] = list(messages) + [assistant_msg] + tool_msgs # Ask the SSE stream to include a final `usage` block so we can # see prompt_tokens (which jumps to thousands when the server # injects search context). Without this, OpenAI-compat streams # omit usage entirely. Kimi follows the same convention. followup_body["stream_options"] = {"include_usage": True} # Keep the tool definition on the second call so the model can # decide to search again mid-turn if needed. Kimi's doc shows # the same tools array on every step. try: async with _http_client.stream( "POST", url, json = followup_body, headers = self._auth_headers(), timeout = self._stream_timeout, ) as response: if response.status_code != 200: error_body = await response.aread() error_text = error_body.decode("utf-8", errors = "replace") logger.error( "Kimi second-call returned %d: %s", response.status_code, error_text[:500], ) yield _error_sse_line( response.status_code, error_text, self.provider_type ) return lines_gen = response.aiter_lines().__aiter__() # Diagnostics: latch usage.prompt_tokens from the final # chunk. The Kimi docs say search results count toward # prompt_tokens, so a big value here is direct evidence # the server actually injected results into context. last_usage: Optional[dict[str, Any]] = None annotation_shapes: set[str] = set() try: while True: try: line = await lines_gen.__anext__() except StopAsyncIteration: break if not line.strip(): continue if line.startswith("data:"): data_str = line[len("data:") :].strip() if data_str and data_str != "[DONE]": try: parsed = _json.loads(data_str) except Exception: parsed = None if isinstance(parsed, dict): usage = parsed.get("usage") if isinstance(usage, dict): last_usage = usage # Scan annotations only for diagnostics — # Kimi today doesn't emit url_citation, but # if a future model version starts to we'll # see the type name in the final log line # and can wire it into the tool_end payload. for choice in parsed.get("choices") or []: if not isinstance(choice, dict): continue for envelope in ( choice.get("delta"), choice.get("message"), ): if not isinstance(envelope, dict): continue for ann in ( envelope.get("annotations") or [] ): if isinstance(ann, dict): annotation_shapes.add( str(ann.get("type") or "?") ) yield line except GeneratorExit: await response.aclose() await lines_gen.aclose() raise finally: logger.info( "Kimi $web_search complete (model=%s, " "search_ctx_tokens=%s, annotation_types=%s, " "prompt_tokens=%s, completion_tokens=%s)", model, first_args_search_tokens, sorted(annotation_shapes) or None, (last_usage or {}).get("prompt_tokens"), (last_usage or {}).get("completion_tokens"), ) await response.aclose() await lines_gen.aclose() except httpx.HTTPError as exc: logger.error("Kimi second-call HTTP error: %s", exc) yield _error_sse_line( 502, f"Error communicating with kimi: {exc}", self.provider_type, ) async def _stream_anthropic( self, messages: list[dict[str, Any]], model: str, temperature: float, top_p: float, max_tokens: Optional[int], top_k: Optional[int] = None, enable_thinking: Optional[bool] = None, reasoning_effort: Optional[str] = None, enabled_tools: Optional[list[str]] = None, enable_prompt_caching: Optional[bool] = None, ) -> AsyncGenerator[str, None]: """ Call the Anthropic Messages API and translate its SSE to OpenAI format. Anthropic SSE event types: content_block_delta → OpenAI chunk with delta.content message_delta → OpenAI chunk with finish_reason message_stop → data: [DONE] (all others skipped) """ import json as _json # Extract system prompt and translate image_url parts to Anthropic format system: Optional[str] = None filtered: list[dict[str, Any]] = [] for msg in messages: if msg.get("role") == "system": content = msg.get("content", "") system = ( content if isinstance(content, str) else "\n".join( p["text"] for p in content if p.get("type") == "text" ) ) continue content = msg.get("content") if isinstance(content, list): # Translate OpenAI image_url parts → Anthropic native image format anthropic_parts: list[dict[str, Any]] = [] for part in content: if part.get("type") == "text": anthropic_parts.append({"type": "text", "text": part["text"]}) elif part.get("type") == "image_url": url = part.get("image_url", {}).get("url", "") if url.startswith("data:"): # data:image/png;base64, → split header and data header, _, b64data = url.partition(",") media_type = ( header.split(";")[0].replace("data:", "") or "image/jpeg" ) anthropic_parts.append( { "type": "image", "source": { "type": "base64", "media_type": media_type, "data": b64data, }, } ) else: # Remote URL — Anthropic supports url source type natively. # See: https://docs.anthropic.com/en/docs/build-with-claude/vision#url-based-images anthropic_parts.append( { "type": "image", "source": { "type": "url", "url": url, }, } ) filtered.append({"role": msg["role"], "content": anthropic_parts}) else: filtered.append(msg) # Claude 4.7 family removed temperature / top_p / top_k entirely. # The earlier guard only handled top_k; temperature is now also # rejected with 400 "temperature is deprecated for this model". # Latch the match once and reuse it everywhere temperature or # top_k would otherwise be set — including the thinking-mode # override below, which used to force temperature=1. sampling_removed = bool(_ANTHROPIC_4_7_SAMPLING_REMOVED.match(model)) body: dict[str, Any] = { "model": model, "messages": filtered, "max_tokens": max_tokens or 1024, # required by Anthropic "stream": True, } if not sampling_removed: body["temperature"] = temperature if top_k is not None and top_k > 0 and not sampling_removed: body["top_k"] = top_k # Anthropic only caches a prefix when at least one cache_control # marker is attached to it — the frontend defaults # enable_prompt_caching to True for Anthropic, so treat `None` the # same as True here (callers that don't set the flag still get # caching). Pass False explicitly to opt out. prompt_caching_enabled = enable_prompt_caching is not False if system: if prompt_caching_enabled: # System block is the most stable prefix across turns, so # it gets its own breakpoint. Skipped when system is # empty — there's nothing to cache, and an empty marker # is a no-op. body["system"] = [ { "type": "text", "text": system, "cache_control": {"type": "ephemeral"}, } ] else: body["system"] = system if prompt_caching_enabled and filtered: # Second breakpoint at the end of the conversation. Anthropic # caches the longest matching prefix up to a cache_control # marker; placing one on the latest message means turn N+1 # rehydrates everything up through turn N from cache instead # of recomputing it. This is what makes caching actually work # when the system prompt is empty or shorter than Anthropic's # ~1024-token cache floor — the conversation history carries # the bulk of the input tokens. Anthropic allows up to 4 # breakpoints per request; we use at most 2 (system + tail). last_msg = filtered[-1] content = last_msg.get("content") if isinstance(content, str): last_msg["content"] = [ { "type": "text", "text": content, "cache_control": {"type": "ephemeral"}, } ] elif isinstance(content, list) and content: # Don't mutate the caller's list. Rebuild the tail with # cache_control attached to the final block so an # upstream image-bearing turn still cleanly slots into # the cache as part of the conversational prefix. head = list(content[:-1]) tail = content[-1] if isinstance(tail, dict): head.append({**tail, "cache_control": {"type": "ephemeral"}}) else: head.append(tail) last_msg["content"] = head thinking_spec = _anthropic_thinking_spec(model) allowed_efforts = ( thinking_spec.efforts if thinking_spec else ("none", "low", "medium", "high") ) effort = reasoning_effort if reasoning_effort in allowed_efforts else None # Claude 4.6 Opus/Sonnet accept top-tier adaptive effort as "max" only; # "xhigh" is rejected (supported on Claude 4.7). Map our shared "xhigh" # semantic to "max" for 4.6 outbound requests while still accepting # both in ``allowed_efforts`` for persisted / cross-provider UI state. if effort == "xhigh" and model.startswith( ("claude-opus-4-6", "claude-sonnet-4-6") ): effort = "max" if effort is None: if enable_thinking is False: effort = "none" elif enable_thinking is True: effort = "medium" # Normalize one semantic Thinking control into Anthropic's two model-era # APIs: adaptive effort on Claude 4.6/4.7, manual budget_tokens on 4.5. if effort and effort != "none": # Anthropic rejects top_k whenever thinking is enabled. body.pop("top_k", None) # Earlier families (4.5/4.6) require temperature=1 when # thinking is enabled and forbid top_p in the same request: # "temperature and top_p cannot both be specified for this # model. Please use only one." # On Claude 4.7, temperature was removed entirely — sending # any value (including 1) returns 400 — so skip the override # there and let the model use its default sampling. if not sampling_removed: body["temperature"] = 1 body.pop("top_p", None) if thinking_spec and thinking_spec.kind == "adaptive": # `display` defaults to "omitted" on Claude Opus 4.7 (per the # adaptive-thinking docs) — without an explicit opt-in the # API emits an empty thinking block plus a signature_delta, # so our SSE handler would surface a stray # and the reasoning panel would stay blank. Force # "summarized" so 4.7 streams thinking_delta events like # 4.6 does. On 4.6 / Sonnet 4.6 this is the default, so # setting it explicitly is harmless. body["thinking"] = {"type": "adaptive", "display": "summarized"} # Per the Messages API reference, the effort knob for # adaptive thinking lives under `output_config.effort` — # NOT as a top-level field. Sending `effort: ...` directly # produces a 400 "effort: Extra inputs are not permitted". # Allowed values: low | medium | high | xhigh | max. See: # https://platform.claude.com/docs/en/api/messages body["output_config"] = {"effort": effort} elif thinking_spec and thinking_spec.kind == "manual": budget_tokens = {"low": 1024, "medium": 2048, "high": 4096}[effort] body["thinking"] = { "type": "enabled", "budget_tokens": budget_tokens, } # Anthropic requires max_tokens to be strictly greater than # thinking.budget_tokens on the manual-thinking path. if body.get("max_tokens", 0) <= budget_tokens: body["max_tokens"] = budget_tokens + 1024 # Anthropic server-side web_search — see # https://docs.anthropic.com/en/docs/agents-and-tools/tool-use/web-search-tool # The tool type is date-pinned (web_search_20250305 today) and # Anthropic dispatches search calls server-side, returning # server_tool_use + web_search_tool_result blocks in the SSE # stream, plus url-citation annotations on text deltas. We # translate all of that into our local _toolEvent shape so the # chat UI renders web_search exactly like OpenAI's path. if enabled_tools and "web_search" in enabled_tools: anthropic_tools = list(body.get("tools") or []) anthropic_tools.append( { "type": "web_search_20250305", "name": "web_search", "max_uses": 5, } ) body["tools"] = anthropic_tools # Anthropic server-side code execution — see # https://platform.claude.com/docs/en/agents-and-tools/tool-use/code-execution-tool # `code_execution_20250825` runs Python + bash + str_replace # file edits inside a 5 GB sandboxed container per request, with # no internet access. The tool entry itself takes no extra # parameters; on the SSE stream Anthropic emits two sub-tool # names — `bash_code_execution` and # `text_editor_code_execution` — wrapped in the standard # server_tool_use / *_tool_result block shape. The matching # beta header (`code-execution-2025-08-25`) is set further down # in this function alongside the request headers. # v1 wires the tool only; file uploads (container_upload # content blocks and generated-file retrieval via the Files # API) are a deliberate follow-up. code_execution_enabled = bool( enabled_tools and "code_execution" in enabled_tools ) if code_execution_enabled: anthropic_tools = list(body.get("tools") or []) anthropic_tools.append( { "type": "code_execution_20250825", "name": "code_execution", } ) body["tools"] = anthropic_tools url = f"{self.base_url}/messages" completion_id = f"chatcmpl-anthropic-{model.replace('/', '-')}" # Log the outgoing config keys (not the messages themselves) so we # can prove which thinking/effort fields actually reached the wire. # If Anthropic skips reasoning despite a configured effort, this # tells us whether we sent the field or dropped it on the floor. logger.info( "Anthropic request shape (model=%s, has_thinking=%s, thinking=%s, " "output_config=%s, temperature=%s, has_top_p=%s, has_top_k=%s, " "max_tokens=%s)", model, "thinking" in body, body.get("thinking"), body.get("output_config"), body.get("temperature"), "top_p" in body, "top_k" in body, body.get("max_tokens"), ) _finish_reason_map = { "end_turn": "stop", "max_tokens": "length", "stop_sequence": "stop", } logger.info("Proxying Anthropic Messages API to %s (model=%s)", url, model) request_headers = self._auth_headers() if code_execution_enabled: # Anthropic accepts comma-separated beta features in a single # `anthropic-beta` header. Merge our flag onto whatever the # registry's extra_headers contributed (currently nothing on # the beta axis, just anthropic-version) so future betas # added at the registry level keep working. existing_beta = request_headers.get("anthropic-beta", "").strip() beta_parts = ( [p.strip() for p in existing_beta.split(",") if p.strip()] if existing_beta else [] ) if "code-execution-2025-08-25" not in beta_parts: beta_parts.append("code-execution-2025-08-25") request_headers["anthropic-beta"] = ",".join(beta_parts) try: async with _http_client.stream( "POST", url, json = body, headers = request_headers, timeout = self._stream_timeout, ) as response: if response.status_code != 200: error_body = await response.aread() error_text = error_body.decode("utf-8", errors = "replace") logger.error( "Anthropic returned %d: %s", response.status_code, error_text[:500], ) yield _error_sse_line( response.status_code, error_text, self.provider_type ) return # NOTE: same manual __anext__ loop as stream_chat_completion — see comment there. lines_gen = response.aiter_lines().__aiter__() thinking_open = False # Diagnostic counters for the next time the user reports # "no thinking content" — distinguishes "Anthropic never sent # thinking_delta" from "frontend didn't render the chunks". event_counts: dict[str, int] = {} # web_search state. Anthropic emits the query inside an # `input_json_delta` stream on a `server_tool_use` content # block, then a separate `web_search_tool_result` block # with the URL list. Unlike OpenAI we get per-call results # directly, so each tool card carries its own citations. # `current_server_tool_use`: {id, name, partial_json_buffer} # `current_result_block`: {tool_use_id, results} # Both go to None when the matching content_block_stop fires. current_server_tool_use: Optional[dict[str, Any]] = None current_result_block: Optional[dict[str, Any]] = None web_search_calls: dict[str, dict[str, Any]] = {} # code_execution state. Anthropic's # `code_execution_20250825` tool emits the same # server_tool_use → *_tool_result block shape as # web_search, but the server_tool_use carries one of # two sub-tool names (`bash_code_execution` or # `text_editor_code_execution`) and the result block # type matches (`bash_code_execution_tool_result` / # `text_editor_code_execution_tool_result`). Kept # parallel to web_search state so the two paths don't # collide when both pills are on in the same turn. current_code_exec_use: Optional[dict[str, Any]] = None current_code_exec_result: Optional[dict[str, Any]] = None code_execution_calls: dict[str, dict[str, Any]] = {} # Counts surfaced in the final log line so reports of # "Code execution did nothing" can be triaged at a # glance. generated_files_count is interesting for the # future Files API PR — when bash creates files inside # the container, they show up as file_id entries on # bash_code_execution_result.content, and v1 drops # them. Track the count so we know how often it would # have mattered. code_execution_generated_files = 0 # Cache usage tracking. message_start carries the input # accounting (incl. cache_creation_input_tokens and # cache_read_input_tokens); message_delta carries cumulative # output_tokens. Both are surfaced in the "stream complete" # log so prompt caching can be verified per-request without # opening the Anthropic dashboard. last_usage: dict[str, Any] = {} def _content_chunk(text: str) -> str: chunk = { "id": completion_id, "object": "chat.completion.chunk", "choices": [ { "index": 0, "delta": {"content": text}, "finish_reason": None, } ], } return f"data: {_json.dumps(chunk)}" def _emit_tool_event(payload: dict[str, Any]) -> str: chunk = { "id": completion_id, "object": "chat.completion.chunk", "choices": [ { "index": 0, "delta": {}, "finish_reason": None, } ], "_toolEvent": payload, } return f"data: {_json.dumps(chunk)}" def _format_web_search_results( results: list[Any], ) -> str: blocks: list[str] = [] for r in results: if not isinstance(r, dict): continue if r.get("type") != "web_search_result": continue url = r.get("url", "") title = r.get("title") or url if not url: continue blocks.append(f"Title: {title}\nURL: {url}") return "\n---\n".join(blocks) def _format_code_execution_result( inner: dict[str, Any], ) -> str: """Render an Anthropic code-execution result block as the preformatted text payload the frontend's CodeExecutionToolUI displays inside a
. Handles
                    bash, text_editor (view/create/str_replace), and the
                    matching error variants.
                    """
                    inner_type = inner.get("type") or ""
                    if inner_type.endswith("_error"):
                        return f"Error: {inner.get('error_code', 'unknown')}"
                    if inner_type == "bash_code_execution_result":
                        stdout = inner.get("stdout") or ""
                        stderr = inner.get("stderr") or ""
                        return_code = inner.get("return_code")
                        parts: list[str] = []
                        if stdout:
                            parts.append(stdout)
                        if stderr:
                            parts.append(f"--- stderr ---\n{stderr}")
                        if isinstance(return_code, int) and return_code != 0:
                            parts.append(f"return_code: {return_code}")
                        return "\n".join(parts) if parts else "(no output)"
                    if inner_type == "text_editor_code_execution_result":
                        # view: file content; create: is_file_update flag;
                        # str_replace: diff `lines` list. The matching
                        # server_tool_use carries the command + path, but
                        # that's encoded into the tool_start arguments
                        # already — here we only format the result body.
                        if "lines" in inner and isinstance(inner.get("lines"), list):
                            return "\n".join(str(line) for line in inner["lines"])
                        if "is_file_update" in inner:
                            return (
                                "Updated" if inner.get("is_file_update") else "Created"
                            )
                        content_field = inner.get("content")
                        if isinstance(content_field, str):
                            return content_field
                        return "(file operation complete)"
                    return "(code execution complete)"

                try:
                    while True:
                        try:
                            line = await lines_gen.__anext__()
                        except StopAsyncIteration:
                            break
                        if not line or line.startswith("event:"):
                            continue
                        if not line.startswith("data:"):
                            continue

                        data_str = line[len("data:") :].strip()
                        if not data_str:
                            continue

                        try:
                            event = _json.loads(data_str)
                        except _json.JSONDecodeError:
                            continue

                        event_type = event.get("type")
                        if event_type == "content_block_delta":
                            delta_kind = (event.get("delta") or {}).get("type")
                            key = f"{event_type}:{delta_kind}"
                        else:
                            key = event_type or ""
                        event_counts[key] = event_counts.get(key, 0) + 1

                        # message_start carries the input-side usage block
                        # including cache_creation_input_tokens and
                        # cache_read_input_tokens. message_delta updates
                        # output_tokens (and may overwrite the input fields
                        # with final values). Merge both into last_usage.
                        if event_type == "message_start":
                            start_usage = (event.get("message") or {}).get("usage")
                            if isinstance(start_usage, dict):
                                last_usage.update(start_usage)

                        if event_type == "content_block_start":
                            content_block = event.get("content_block") or {}
                            block_type = content_block.get("type")
                            block_name = content_block.get("name")
                            if (
                                block_type == "server_tool_use"
                                and block_name == "web_search"
                            ):
                                tool_use_id = content_block.get("id", "") or (
                                    f"ws_{len(web_search_calls)}"
                                )
                                current_server_tool_use = {
                                    "id": tool_use_id,
                                    "buffer": "",
                                }
                                web_search_calls[tool_use_id] = {
                                    "query": "",
                                    "results": [],
                                }
                            elif block_type == "web_search_tool_result":
                                tool_use_id = content_block.get("tool_use_id", "")
                                # Anthropic sometimes ships the full results
                                # list on the start event; sometimes deltas
                                # follow. Capture whatever is present and
                                # finalize on content_block_stop.
                                content = content_block.get("content") or []
                                current_result_block = {
                                    "tool_use_id": tool_use_id,
                                    "results": list(content)
                                    if isinstance(content, list)
                                    else [],
                                }
                            elif block_type == "server_tool_use" and block_name in (
                                "bash_code_execution",
                                "text_editor_code_execution",
                            ):
                                tool_use_id = content_block.get("id", "") or (
                                    f"ce_{len(code_execution_calls)}"
                                )
                                kind = (
                                    "bash"
                                    if block_name == "bash_code_execution"
                                    else "text_editor"
                                )
                                current_code_exec_use = {
                                    "id": tool_use_id,
                                    "kind": kind,
                                    "buffer": "",
                                }
                                code_execution_calls[tool_use_id] = {
                                    "kind": kind,
                                    "arguments": {},
                                    "result": None,
                                }
                            elif block_type in (
                                "bash_code_execution_tool_result",
                                "text_editor_code_execution_tool_result",
                            ):
                                # Anthropic ships the full result content
                                # on the start event for code-exec result
                                # blocks (unlike web_search, which can
                                # split across deltas). Capture it and
                                # finalize on content_block_stop so the
                                # ordering matches the web_search path.
                                tool_use_id = content_block.get("tool_use_id", "")
                                inner = content_block.get("content") or {}
                                current_code_exec_result = {
                                    "tool_use_id": tool_use_id,
                                    "inner": inner if isinstance(inner, dict) else {},
                                }

                        elif event_type == "content_block_delta":
                            delta = event.get("delta", {})
                            delta_type = delta.get("type")
                            if delta_type == "thinking_delta":
                                # Anthropic streams extended-thinking content as
                                # thinking_delta events on a separate content
                                # block. Wrap as inline ... so
                                # the frontend's parseAssistantContent lifts it
                                # into the reasoning panel — same pattern as
                                # the OpenAI Responses path.
                                thinking_text = delta.get("thinking", "")
                                if thinking_text:
                                    if not thinking_open:
                                        thinking_text = f"{thinking_text}"
                                        thinking_open = True
                                    yield _content_chunk(thinking_text)
                            elif delta_type == "text_delta":
                                # First text after a thinking block closes the
                                #  tag we opened above. Anthropic emits
                                # a content_block_stop between blocks, but
                                # closing on the text_delta transition is more
                                # forgiving if events arrive out of order.
                                if thinking_open:
                                    yield _content_chunk("")
                                    thinking_open = False
                                text = delta.get("text", "")
                                if text:
                                    yield _content_chunk(text)
                                # Citations on text deltas are attached
                                # per-call by Anthropic via the
                                # `web_search_tool_result` block; we don't
                                # need to scrape them off the text events.
                            elif delta_type == "input_json_delta":
                                # Streamed partial_json carrying tool inputs
                                # — the search query for web_search, or the
                                # command/path/etc. for code execution.
                                # Route to whichever buffer is open. The two
                                # state slots are exclusive in practice
                                # (Anthropic doesn't interleave tool input
                                # streams), but checking both keeps the
                                # dispatch robust if that ever changes.
                                partial = delta.get("partial_json", "")
                                if current_server_tool_use is not None:
                                    current_server_tool_use["buffer"] += partial
                                elif current_code_exec_use is not None:
                                    current_code_exec_use["buffer"] += partial
                            # signature_delta and any other delta types are
                            # intentionally skipped — they carry trust /
                            # verification metadata, not user-visible content.

                        elif event_type == "content_block_stop":
                            if current_server_tool_use is not None:
                                # End of the server_tool_use block — parse the
                                # accumulated input_json into a query and
                                # emit tool_start. The matching tool_end fires
                                # later when the web_search_tool_result block
                                # closes with the actual results.
                                buffer = current_server_tool_use["buffer"]
                                query = ""
                                if buffer:
                                    try:
                                        parsed = _json.loads(buffer)
                                        if isinstance(parsed, dict):
                                            q = parsed.get("query", "")
                                            if isinstance(q, str):
                                                query = q
                                    except Exception:
                                        query = ""
                                tool_use_id = current_server_tool_use["id"]
                                if tool_use_id in web_search_calls:
                                    web_search_calls[tool_use_id]["query"] = query
                                yield _emit_tool_event(
                                    {
                                        "type": "tool_start",
                                        "tool_name": "web_search",
                                        "tool_call_id": tool_use_id,
                                        "arguments": (
                                            {"query": query} if query else {}
                                        ),
                                    }
                                )
                                current_server_tool_use = None
                            elif current_result_block is not None:
                                # End of a web_search_tool_result — emit
                                # tool_end carrying the search results as
                                # Title:/URL: blocks. parseSourcesFromResult
                                # on the frontend lifts these into source
                                # pills at message tail.
                                tool_use_id = current_result_block["tool_use_id"]
                                results = current_result_block["results"]
                                if tool_use_id in web_search_calls:
                                    web_search_calls[tool_use_id]["results"] = results
                                result_text = _format_web_search_results(results)
                                yield _emit_tool_event(
                                    {
                                        "type": "tool_end",
                                        "tool_call_id": tool_use_id,
                                        "result": (result_text or "(search complete)"),
                                    }
                                )
                                current_result_block = None
                            elif current_code_exec_use is not None:
                                # End of a code-execution server_tool_use —
                                # parse the buffered input_json into a
                                # {command, path, ...} dict and emit
                                # tool_start. The matching tool_end fires
                                # on the result block's content_block_stop.
                                buffer = current_code_exec_use["buffer"]
                                parsed_args: dict[str, Any] = {}
                                if buffer:
                                    try:
                                        parsed_obj = _json.loads(buffer)
                                        if isinstance(parsed_obj, dict):
                                            parsed_args = parsed_obj
                                    except Exception:
                                        parsed_args = {}
                                tool_use_id = current_code_exec_use["id"]
                                kind = current_code_exec_use["kind"]
                                emit_args = {"kind": kind, **parsed_args}
                                if tool_use_id in code_execution_calls:
                                    code_execution_calls[tool_use_id]["arguments"] = (
                                        emit_args
                                    )
                                yield _emit_tool_event(
                                    {
                                        "type": "tool_start",
                                        "tool_name": "code_execution",
                                        "tool_call_id": tool_use_id,
                                        "arguments": emit_args,
                                    }
                                )
                                current_code_exec_use = None
                            elif current_code_exec_result is not None:
                                # End of a code-execution result block —
                                # format the inner result into the text
                                # payload CodeExecutionToolUI renders.
                                tool_use_id = current_code_exec_result["tool_use_id"]
                                inner = current_code_exec_result["inner"]
                                # Track generated-file count for the
                                # follow-up Files API PR. v1 drops them.
                                if isinstance(inner, dict):
                                    file_blocks = inner.get("content")
                                    if isinstance(file_blocks, list):
                                        for entry in file_blocks:
                                            if isinstance(entry, dict) and entry.get(
                                                "file_id"
                                            ):
                                                code_execution_generated_files += 1
                                result_text = _format_code_execution_result(
                                    inner if isinstance(inner, dict) else {}
                                )
                                if tool_use_id in code_execution_calls:
                                    code_execution_calls[tool_use_id]["result"] = (
                                        result_text
                                    )
                                yield _emit_tool_event(
                                    {
                                        "type": "tool_end",
                                        "tool_call_id": tool_use_id,
                                        "result": result_text,
                                    }
                                )
                                current_code_exec_result = None
                            elif thinking_open:
                                # Close the  tag when the thinking block
                                # ends, in case no text_delta follows (e.g.
                                # display=omitted on Claude 4.7, or thinking-
                                # only turns).
                                yield _content_chunk("")
                                thinking_open = False

                        elif event_type == "message_delta":
                            delta_usage = event.get("usage")
                            if isinstance(delta_usage, dict):
                                last_usage.update(delta_usage)
                            stop_reason = event.get("delta", {}).get("stop_reason")
                            if stop_reason:
                                if thinking_open:
                                    yield _content_chunk("")
                                    thinking_open = False
                                chunk = {
                                    "id": completion_id,
                                    "object": "chat.completion.chunk",
                                    "choices": [
                                        {
                                            "index": 0,
                                            "delta": {},
                                            "finish_reason": _finish_reason_map.get(
                                                stop_reason, "stop"
                                            ),
                                        }
                                    ],
                                }
                                yield f"data: {_json.dumps(chunk)}"

                        elif event_type == "message_stop":
                            if thinking_open:
                                yield _content_chunk("")
                                thinking_open = False
                            yield "data: [DONE]"
                            await (
                                response.aclose()
                            )  # set PoolByteStream._closed=True FIRST
                            break
                except GeneratorExit:
                    await response.aclose()  # set PoolByteStream._closed=True FIRST
                    await lines_gen.aclose()  # now safe — aclose() is a no-op
                    raise
                finally:
                    # Surface per-event-type counts + web_search summary so
                    # reports of "no reasoning panel content" / "Search
                    # didn't do anything" can be triaged at a glance.
                    web_search_requested = bool(
                        enabled_tools and "web_search" in enabled_tools
                    )
                    web_search_invocations = len(web_search_calls)
                    total_results = sum(
                        len(sc.get("results") or []) for sc in web_search_calls.values()
                    )
                    queries = [
                        sc["query"]
                        for sc in web_search_calls.values()
                        if sc.get("query")
                    ]
                    # cache_read_input_tokens > 0 on turn N proves the
                    # cache_control marker on the system block is doing
                    # its job — turn 1 will show cache_creation > 0
                    # instead. cache_creation tokens are billed at a
                    # small premium; cache_read tokens are billed at a
                    # discount.
                    code_execution_invocations = len(code_execution_calls)
                    code_execution_results = sum(
                        1
                        for c in code_execution_calls.values()
                        if c.get("result") is not None
                    )
                    logger.info(
                        "Anthropic stream complete (model=%s, "
                        "web_search_requested=%s, web_search_invocations=%s, "
                        "results=%s, queries=%s, "
                        "code_execution_requested=%s, "
                        "code_execution_invocations=%s, "
                        "code_execution_results=%s, "
                        "code_execution_generated_files=%s, "
                        "input_tokens=%s, output_tokens=%s, "
                        "cache_creation_input_tokens=%s, "
                        "cache_read_input_tokens=%s, events=%s)",
                        model,
                        web_search_requested,
                        web_search_invocations,
                        total_results,
                        queries,
                        code_execution_enabled,
                        code_execution_invocations,
                        code_execution_results,
                        code_execution_generated_files,
                        last_usage.get("input_tokens"),
                        last_usage.get("output_tokens"),
                        last_usage.get("cache_creation_input_tokens"),
                        last_usage.get("cache_read_input_tokens"),
                        event_counts,
                    )
                    await response.aclose()
                    await lines_gen.aclose()

        except httpx.ConnectError as exc:
            logger.error("Connection error to %s: %s", self.provider_type, exc)
            yield _error_sse_line(
                502,
                f"Failed to connect to {self.provider_type}: {exc}",
                self.provider_type,
            )
        except httpx.ReadTimeout as exc:
            logger.error("Read timeout from %s: %s", self.provider_type, exc)
            yield _error_sse_line(
                504,
                f"Timeout waiting for {self.provider_type} response",
                self.provider_type,
            )
        except httpx.HTTPError as exc:
            logger.error("HTTP error from %s: %s", self.provider_type, exc)
            yield _error_sse_line(
                502,
                f"Error communicating with {self.provider_type}: {exc}",
                self.provider_type,
            )

    async def _stream_openai_responses(
        self,
        messages: list[dict[str, Any]],
        model: str,
        temperature: float,
        top_p: float,
        max_tokens: Optional[int],
        enable_thinking: Optional[bool],
        reasoning_effort: Optional[str],
        enabled_tools: Optional[list[str]] = None,
        enable_prompt_caching: Optional[bool] = None,
        openai_code_exec_container_id: Optional[str] = None,
    ) -> AsyncGenerator[str, None]:
        """
        Call OpenAI's /v1/responses endpoint and translate its SSE stream back
        into OpenAI Chat Completions chunk format.

        The Responses API uses a different request shape (``input`` instead of
        ``messages``, ``instructions`` for system prompts, ``max_output_tokens``
        for the budget) and emits event-typed SSE frames (e.g.
        ``response.output_text.delta``) rather than chat-completion chunks.
        ``presence_penalty`` / ``top_k`` are not part of the Responses contract
        and are dropped here intentionally.
        """
        import json as _json

        # Split system messages out into a single `instructions` string and
        # translate user/assistant messages into the Responses input shape.
        instructions_parts: list[str] = []
        input_items: list[dict[str, Any]] = []
        for msg in messages:
            role = msg.get("role")
            content = msg.get("content", "")

            if role == "system":
                if isinstance(content, str):
                    if content:
                        instructions_parts.append(content)
                elif isinstance(content, list):
                    for part in content:
                        if part.get("type") == "text" and part.get("text"):
                            instructions_parts.append(part["text"])
                continue

            if isinstance(content, str):
                input_items.append({"role": role, "content": content})
                continue

            if isinstance(content, list):
                translated_parts: list[dict[str, Any]] = []
                for part in content:
                    part_type = part.get("type")
                    if part_type == "text":
                        translated_parts.append(
                            {"type": "input_text", "text": part.get("text", "")}
                        )
                    elif part_type == "image_url":
                        url = part.get("image_url", {}).get("url", "")
                        if url:
                            # Responses takes image_url as a flat string (both
                            # https:// URLs and data: URLs are accepted).
                            translated_parts.append(
                                {"type": "input_image", "image_url": url}
                            )
                if translated_parts:
                    input_items.append({"role": role, "content": translated_parts})

        # NOTE: gpt-5.x / o3 / gpt-4.5 are reasoning-class models. They reject
        # temperature and top_p with `Unsupported parameter` 400s on
        # /v1/responses (and on /v1/chat/completions for the same families).
        # The PROVIDER_REGISTRY['openai'] model_id_allowlist already scopes
        # the picker to those families, so we never need to send sampling
        # knobs here. ``reasoning.effort`` defaults to "medium" server-side
        # if omitted — surface it in a future commit if a knob is wanted.
        del temperature, top_p  # explicit drop — params are accepted for
        # API symmetry with the other stream methods but not forwarded.

        body: dict[str, Any] = {
            "model": model,
            "input": input_items,
            "stream": True,
        }
        # `summary: "auto"` is what makes /v1/responses emit reasoning
        # summary events — without it OpenAI returns no thinking text on
        # most reasoning models, the SSE handler has no 
        # to wrap, and the chat reasoning panel stays blank. Always pair
        # an explicit effort with summary except for the explicit "off"
        # case (effort: "none"), where summaries are pointless.
        summary_unsupported = bool(
            _OPENAI_REASONING_SUMMARY_UNSUPPORTED.match(model.strip().lower())
        )
        if reasoning_effort in (
            "minimal",
            "low",
            "medium",
            "high",
            "max",
            "xhigh",
        ):
            body["reasoning"] = {"effort": reasoning_effort}
            if not summary_unsupported:
                body["reasoning"]["summary"] = "auto"
        elif reasoning_effort == "none" or enable_thinking is False:
            body["reasoning"] = {"effort": "none"}
        elif enable_thinking is True:
            body["reasoning"] = {"effort": "medium"}
            if not summary_unsupported:
                body["reasoning"]["summary"] = "auto"
        if instructions_parts:
            body["instructions"] = "\n\n".join(instructions_parts)
        if max_tokens is not None:
            body["max_output_tokens"] = max_tokens

        # Prompt caching on /v1/responses is automatic and free, but the
        # default in-memory policy only survives ~5-10 min of inactivity
        # (up to ~1 hr). Opt into the 24-hour retention policy so a chat
        # left idle overnight still hits the cache on the next turn.
        # Pricing is identical to in_memory per OpenAI's docs.
        #
        # Gated on the base URL because ollama / llama.cpp / "custom"
        # presets all collapse to provider_type="openai" in
        # toExternalBackendProviderType, so they also land in this
        # helper. Those servers expose /v1/responses-shaped routes in
        # some configurations but don't implement
        # prompt_cache_retention — sending the field unconditionally
        # would 400 them. Match the public OpenAI host strictly so the
        # field only goes to OpenAI cloud. Studio's openai model picker
        # is registry-scoped to gpt-5.x / o3 / gpt-4.5, all of which
        # accept this parameter (gpt-5.5+ already defaults to "24h" and
        # rejects "in_memory", so it's a safe no-op there).
        is_openai_cloud = "api.openai.com" in (self.base_url or "")
        if is_openai_cloud and enable_prompt_caching is not False:
            body["prompt_cache_retention"] = "24h"

        # OpenAI server-side tools — see
        #   https://developers.openai.com/api/docs/guides/tools
        #   https://developers.openai.com/api/docs/guides/tools-shell
        # The frontend's Search/Code buttons map to the unified
        # enabled_tools shorthand; translate that into the Responses-API
        # tool schema. Other built-in tools (file_search,
        # code_interpreter, image_generation, computer_use_preview) can
        # be added with the same pattern when we surface their toggles.
        code_execution_enabled_openai = bool(
            enabled_tools and "code_execution" in enabled_tools and is_openai_cloud
        )
        if enabled_tools:
            tools_array: list[dict[str, Any]] = []
            if "web_search" in enabled_tools:
                tools_array.append({"type": "web_search"})
            if code_execution_enabled_openai:
                # `container_auto` lets OpenAI auto-create a fresh
                # container per request; we capture the resulting
                # container_id off the SSE stream and the chat-adapter
                # persists it onto the thread record. Subsequent turns
                # in the same thread pass it back as
                # `openai_code_exec_container_id`, which we translate to
                # `container_reference` here so the model sees
                # filesystem state from prior turns. Container expires
                # after ~20 min of inactivity per OpenAI's default
                # policy — a stale id 400s, the chat-adapter clears it
                # via container_invalidated, and the next turn falls
                # back to auto-create.
                shell_env: dict[str, Any]
                if openai_code_exec_container_id:
                    shell_env = {
                        "type": "container_reference",
                        "container_id": openai_code_exec_container_id,
                    }
                else:
                    shell_env = {"type": "container_auto"}
                tools_array.append({"type": "shell", "environment": shell_env})
            if tools_array:
                body["tools"] = tools_array

        url = f"{self.base_url}/responses"
        completion_id = f"chatcmpl-openai-{model.replace('/', '-')}"

        logger.info("Proxying OpenAI Responses API to %s (model=%s)", url, model)

        try:
            async with _http_client.stream(
                "POST",
                url,
                json = body,
                headers = self._auth_headers(),
                timeout = self._stream_timeout,
            ) as response:
                if response.status_code != 200:
                    error_body = await response.aread()
                    error_text = error_body.decode("utf-8", errors = "replace")
                    logger.error(
                        "OpenAI Responses returned %d: %s",
                        response.status_code,
                        error_text[:500],
                    )
                    # Detect stale-container errors so the frontend can
                    # drop its persisted id. OpenAI doesn't pin an
                    # error code in the public docs for this case, so
                    # match a couple of likely substrings. If we sent
                    # a container_reference and the response is 4xx
                    # with any hint of "container not found / expired",
                    # emit container_invalidated; the next turn will
                    # fall back to container_auto.
                    if (
                        openai_code_exec_container_id
                        and 400 <= response.status_code < 500
                    ):
                        lowered = error_text.lower()
                        if "container" in lowered and (
                            "expired" in lowered
                            or "not_found" in lowered
                            or "not found" in lowered
                            or "no such container" in lowered
                        ):
                            yield (
                                f"data: "
                                f"{_json.dumps({'id': completion_id, 'object': 'chat.completion.chunk', 'choices': [{'index': 0, 'delta': {}, 'finish_reason': None}], '_toolEvent': {'type': 'container_invalidated'}})}"
                            )
                    yield _error_sse_line(
                        response.status_code, error_text, self.provider_type
                    )
                    return

                # NOTE: same manual __anext__ loop as stream_chat_completion —
                # see comment there for the GeneratorExit / aclose ordering.
                lines_gen = response.aiter_lines().__aiter__()
                done_emitted = False
                reasoning_open = False
                reasoning_emitted = False
                # Latched from response.completed / response.incomplete so
                # the final log can surface input_tokens_details.cached_tokens —
                # the field that proves prompt_cache_retention="24h" is
                # actually hitting OpenAI's cache instead of recomputing
                # the prefix every turn.
                last_usage: Optional[dict[str, Any]] = None
                # Per-call state for OpenAI's server-side web_search tool. Mapped
                # back into our local _toolEvent shape so the existing chat-UI
                # renderer surfaces web_search the same way it does for local
                # tool calls: a "Searching…" tool-call card, then a `tool_end`
                # carrying citations formatted as
                #   Title: …\nURL: …\nSnippet: …\n---\n…
                # blocks (which the frontend's parseSourcesFromResult lifts
                # into source content parts at end of stream).
                # web_search_calls preserves insertion order so we can apply
                # the aggregated citation list onto the *last* call's
                # tool_end — that's the one the frontend's source-pill
                # extraction reads (parseSourcesFromResult flatMaps every
                # web_search result, so a single non-empty result is enough
                # to surface all sources at message tail).
                # OpenAI emits url_citation annotations on text deltas, not
                # per call — there's no wire field linking a citation back
                # to a specific search invocation. Hence the shared list.
                # web_search_calls: { item_id -> {query} }
                web_search_calls: dict[str, dict[str, Any]] = {}
                all_url_citations: list[dict[str, str]] = []
                # Shell-tool (code execution) state. OpenAI emits
                # `shell_call` items (model requesting a command list)
                # paired with `shell_call_output` items (execution
                # results). We mirror the Anthropic code-execution UX
                # by emitting one `_toolEvent` tool_start per
                # shell_call and one tool_end per shell_call_output;
                # they're linked via `shell_call_output.call_id`
                # matching `shell_call.id`. Items are independent of
                # web_search (different keyed map).
                # shell_calls: { call_id -> {commands, output} }
                shell_calls: dict[str, dict[str, Any]] = {}
                # Container id captured from the response stream. When
                # it differs from the inbound id, emit a synthetic
                # `container_ready` _toolEvent so the frontend can
                # persist it onto the thread record for the next turn.
                # Where OpenAI surfaces it is documented loosely; we
                # probe two known fields (response.container_id on
                # response.completed, item.environment.container_id on
                # shell_call output items) and latch the first one we
                # see.
                latched_container_id: Optional[str] = None
                container_id_emitted = False

                def _emit_tool_event(payload: dict[str, Any]) -> str:
                    chunk = {
                        "id": completion_id,
                        "object": "chat.completion.chunk",
                        "choices": [
                            {
                                "index": 0,
                                "delta": {},
                                "finish_reason": None,
                            }
                        ],
                        "_toolEvent": payload,
                    }
                    return f"data: {_json.dumps(chunk)}"

                def _format_shell_output(output: Any) -> str:
                    """Render an OpenAI `shell_call_output.output` list
                    as the preformatted text payload the frontend's
                    CodeExecutionToolUI displays inside a 
. Each
                    entry has stdout/stderr/outcome — concatenate them
                    with a separator block per entry and append
                    `return_code` / `(timeout)` annotations only when
                    they convey information beyond "succeeded".
                    """
                    if not isinstance(output, list):
                        return ""
                    parts: list[str] = []
                    for entry in output:
                        if not isinstance(entry, dict):
                            continue
                        stdout = entry.get("stdout") or ""
                        stderr = entry.get("stderr") or ""
                        outcome = entry.get("outcome") or {}
                        chunk_parts: list[str] = []
                        if stdout:
                            chunk_parts.append(stdout)
                        if stderr:
                            chunk_parts.append(f"--- stderr ---\n{stderr}")
                        if isinstance(outcome, dict):
                            outcome_type = outcome.get("type")
                            if outcome_type == "exit":
                                exit_code = outcome.get("exit_code")
                                if isinstance(exit_code, int) and exit_code != 0:
                                    chunk_parts.append(f"return_code: {exit_code}")
                            elif outcome_type == "timeout":
                                chunk_parts.append("(timeout)")
                        if chunk_parts:
                            parts.append("\n".join(chunk_parts))
                    return (
                        "\n--- next command ---\n".join(parts)
                        if parts
                        else "(no output)"
                    )

                def _record_url_citation(payload: dict[str, Any]) -> None:
                    """Append a url_citation onto the shared all_url_citations
                    list. Dedup by URL — the same source can be cited multiple
                    times across deltas. We do NOT try to attribute citations
                    to individual web_search_call invocations because OpenAI's
                    annotation events don't carry that linkage."""
                    if payload.get("type") != "url_citation":
                        return
                    url = payload.get("url", "")
                    if not url:
                        return
                    if any(c["url"] == url for c in all_url_citations):
                        return
                    title = payload.get("title") or url
                    snippet = payload.get("snippet") or payload.get("quote") or ""
                    all_url_citations.append(
                        {
                            "url": url,
                            "title": title,
                            "snippet": snippet,
                        }
                    )

                def _extract_reasoning_text(payload: Any) -> str:
                    if payload is None:
                        return ""
                    if isinstance(payload, str):
                        return payload
                    if isinstance(payload, list):
                        out: list[str] = []
                        for item in payload:
                            text = _extract_reasoning_text(item)
                            if text:
                                out.append(text)
                        return "".join(out)
                    if isinstance(payload, dict):
                        # OpenAI responses may carry reasoning summaries in
                        # different envelope fields across event variants.
                        for key in ("text", "delta", "content", "summary"):
                            if key in payload:
                                text = _extract_reasoning_text(payload.get(key))
                                if text:
                                    return text
                        if payload.get("type") == "summary_text":
                            return _extract_reasoning_text(payload.get("text"))
                    return ""

                def _chunk_with_text(text: str) -> str:
                    chunk = {
                        "id": completion_id,
                        "object": "chat.completion.chunk",
                        "choices": [
                            {
                                "index": 0,
                                "delta": {"content": text},
                                "finish_reason": None,
                            }
                        ],
                    }
                    return f"data: {_json.dumps(chunk)}"

                try:
                    while True:
                        try:
                            line = await lines_gen.__anext__()
                        except StopAsyncIteration:
                            break
                        if not line or line.startswith("event:"):
                            continue
                        if not line.startswith("data:"):
                            continue

                        data_str = line[len("data:") :].strip()
                        if not data_str:
                            continue
                        if data_str == "[DONE]":
                            if not done_emitted:
                                yield "data: [DONE]"
                                done_emitted = True
                            break

                        try:
                            event = _json.loads(data_str)
                        except _json.JSONDecodeError:
                            continue

                        event_type = event.get("type")

                        if event_type == "response.output_text.delta":
                            delta_text = event.get("delta", "")
                            if delta_text:
                                if reasoning_open:
                                    yield _chunk_with_text("")
                                    reasoning_open = False
                                yield _chunk_with_text(delta_text)
                            # Some API versions inline url citations on the
                            # delta event itself rather than as a separate
                            # response.output_text.annotation.added event.
                            for ann in event.get("annotations") or []:
                                if isinstance(ann, dict):
                                    _record_url_citation(ann)

                        elif event_type == "response.output_text.annotation.added":
                            ann = event.get("annotation")
                            if isinstance(ann, dict):
                                _record_url_citation(ann)

                        elif event_type == "response.output_item.added":
                            # Track the call early but do NOT emit tool_start
                            # yet — action.query is not reliably populated on
                            # added across OpenAI API versions, and the
                            # frontend's tool_start is a one-shot push (no
                            # update mechanism). Wait for output_item.done.
                            item = event.get("item", {})
                            if (
                                isinstance(item, dict)
                                and item.get("type") == "web_search_call"
                            ):
                                item_id = item.get("id", "") or (
                                    f"ws_{len(web_search_calls)}"
                                )
                                web_search_calls.setdefault(item_id, {"query": ""})
                            # Shell-tool: register the call eagerly so
                            # the matching shell_call_output can link
                            # back even if `done` arrives out of order.
                            # Also probe for container_id on the
                            # environment field — when container_auto
                            # auto-creates one, this is the first place
                            # the new id might surface (OpenAI doesn't
                            # promise this in docs, but the field is
                            # cheap to scan and lets us emit
                            # container_ready earlier than
                            # response.completed).
                            if (
                                isinstance(item, dict)
                                and item.get("type") == "shell_call"
                            ):
                                item_id = item.get("id", "") or (
                                    f"sc_{len(shell_calls)}"
                                )
                                shell_calls.setdefault(
                                    item_id,
                                    {"commands": [], "output": None},
                                )
                                env = item.get("environment")
                                if isinstance(env, dict):
                                    probe = env.get("container_id") or env.get("id")
                                    if (
                                        isinstance(probe, str)
                                        and probe.startswith("cntr_")
                                        and latched_container_id is None
                                    ):
                                        latched_container_id = probe

                        elif event_type == "response.output_item.done":
                            item = event.get("item", {})
                            if not isinstance(item, dict):
                                continue
                            if item.get("type") == "reasoning":
                                summary_text = _extract_reasoning_text(
                                    item.get("summary")
                                )
                                if summary_text and not reasoning_emitted:
                                    if not reasoning_open:
                                        summary_text = f"{summary_text}"
                                        reasoning_open = True
                                    yield _chunk_with_text(summary_text)
                                    reasoning_emitted = True
                            elif item.get("type") == "web_search_call":
                                # done is the canonical place to read the
                                # query, so emit both tool_start and tool_end
                                # here. Frontend then renders a card per call
                                # with the proper "Searching: " label.
                                # Citations are aggregated separately and the
                                # *last* call's result is overwritten at
                                # response.completed with the citation list
                                # (so the source-pill extraction at message
                                # tail surfaces them once).
                                item_id = item.get("id", "") or (
                                    f"ws_{len(web_search_calls)}"
                                )
                                action = item.get("action")
                                query = (
                                    action.get("query", "")
                                    if isinstance(action, dict)
                                    else ""
                                )
                                web_search_calls[item_id] = {"query": query}
                                yield _emit_tool_event(
                                    {
                                        "type": "tool_start",
                                        "tool_name": "web_search",
                                        "tool_call_id": item_id,
                                        "arguments": (
                                            {"query": query} if query else {}
                                        ),
                                    }
                                )
                                yield _emit_tool_event(
                                    {
                                        "type": "tool_end",
                                        "tool_call_id": item_id,
                                        # Empty result — the last call gets
                                        # overwritten with citations at
                                        # response.completed.
                                        "result": "",
                                    }
                                )
                            elif item.get("type") == "shell_call":
                                # OpenAI ships the commands array on the
                                # action field. Join them onto one
                                # command string for the tool card —
                                # the renderer is shared with Anthropic
                                # bash, which only carries a single
                                # `command`. Multiple commands in one
                                # shell_call get joined with newlines so
                                # they still render as one card.
                                item_id = item.get("id", "") or (
                                    f"sc_{len(shell_calls)}"
                                )
                                action = item.get("action") or {}
                                commands = (
                                    action.get("commands")
                                    if isinstance(action, dict)
                                    else None
                                ) or []
                                joined_command = (
                                    "\n".join(str(c) for c in commands)
                                    if isinstance(commands, list)
                                    else ""
                                )
                                shell_calls.setdefault(
                                    item_id,
                                    {"commands": [], "output": None},
                                )
                                shell_calls[item_id]["commands"] = (
                                    list(commands) if isinstance(commands, list) else []
                                )
                                yield _emit_tool_event(
                                    {
                                        "type": "tool_start",
                                        "tool_name": "code_execution",
                                        "tool_call_id": item_id,
                                        "arguments": {
                                            "kind": "bash",
                                            "command": joined_command,
                                        },
                                    }
                                )
                            elif item.get("type") == "shell_call_output":
                                # `call_id` links back to the shell_call's
                                # `id`, which is what we used as the
                                # tool_call_id on tool_start. Match on
                                # call_id when present so the matching
                                # card transitions to complete.
                                call_id = item.get("call_id") or item.get("id") or ""
                                output = item.get("output") or []
                                if call_id in shell_calls:
                                    shell_calls[call_id]["output"] = output
                                result_text = _format_shell_output(output)
                                yield _emit_tool_event(
                                    {
                                        "type": "tool_end",
                                        "tool_call_id": call_id,
                                        "result": result_text,
                                    }
                                )

                        elif isinstance(event_type, str) and "reasoning" in event_type:
                            reasoning_delta = _extract_reasoning_text(event)
                            if reasoning_delta:
                                if not reasoning_open:
                                    reasoning_delta = f"{reasoning_delta}"
                                    reasoning_open = True
                                yield _chunk_with_text(reasoning_delta)
                                reasoning_emitted = True

                        elif event_type == "response.completed":
                            completed_usage = (event.get("response") or {}).get("usage")
                            if isinstance(completed_usage, dict):
                                last_usage = completed_usage
                            if reasoning_open:
                                yield _chunk_with_text("")
                                reasoning_open = False
                            # Probe response.container_id (top-level) and
                            # response.container.id for the shell-tool
                            # container id. OpenAI's docs don't pin the
                            # exact field, so we scan both. Emit
                            # `container_ready` only when the value
                            # differs from the inbound one — no churn on
                            # reuse.
                            response_obj = event.get("response") or {}
                            if isinstance(response_obj, dict):
                                probe_id = response_obj.get("container_id")
                                if not probe_id:
                                    container_field = response_obj.get("container")
                                    if isinstance(container_field, dict):
                                        probe_id = container_field.get("id")
                                if (
                                    isinstance(probe_id, str)
                                    and probe_id.startswith("cntr_")
                                    and latched_container_id is None
                                ):
                                    latched_container_id = probe_id
                            if (
                                latched_container_id
                                and not container_id_emitted
                                and latched_container_id
                                != openai_code_exec_container_id
                            ):
                                yield _emit_tool_event(
                                    {
                                        "type": "container_ready",
                                        "container_id": latched_container_id,
                                    }
                                )
                                container_id_emitted = True
                            # Apply the aggregated citation list onto the
                            # *last* web_search call by overwriting its
                            # tool_end result. The frontend's
                            # parseSourcesFromResult flatMaps every
                            # web_search tool-call result, so a single
                            # non-empty result is enough to surface the
                            # whole source-pill set at the message tail —
                            # no need to fan out across every card (which
                            # would just duplicate the same pills).
                            if web_search_calls and all_url_citations:
                                last_id = list(web_search_calls.keys())[-1]
                                blocks: list[str] = []
                                for cit in all_url_citations:
                                    line = (
                                        f"Title: {cit['title']}\n" f"URL: {cit['url']}"
                                    )
                                    if cit.get("snippet"):
                                        line += f"\nSnippet: {cit['snippet']}"
                                    blocks.append(line)
                                yield _emit_tool_event(
                                    {
                                        "type": "tool_end",
                                        "tool_call_id": last_id,
                                        "result": "\n---\n".join(blocks),
                                    }
                                )
                            chunk = {
                                "id": completion_id,
                                "object": "chat.completion.chunk",
                                "choices": [
                                    {
                                        "index": 0,
                                        "delta": {},
                                        "finish_reason": "stop",
                                    }
                                ],
                            }
                            yield f"data: {_json.dumps(chunk)}"

                        elif event_type == "response.incomplete":
                            incomplete_usage = (event.get("response") or {}).get(
                                "usage"
                            )
                            if isinstance(incomplete_usage, dict):
                                last_usage = incomplete_usage
                            if reasoning_open:
                                yield _chunk_with_text("")
                                reasoning_open = False
                            # Same backfill as response.completed — apply
                            # whatever citations we managed to gather
                            # before truncation onto the last call. All
                            # earlier tool cards already have their proper
                            # query + empty placeholder result from the
                            # output_item.done emissions above.
                            if web_search_calls and all_url_citations:
                                last_id = list(web_search_calls.keys())[-1]
                                blocks = []
                                for cit in all_url_citations:
                                    line = (
                                        f"Title: {cit['title']}\n" f"URL: {cit['url']}"
                                    )
                                    if cit.get("snippet"):
                                        line += f"\nSnippet: {cit['snippet']}"
                                    blocks.append(line)
                                yield _emit_tool_event(
                                    {
                                        "type": "tool_end",
                                        "tool_call_id": last_id,
                                        "result": "\n---\n".join(blocks),
                                    }
                                )
                            chunk = {
                                "id": completion_id,
                                "object": "chat.completion.chunk",
                                "choices": [
                                    {
                                        "index": 0,
                                        "delta": {},
                                        "finish_reason": "length",
                                    }
                                ],
                            }
                            yield f"data: {_json.dumps(chunk)}"

                        elif event_type in ("response.failed", "error"):
                            # Surface the failure to the client; let the
                            # outer route emit [DONE] as part of its cleanup.
                            error_payload = event.get("response", {}).get(
                                "error", {}
                            ) or {
                                "message": event.get("message", "Unknown error"),
                                "code": event.get("code"),
                            }
                            yield _error_sse_line(
                                502,
                                _json.dumps(error_payload),
                                self.provider_type,
                            )
                            break
                except GeneratorExit:
                    await response.aclose()
                    await lines_gen.aclose()
                    raise
                finally:
                    # Summarise what the model actually did this turn so
                    # support reports of "I clicked Search and got nothing"
                    # can be triaged at a glance: was the tool requested,
                    # did OpenAI invoke it, and how many sources came back?
                    web_search_requested = bool(
                        enabled_tools and "web_search" in enabled_tools
                    )
                    web_search_invocations = len(web_search_calls)
                    total_citations = len(all_url_citations)
                    queries = [
                        sc["query"]
                        for sc in web_search_calls.values()
                        if sc.get("query")
                    ]
                    # cached_input_tokens > 0 on turn N proves
                    # prompt_cache_retention="24h" is letting the previous
                    # turn's prefix hit the cache instead of being
                    # recomputed. On /v1/responses the field is nested as
                    # usage.input_tokens_details.cached_tokens (not
                    # prompt_tokens_details, which is the /v1/chat/completions
                    # shape).
                    cached_input_tokens = None
                    if isinstance(last_usage, dict):
                        details = last_usage.get("input_tokens_details")
                        if isinstance(details, dict):
                            cached_input_tokens = details.get("cached_tokens")
                    code_execution_requested = code_execution_enabled_openai
                    code_execution_invocations = len(shell_calls)
                    code_execution_results = sum(
                        1 for sc in shell_calls.values() if sc.get("output") is not None
                    )
                    logger.info(
                        "OpenAI Responses stream complete (model=%s, "
                        "web_search_requested=%s, web_search_invocations=%s, "
                        "citations=%s, queries=%s, reasoning_emitted=%s, "
                        "code_execution_requested=%s, "
                        "code_execution_invocations=%s, "
                        "code_execution_results=%s, "
                        "container_id_in=%s, container_id_out=%s, "
                        "input_tokens=%s, output_tokens=%s, "
                        "cached_input_tokens=%s)",
                        model,
                        web_search_requested,
                        web_search_invocations,
                        total_citations,
                        queries,
                        reasoning_emitted,
                        code_execution_requested,
                        code_execution_invocations,
                        code_execution_results,
                        openai_code_exec_container_id,
                        latched_container_id,
                        (last_usage or {}).get("input_tokens"),
                        (last_usage or {}).get("output_tokens"),
                        cached_input_tokens,
                    )
                    await response.aclose()
                    await lines_gen.aclose()

        except httpx.ConnectError as exc:
            logger.error("Connection error to %s: %s", self.provider_type, exc)
            yield _error_sse_line(
                502,
                f"Failed to connect to {self.provider_type}: {exc}",
                self.provider_type,
            )
        except httpx.ReadTimeout as exc:
            logger.error("Read timeout from %s: %s", self.provider_type, exc)
            yield _error_sse_line(
                504,
                f"Timeout waiting for {self.provider_type} response",
                self.provider_type,
            )
        except httpx.HTTPError as exc:
            logger.error("HTTP error from %s: %s", self.provider_type, exc)
            yield _error_sse_line(
                502,
                f"Error communicating with {self.provider_type}: {exc}",
                self.provider_type,
            )

    async def chat_completion(
        self,
        messages: list[dict[str, Any]],
        model: str,
        temperature: float = 0.7,
        top_p: float = 0.95,
        max_tokens: Optional[int] = None,
        presence_penalty: float = 0.0,
    ) -> dict[str, Any]:
        """Non-streaming chat completion. Returns the full response dict.

        Note: only valid for OpenAI-compatible providers. Anthropic requires its
        own Messages API; use stream_chat_completion (with stream=False) instead
        if a non-streaming Anthropic path is needed in the future.
        """
        body: dict[str, Any] = {
            "model": model,
            "messages": messages,
            "stream": False,
            "temperature": temperature,
            "top_p": top_p,
            "presence_penalty": presence_penalty,
        }
        if max_tokens is not None:
            if self.provider_type == "openai":
                body["max_completion_tokens"] = max_tokens
            else:
                body["max_tokens"] = max_tokens

        response = await _http_client.post(
            f"{self.base_url}/chat/completions",
            json = body,
            headers = self._auth_headers(),
            timeout = self._timeout,
        )
        response.raise_for_status()
        return response.json()

    async def list_models(self) -> list[dict[str, Any]]:
        """
        Call GET /models on the provider to discover available models.

        Returns a list of model dicts with at least 'id' and optionally
        'created', 'owned_by', etc.

        All supported providers expose a /models endpoint:
        - OpenAI-compatible: standard {"data": [...]} response
        - Anthropic: https://api.anthropic.com/v1/models — same {"data": [...]} shape
        """
        try:
            response = await _http_client.get(
                f"{self.base_url}/models",
                headers = self._auth_headers(),
                timeout = self._timeout,
            )
            response.raise_for_status()
            data = response.json()
            # OpenAI format: {"data": [{"id": "...", ...}, ...]}
            models = data.get("data", [])
            return models
        except httpx.HTTPError as exc:
            logger.error("Failed to list models from %s: %s", self.provider_type, exc)
            raise

    async def verify_models_endpoint_lightweight(self) -> None:
        """
        Confirm GET /models returns 200 without buffering the full response body.

        Used for providers with enormous catalogs (e.g. OpenRouter, Hugging Face router)
        where downloading the full JSON would be prohibitive.
        """
        url = f"{self.base_url}/models"
        try:
            async with _http_client.stream(
                "GET",
                url,
                headers = self._auth_headers(),
                timeout = self._timeout,
            ) as response:
                if response.status_code != 200:
                    response.raise_for_status()
                async for _chunk in response.aiter_bytes(chunk_size = 2048):
                    break
        except httpx.HTTPError as exc:
            logger.error(
                "Lightweight /models check failed for %s: %s",
                self.provider_type,
                exc,
            )
            raise

    def _container_headers(self) -> dict[str, str]:
        """Auth headers plus the OpenAI-Beta opt-in for /v1/containers.

        OpenAI's containers API requires ``OpenAI-Beta: containers=v1``.
        Without it, DELETE silently no-ops: the API returns 200 with a
        ``{"deleted": true}`` body but does not actually remove the
        container (verified 2026-05-15). The header is required for
        list / create / delete to behave consistently.
        """
        headers = self._auth_headers()
        headers["OpenAI-Beta"] = "containers=v1"
        return headers

    async def list_openai_containers(self) -> list[dict[str, Any]]:
        """
        GET /v1/containers on the user's OpenAI account.

        Returns the raw container records (id, name, created_at,
        last_active_at, expires_after, status). The route layer
        reshapes these into the UI summary shape.

        Only valid against api.openai.com — non-cloud OpenAI-compat
        servers don't implement /v1/containers and would 404 here.
        Caller is responsible for the is_openai_cloud guard.
        """
        response = await _http_client.get(
            f"{self.base_url}/containers",
            headers = self._container_headers(),
            timeout = self._timeout,
        )
        response.raise_for_status()
        data = response.json()
        containers = data.get("data") if isinstance(data, dict) else None
        result = list(containers) if isinstance(containers, list) else []
        logger.info(
            "openai_container_list.response count=%s items=%s",
            len(result),
            [
                {"id": c.get("id"), "status": c.get("status")}
                for c in result
                if isinstance(c, dict)
            ],
        )
        return result

    async def create_openai_container(
        self,
        name: str,
        ttl_minutes: int,
    ) -> dict[str, Any]:
        """
        POST /v1/containers with ``expires_after.anchor="last_active_at"``.
        ``ttl_minutes`` is the idle timeout — every API call that
        touches the container resets the timer.
        """
        body = {
            "name": name,
            "expires_after": {
                "anchor": "last_active_at",
                "minutes": ttl_minutes,
            },
        }
        response = await _http_client.post(
            f"{self.base_url}/containers",
            json = body,
            headers = self._container_headers(),
            timeout = self._timeout,
        )
        response.raise_for_status()
        return response.json()

    async def delete_openai_container(self, container_id: str) -> None:
        """DELETE /v1/containers/{id}. 404s are surfaced as HTTPError.

        Uses a fresh httpx client (not the shared ``_http_client``) so
        connection-pool state from earlier chat requests cannot
        interfere — observed in the wild that DELETEs over the shared
        pool returned ``deleted: true`` while the container persisted
        in subsequent /containers list calls, even though the same
        DELETE issued from a fresh client genuinely removed it.

        Verifies the response body reports ``deleted: true``. OpenAI
        returns a 2xx ``deleted: true`` body even when the request is
        silently rejected (e.g. missing OpenAI-Beta header), so a
        status-only check is not sufficient.
        """
        url = f"{self.base_url}/containers/{container_id}"
        headers = self._container_headers()
        logger.info(
            "openai_container_delete.outbound url=%s has_auth=%s openai_beta=%s",
            url,
            "Authorization" in headers,
            headers.get("OpenAI-Beta"),
        )
        async with httpx.AsyncClient(timeout = self._timeout) as fresh_client:
            response = await fresh_client.delete(url, headers = headers)
        logger.info(
            "openai_container_delete.response status=%s cf_ray=%s "
            "request_id=%s organization=%s project=%s processing_ms=%s body=%s",
            response.status_code,
            response.headers.get("cf-ray"),
            response.headers.get("x-request-id"),
            response.headers.get("openai-organization"),
            response.headers.get("openai-project"),
            response.headers.get("openai-processing-ms"),
            response.text[:300],
        )
        response.raise_for_status()
        try:
            payload = response.json()
        except ValueError:
            payload = None
        if not (isinstance(payload, dict) and payload.get("deleted") is True):
            raise httpx.HTTPError(
                f"OpenAI did not confirm container deletion: {response.text[:200]}"
            )

    async def close(self) -> None:
        """No-op — the underlying client is shared across requests."""


def _error_sse_line(status_code: int, message: str, provider_type: str) -> str:
    """Format an error as an SSE data line in OpenAI error format."""
    import json

    error_obj = {
        "error": {
            "message": message,
            "type": "provider_error",
            "code": str(status_code),
            "provider": provider_type,
        }
    }
    return f"data: {json.dumps(error_obj)}"