cendor-sdk 1.0.0

First public release on PyPI — the governed agent loop, multi-agent orchestration, ecosystem
interop, and production hardening + governed eval, consolidated with the gap-analysis remediation,
retrieval (RAG), and rolling memory. Requires cendor-core>=1.3 (Hugging Face detection).

Fixed

• Tool calling on Gemini & Bedrock (P0) — build_kwargs dropped assistant tool_calls and all
  tool results, so multi-turn tool loops silently broke on those providers. Added
  _canonical_to_gemini / _canonical_to_bedrock (functionCall/functionResponse; toolUse/toolResult),
  mirroring the Anthropic translator, with round-trip tests.

Added — gap-analysis remediation (see plan/SDK_FIT_GAP_ANALYSIS.md)

• Provider param passthrough — Agent.extra merges arbitrary request kwargs (tool_choice,
  reasoning_effort, top_p, stop, seed, response_format, extra_body, …) into every call.
• o-series temperature guard — the OpenAI provider omits temperature for o1/o3/o4 models
  (which reject it) instead of erroring.
• Robust structured output — an output_type now derives a JSON Schema used via each provider's
  native structured-output feature: OpenAI json_schema, Ollama format, Gemini response_schema;
  Anthropic/Bedrock/Responses embed the schema in the JSON nudge. Dataclass / Pydantic / dict.
• Streaming — run.stream(agent, input) / run.astream(...) yield TextDelta / ToolCallEvent
  / ToolResultEvent and a terminal RunComplete(result). Native reassembly for the OpenAI family +
  Ollama (tool-call deltas included); whole-response fallback for the rest. Single-agent.
• Multimodal input — message content may be a parts list (text + image_url); OpenAI-family
  passes it through, Anthropic/Gemini translate to image blocks (base64/url), Bedrock keeps the text.
• Unpriced-model cost governance — register_model_price(model, input=…, output=…) registers a
  rate so cost/USD budgets bind on HF/Azure-deployment/Foundry-Local/custom ids; configure
  (on_unpriced="raise") re-exported.
• Embeddings — embed(model, inputs) / aembed(...) return vectors and emit a governed
  LLMCall (tokens/cost/audit captured) for OpenAI-family providers — RAG calls become first-class.
• RAG seam — VectorIndex (a dependency-free in-memory cosine index over embed()) +
  Agent(retriever=…), which injects retrieved context as a system message before the call
  ("always-on" RAG). Bring your own embedder via embedder=, or plug a real vector DB as a
  retriever. The SDK governs retrieval; it does not store your vectors.
• Summarizing memory — SummarizingSession(model=… | summarizer=…, max_messages=…, keep_recent=…) folds old turns into a durable summary note when a conversation grows, keeping
  recent turns verbatim so memory stays bounded but the gist persists (beyond context_budget
  trimming). llm_summarizer(model) builds a governed summarizer; pass any callable for offline
  summaries.
• Multi-agent checkpointing — run([...], checkpoint=…) persists the trajectory per turn/segment
  and resumes a crashed team run.
• MCP prompts — load_mcp_prompts(session) + get_mcp_prompt(session, name, args) (renders to
  canonical messages); load_mcp_resources now exported too.
• Concurrent tool execution — a turn's multiple tool calls run via asyncio.gather in async runs.
• max_turns signal — Result.incomplete is True when a run ends with no final answer.
• Nested tool params — tool schema generation expands nested dataclass / Pydantic parameters.
• Richer OTel spans — span_tree now records latency, finish reason, reasoning tokens, and tool
  argument names.
• Eval judge — EvalCase.judge adds a pluggable semantic / LLM-judge scorer.

Added

• Hugging Face provider (provider="huggingface" / "hf", extra [huggingface]) — wraps
  huggingface_hub.InferenceClient.chat_completion (OpenAI-shaped response, so request formatting
  and normalization are reused). Token via api_key= or HF_TOKEN/HUGGINGFACEHUB_API_TOKEN;
  base_url= targets a dedicated Inference Endpoint; HF_PROVIDER routes through a specific
  inference provider. cendor-core's instrument now detects chat_completion structurally and
  attributes the LLMCall to huggingface, so budgets/guard/audit apply.
• Azure AI Foundry provider (provider="azure" / "azure_openai" / "foundry", extra
  [azure]) — connects to Foundry deployments via the standard openai SDK on the
  /openai/v1/ endpoint, per Microsoft's current guidance (the AzureOpenAI client and
  azure-ai-inference are being retired). model is the Foundry deployment name; base_url
  (or AZURE_OPENAI_ENDPOINT) is normalized to the v1 route; api_key falls back to
  AZURE_OPENAI_API_KEY/AZURE_INFERENCE_CREDENTIAL. Entra ID via a bearer-token api_key or a
  bring-your-own client=. See docs/sdk.md → Connecting to Hugging Face & Azure AI Foundry.
• Azure Foundry Responses API (provider="azure_responses" / "foundry_responses") — drives
  the OpenAI Responses API over a Foundry endpoint for OpenAI-family deployments; same Foundry-aware
  construction as provider="azure" (Chat Completions).
• Foundry Local provider (provider="foundry_local" / "foundry-local", extra
  [foundry-local]) — Microsoft's on-device runtime over its local OpenAI-compatible REST server
  (the local counterpart to Ollama). Endpoint via base_url= or FOUNDRY_LOCAL_ENDPOINT (e.g.
  foundry_local.FoundryLocalManager(alias).endpoint); no key required. model is the resolved
  Foundry Local model id.
• Examples: examples/foundry_agent.py (Azure Foundry + Foundry Local) and
  examples/huggingface_agent.py, both offline.

Consolidation & docs (Phases 1–4)

Consolidates Phases 1–4 into a stable, documented release (plan §12): the governed agent loop,
multi-agent orchestration, ecosystem interop, and production hardening + governed eval — complete,
provider-agnostic, local-first, and tested offline.

• Full docs pass: docs/index.md links every page and lists the public API surface; the two-door
  framing (libraries primary, SDK secondary) is consistent across the README and docs.
• Marked Development Status :: 5 - Production/Stable.
• Verification: uv run pytest green (91 tests, no network), ruff check/ruff format clean,
  mypy clean, namespace-guard clean, import cendor.sdk works.