eq-chatbot-core 3.2.0__tar.gz → 3.3.0__tar.gz

eq_chatbot_core-3.3.0/AGENTS.md

@@ -0,0 +1,225 @@
+# AGENTS.md
+
+This file provides guidance to AI coding agents when working with code in this repository.
+
+## Project Overview
+
+Standalone Python library for LLM chatbot integration. Provides unified API across multiple providers (OpenAI, Anthropic, LangDock, OpenRouter, local servers). Originally extracted from v18-chatbot Odoo project for independent PyPI publishing.
+
+## Commands
+
+### Development Setup
+
+```bash
+uv venv && source .venv/bin/activate  # or: venv+
+uv pip install -e ".[dev,pdf,security]"
+```
+
+### Testing
+
+```bash
+# Run all unit tests (fast, mocked)
+pytest tests/unit/ -v
+
+# Run all tests with coverage
+pytest tests/ -v --cov=eq_chatbot_core --cov-report=html
+
+# Run single test file
+pytest tests/unit/test_openai.py -v
+
+# Run single test function
+pytest tests/unit/test_openai.py::TestOpenAIProviderInit::test_basic_init -v
+
+# Run tests by marker
+pytest -m unit -v                    # Unit tests only (mocked)
+pytest -m integration -v             # Integration tests (real API calls)
+pytest -m local -v                   # Local server tests
+
+# Skip expensive tests
+SKIP_LIVE_TESTS=true pytest tests/ -v
+```
+
+### Linting
+
+```bash
+ruff check src/ && ruff format src/ --check   # Check only
+ruff check src/ --fix && ruff format src/     # Auto-fix
+mypy src/                               # Type checking
+```
+
+### Building
+
+```bash
+python -m build                         # Build package
+twine check dist/*                      # Verify package
+```
+
+## Architecture
+
+### Provider Factory Pattern
+
+All providers use a factory function with consistent interface:
+
+```python
+from eq_chatbot_core.providers import get_provider
+
+# Cloud providers
+provider = get_provider("openai", api_key="sk-...")
+provider = get_provider("anthropic", api_key="sk-ant-...")
+provider = get_provider("langdock", api_key="ld-...")
+provider = get_provider("openrouter", api_key="sk-or-...")
+provider = get_provider("mammouth", api_key="mm-...")
+provider = get_provider("litellm", api_key="...", base_url="https://gateway/v1")  # OpenAI-compatible gateway
+provider = get_provider("ionos", api_key="...")  # IONOS AI Model Hub (EU-hosted, base_url has a default)
+provider = get_provider("melious", api_key="sk-mel-...")  # Melious.ai (sovereign EU-hosted, base_url has a default)
+provider = get_provider("privatemode")  # Privatemode.ai (end-to-end encrypted; needs the local privatemode-proxy on :8080)
+
+# Local providers (LM Studio, Ollama)
+provider = get_provider("lm_studio")  # defaults to localhost:1234
+provider = get_provider("ollama")      # defaults to localhost:11434
+provider = get_provider("local", base_url="http://custom:port/v1")
+```
+
+### Provider Base Class
+
+All providers inherit from `BaseLLMProvider` and must implement:
+
+```python
+class BaseLLMProvider(ABC):
+    provider_name: str              # Property: "openai", "anthropic", etc.
+    default_model: str              # Property: Default model ID
+
+    def chat_completion(messages, model, temperature, max_tokens, tools) -> LLMResponse
+    def stream_completion(messages, model, ...) -> Iterator[StreamChunk]
+    def list_models() -> list[dict]
+```
+
+### Response Types
+
+- `LLMResponse`: Complete response with content, token counts, tool_calls
+- `StreamChunk`: Streaming delta with is_final flag and accumulated tool_calls
+- `ModelInfo`: Model metadata (id, context_length, supports_vision, etc.)
+
+### Exception Hierarchy
+
+```
+ProviderError (base)
+├── RateLimitError     # 429 errors, has retry_after
+├── AuthenticationError # 401/403 errors
+├── ContextLengthError  # Token limit exceeded
+└── OverloadedError     # 529/503 transient errors (retryable)
+```
+
+### Module Structure
+
+```
+src/eq_chatbot_core/
+├── providers/              # LLM adapters
+│   ├── base.py             # BaseLLMProvider, response types, exceptions
+│   ├── openai_provider.py  # OpenAI
+│   ├── anthropic_provider.py
+│   ├── langdock_provider.py # LangDock gateway (EU/US regions)
+│   ├── openrouter_provider.py # OpenRouter (400+ models)
+│   ├── litellm_provider.py  # LiteLLM / any OpenAI-compatible gateway
+│   ├── ionos_provider.py    # IONOS AI Model Hub (EU-hosted, OpenAI-compatible)
+│   ├── melious_provider.py  # Melious.ai (sovereign EU-hosted, OpenAI-compatible)
+│   ├── privatemode_provider.py # Privatemode.ai (E2E-encrypted, via local attesting proxy)
+│   └── local_provider.py   # LM Studio, Ollama (OpenAI-compatible)
+├── security/
+│   ├── encryption.py       # FernetEncryption for API key storage
+│   ├── injection.py        # Prompt injection detection
+│   ├── rate_limit.py       # Token bucket rate limiter
+│   └── file_validator.py   # MIME type validation (requires [security])
+├── rag/
+│   ├── chunker.py          # Text chunking strategies
+│   ├── embedder.py         # Embedding generation
+│   ├── retriever.py        # Qdrant vector retrieval
+│   └── context_manager.py  # RAG context assembly
+├── mcp/
+│   ├── client.py           # MCP client (legacy HTTP/SSE and stdio transports)
+│   └── streamable_http.py  # MCP client (Streamable HTTP, the current remote transport)
+├── services/
+│   ├── error_handler.py    # Centralized error handling
+│   └── knowledge_service.py # Knowledge export for vector DBs
+├── utils/
+│   ├── url_validation.py   # SSRF / DNS-rebinding guard (shared by all providers)
+│   └── pdf.py              # PDF to image (requires [pdf])
+├── cli.py                  # Click CLI: eq-chatbot
+└── version.py              # Version string
+```
+
+## Testing Patterns
+
+### Unit Test Structure
+
+Unit tests mock SDK clients at module level to avoid API calls:
+
+```python
+import sys
+from unittest.mock import MagicMock
+
+# Mock before importing provider
+mock_openai_module = MagicMock()
+sys.modules["openai"] = mock_openai_module
+
+from eq_chatbot_core.providers.openai_provider import OpenAIProvider
+```
+
+### Test Markers
+
+- `@pytest.mark.unit` - Mocked tests, no external dependencies
+- `@pytest.mark.integration` - Real API calls, requires keys
+- `@pytest.mark.local` - Requires running LM Studio/Ollama
+- `@pytest.mark.expensive` - Uses expensive models, skipped in CI
+- `@pytest.mark.slow` - Long-running tests
+
+### Test Configuration
+
+API keys come from `~/.config/eq-chatbot/config.toml` (`[providers.<name>].api_key`),
+the same file the CLI uses. `conftest.py` mirrors them into `<NAME>_API_KEY` at
+import; a real environment variable wins. `tests/.env.test` was removed in 3.1.0 —
+it lived inside the repository, so the keys were one `git add -f` from publication.
+
+- Keys: `OPENAI_API_KEY`, `ANTHROPIC_API_KEY`, `LANGDOCK_API_KEY`, `MAMMOUTH_API_KEY`, `IONOS_API_KEY`, `MELIOUS_API_KEY`, `PRIVATEMODE_API_KEY`, `LITELLM_API_KEY`, `OPENROUTER_API_KEY`
+- `SKIP_LIVE_TESTS=true` - Skip integration tests (default false)
+- `SKIP_LOCAL_TESTS=true` - Skip local server tests (default false)
+- `<PROVIDER>_TEST_MODEL` - Override a model choice; defaults in `tests/model_registry.py`
+
+### Fixtures (conftest.py)
+
+- `test_config` - Session config from environment
+- `mock_openai_client`, `mock_anthropic_client` - Pre-configured mocks
+- `sample_messages`, `minimal_test_messages` - Test data
+- `openai_provider`, `anthropic_provider` - Live provider instances
+
+## Git Commit Conventions
+
+- `[ADD]` - New features
+- `[CHG]` - Modifications
+- `[FIX]` - Bug fixes
+
+## Dependencies
+
+### Core (always installed)
+
+openai (>=3), anthropic (>=1), httpx2, pydantic, cryptography, tiktoken, qdrant-client, click
+
+One HTTP client library: `httpx2` (Pydantic's maintained continuation of httpx) carries this
+library's own requests, the OpenAI SDK and — since anthropic 1.0.0 — the Anthropic SDK too.
+The separate `httpx<1` dependency is gone; anthropic 1.0.0 rejects an httpx client with
+`TypeError: Expected an instance of httpx2.Client`, so the floor is `>=1.0.0`.
+`build_pinned_transport_for_url()` still takes an `http=` argument and works against either
+module, which is why one guard implementation covers a future SDK that diverges again.
+
+That release also removed `temperature`, `top_p` and `top_k` from `messages.create()` and
+`messages.stream()`. Never write `params["temperature"]` for an Anthropic call — use
+`apply_anthropic_temperature()` from `providers/temperature_constraints.py`, which clamps and
+routes the value into `extra_body`.
+
+### Optional
+
+- `[dev]` - pytest, ruff, mypy, twine, pytest-cov, pytest-asyncio
+- `[security]` - puremagic (MIME validation)
+- `[pdf]` - pymupdf (PDF to image conversion)
+- `[rag]` - qdrant-client (Qdrant vector retrieval; optional since v3.0.0)
+- `[local]` - sentence-transformers (local embeddings)

{eq_chatbot_core-3.2.0 → eq_chatbot_core-3.3.0}/CHANGELOG.md

@@ -5,6 +5,143 @@ All notable changes to eq-chatbot-core will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [3.3.0] - 2026-09-08
+
+### Added
+
+- **MCP client for the Streamable HTTP transport** (`mcp/streamable_http.py`,
+  `StreamableHTTPMCPClient`, factory `get_mcp_client(transport="http", ...)`). Until now the
+  library only spoke the HTTP+SSE transport of the 2024-11-05 spec (`GET /sse` plus a
+  server-announced POST URL). Current server SDKs - .NET, TypeScript, Python - offer Streamable
+  HTTP (spec 2025-03-26 and later) for remote operation: one POST endpoint such as
+  `http://host:5100/mcp`, answered with JSON or with an SSE stream. Against such a server the
+  old client ran into a connect timeout before it ever sent `initialize`. The new client reads
+  both answer shapes, carries a server-assigned `Mcp-Session-Id`, re-initialises once when the
+  server reports the session as gone (HTTP 404), and sends `MCP-Protocol-Version` after the
+  negotiation; stateless servers need none of that. `list_tools()` follows `nextCursor`
+  pagination; a tool reporting `isError: true` comes back as `success=False` with its text.
+  Same public surface as the SSE client, so callers switch without code changes. Verified
+  live against the official Python SDK 1.x in both modes (stateful with SSE answers, stateless
+  with JSON answers). `sse` and `stdio` are unchanged; the docs had promised
+  `transport="http"` all along, which the factory never accepted.
+- **LAN mode for both HTTP transports** - `allow_private_ranges=True` on `get_mcp_client()`,
+  `MCPClient` and `StreamableHTTPMCPClient`. The SSRF check of the server URL always ran in
+  strict mode for MCP servers: private addresses, loopback and unresolvable hostnames were
+  rejected, so an MCP server on the company intranet was unreachable by construction even
+  though `validate_url()` has had a LAN mode for local LLM servers for a long time. In LAN
+  mode private and loopback targets become reachable and a hostname only the on-prem resolver
+  knows is accepted without an IP pin; link-local (cloud-metadata endpoint), reserved and
+  multicast addresses stay blocked. The SSE client applies the same mode to the POST URL it
+  receives in the `endpoint` event. Strict remains the default.
+
+Tests: 1913 unit tests (52 new in `tests/unit/test_mcp_streamable_http.py`), mypy --strict clean.
+
+## [3.2.1] - 2026-08-24
+
+### Fixed
+
+- **`.xlsx` conversion failed with an error about pandas.** The `[docs]` extra pulls
+  `markitdown[xlsx]`, which asks for a bare `openpyxl` and a bare `pandas` - no version on
+  either. The floor `openpyxl>=3.1.5` exists only inside pandas' own `excel` extra, and
+  markitdown does not request that extra, so nothing in the chain ever states it. A resolver
+  is therefore free to leave an older `openpyxl` in place, and Odoo pins exactly 3.1.2, so the
+  combination is not hypothetical - it is the normal case inside an Odoo install. Every
+  spreadsheet ended in:
+
+      XlsxConverter threw ImportError with message: Pandas requires version '3.1.5' or
+      newer of 'openpyxl' (version '3.1.2' currently installed).
+
+  Downstream that surfaces as "no text could be extracted", which names neither the library
+  nor the version. The `[docs]` extra now declares `openpyxl>=3.1.5,<4.0.0` itself.
+
+  Only `.xlsx` was affected - `.docx` goes through mammoth, `.pptx` through python-pptx and
+  `.pdf` through pymupdf, none of which route via pandas. Verified against all ten supported
+  extensions after the change.
+
+## [3.2.0] - 2026-08-23
+
+### Breaking
+
+- **`anthropic` floor raised to >= 1.0.0; `httpx` is no longer a core dependency.**
+  The Anthropic SDK moved to `httpx2` in 1.0.0 and now rejects an `httpx` client outright
+  (`TypeError: Expected an instance of httpx2.Client`). Since `pyproject.toml` allowed
+  `<2.0.0`, a `pip install -U` would have silently broken the Anthropic provider. Both SDK
+  clients run on `httpx2` now, and the separate `httpx>=0.27,<1` dependency is gone — it
+  existed solely because of that one incompatibility.
+
+  The same SDK release **removed** `temperature`, `top_p` and `top_k` from `messages.create()`
+  and `messages.stream()`; passing one is a `TypeError`, not a warning. The SDK notes call this
+  "some minor breaking changes" — in practice every temperature-setting call failed. The new
+  helper `providers.temperature_constraints.apply_anthropic_temperature()` clamps the value and
+  routes it into `extra_body`. **Never write `params["temperature"]` for an Anthropic call again.**
+
+### Fixed
+
+- **GPT-5 temperature support is not uniform — every call to `gpt-5`, `gpt-5.5` and the whole
+  current `gpt-5.6` tier failed with HTTP 400.** The constraint table carried a single permissive
+  `gpt-5` prefix declaring the entire family temperature-capable. Measured live on 23.08.2026:
+  `gpt-5` and `gpt-5-mini` refuse ("Only the default (1) value is supported"), `gpt-5.1`/`5.2`/`5.4`
+  accept, `gpt-5.5` and `gpt-5.6` refuse again. The generic prefix now says **no** and the
+  accepting generations are listed individually, so an untested future generation inherits "no
+  temperature" — omitting it never breaks, sending it does.
+
+- **`get_temperature_constraints()` fell through to the permissive default for a prefixed model.**
+  `openai/gpt-5.6-luna` matched no entry, got 0.0-2.0, and was therefore sent a parameter the model
+  rejects. The function now strips the provider prefix itself instead of trusting every caller to.
+
+- **LangDock agents really stream now.** `_agent_stream_completion` sent `"stream": false` and
+  yielded the finished answer as a single chunk — no streaming, and LangDock aborts non-streaming
+  requests after 100 seconds with HTTP 524, which an agent with tools or knowledge search reaches
+  easily. Both agent paths now go through `_agent_request_events()` and the AI SDK 5 data-stream
+  protocol; the synchronous path streams internally and assembles the answer itself. The parser
+  splits on the `data: ` marker rather than on line boundaries (LangDock glues events together) and
+  skips unparseable blocks instead of letting one kill a running answer.
+
+- **The LangDock Google backend was entirely broken.** `list_models()` served a hand-maintained
+  catalogue (`gemini-2.5-flash`, `gemini-2.5-pro`) the gateway had retired — including the backend
+  default model, so every call without an explicit model ended in 400. The list now comes live from
+  the same endpoint the error came from. On top of that the streaming path silently yielded
+  **nothing**: the parser looks for SSE lines, but without `?alt=sse` LangDock answers with a JSON
+  array — no match, no chunks, no error. The parameter is now set.
+
+- **Google and Codestral errors swallowed their own cause.** Both logged the response body and then
+  called `raise_for_status()`, so the caller saw only "Client error '400 Bad Request'" while the
+  actual reason — a retired model, a missing entitlement — sat in a log line. The shared
+  `_raise_http_error()` now appends the body without losing the typed mapping (401 stays
+  `AuthenticationError`, 429 `RateLimitError`).
+
+- **The Agent API's HTTP 400 is translated.** It covers three entirely different cases: agent
+  unknown, agent not shared with the API key, or no model resolvable at all. The last happens when
+  the agent's model is set to "Auto" in LangDock — the auto-router works only in the chat UI, and a
+  top-level `model` demonstrably does not override it. The message now names the remedy in plain
+  words.
+
+### Changed
+
+- **Every default model raised to the current generation.** Four were unusable or outdated:
+  `anthropic` pointed at `claude-sonnet-4-20250514` and `melious` at `minimax-428b-m3` — both
+  withdrawn — while `openai`, `mammouth` and `openrouter` still named `gpt-4o`. Now:
+  `gpt-5.6-luna`, `claude-sonnet-5`, `openai/gpt-5.6-luna`, `nemotron-3-nano-30b-a3b`. Three of
+  LangDock's four backend defaults were dead (`gpt-4o`, `claude-sonnet-4-20250514`,
+  `gemini-2.5-flash`).
+
+  A default model is the one unavoidable constant — everything else is resolved live. The new live
+  guards `TestProviderDefaultsAreLive` and `TestLangDockDefaultsAreLive` fail the moment a provider
+  stops serving one, and additionally check that it actually answers: being listed is not enough.
+
+- **Dependency floors raised to the current release** (21 entries), among them
+  `sentence-transformers` 3 -> 6, `websockets` 13 -> 17.0.1, `pytest-asyncio` 0.24 -> 1.4,
+  `pytest` 8 -> 9.1.1, `mypy` 1.15 -> 2.3.1, `pytest-cov` 6 -> 7.1, `qdrant-client` 1.12 -> 1.19,
+  `fastapi` 0.115 -> 0.141, `uvicorn` 0.32 -> 0.52. An old floor is an open vulnerability waiting
+  for a fresh install. The suite runs green on all new majors without a code change.
+
+- **Live coverage now spans all LangDock backends.** Only `openai` and `anthropic` were tested;
+  the Google backend speaks an entirely different protocol and was never touched live — which is
+  precisely why its defect went unnoticed. The Codestral path has no coverage any more: Mistral is
+  not of interest via LangDock.
+
+**Test suite:** 1858 unit tests, 51 live tests (0 failures), mypy `--strict` clean.
+
 ## [3.1.1] - 2026-08-20
 
 ### Fixed

{eq_chatbot_core-3.2.0 → eq_chatbot_core-3.3.0}/CLAUDE.md

@@ -136,7 +136,8 @@ src/eq_chatbot_core/
 │   ├── retriever.py        # Qdrant vector retrieval
 │   └── context_manager.py  # RAG context assembly
 ├── mcp/
-│   └── client.py           # MCP client (HTTP/SSE and stdio transports)
+│   ├── client.py           # MCP client (legacy HTTP/SSE and stdio transports)
+│   └── streamable_http.py  # MCP client (Streamable HTTP, the current remote transport)
 ├── services/
 │   ├── error_handler.py    # Centralized error handling
 │   └── knowledge_service.py # Knowledge export for vector DBs

{eq_chatbot_core-3.2.0 → eq_chatbot_core-3.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: eq-chatbot-core
-Version: 3.2.0
+Version: 3.3.0
 Summary: Core library for LLM chatbot integration with multi-provider support (OpenAI, Anthropic, LangDock, OpenRouter, Mammouth, Azure, Vertex AI, LiteLLM, IONOS, Local)
 Project-URL: Homepage, https://www.ownerp.com
 Project-URL: Documentation, https://github.com/equitania/eq-chatbot-core
@@ -39,6 +39,7 @@ Requires-Dist: ruff<1.0.0,>=0.16.4; extra == 'dev'
 Requires-Dist: twine<8.0.0,>=7.0.0; extra == 'dev'
 Provides-Extra: docs
 Requires-Dist: markitdown[docx,pdf,pptx,xlsx]<0.2.0,>=0.1.7; extra == 'docs'
+Requires-Dist: openpyxl<4.0.0,>=3.1.5; extra == 'docs'
 Requires-Dist: pymupdf<2.0.0,>=1.28.2; extra == 'docs'
 Provides-Extra: image
 Requires-Dist: pillow<13.0,>=12.3.0; extra == 'image'
@@ -85,7 +86,7 @@ Originally extracted from an Odoo 18 chatbot integration; works standalone witho
 - **Temperature Safety** — automatic model-specific temperature clamping
 - **Security** — Fernet encryption, prompt-injection detection (direct user input + indirect tool/RAG content), file-upload validation, race-free token-bucket rate limiting
 - **RAG Pipeline** — chunking, embeddings (incl. Melious.ai and IONOS embedders), Qdrant-backed retrieval, context-window management
-- **MCP Client** — HTTP/SSE and stdio transports, hardened against DNS rebinding, SSRF, and subprocess env injection
+- **MCP Client** — Streamable HTTP, legacy HTTP/SSE and stdio transports, hardened against DNS rebinding, SSRF, and subprocess env injection; LAN mode for servers on the intranet
 - **CLI Tool** — provider testing, model discovery, programmatic JSON I/O chat
 - **Text-to-Image Generation** (v1.14.0) — `eq-chatbot image` (single PNG) and `eq-chatbot listing-assets` (batch from a recipe); OpenAI `gpt-image-1` and OpenRouter image models
 - **HTTP/SSE Server Mode** (v1.7.0) — run as a local sidecar (`eq-chatbot serve`) for cross-language integrations (Avalonia/.NET, Electron, native mobile)
@@ -162,7 +163,7 @@ For more — streaming, other providers, error handling — see [docs/providers.
 | CLI commands | [docs/cli.md](docs/cli.md#english) |
 | HTTP/SSE server mode | [docs/server-mode.md](docs/server-mode.md#english) |
 | Security (encryption, injection, files, rate limit) | [docs/security.md](docs/security.md#english) |
-| MCP client (HTTP/SSE + stdio) | [docs/mcp.md](docs/mcp.md#english) |
+| MCP client (Streamable HTTP + SSE + stdio) | [docs/mcp.md](docs/mcp.md#english) |
 | RAG pipeline (chunking, embedding, retrieval) | [docs/rag.md](docs/rag.md#english) |
 | Testing (markers, integration setup, cost-aware patterns) | [docs/testing.md](docs/testing.md) |
 
@@ -289,7 +290,7 @@ Ursprünglich aus einer Odoo-18-Chatbot-Integration extrahiert; funktioniert sta
 - **Temperature-Sicherheit** — automatisches modellspezifisches Temperature-Clamping
 - **Sicherheit** — Fernet-Verschlüsselung, Prompt-Injection-Erkennung (direkte Nutzereingaben + indirekte Tool-/RAG-Inhalte), File-Upload-Validierung, Race-freies Token-Bucket-Rate-Limiting
 - **RAG-Pipeline** — Chunking, Embeddings (inkl. Melious.ai- und IONOS-Embedder), Qdrant-basiertes Retrieval, Context-Window-Management
-- **MCP-Client** — HTTP/SSE und stdio Transports, gehärtet gegen DNS-Rebinding, SSRF und Subprocess-Env-Injection
+- **MCP-Client** — Streamable HTTP, älteres HTTP/SSE und stdio als Transports, gehärtet gegen DNS-Rebinding, SSRF und Subprocess-Env-Injection; LAN-Modus für Server im Intranet
 - **CLI-Tool** — Provider-Tests, Modell-Discovery, programmatische JSON-I/O-Chat-Calls
 - **Text-zu-Bild-Generierung** (v1.14.0) — `eq-chatbot image` (einzelnes PNG) und `eq-chatbot listing-assets` (Batch aus einer Recipe); OpenAI `gpt-image-1` und OpenRouter-Bildmodelle
 - **HTTP/SSE-Server-Mode** (v1.7.0) — lokaler Sidecar (`eq-chatbot serve`) für Cross-Language-Integrationen (Avalonia/.NET, Electron, native Mobile)
@@ -370,7 +371,7 @@ Für mehr — Streaming, andere Provider, Error-Handling — siehe [docs/provide
 | CLI-Befehle | [docs/cli.md](docs/cli.md#deutsch) |
 | HTTP/SSE-Server-Mode | [docs/server-mode.md](docs/server-mode.md#deutsch) |
 | Security (Verschlüsselung, Injection, Files, Rate-Limit) | [docs/security.md](docs/security.md#deutsch) |
-| MCP-Client (HTTP/SSE + stdio) | [docs/mcp.md](docs/mcp.md#deutsch) |
+| MCP-Client (Streamable HTTP + SSE + stdio) | [docs/mcp.md](docs/mcp.md#deutsch) |
 | RAG-Pipeline (Chunking, Embedding, Retrieval) | [docs/rag.md](docs/rag.md#deutsch) |
 | Testing (Marker, Integration-Setup, Cost-Aware-Patterns) | [docs/testing.md](docs/testing.md) |
 

{eq_chatbot_core-3.2.0 → eq_chatbot_core-3.3.0}/README.md

@@ -25,7 +25,7 @@ Originally extracted from an Odoo 18 chatbot integration; works standalone witho
 - **Temperature Safety** — automatic model-specific temperature clamping
 - **Security** — Fernet encryption, prompt-injection detection (direct user input + indirect tool/RAG content), file-upload validation, race-free token-bucket rate limiting
 - **RAG Pipeline** — chunking, embeddings (incl. Melious.ai and IONOS embedders), Qdrant-backed retrieval, context-window management
-- **MCP Client** — HTTP/SSE and stdio transports, hardened against DNS rebinding, SSRF, and subprocess env injection
+- **MCP Client** — Streamable HTTP, legacy HTTP/SSE and stdio transports, hardened against DNS rebinding, SSRF, and subprocess env injection; LAN mode for servers on the intranet
 - **CLI Tool** — provider testing, model discovery, programmatic JSON I/O chat
 - **Text-to-Image Generation** (v1.14.0) — `eq-chatbot image` (single PNG) and `eq-chatbot listing-assets` (batch from a recipe); OpenAI `gpt-image-1` and OpenRouter image models
 - **HTTP/SSE Server Mode** (v1.7.0) — run as a local sidecar (`eq-chatbot serve`) for cross-language integrations (Avalonia/.NET, Electron, native mobile)
@@ -102,7 +102,7 @@ For more — streaming, other providers, error handling — see [docs/providers.
 | CLI commands | [docs/cli.md](docs/cli.md#english) |
 | HTTP/SSE server mode | [docs/server-mode.md](docs/server-mode.md#english) |
 | Security (encryption, injection, files, rate limit) | [docs/security.md](docs/security.md#english) |
-| MCP client (HTTP/SSE + stdio) | [docs/mcp.md](docs/mcp.md#english) |
+| MCP client (Streamable HTTP + SSE + stdio) | [docs/mcp.md](docs/mcp.md#english) |
 | RAG pipeline (chunking, embedding, retrieval) | [docs/rag.md](docs/rag.md#english) |
 | Testing (markers, integration setup, cost-aware patterns) | [docs/testing.md](docs/testing.md) |
 
@@ -229,7 +229,7 @@ Ursprünglich aus einer Odoo-18-Chatbot-Integration extrahiert; funktioniert sta
 - **Temperature-Sicherheit** — automatisches modellspezifisches Temperature-Clamping
 - **Sicherheit** — Fernet-Verschlüsselung, Prompt-Injection-Erkennung (direkte Nutzereingaben + indirekte Tool-/RAG-Inhalte), File-Upload-Validierung, Race-freies Token-Bucket-Rate-Limiting
 - **RAG-Pipeline** — Chunking, Embeddings (inkl. Melious.ai- und IONOS-Embedder), Qdrant-basiertes Retrieval, Context-Window-Management
-- **MCP-Client** — HTTP/SSE und stdio Transports, gehärtet gegen DNS-Rebinding, SSRF und Subprocess-Env-Injection
+- **MCP-Client** — Streamable HTTP, älteres HTTP/SSE und stdio als Transports, gehärtet gegen DNS-Rebinding, SSRF und Subprocess-Env-Injection; LAN-Modus für Server im Intranet
 - **CLI-Tool** — Provider-Tests, Modell-Discovery, programmatische JSON-I/O-Chat-Calls
 - **Text-zu-Bild-Generierung** (v1.14.0) — `eq-chatbot image` (einzelnes PNG) und `eq-chatbot listing-assets` (Batch aus einer Recipe); OpenAI `gpt-image-1` und OpenRouter-Bildmodelle
 - **HTTP/SSE-Server-Mode** (v1.7.0) — lokaler Sidecar (`eq-chatbot serve`) für Cross-Language-Integrationen (Avalonia/.NET, Electron, native Mobile)
@@ -310,7 +310,7 @@ Für mehr — Streaming, andere Provider, Error-Handling — siehe [docs/provide
 | CLI-Befehle | [docs/cli.md](docs/cli.md#deutsch) |
 | HTTP/SSE-Server-Mode | [docs/server-mode.md](docs/server-mode.md#deutsch) |
 | Security (Verschlüsselung, Injection, Files, Rate-Limit) | [docs/security.md](docs/security.md#deutsch) |
-| MCP-Client (HTTP/SSE + stdio) | [docs/mcp.md](docs/mcp.md#deutsch) |
+| MCP-Client (Streamable HTTP + SSE + stdio) | [docs/mcp.md](docs/mcp.md#deutsch) |
 | RAG-Pipeline (Chunking, Embedding, Retrieval) | [docs/rag.md](docs/rag.md#deutsch) |
 | Testing (Marker, Integration-Setup, Cost-Aware-Patterns) | [docs/testing.md](docs/testing.md) |
 

{eq_chatbot_core-3.2.0 → eq_chatbot_core-3.3.0}/RELEASE_NOTES.md

@@ -1,5 +1,61 @@
 # Release Notes
 
+## Version 3.3.0 (08.09.2026)
+
+### [ADD]
+
+- **MCP-Client für den Streamable-HTTP-Transport** (`mcp/streamable_http.py`,
+  `StreamableHTTPMCPClient`, Fabrik `get_mcp_client(transport="http", …)`). Bisher sprach die
+  Bibliothek nur den HTTP+SSE-Transport der Spec 2024-11-05 (`GET /sse`, vom Server gemeldete
+  POST-URL). Aktuelle Server-SDKs — .NET, TypeScript, Python — bieten für den Fernbetrieb
+  Streamable HTTP (Spec ab 2025-03-26): ein einziger POST-Endpunkt wie `http://host:5100/mcp`,
+  Antwort als JSON oder als SSE-Stream. Gegen so einen Server lief der alte Client in einen
+  Verbindungs-Timeout, bevor überhaupt ein `initialize` gesendet wurde. Der neue Client
+  liest beide Antwortformen, führt eine vom Server vergebene `Mcp-Session-Id` mit und
+  initialisiert einmal neu, wenn der Server die Sitzung mit HTTP 404 als abgelaufen meldet;
+  zustandslose Server brauchen nichts davon. Nach der Aushandlung sendet er
+  `MCP-Protocol-Version`. `list_tools()` folgt der `nextCursor`-Blätterung; ein Werkzeug mit
+  `isError: true` kommt als `success=False` samt Text zurück. Öffentliche Oberfläche wie beim
+  SSE-Client, daher ohne Änderung am Aufrufer austauschbar. Live geprüft gegen das offizielle
+  Python-SDK 1.x in beiden Betriebsarten (zustandsbehaftet mit SSE-Antworten, zustandslos mit
+  JSON-Antworten). Die Transportnamen `sse` und `stdio` bleiben unverändert; die Doku hatte
+  bereits `transport="http"` versprochen, das die Fabrik bis jetzt gar nicht kannte.
+- **LAN-Modus für beide HTTP-Transports** — `allow_private_ranges=True` an
+  `get_mcp_client()`, `MCPClient` und `StreamableHTTPMCPClient`. Die SSRF-Prüfung der
+  Server-URL lief für MCP-Server immer im strikten Modus: private Adressen, Loopback und
+  nicht auflösbare Hostnamen wurden abgewiesen — ein MCP-Server im Firmen-Intranet war damit
+  grundsätzlich unerreichbar, obwohl `validate_url()` den LAN-Modus für lokale LLM-Server
+  längst kannte. Im LAN-Modus werden private und Loopback-Ziele erreichbar und ein nur intern
+  auflösbarer Hostname ohne IP-Pin akzeptiert; Link-Local (Cloud-Metadaten-Endpunkt),
+  reservierte und Multicast-Adressen bleiben gesperrt. Beim SSE-Client gilt derselbe Modus
+  auch für die im `endpoint`-Event gemeldete POST-URL. Standard bleibt strikt.
+
+Tests: 1913 Unit-Tests (52 neu in `tests/unit/test_mcp_streamable_http.py`), mypy --strict
+sauber.
+
+## Version 3.2.1 (24.08.2026)
+
+### [FIX]
+
+- **Die Umwandlung von `.xlsx` scheiterte mit einer Meldung über pandas.** Das `[docs]`-Extra
+  zieht `markitdown[xlsx]`, und das verlangt schlicht „openpyxl" und „pandas" — ohne jede
+  Fassungsangabe. Die Untergrenze `openpyxl>=3.1.5` steht allein im `excel`-Extra von pandas,
+  das markitdown nicht anfordert; in der ganzen Kette nennt sie also niemand. Ein Auflöser darf
+  damit eine ältere Fassung stehen lassen, und Odoo legt sich genau auf 3.1.2 fest — die
+  Kombination ist demnach kein Sonderfall, sondern der Regelfall in einer Odoo-Installation.
+  Jede Tabellenkalkulation endete mit:
+
+      XlsxConverter threw ImportError with message: Pandas requires version '3.1.5' or
+      newer of 'openpyxl' (version '3.1.2' currently installed).
+
+  Beim Aufrufer kommt davon „Es konnte kein Text extrahiert werden" an — eine Meldung, die
+  weder die Bibliothek noch die Fassung nennt. Das `[docs]`-Extra verlangt jetzt selbst
+  `openpyxl>=3.1.5,<4.0.0`.
+
+  Betroffen war ausschließlich `.xlsx`: `.docx` läuft über mammoth, `.pptx` über python-pptx
+  und `.pdf` über pymupdf, keines davon über pandas. Nach der Änderung gegen alle zehn
+  unterstützten Endungen geprüft.
+
 ## Version 3.2.0 (23.08.2026)
 
 ### ⚠️ Breaking