…I intact

Swap the per-provider SDK internals of ai/agent.py (_run/_stream over the
anthropic/openai/google SDKs) for a single LangChain/LangGraph backend:
init_chat_model builds the chat model and create_agent drives the tool loop,
with the final AIMessage mapped back to AgentResponse. The user-facing surface
is unchanged — prompt/stream/fake/assert_prompted/assert_not_prompted/reset,
the lifecycle hooks, and the decorators keep identical signatures.

- _build_model() is the seam tests patch to inject a fake chat model.
- _build_messages() now renders attachments via Document.to_langchain_block().
- Add Document.to_langchain_block(): inline text, base64 image/file blocks.
- Add ai/fakes.py fake_chat_model(): replays scripted AIMessage turns through a
  GenericFakeChatModel (bind_tools no-op) so the real create_agent loop runs
  offline; exported from the ai package root.
- New optional [langgraph] extra (langchain + langchain-core + langgraph).

The 23 tests in tests/ai/test_agent_fake.py stay green and unmodified (fake()
short-circuits before the backend). Adds tests/ai/test_agent_langgraph_backend.py
exercising the real loop offline: simple reply, full tool-calling loop, usage
mapping, attachment blocks, provider mapping, and streaming.

…ent through it

Turn ai/config.py into a config package and resolve models/providers through a
new Lab helper instead of hardcoded dicts on the Agent.

- ai/config/: split provider dataclasses (config.py) from the top-level AIConfig
  (ai.py), add config/__init__.py re-exporting them, and give each provider a
  models map keyed by modality (default / default_image / default_audio /
  default_transcribe). Fix the draft's circular import (AIConfig imported the
  provider configs from the package root mid-init) and the placeholder model
  values (google text default, elevenlabs models).
- AIConfig selects the default provider per modality: default (text),
  default_image, default_audio, default_transcribe. image.py/audio.py now read
  default_image/default_audio (was image_provider/audio_provider).
- ai/lab.py: Lab(StrEnum) + ModelType resolve the provider, default model, and
  the "<langchain-provider>:<model>" URL from Config (google → google_genai).
- Agent: _resolve_model() and _build_model() now go through Lab; removed the
  stale _DEFAULT_MODELS/_LANGCHAIN_PROVIDERS references and the dead
  _execute_tool() (create_agent runs tools itself).
- Tests: drop the _build_model monkeypatch helper; the backend tests now patch
  the real langchain.chat_models.init_chat_model seam via pytest monkeypatch.
  Add test_lab.py; update image/audio provider-selection mocks.

The 23 tests in tests/ai/test_agent_fake.py stay green and unmodified.

…l loop

Replace the LangGraph create_agent backend with a plain init_chat_model call
driven by a Runner that resolves and executes tool calls itself.

- runner.py: Runner(model, tools, max_steps) binds tools, invokes the model,
  executes requested tool calls, feeds results back, loops to a final answer;
  StreamRunner yields content tokens through the same loop. Fully typed.
- Agent._run/_stream delegate to Runner/StreamRunner (threading _max_steps);
  no create_agent.
- System message is declarative via instructions()/_instructions — removed the
  per-call system= and messages= arguments from prompt()/stream().
- Resolve provider/model through Lab directly (dropped the _lab() helper) and
  import Lab at module top.
- Config split into ai/config/{ai,config}.py.

KNOWN RED: ai/config/__init__.py is absent, so 'from fastapi_startkit.ai.config
import AIConfig' fails — the AI test suite does not collect and AIProvider import
breaks. Backend tests also still assume the old create_agent result shape and the
tuple return of _build_messages. Follow-ups.

Add a class-level testing harness so an agent can be faked or recorded
without a real model provider:

- Agent.fake({...}) and Agent.record(path) return an AgentBinding usable
  as a context manager or test decorator. The binding swaps a stand-in
  into the service container under the agent's class name and auto-resets
  on exit, so even a controller's own ChatAgent().prompt(...) is covered.
- FakeAgent answers from glob patterns; RecordingAgent records the real
  reply to JSON once, then replays it. Both expose assert_prompted().
- Agent.make()/faked() resolve the bound stand-in for assertions.
- prompt()/stream() delegate to an active binding; the in-process
  agent.fake({...}) instance API is preserved via a dual-purpose accessor.
- Lab.ModelType carries the models-map key as a static mapping.
- Import AIConfig from the fastapi_startkit.ai namespace in tests and the
  AI facade stub, matching the provider registration.

Tests: full suite 1541 passed, 7 skipped; ruff clean.

Convert bare assert statements to self.assertEqual in the example/agents
feature tests, matching the unittest.IsolatedAsyncioTestCase base. Keep the
async test methods and the @ChatAgent.fake / @ChatAgent.record decorators
intact.

…sertions

test(agents): use unittest assertions in example feature tests

Add direct tests for ai/pipeline.py — the Response deferred-callback
mechanism and build_pipeline onion that back agent middleware. Locks down
that the .then after-hook fires exactly once on both the buffered (await)
and streaming (async for) consumption paths, modelled on the AgentLogger
request/response logging pattern. Brings pipeline.py to 100% coverage.