caplog.at_level(level, logger=NAME) only sets the level on the named
logger; pytest's caplog handler is attached to root, so any message
that exceeds the root logger's effective level still gets filtered out
before reaching the handler. CI's root logger config blocked the
adapter logger's WARNING/DEBUG records — the messages emitted (they
appear in the test output), caplog.records was empty, the
``any(...)`` assertions failed.

Drop the logger= argument so at_level() adjusts the root logger and
caplog reliably captures every record at that severity regardless of
runner config. The tests only assert on the message content, not the
emitting logger, so the change is behaviour-preserving.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

caplog wasn't catching adapter logger records on the Linux CI runners
even though the message reached stderr via lastResort. Some interaction
between caplog's root-attached LogCaptureHandler and the package's
propagation chain on those runners — fighting it isn't worth it.

Switch the warning/debug assertions to mock.patch the adapter module's
logger and verify the calls directly. The behaviour under test
(exception swallowed, warning emitted, debug breadcrumb on the
attach path) is the same; the assertion is just more direct and
doesn't depend on logging handler topology.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

CI's mypy config has disallow_any_generics=true and warn_unused_ignores=true,
which the test scaffolding tripped on every inline helper class:

- bare ``list`` / ``dict`` annotations needed parameterisation
- ``# type: ignore[no-untyped-def]`` on the duck-typed agent methods
  weren't actually needed (mypy lets untyped functions slide outside
  the package boundary)

Annotate ``list[Any]`` / ``dict[str, Any]`` and drop the noisy
type-ignore comments.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…f-blocks cleanly

Two bugs in ``GovernanceCallbackHandler.on_chat_model_start`` /
``on_llm_start`` made the BEFORE_MODEL evaluation noisy and inaccurate
for multi-turn agents:

1. ``str(msg.content)`` produced ``[{'type': 'text', 'text': ...},
   {'type': 'function_call', 'arguments': '{...}'}]`` dict-repr garble
   for list-of-blocks content (multimodal, OpenAI function-call,
   Anthropic tool_use, Claude extended thinking). Regex rules would
   still match by accident, but field-precise rules couldn't navigate
   the shape, quote escapes broke ``\b`` boundary anchors, and the
   meaningful text was buried under dict syntax.

2. Every LLM call concatenated **every** message in the prompt stack.
   For a multi-turn chat the model receives the full history each
   call; the adapter was passing that whole blob to
   ``evaluate_before_model``. A commitment-language violation in
   turn 3's user message kept re-firing on turns 4, 5, 6, ... because
   that text stayed in the prompt for context. Symptom dual of the
   BEFORE_AGENT bug fixed in uipath-runtime: there it under-fired
   (truncation dropped the latest message); here it over-fired
   (history kept re-triggering).

Fix:
- ``on_chat_model_start`` now uses ``_latest_message_input(messages)``
  which takes the last entry of the last batched prompt — the new
  content the LLM is about to respond to. Mirrors the BEFORE_AGENT
  ``latest_only=True`` contract added in uipath-runtime.
- List-of-blocks content is walked via the existing
  ``_extract_block_text`` helper (text + arguments + thinking + input)
  so structured shapes produce clean text, no dict-repr noise.
- ``on_llm_start`` (non-chat completion path) similarly takes only
  ``prompts[-1]`` — batched non-chat calls would otherwise re-scan
  earlier prompts on every callback.
- Both paths cap the extracted blob at ``_BEFORE_MODEL_TEXT_CAP =
  64000`` (matches the runtime side's ``_GOVERNANCE_TEXT_CAP``).

The LLM call is unaffected — ``messages`` / ``prompts`` are read-only
in the callback. Only ``model_input`` (what the evaluator scans)
changes.

Tests:
- Updated ``test_on_llm_start_invokes_evaluator_with_latest_prompt``
  and ``test_on_chat_model_start_latest_message_only`` /
  ``test_on_chat_model_start_dict_messages_latest_only`` to assert
  the new latest-only contract.
- New ``test_on_chat_model_start_list_of_blocks_content`` pins the
  function-call block extraction and the "no dict-repr noise"
  invariant.
- New ``test_on_chat_model_start_caps_model_input`` /
  ``test_on_chat_model_start_empty_messages`` /
  ``test_on_chat_model_start_empty_inner_batch`` cover the cap and
  empty-stack edges.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>