GeezerVison

Adds an additive on_final_ai_message=None parameter to
convert_langgraph_to_agentex_events so callers can capture AIMessage
usage_metadata without re-traversing the stream. No behavior change when
omitted. Also adds a DeprecationWarning to create_langgraph_tracing_handler
and its module docstring, pointing to the unified harness surface, and
updates the sync module docstring with the preferred unified path.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Implements LangGraphTurn (HarnessTurn protocol) that wraps a LangGraph
astream() event stream and captures usage from AIMessage.usage_metadata
via the on_final_ai_message callback. Implements langgraph_usage_to_turn_usage
that maps all UsageMetadata fields (input/output/total/cache_read/reasoning)
onto the framework-agnostic TurnUsage model. Zero token counts are preserved.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…pre-refactor)

Records the current bespoke behavior as a contract test. After Task 4 rewrites
the internals to use UnifiedEmitter + LangGraphTurn, these tests must still pass
to confirm behavioral parity.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…urface

Replaces the bespoke Redis-streaming loop with UnifiedEmitter.auto_send_turn(
LangGraphTurn(...)), matching the pattern established for pydantic-ai. Public
signature preserved identically. Behavioral difference: tool calls/responses
are now posted via streaming_task_message_context (not adk.messages.create),
and final_text accumulates all text across the turn. Updates the characterization
test to document these unified-surface semantics.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Verifies yield_turn(LangGraphTurn) produces identical events to direct
iteration, and documents the AGX1-377 behavior (LangGraph Full tool events
don't produce SpanDeriver spans today; cross-channel equivalence comes with
AGX1-373).

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…-step)

Registers LangGraph-specific conformance fixtures with the shared harness
conformance runner. Documents the AGX1-377 behavior (tool requests are Full
events, not Start+Done). Span derivation is deterministic for all 4 fixtures.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…ral channels

Adds 18 offline integration tests across the three delivery channels using
fake LangGraph event streams and fake streaming backends. Documents the
AGX1-377 behavior (Full events don't produce tool spans). Notes the usage
capture timing: turn.usage() is the authoritative post-iteration value since
auto_send_turn evaluates usage eagerly before events are consumed.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Task 9: add 3 deployable tutorial agents that demonstrate the unified
harness surface side-by-side with the bespoke reference examples:
- examples/tutorials/00_sync/harness_langgraph/ (s-harness-langgraph)
  uses UnifiedEmitter.yield_turn(LangGraphTurn(stream))
- examples/tutorials/10_async/00_base/harness_langgraph/ (a-harness-langgraph)
  uses UnifiedEmitter.auto_send_turn(LangGraphTurn(stream))
- examples/tutorials/10_async/10_temporal/harness_langgraph/ (at-harness-langgraph)
  follows 130_langgraph pattern (LangGraphPlugin + emit_langgraph_messages)

Task 10: enable live-matrix CI job in harness-integration.yml with a
3-way matrix over [sync, async, temporal] running offline integration tests.
Also add test_harness_langgraph_*.py to PR path triggers.

Task 11 (pyright fixes): annotate convert_langgraph_to_agentex_events and
_generate_events with AsyncGenerator return types so pyright infers them as
async generators rather than coroutines. Add start_time to Span construction
in test_langgraph_sync_unified.py fake tracing backend.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>

…racing handler (PR 5/6)

AGX1-378: wire workflow_now_if_in_workflow() into stream_langgraph_events so
Temporal callers get deterministic message timestamps, matching the pattern
used by the openai/litellm providers.

Deprecation alignment: remove runtime warnings.warn from
create_langgraph_tracing_handler (and unused import warnings) to match PR 4/6
pydantic-ai convention. Deprecation remains in docstrings on module, class,
and function. Callers under -W error are no longer broken.

Test alignment after rebase onto unified-harness-surface (b4b8b33):
- FakeStreamingModule.streaming_task_message_context in test_langgraph_async.py
  and test_pydantic_ai_async.py updated to accept **kw (foundation now passes
  created_at).
- Three "no tool spans for Full events" tests updated to assert the new
  SpanDeriver behaviour: Full(ToolRequestContent) opens a span,
  Full(ToolResponseContent) closes it.
- Two "accumulates all text" multi-step tests corrected to last-segment
  semantics (auto_send resets final_text_parts on each new Start(TextContent)).

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…-373)

Rewrites test_langgraph_conformance.py to use the cross-channel runner
from PR #414 (run_cross_channel_conformance, LogicalDelivery) instead of
the simpler derive_all-only API it was written against.

The four fixtures (text-only, single-tool, reasoning, multi-step) are
retained as canonical StreamTaskMessage* sequences. Each is now exercised
by test_cross_channel_equivalence (yield_events vs auto_send logical
deliveries and span signals) plus the backward-compat
test_span_derivation_is_deterministic guard.

LangGraph tool requests arrive as Full events from the "updates" stream;
auto_send handles them via open+close, yielding the same LogicalDelivery
on both channels. No coalesce_tool_requests option is needed.

Co-Authored-By: Claude Sonnet 4.6 <[email protected]>

_capture overwrote self._usage on every AIMessage, so a multi-step turn (text
-> tool decision -> final text) reported only the last LLM call's tokens and
silently dropped the rest — undercounting in any billing/monitoring that reads
turn.usage(). Accumulate additively across calls via _accumulate_turn_usage
(None+None stays None; real 0 contributes 0). Add a test asserting summed
input/output/total/cache/reasoning tokens across two AIMessages.

The separate 06-18 "TurnResult.usage empty via auto_send_turn" comment is
resolved by the foundation (emitter reads turn.usage() after stream exhaustion).

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…e [greptile]

The sync harness_langgraph tutorial set turn_span.output to
{"final_output": turn.usage().model_dump()} — token metrics under a key that
means the assistant's text, producing misleading AGENT_WORKFLOW trace data
versus the async tutorial. Accumulate text deltas during the yield loop (as the
030_langgraph tutorial does) and store {"final_output": final_text, "usage":
...} so the final output is the text and usage stays available under its own key.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…nt [greptile]

The sync converter opened a reasoning stream with a StreamTaskMessageStart
wrapping TextContent, even though the deltas are ReasoningContentDelta — so the
Start's content type contradicted its payload (and downstream span/UI handling
of reasoning). Emit ReasoningContent for the reasoning Start. Add a test
asserting the reasoning block opens a ReasoningContent Start + ReasoningContentDelta.

Also refresh the stale test_harness_langgraph_sync.py module docstring: it
claimed the SpanDeriver does NOT produce tool spans for LangGraph Full events,
but the foundation now opens/closes tool spans from Full(ToolRequest/Response)
(asserted by test_tracer_produces_tool_spans_for_full_events).

Note: the "TurnResult.usage pre-iteration snapshot" item is already resolved at
root — UnifiedEmitter.auto_send_turn reads turn.usage() after auto_send drains
the stream.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

- _langgraph_sync.py: reasoning StreamTaskMessageStart now sets style="active".
  The AgentEx server's StreamTaskMessageStartEntity rejects reasoning.style=None
  (enum), which killed the sync stream — breaking harness_langgraph and the
  pre-existing 030_langgraph that share this emitter.
- temporal harness tools.py: give get_weather a native async coroutine so
  tools_node's `await tool.ainvoke(...)` runs on the workflow loop instead of
  LangChain's run_in_executor fallback (NotImplementedError in the deterministic
  Temporal workflow sandbox).
- test_langgraph_sync.py: assert the reasoning Start carries a non-null style.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

The shared TurnUsage.num_llm_calls became Optional (None = "not reported")
when it landed on next. The openai / pydantic-ai turn tests still asserted
== 0 for the default (no-usage / pre-exhaustion) construction path, which now
yields None — matching the token fields. Update those three assertions to
`is None`; real-zero cases (provider reported 0) are unchanged.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>