GeezerVison

Approach A (Agentex event stream as canonical source of truth): one tap per
harness feeds shared yield/auto-send delivery adapters and a span-deriving
tracing tap. Additive backwards-compat, stacked PRs <1000 lines, conformance +
live-matrix testing (3 test agents per harness: sync/async/temporal).

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

… golden-agent integration

- Make tracing-tap span derivation explicit (tool open on Done of a
  ToolRequestContent index, close on matching ToolResponseContent by
  tool_call_id; parallel-safe; reasoning start->done). Flag missing
  is_error on ToolResponseContent as an additive upstream decision.
- Add first-class TurnUsage/TurnResult shape (aligned to llm_metrics token
  taxonomy) attached to the turn span via span(data=) and reused for metrics.
- Document golden-agent integration: all SGP/sandbox/secret/MCP coupling
  stays in the agent; only parsing/streaming/tracing/usage move to SDK taps +
  emitter; sandbox-setup events chain before the harness stream.

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

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

… 1-3)

Bite-sized TDD tasks: foundation types, pure SpanDeriver, SpanTracer adapter,
yield + auto_send delivery, UnifiedEmitter facade, conformance scaffold + CI
job. Migration/parser PRs (4-9) listed as follow-on plans.

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

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

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

… signals

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

… handling in SpanDeriver

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

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

…sts for SpanTracer

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

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

…on early close

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

…reaming + tracing)

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

… + cover error/finally paths in auto_send

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

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

…send_turn + doc tracer modes

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

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

…egistry semantics

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

…he package

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

…or consistency

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

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

…reportImplicitOverride)

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

…MissingImports)

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

…ast-segment, created_at (AGX1-377, AGX1-378)

auto_send.py:
- Replace single current_ctx with ctx_map[index] so parallel streams route correctly
- Open a streaming context for ALL content types on Start (not just text/reasoning),
  fixing tool_request/tool_response stream delivery (AGX1-377)
- Reset final_text_parts on each new Start(TextContent) and on Full(TextContent)
  so multi-step turns return the LAST text segment, not the full accumulation
- Add created_at: datetime | None param; forward to every
  streaming_task_message_context call (AGX1-378)

span_derivation.py:
- _on_full: handle Full(ToolRequestContent) by opening a tool span keyed by
  tool_call_id if not already open; adds LangGraph full-event harness support

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

…ting, last-segment, created_at, Full ToolRequest spans

test_auto_send.py:
- Fix test 2: remove bare Start(ToolRequestContent) from events (old behavior was
  that Start did not open a ctx; new behavior does, so test was updated to use
  Full-only events that still verify the two-context behavior)
- Extend _FakeStreaming to record created_at on each context call
- Add test 6: streamed tool_request opens a ctx + routes deltas (AGX1-377 core)
- Add test 7: interleaved indexes route deltas to correct per-index contexts
- Add test 8: multi-step turns return the LAST text segment only
- Add test 9: Full(TextContent) contributes its content to final_text
- Add test 10: created_at is forwarded to every streaming context call (AGX1-378)

test_span_derivation.py:
- Add test_full_tool_request_opens_span: Full(ToolRequestContent) opens a span
- Add test_full_tool_request_and_response_paired: paired Full request+response
  produces a complete OpenSpan+CloseSpan
- Add test_full_tool_request_does_not_double_open: idempotent; a Full for an
  already-open tool_call_id is a no-op

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

…n (AGX1-378)

So migration helpers can restore the deterministic first-message timestamp on
the temporal path. Default None preserves current behavior.

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

…turn

Bug: auto_send_turn passed usage=turn.usage() eagerly, capturing the empty
default before the stream was consumed — so TurnResult.usage came back empty
for every auto_send harness (real turns populate usage during consumption).
Now drives delivery first, then reads usage. Adds a contract-honoring
regression test.

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

_close_all() iterated open contexts and closed them sequentially with no
per-context error guard, so a failure on the first context (e.g. a backend
hiccup during teardown) abandoned every remaining open context and their task
messages were never finalized. Wrap each close() in its own try/except (warn
and continue) so all contexts are attempted regardless of individual failures.

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

These were brainstorming/spec artifacts not meant to ship; per review
feedback they don't belong in the foundation PR. Also de-references the
now-removed docs path in the harness-integration TODO comment.

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

…ntent.is_error

Now that the SDK carries ToolResponseContent.is_error (AGX1-371), thread it
through the harness surface: CloseSpan gains an is_error field, the span
deriver populates it when closing a tool span on a ToolResponseContent, and
the tracer records the status on span.data (Span has no dedicated error
field; None means the harness reported no status, so data is left untouched).

This preserves tool-error tracing fidelity for harnesses that report tool
failures (e.g. golden agent's ToolCompleted.is_error) once they move onto the
unified surface, instead of silently dropping the status.

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