[DO NOT MERGE] Dev/sdk 0.14.0.dev0

[NiteshDhanpal]

⚠️ DO NOT MERGE — dev artifact. Pre-release version bump to build a throwaway wheel for load testing. Not for release.

What

Bumps adk/pyproject.toml (agentex-sdk) 0.13.2 → 0.14.0.dev0 so we can uv build a clearly-versioned pre-release wheel off next.

Why

next carries two tracing fixes that aren't in any published agentex-sdk (latest is 0.13.2):

• fix(tracing): fail open temporal span activities #437 — Temporal adk.tracing.span START/END activities fail open (a span-activity timeout no longer fails the workflow; emits agentex.tracing.temporal_span_activity.dropped).
• feat(tracing): skip Agentex span-start write by default (end-only ingest) #438 — Agentex span-start write skipped by default (single INSERT on end instead of INSERT+UPDATE).

We need these in the rocket_mock_async_agent load-test image now, ahead of a real 0.14.x.

How it's consumed

The wheel built from this branch is vendored into the agent image — see agentex-agents #1736. Nothing here is published to PyPI; .dev0 sorts below a future real 0.14.0.

Cleanup

Close once 0.14.x is cut through the normal release flow.

Greptile Summary

This is a DO NOT MERGE dev artifact cutting agentex-sdk 0.14.0.dev0 off the next branch to produce a vendored wheel for load testing the rocket_mock_async_agent. It bundles two targeted fixes and a large new unified harness surface.

• PR fix(tracing): fail open temporal span activities #437 (fail-open tracing): TracingModule.start_span / end_span now catch ActivityError and TemporalTimeoutError in Temporal workflows, re-raise cancellations, emit the agentex.tracing.temporal_span_activity.dropped Datadog counter, and return None / the original span instead of failing the workflow. state_machine.py is updated to guard end_span against the newly possible None return.
• PR feat(tracing): skip Agentex span-start write by default (end-only ingest) #438 (single-INSERT tracing): AgentexTracingProcessor gains a skip-start flag (env AGENTEX_TRACING_SKIP_AGENTEX_SPAN_START, default ON) that turns the previous INSERT-on-start + UPDATE-on-end double-write into a single spans.create on end; the decision is captured once at __init__ to prevent mid-span splits.
• Unified harness surface (AGX1-375): Adds span_derivation.py, tracer.py, types.py, auto_send.py, emitter.py, and yield_delivery.py under src/agentex/lib/core/harness/, along with new *Turn adapters for LangGraph, pydantic-ai, Claude Code, and Codex; all re-exported from adk.__init__.

Confidence Score: 4/5

Safe to build from for load-testing purposes; both targeted tracing fixes are structurally sound and the state_machine guard is correct.

The two core fixes are narrow and well-tested. The unified harness surface is new code with good test coverage and explicit docstrings for known limitations. Two minor inefficiencies — a per-call counter allocation in the dropped-metric helper and a noisy init-time logger.info on every processor construction — don't affect correctness but are worth addressing before the real 0.14.x release.

agentex_tracing_processor.py (init-time log noise), tracing.py (counter allocation per drop event). All harness surface files look clean.

Important Files Changed

Filename	Overview
adk/pyproject.toml	Version bump 0.13.2 → 0.14.0.dev0 to produce a clearly-versioned pre-release wheel for load testing.
src/agentex/lib/adk/_modules/tracing.py	Adds fail-open handling for Temporal span activities: ActivityError / TemporalTimeoutError are caught, cancellations re-raised, and a Datadog counter emitted on drop. The span context manager already guards end_span with if span:, so the new None return from start_span is handled correctly.
src/agentex/lib/core/tracing/processors/agentex_tracing_processor.py	Adds AGENTEX_TRACING_SKIP_AGENTEX_SPAN_START env flag (default ON) to collapse the per-span INSERT+UPDATE into a single INSERT on end. The skip decision is correctly captured once at init to prevent mid-span splits. The logger.info on init fires on every new processor instance, which may be noisy in high-churn environments.
src/agentex/lib/core/harness/span_derivation.py	New pure stateful reducer: derives OpenSpan/CloseSpan signals from the canonical StreamTaskMessage* stream; handles both Full-only (LangGraph) and Start+Delta+Done (pydantic-ai, Claude Code) harnesses. flush() closes dangling opens as incomplete on stream end.
src/agentex/lib/core/harness/tracer.py	New adapter: consumes SpanSignals and opens/closes adk.tracing child spans. Duplicate OpenSpan keys silently orphan the prior span (acknowledged in the docstring). All errors are caught and logged; tracing never breaks delivery.
src/agentex/lib/core/harness/auto_send.py	New async delivery path: drives canonical stream to adk.streaming contexts, derives spans as a side effect. The finally block flushes remaining open contexts and span signals even on exception.
src/agentex/lib/core/harness/emitter.py	New UnifiedEmitter facade: ties trace context + delivery mode (yield vs auto-send). usage=turn.usage() is correctly read after auto_send consumes the stream, not before.
src/agentex/lib/sdk/state_machine/state_machine.py	Fixes AttributeError introduced by the fail-open change: initialises span = None before the tracing branch and guards end_span with span is not None.
src/agentex/lib/adk/_modules/_langgraph_async.py	Rewrites stream_langgraph_events on LangGraphTurn+UnifiedEmitter. Hardcodes trace_id=None (tracing disabled), consistent with the original implementation. Public signature is preserved.
src/agentex/lib/core/harness/types.py	New type definitions for the unified harness surface: OpenSpan, CloseSpan, TurnUsage, TurnResult, HarnessTurn Protocol.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    subgraph fix437 ["Fix #437 – Temporal fail-open"]
        A["ActivityHelpers.execute_activity()"] -->|"ActivityError / TimeoutError"| B{"is_cancelled?"}
        B -->|Yes| C["re-raise → fail workflow"]
        B -->|No| D["log warning + emit counter"]
        D -->|start_span| E["return None"]
        D -->|end_span| F["return original span"]
        A -->|Success| G["return Span"]
    end

    subgraph fix438 ["Fix #438 – Single-INSERT tracing"]
        H["on_span_start()"] --> I{"_skip_span_start (default ON)"}
        I -->|True| J["no-op"]
        I -->|False| K["spans.create(start kwargs)"]
        L["on_span_end()"] --> M{"_skip_span_start"}
        M -->|True| N["spans.create(full kwargs)"]
        M -->|False| O["spans.update(end kwargs)"]
    end

    subgraph harness ["Unified harness surface (AGX1-375)"]
        P["HarnessTurn"] --> Q["convert_*_to_agentex_events()"]
        Q --> R["SpanDeriver.observe()"]
        R -->|"OpenSpan / CloseSpan"| S["SpanTracer.handle()"]
        S --> T["adk.tracing.start_span / end_span"]
        Q --> U{"Delivery mode"}
        U -->|"Async / Temporal"| V["auto_send() → adk.streaming"]
        U -->|"Sync HTTP ACP"| W["yield_events() → HTTP yield"]
        V --> X["TurnResult"]
        W --> X
    end

Loading

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    subgraph fix437 ["Fix #437 – Temporal fail-open"]
        A["ActivityHelpers.execute_activity()"] -->|"ActivityError / TimeoutError"| B{"is_cancelled?"}
        B -->|Yes| C["re-raise → fail workflow"]
        B -->|No| D["log warning + emit counter"]
        D -->|start_span| E["return None"]
        D -->|end_span| F["return original span"]
        A -->|Success| G["return Span"]
    end

    subgraph fix438 ["Fix #438 – Single-INSERT tracing"]
        H["on_span_start()"] --> I{"_skip_span_start (default ON)"}
        I -->|True| J["no-op"]
        I -->|False| K["spans.create(start kwargs)"]
        L["on_span_end()"] --> M{"_skip_span_start"}
        M -->|True| N["spans.create(full kwargs)"]
        M -->|False| O["spans.update(end kwargs)"]
    end

    subgraph harness ["Unified harness surface (AGX1-375)"]
        P["HarnessTurn"] --> Q["convert_*_to_agentex_events()"]
        Q --> R["SpanDeriver.observe()"]
        R -->|"OpenSpan / CloseSpan"| S["SpanTracer.handle()"]
        S --> T["adk.tracing.start_span / end_span"]
        Q --> U{"Delivery mode"}
        U -->|"Async / Temporal"| V["auto_send() → adk.streaming"]
        U -->|"Sync HTTP ACP"| W["yield_events() → HTTP yield"]
        V --> X["TurnResult"]
        W --> X
    end

Loading

Prompt To Fix All With AI

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
src/agentex/lib/core/tracing/processors/agentex_tracing_processor.py:157-165
**Init-time logger.info fires on every new processor instance**

Both `AgentexSyncTracingProcessor.__init__` and `AgentexAsyncTracingProcessor.__init__` call `logger.info(...)` to announce the skip-start decision. In practice the async processor is re-constructed per event loop (the existing `_build_client` / `WeakKeyDictionary` pattern), so in a high-throughput worker this log line can fire thousands of times. Consider using `logger.debug` or logging once at module load time via a module-level call to `_skip_span_start_enabled()`.

### Issue 2 of 2
src/agentex/lib/adk/_modules/tracing.py:34-42
**Counter object recreated on every drop event**

`workflow.metric_meter().create_counter(...)` is called on every invocation rather than reusing a cached counter. Temporal's SDK recommends creating metric objects once and reusing them — each `create_counter` call allocates a new object and may incur SDK-internal bookkeeping. A simple module-level or class-level cache (e.g. using `functools.lru_cache` or a `_counter` sentinel on the function) would avoid the repeated allocation on busy workflows.

_{Reviews (1): Last reviewed commit: "chore(sdk): bump agentex-sdk to 0.14.0.d..." | Re-trigger Greptile}

[github-actions]

This PR is targeting main, but PRs should target the next branch by default.

The main branch is reserved for release-please and Stainless automation. To resolve, pick one of:

• Re-target the PR to next (recommended). On the PR page, click Edit next to the title and change the base branch to next.
• Add the target-main label if this is an intentional exception (e.g. an urgent hotfix). The check will re-run and pass.

See CONTRIBUTING.md for the full branch model.