chore: release main

[stainless-app]

✨ Stainless prepared a new release

agentex-client: 0.15.0

0.15.0 (2026-06-23)

Full Changelog: agentex-client-v0.14.0...agentex-client-v0.15.0

Features

• claude-code: stream-json parser tap for the unified harness surface (#420) (904339c)
• codex: event-stream parser tap for the unified harness surface (#421) (9b2b031)
• harness: public adk facade + docs for the unified harness surface (PR 9) (#423) (fa60632)
• harness: unified harness surface — foundation (span derivation, delivery adapters, emitter) (#412) (a9cacf4)
• langgraph: migrate LangGraph harness onto unified surface (#417) (d344228)
• openai-agents: migrate onto the unified harness surface (#416) (d10e151)
• pydantic-ai: migrate onto unified harness surface (PR4) (#415) (5ec62c2)
• streaming: stream tool call argument deltas in TemporalStreamingModel (#355) (c8de1d4)

Bug Fixes

• harness: assert cross-channel (yield vs auto-send) conformance equivalence [AGX1-373] (#414) (694960f)

agentex-sdk: 0.14.0

0.14.0 (2026-06-23)

Full Changelog: agentex-sdk-v0.13.2...agentex-sdk-v0.14.0

Features

• harness: public adk facade + docs for the unified harness surface (PR 9) (#423) (fa60632)

---

This pull request is managed by Stainless's GitHub App.

The semver version number is based on included commit messages. Alternatively, you can manually set the version number in the title of this pull request.

For a better experience, it is recommended to use either rebase-merge or squash-merge when merging this pull request.

🔗 Stainless website
📚 Read the docs
🙋 Reach out for help or questions

Greptile Summary

This is the 0.15.0 / 0.14.0 release PR, cutting changes already merged to main. The core addition is a unified harness surface (src/agentex/lib/core/harness/) that gives every agent harness a shared path for streaming delivery, span derivation, and turn-usage reporting via UnifiedEmitter, SpanDeriver, and SpanTracer. On top of that foundation, two new tap modules are introduced for parsing the Claude Code stream-json format and the Codex exec --json event stream, and the OpenAI Agents SDK is adapted via a new OpenAITurn wrapper. The TemporalStreamingModel is extended to stream tool-call argument deltas incrementally instead of buffering them silently.

• New harness core: UnifiedEmitter ties a canonical StreamTaskMessage* stream to either yield (sync HTTP ACP) or auto-send (async/Temporal) delivery, with span derivation as a side effect.
• New taps: convert_claude_code_to_agentex_events and convert_codex_to_agentex_events are added and exported from agentex.lib.adk, along with ClaudeCodeTurn and CodexTurn implementing the HarnessTurn protocol.
• Temporal streaming: TemporalStreamingModel now opens a streaming_task_message_context per function call and streams ToolRequestDelta updates; a defensive cleanup loop closes any orphaned contexts on truncated streams.

Confidence Score: 5/5

Safe to merge — all changes are additive new harness infrastructure and tap implementations built on already-merged commits; no existing call sites are broken.

The new harness surface introduces well-isolated, well-tested modules. The only issues found are documentation inconsistencies that do not affect runtime behavior.

adk/docs/harness.md and src/agentex/lib/adk/_modules/_codex_sync.py both have stale documentation that should be updated.

Important Files Changed

Filename	Overview
adk/docs/harness.md	New harness surface docs; the tap table and the sentence about AGX1-420/421 are already stale — claude-code and codex taps ship in this very release.
src/agentex/lib/adk/_modules/_codex_sync.py	New codex event-stream parser tap; module-level canonical-mapping table for agent_message is inconsistent with the actual implementation (missing StreamTaskMessageStart).
src/agentex/lib/adk/_modules/_claude_code_sync.py	New claude-code stream-json parser tap; well-documented dedup logic for thinking blocks, per-block streaming state correctly reset at message boundaries.
src/agentex/lib/core/harness/emitter.py	New UnifiedEmitter facade: cleanly separates yield vs auto-send delivery, correctly defers turn.usage() read until after stream exhaustion.
src/agentex/lib/core/harness/auto_send.py	Canonical stream to adk.streaming side-effect delivery; index-keyed context map, guarded cleanup, _close_all teardown — logic looks correct.
src/agentex/lib/core/harness/span_derivation.py	Pure stateful reducer from stream events to span open/close signals; correctly handles Start+Done (streaming) and Full (batch) tool-request patterns.
src/agentex/lib/adk/providers/_modules/openai_turn.py	New OpenAITurn implementing HarnessTurn; usage is correctly aggregated from raw_responses only after stream exhaustion.
src/agentex/lib/core/temporal/plugins/openai_agents/models/temporal_streaming_model.py	Adds per-function-call streaming context + ToolRequestDelta delivery; defensive orphan-close loop in finally path; existing stream_update(Full) pattern for reasoning already in use.
tests/lib/core/harness/conformance/runner.py	New cross-channel conformance engine; cleanly compares yield vs auto_send logical deliveries and span signals per fixture.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["Harness Tap\n(pydantic-ai / LangGraph /\nOpenAI / claude-code / codex)"] -->|"canonical StreamTaskMessage* stream"| B["HarnessTurn\n(events + usage())"]
    B --> C["UnifiedEmitter"]
    C -->|"yield_turn()"| D["yield_events()\n(passthrough + tracing)"]
    C -->|"auto_send_turn()"| E["auto_send()\n(adk.streaming contexts + tracing)"]
    D --> F["Caller yields events\nover HTTP response"]
    E --> G["Redis / Task stream\n(adk.streaming)"]
    D --> H["SpanDeriver to SpanTracer"]
    E --> H
    H --> I["adk.tracing\n(span open/close)"]

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
    A["Harness Tap\n(pydantic-ai / LangGraph /\nOpenAI / claude-code / codex)"] -->|"canonical StreamTaskMessage* stream"| B["HarnessTurn\n(events + usage())"]
    B --> C["UnifiedEmitter"]
    C -->|"yield_turn()"| D["yield_events()\n(passthrough + tracing)"]
    C -->|"auto_send_turn()"| E["auto_send()\n(adk.streaming contexts + tracing)"]
    D --> F["Caller yields events\nover HTTP response"]
    E --> G["Redis / Task stream\n(adk.streaming)"]
    D --> H["SpanDeriver to SpanTracer"]
    E --> H
    H --> I["adk.tracing\n(span open/close)"]

Loading

_{Reviews (4): Last reviewed commit: "chore: release main" | Re-trigger Greptile}

[greptile-apps]

ReasoningSummaryDelta not accumulated in payload comparison — both _yield_logical_deliveries (here) and _auto_send_logical_deliveries (line 394) handle TextDelta and ReasoningContentDelta but silently drop ReasoningSummaryDelta events. A fixture that streams reasoning summaries as deltas would have both channels produce an identical (truncated) payload, so a regression where one channel drops summary deltas would pass the cross-channel assertion undetected.

Prompt To Fix With AI

This is a comment left during a code review.
Path: tests/lib/core/harness/conformance/runner.py
Line: 174-179

Comment:
`ReasoningSummaryDelta` not accumulated in payload comparison — both `_yield_logical_deliveries` (here) and `_auto_send_logical_deliveries` (line 394) handle `TextDelta` and `ReasoningContentDelta` but silently drop `ReasoningSummaryDelta` events. A fixture that streams reasoning summaries as deltas would have both channels produce an identical (truncated) payload, so a regression where one channel drops summary deltas would pass the cross-channel assertion undetected.

How can I resolve this? If you propose a fix, please make it concise.