chore: release main

[stainless-app]

✨ Stainless prepared a new release

agentex-client: 0.16.0

0.16.0 (2026-06-24)

Full Changelog: agentex-client-v0.15.0...agentex-client-v0.16.0

⚠ BREAKING CHANGES

• harness: consolidate the Pydantic-AI harness + remove tracing handler (#431)
• harness: consolidate the LangGraph harness + remove tracing handler (#430)

Features

• cli: add claude-code init templates (sync / async / temporal) (#435) (fd9bc4a)
• cli: add codex init templates (sync / async / temporal) (#436) (0fadfd7)
• cli: add default-openai-agents init template (async base) (#434) (624e9c8)
• openai-temporal: render hosted/server-side tool calls in TemporalStreamingModel (#442) (5dce9f0)

Bug Fixes

• cli: harden init templates per Greptile feedback (suite-wide) (#444) (2d85eb0)

Refactors

• harness: consolidate the LangGraph harness + remove tracing handler (#430) (a3fb5ad)
• harness: consolidate the Pydantic-AI harness + remove tracing handler (#431) (48c3da8)
• harness: move OpenAI harness into adk/_modules + facade export (#432) (58bdb16)

agentex-sdk: 0.15.0

0.15.0 (2026-06-24)

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

⚠ BREAKING CHANGES

• harness: consolidate the LangGraph harness + remove tracing handler (#430)

Refactors

• harness: consolidate the LangGraph harness + remove tracing handler (#430) (a3fb5ad)

---

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

• Prepares the Agentex client and SDK release with version and changelog updates.
• Consolidates LangGraph, Pydantic-AI, and OpenAI ADK harness modules and turn exports.
• Adds Claude Code, Codex, and OpenAI Agents CLI init templates across sync, async, and Temporal modes.
• Extends Temporal OpenAI streaming support for hosted/server-side tool calls and expands harness tests.

Confidence Score: 3/5

Merge safety is reduced by OpenAI streaming conversion edge cases and a CLI cancellation path that can create files after cancellation.

The changes are broad but the main risks are concentrated in the OpenAI sync converter and init command prompt flow.

src/agentex/lib/adk/_modules/_openai_sync.py; src/agentex/lib/cli/commands/init.py

T-Rex Logs

What T-Rex did

• **Verification: Text reuses index**

• **Result**: Blocked

• Executed a focused direct-converter repro harness with fake OpenAI reasoning and text stream events, but runtime import could not load the reviewed converter file from the repository checkout.

• The captured run output shows the repro stopped before exercising the converter because src/agentex/lib/adk/_modules/_openai_sync.py was missing at execution time.

• **Verification: Reasoning streams hang**

• **Result**: Blocked

• I attempted to run a focused converter harness, but the repository checkout no longer contains src/agentex/lib/adk/_modules/_openai_sync.py, so the changed code could not be executed.

• Installing the ADK package was also blocked because the available interpreter is Python 3.11.6 while agentex-sdk requires Python >=3.12.

• **Verification: Bad args abort**

• **Result**: Blocked

• I created a focused local Python harness with faked OpenAI response classes to exercise both the ResponseFunctionToolCall branch and the generic arguments branch with malformed JSON followed by later tool output and final text.

• Running the harness with PYTHONPATH=src was blocked before converter execution because the reviewed file src/agentex/lib/adk/_modules/_openai_sync.py is missing from the working tree, so the runtime path could not be loaded.

• **Verification: Cancel creates project**

• **Result**: Reproduced

• Executed a focused Python harness against the actual init command with mocked questionary prompts where earlier answers were valid and the package-manager select returned None.

• The run showed the temp directory was empty before init and contained a generated cancel_agent project afterward.

• The captured CLI output showed the package-manager prompt returned None, followed by project structure creation and the success message.

• The generated files included requirements.txt and Dockerfile, proving the cancellation flowed through as the non-uv path.

• `trex-artifacts/release-metadata-01-before.txt` contains the executed base run with command, working directory, full parsed metadata, and exit code 0.

• The required after artifact is missing because the runtime command facility failed before it could be generated.

• Per validation instructions, I am marking this inconclusive rather than substituting static inspection or prose for executed head evidence.

• Successful setup command output showed both requested worktrees were created.

• Initial package-import smoke command failed with `ModuleNotFoundError: No module named 'httpx'`, before exercising changed template code.

• Subsequent attempts to run the adjusted harness and even a trivial `pwd; ls` command returned `Not connected`, preventing executable proof capture in `trex-artifacts`.

• No valid proof artifacts were produced for this request.

• Blocker: shell execution backend disconnected before the focused contract script and pytest commands could complete, preventing executed before/after comparison evidence.

• `trex-artifacts/temporal-hosted-tools-01-before.txt` contains the actual captured base command output.

• Missing required comparable after artifact: command execution became unavailable before the head run could be performed.

• Blocker: `multibash` returned `Not connected` for simple commands including `bash -lc 'echo hi'`, preventing further runtime validation.

_{Ran code and verified through T-Rex}

Comments Outside Diff (1)

1. src/agentex/lib/cli/commands/init.py, line 266-274 (link)

   Cancel creates project

   Earlier prompts return immediately when the user cancels, but this prompt does not. If the user cancels at the package-manager prompt, use_uv becomes None, then flows through as falsy and creates the non-uv template with a success message. Add the same cancel guard before building answers.

   Artifacts

   
   

   Repro: focused init cancellation harness

   • Contains supporting evidence from the run (text/x-python; charset=utf-8).

   Repro: harness output showing project creation after package-manager cancellation

   • Keeps the command output available without making the summary code-heavy.

   

   _{Ran code and verified through T-Rex}

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: src/agentex/lib/cli/commands/init.py
   Line: 266-274
   
   Comment:
   **Cancel creates project**
   
   Earlier prompts return immediately when the user cancels, but this prompt does not. If the user cancels at the package-manager prompt, `use_uv` becomes `None`, then flows through as falsy and creates the non-uv template with a success message. Add the same cancel guard before building `answers`.
   
   How can I resolve this? If you propose a fix, please make it concise.

   

Prompt To Fix All With AI

Fix the following 4 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 4
src/agentex/lib/adk/_modules/_openai_sync.py:299-305
**Text reuses index**

The first text item reuses the current `message_index` unless `seen_tool_output` is true. For a reasoning-model stream, the reasoning delta above creates an index first; when the answer text arrives with a new `item_id`, it is mapped to that same index. This can send two `Start` events with the same index, route text deltas into the still-open reasoning context, or overwrite the reasoning context in `auto_send`'s `ctx_map`, leaving the reasoning message unfinalized and attaching the final answer to the wrong message. Reserve a fresh index for every new text `item_id`, not only after tool output, or close and advance the reasoning item before starting text.

### Issue 2 of 4
src/agentex/lib/adk/_modules/_openai_sync.py:178-184
**Reasoning streams hang**

Reasoning messages are opened with `StreamTaskMessageStart`, but this branch skips the matching `Done`. `UnifiedEmitter.auto_send` only closes contexts on `StreamTaskMessageDone`; otherwise it closes them during final teardown, and `SpanDeriver.flush()` can mark the reasoning span incomplete. When an OpenAI reasoning or summary item is emitted through `OpenAITurn(result=...)`, sync/yield consumers never receive a normal done event for the reasoning message. Emit `StreamTaskMessageDone` when a reasoning content or summary output item completes, since the accumulator already rebuilds `ReasoningContent` from reasoning deltas.

### Issue 3 of 4
src/agentex/lib/adk/_modules/_openai_sync.py:73-77
**Bad args abort**

This `json.loads()` runs on the OpenAI streaming conversion path without a guard. If the Agents SDK surfaces malformed, truncated, or provider-specific raw function arguments, `convert_openai_to_agentex_events` / `OpenAITurn.events` raises and stops the whole turn before later tool output or final text can be delivered. The Temporal streaming path already catches `JSONDecodeError` and falls back to `{}`; this converter should use the same defensive parsing here and in the generic arguments branch, for example by using `{}` or preserving the raw string under `_raw`.

### Issue 4 of 4
src/agentex/lib/cli/commands/init.py:266-274
**Cancel creates project**

Earlier prompts return immediately when the user cancels, but this prompt does not. If the user cancels at the package-manager prompt, `use_uv` becomes `None`, then flows through as falsy and creates the non-uv template with a success message. Add the same cancel guard before building `answers`.

_{Reviews (8): Last reviewed commit: "chore: release main" | Re-trigger Greptile}

Greptile also left 3 inline comments on this PR.

[greptile-apps]

Text reuses index

The first text item reuses the current message_index unless seen_tool_output is true. For a reasoning-model stream, the reasoning delta above creates an index first; when the answer text arrives with a new item_id, it is mapped to that same index. This can send two Start events with the same index, route text deltas into the still-open reasoning context, or overwrite the reasoning context in auto_send's ctx_map, leaving the reasoning message unfinalized and attaching the final answer to the wrong message. Reserve a fresh index for every new text item_id, not only after tool output, or close and advance the reasoning item before starting text.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agentex/lib/adk/_modules/_openai_sync.py
Line: 299-305

Comment:
**Text reuses index**

The first text item reuses the current `message_index` unless `seen_tool_output` is true. For a reasoning-model stream, the reasoning delta above creates an index first; when the answer text arrives with a new `item_id`, it is mapped to that same index. This can send two `Start` events with the same index, route text deltas into the still-open reasoning context, or overwrite the reasoning context in `auto_send`'s `ctx_map`, leaving the reasoning message unfinalized and attaching the final answer to the wrong message. Reserve a fresh index for every new text `item_id`, not only after tool output, or close and advance the reasoning item before starting text.

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

[greptile-apps]

Reasoning streams hang

Reasoning messages are opened with StreamTaskMessageStart, but this branch skips the matching Done. UnifiedEmitter.auto_send only closes contexts on StreamTaskMessageDone; otherwise it closes them during final teardown, and SpanDeriver.flush() can mark the reasoning span incomplete. When an OpenAI reasoning or summary item is emitted through OpenAITurn(result=...), sync/yield consumers never receive a normal done event for the reasoning message. Emit StreamTaskMessageDone when a reasoning content or summary output item completes, since the accumulator already rebuilds ReasoningContent from reasoning deltas.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agentex/lib/adk/_modules/_openai_sync.py
Line: 178-184

Comment:
**Reasoning streams hang**

Reasoning messages are opened with `StreamTaskMessageStart`, but this branch skips the matching `Done`. `UnifiedEmitter.auto_send` only closes contexts on `StreamTaskMessageDone`; otherwise it closes them during final teardown, and `SpanDeriver.flush()` can mark the reasoning span incomplete. When an OpenAI reasoning or summary item is emitted through `OpenAITurn(result=...)`, sync/yield consumers never receive a normal done event for the reasoning message. Emit `StreamTaskMessageDone` when a reasoning content or summary output item completes, since the accumulator already rebuilds `ReasoningContent` from reasoning deltas.

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

[greptile-apps]

Bad args abort

This json.loads() runs on the OpenAI streaming conversion path without a guard. If the Agents SDK surfaces malformed, truncated, or provider-specific raw function arguments, convert_openai_to_agentex_events / OpenAITurn.events raises and stops the whole turn before later tool output or final text can be delivered. The Temporal streaming path already catches JSONDecodeError and falls back to {}; this converter should use the same defensive parsing here and in the generic arguments branch, for example by using {} or preserving the raw string under _raw.

Prompt To Fix With AI

This is a comment left during a code review.
Path: src/agentex/lib/adk/_modules/_openai_sync.py
Line: 73-77

Comment:
**Bad args abort**

This `json.loads()` runs on the OpenAI streaming conversion path without a guard. If the Agents SDK surfaces malformed, truncated, or provider-specific raw function arguments, `convert_openai_to_agentex_events` / `OpenAITurn.events` raises and stops the whole turn before later tool output or final text can be delivered. The Temporal streaming path already catches `JSONDecodeError` and falls back to `{}`; this converter should use the same defensive parsing here and in the generic arguments branch, for example by using `{}` or preserving the raw string under `_raw`.

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