feat(cli): add claude-code init templates (sync / async / temporal)

[declan-scale]

Summary

Ninth slice of #425. Adds the claude-code agentex init templates across all three tiers.

• New default-claude-code, sync-claude-code, temporal-claude-code template dirs.
• Wires DEFAULT_CLAUDE_CODE, SYNC_CLAUDE_CODE, TEMPORAL_CLAUDE_CODE into init.py (enum, project-files map, prompts).
• Scaffolded code imports ClaudeCodeTurn from the agentex.lib.adk facade.

Test plan

• pytest tests/lib/cli/ -k "init or template" — 23 passed (all render to valid Python)

Notes

Stacked on #434. Retarget to next after the chain merges.

🤖 Generated with Claude Code

Greptile Summary

• Adds Claude Code agentex init templates for default, sync, and Temporal project tiers.
• Wires the new Claude Code template choices into CLI init prompts and project-file generation.
• Expands ADK/harness unification and test coverage across OpenAI, Claude Code, Codex, LangGraph, and Pydantic AI integrations.

Confidence Score: 4/5

Merge should wait for fixes to the OpenAI sync stream event lifecycle and the generated Claude Code Dockerfile.

The changed areas are covered by focused checks, and the remaining concerns are localized to stream event ordering and template rendering.

src/agentex/lib/adk/_modules/_openai_sync.py and src/agentex/lib/cli/templates/default-claude-code/Dockerfile.j2

T-Rex Logs

What T-Rex did

• Ran a focused pytest harness against the real sync OpenAI converter with mocked OpenAI SDK events for a reasoning delta followed by an output item done, and observed that the reasoning start event at index 1 did not have a corresponding done event.
• Ran a focused pytest harness against the real _openai_sync.py converter with a mocked OpenAI stream containing a reasoning delta followed by a normal answer text delta, and confirmed that the reasoning and text start indices reused index 1.
• Rendered the default-claude-code Dockerfile.j2 template into a temporary project and validated that the next Dockerfile command would fail parsing due to a broken line continuation in the template.
• Checked the Claude Code scaffold process and confirmed that after scaffolding, template directories and Jinja files exist, init wiring passes, and create_project_structure runs for all three tiers with the generated Claude Code sources importing the facade.

_{Ran code and verified through T-Rex}

Comments Outside Diff (3)

1. General comment

   Claude Code scaffolds omit expected root files

   • Bug
     • The new Claude Code templates are selectable and create projects, but the generated structures do not match the requested contract. default-claude-code and sync-claude-code generate project/acp.py plus root files such as manifest.yaml, README.md, pyproject.toml, Dockerfile, etc., but no agentex.json. temporal-claude-code generates project/acp.py, project/activities.py, project/workflow.py, and project/run_worker.py, but the expected contract requires run_worker.py at the project root and also expects agentex.json.
   • Cause
     • create_project_structure writes every file listed in project_files into code_dir (project/), so temporal run_worker.py is placed under project/run_worker.py. The root template map also renders manifest.yaml but does not render an agentex.json root file for these templates.
   • Fix
     • Update the Claude Code scaffold mapping/rendering so temporal-claude-code emits run_worker.py at the scaffold root when that is the intended contract, and add or map the expected agentex.json root template for the new Claude Code templates. If manifest.yaml is the intended replacement for agentex.json, update the contract/tests accordingly.

   _{Ran code and verified through T-Rex}

2. src/agentex/lib/adk/_modules/_openai_sync.py, line 178-184 (link)

   Reasoning never closes

   Reasoning items are opened with StreamTaskMessageStart, but this completion branch skips the matching StreamTaskMessageDone. A reasoning stream that emits deltas and then completes leaves sync ACP clients with an unterminated reasoning message, and tracing has to mark the reasoning span incomplete during final cleanup instead of closing it normally.

   Artifacts

   
   

   Repro: focused pytest harness for mocked sync OpenAI reasoning stream

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

   Repro: verbose pytest output showing missing reasoning done event

   • 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/adk/_modules/_openai_sync.py
   Line: 178-184
   
   Comment:
   **Reasoning never closes**
   
   Reasoning items are opened with `StreamTaskMessageStart`, but this completion branch skips the matching `StreamTaskMessageDone`. A reasoning stream that emits deltas and then completes leaves sync ACP clients with an unterminated reasoning message, and tracing has to mark the reasoning span incomplete during final cleanup instead of closing it normally.
   
   How can I resolve this? If you propose a fix, please make it concise.

   

3. src/agentex/lib/adk/_modules/_openai_sync.py, line 295-306 (link)

   Text reuses reasoning index

   After a reasoning item, the next text item can reuse the reasoning message index. The reasoning handlers increment message_index, but this branch only increments text indices after tool output, so a normal reasoning-then-answer stream can emit the final text Start with the same index as the reasoning message. Downstream consumers key open contexts by index, so the answer can overwrite or alias the reasoning context and mix text output into the reasoning message.

   Artifacts

   
   

   Repro: focused pytest harness for reasoning then text index reuse

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

   Repro: verbose pytest output showing duplicate reasoning and text start indices

   • 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/adk/_modules/_openai_sync.py
   Line: 295-306
   
   Comment:
   **Text reuses reasoning index**
   
   After a reasoning item, the next text item can reuse the reasoning message index. The reasoning handlers increment `message_index`, but this branch only increments text indices after tool output, so a normal reasoning-then-answer stream can emit the final text `Start` with the same index as the reasoning message. Downstream consumers key open contexts by `index`, so the answer can overwrite or alias the reasoning context and mix text output into the reasoning message.
   
   How can I resolve this? If you propose a fix, please make it concise.

   

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/adk/_modules/_openai_sync.py:178-184
**Reasoning never closes**

Reasoning items are opened with `StreamTaskMessageStart`, but this completion branch skips the matching `StreamTaskMessageDone`. A reasoning stream that emits deltas and then completes leaves sync ACP clients with an unterminated reasoning message, and tracing has to mark the reasoning span incomplete during final cleanup instead of closing it normally.

### Issue 2 of 2
src/agentex/lib/adk/_modules/_openai_sync.py:295-306
**Text reuses reasoning index**

After a reasoning item, the next text item can reuse the reasoning message index. The reasoning handlers increment `message_index`, but this branch only increments text indices after tool output, so a normal reasoning-then-answer stream can emit the final text `Start` with the same index as the reasoning message. Downstream consumers key open contexts by `index`, so the answer can overwrite or alias the reasoning context and mix text output into the reasoning message.

_{Reviews (16): Last reviewed commit: "fix(cli): drain claude stdin and stop en..." | Re-trigger Greptile}

[declan-scale]

@greptile review

[declan-scale]

@greptile review

[greptile-apps]

Missing uv lockfile

agentex init renders pyproject.toml for use_uv=True, but it does not create a uv.lock. A new Claude template project built immediately after init therefore fails during Docker build at this COPY ... uv.lock step before uv sync can run. Please either generate the lockfile during init, remove uv.lock from the COPY, or make the Dockerfile support a fresh project without a lockfile.

Artifacts

Repro: script that renders the uv project and checks the Docker COPY sources

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

Repro: command transcript showing no uv.lock and emulated Docker COPY failure

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

Repro: template availability check used during setup

• 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/templates/default-claude-code/Dockerfile-uv.j2
Line: 30-34

Comment:
**Missing uv lockfile**

`agentex init` renders `pyproject.toml` for `use_uv=True`, but it does not create a `uv.lock`. A new Claude template project built immediately after init therefore fails during Docker build at this `COPY ... uv.lock` step before `uv sync` can run. Please either generate the lockfile during init, remove `uv.lock` from the `COPY`, or make the Dockerfile support a fresh project without a lockfile.

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

[declan-scale]

@greptile review

[declan-scale]

@greptile review

[declan-scale]

@greptile review

[declan-scale]

@greptile review

[greptile-apps]

Non-text events crash

SendEventParams.event.content is a TaskMessageContent union, not always text. Valid ACP payloads such as data or tool messages do not have a .content field, so sending one of those events to a generated Claude template raises AttributeError before the user message is recorded or any task error is streamed. Convert or reject non-text content before reading .content; the same assumption is present in the sync and temporal Claude handlers.

Artifacts

Repro: focused generated-template handler invocation script

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

Stack trace captured during the T-Rex run

• Keeps the raw stack trace 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/templates/default-claude-code/project/acp.py.j2
Line: 133-134

Comment:
**Non-text events crash**

`SendEventParams.event.content` is a `TaskMessageContent` union, not always text. Valid ACP payloads such as data or tool messages do not have a `.content` field, so sending one of those events to a generated Claude template raises `AttributeError` before the user message is recorded or any task error is streamed. Convert or reject non-text content before reading `.content`; the same assumption is present in the sync and temporal Claude handlers.

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

[greptile-apps]

Manifest can break

description is user-supplied text from the init prompt, but this template renders it as an unquoted YAML scalar. A description containing YAML-significant text such as Answers questions: with Claude or Claude # helper can make the generated manifest.yaml invalid or truncate the value, so the initialized agent may not load. The same unquoted-description issue exists in sync-claude-code/manifest.yaml.j2; quote the rendered description in both templates.

Artifacts

Repro: focused manifest render and YAML parse script

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

Repro: failing render and YAML parse output

• 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/templates/default-claude-code/manifest.yaml.j2
Line: 68

Comment:
**Manifest can break**

`description` is user-supplied text from the init prompt, but this template renders it as an unquoted YAML scalar. A description containing YAML-significant text such as `Answers questions: with Claude` or `Claude # helper` can make the generated `manifest.yaml` invalid or truncate the value, so the initialized agent may not load. The same unquoted-description issue exists in `sync-claude-code/manifest.yaml.j2`; quote the rendered description in both templates.

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