fix(codex): enforce turn budgets for reasoning and tool-only responses

[ye1457]

Problem and behavior

Codex counted only nonempty agentMessage items as turns. Reasoning/tool-only responses could therefore exceed --max-turns, while multiple text items from one model response consumed multiple turns. The CLI also lacked the Dashboard's budget interruption.

Count completed model responses from cumulative SDK usage updates, deduplicate repeated updates, and use the same recorder in CLI and Dashboard. Multiple tools or text items within one response count once. Request one interrupt at the limit, preserve an already recorded finish, and retain normal SDK/MCP cleanup. turns_used now reports completed model responses. Document the semantics in both languages.

This is the independent fix for #190 (comment); RoboCasa memory changes remain in #190.

Validation

• pytest tests/unit_tests -v -rs: 591 passed, 3 skipped. Skips require the optional RoboTwin/LeRobot dependencies.
• pre-commit run --all-files: passed.
• English and Chinese Sphinx builds with warnings treated as errors: passed.
• Regression coverage: five reasoning/tool-only cycles with budgets 1 and 2, multiple text/tool items, duplicate/replayed usage, successful finish at the limit, CLI cleanup, and Dashboard interruption/close.
• Combined locally with the RoboCasa changes for four real GPT-5.6/xhigh MCP memory-read checks: all passed within the 10-response budget. Raw traces stay local; no robot actions or benchmark reruns were performed.

[read-the-docs-community]

Documentation build overview

📚 RPent-cn | 🛠️ Build #34612359 | 📁 Comparing 8b2d726 against latest (d6876ca)

  🔍 Preview build  

1 file changed

± rst_source/usage/configure_planner.html

[read-the-docs-community]

Documentation build overview

📚 RPent | 🛠️ Build #34612358 | 📁 Comparing 8b2d726 against latest (d6876ca)

  🔍 Preview build  

1 file changed

± rst_source/usage/configure_planner.html