【No.8】Agentic 多协议一致性测试

[AllureCurtain]

Summary

1. Add offline canonical request golden tests for Chat Completions, OpenAI Responses and Anthropic Messages: one semantic conversation is authored natively in each protocol and must project onto an identical messages / tools / chat_template_kwargs triple.
2. Cover text, assistant tool call + tool result, URL and base64 image inputs, tools, thinking enable/disable, assistant reasoning, and system message roles, asserted in three layers: per-protocol golden triple, cross-protocol agreement, and one session state hash.
3. Reject malformed requests — unknown/illegal role, missing or empty tool call id, empty string/list content, invalid image block, reserved template kwargs — asserting the exact message, param and cause type with full field paths such as messages[2].tool_calls[0].id.
4. Tighten relax/agentic/session/state.py where malformed input used to pass validation silently: canonicalize every legal image spelling to one block form (dropping request-only keys such as OpenAI's detail), and require a non-empty tool_calls[].id.

Type of Change

Type	Included	Details
Tests	Yes	Cross-protocol golden tests plus error-path and tolerance pinning.
Bug fix	Yes	Tightens request normalization and validation shared by the three protocol adapters.
Runtime behavior	Limited	No scheduling, lifecycle, training or model execution changes.
Documentation	No
Build or tooling	No	No dependency, CI or build configuration changes.

Test plan

• pytest tests/agentic/test_protocol_golden.py -q — 114 items, CPU only; no network, model service or GPU involved.
• pre-commit run --files on all touched files — every hook passes (ruff, docformatter, gitleaks, whitespace/EOF checks).
• Branch is rebased onto the latest upstream main (0651812).

Conventions

• Test layout follows the existing tests/ per-directory structure (tests/agentic/__init__.py), and golden values are authored inline, consistent with the existing canonicalization tests in tests/test_agentic_rollout.py.
• Reuses check_messages / normalize_tools / normalize_template_kwargs as the task requires; no new dependencies.

Related to #321

[rai-studio-bot]

Nyanpasu 审查看板

审查状态： ✅ 已通过

审查版本： 39bbf7463a8d975b48e3c0d8e704a56e1d3dae84

已完成完整变更审查，未发现需要修改的问题。真实规范化函数的隔离验证 114 项通过；标准 pytest 因缺少 Ray 无法收集，暂无 CI 检查，未运行模型服务或多节点 GPU 集成测试。

没有未解决的审查问题。

^{Powered by Nyanpasu with gpt-6-astra medium, please check the suggestions carefully.}

[rai-studio-bot]

已审查 39bbf746 的完整变更及共享校验的调用路径，未发现需要修改的问题，同意合入。

从当前源码提取真实规范化函数进行隔离验证，114 项测试通过；标准 pytest 因缺少 Ray 在收集阶段停止，因此不视为完整模块测试通过。当前未发现 CI 检查，未运行模型服务或多节点 GPU 集成测试（本环境缺少相应依赖与硬件）。

^{Powered by Nyanpasu with gpt-6-astra medium, please check the suggestions carefully.}

[AllureCurtain]

@SigureMo 任务已完成，请求 review。

[SigureMo]

本赛题已经由 #335 合入并锁定，感谢参与～