Project Rules (auto-synced from codex_standard)

红线 G19；完整手册：docs/_shared/CODEX_PERMISSION_APPROVAL_RUNBOOK.md §10。 当内嵌补丁在读取首个目标前以 helper_unknown_error: setup refresh had errors 失败时：

1. MUST 读取 .codex/.sandbox/setup_error.json 与当日 sandbox 日志并分类；不得把 helper 初始化失败归因于仓库或补丁。
2. 若日志含 write ACE grant failed / SetNamedSecurityInfoW failed: 5，MUST 先使用专用 patch-only 模式完成当前精确变更：codex.ps1 --codex-run-as-apply-patch <complete-utf8-patch>。
3. MUST 在每个补丁后执行目标 diff 与 diff check。
4. MUST NOT 使用任意 shell 写文件、关闭 sandbox、扩大 writable root，或把退出、重启、重开任务作为默认恢复方案。
5. 持久 ACL 修改 MUST 获得用户对精确根目录的显式批准；SID 必须从当前已知良好 writable root 动态识别，禁止硬编码。
6. patch-only 成功只代表当前落盘恢复；只有 Desktop 内嵌补丁复验成功，才可宣称 helper 永久恢复。

Project Rules (auto-synced from codex_standard)

verification

Fixed Test Execution Protocol

Universal protocol for turning "we tested it" into a repeatable, reviewable method.

Iron Law

Do not invent the test method mid-task. Map the change to a fixed path, then execute that path.

Required Order

Every non-trivial verification MUST follow this order:

1. MAP
   • Identify the affected module, contract, and user path.
   • Map them to the concrete test files or validation commands.
   • If no mapping exists, create or update the project test-scenario index first.
2. STATIC
   • Run the smallest static proof first.
   • Examples: import check, lint, source assertion, route/config assertion, schema/contract check.
   • If static evidence already disproves correctness, stop and fix before broader tests.
3. LAYERED
   • Run only the smallest sufficient tests in this order:
     1. Unit / logic
     2. Integration / API / orchestration
     3. E2E / page / browser
   • Expand scope only when the previous layer passes but does not yet cover the risk.
4. RUNTIME
   • Collect runtime evidence tied to the real failure mode.
   • Health checks, page entry checks, or smoke endpoints are supporting evidence only, not final proof by themselves.

Acceptance Record

A verification is valid only if it records all four:

1. Change path — what changed
2. Test path — which test file or command covers it
3. Risk alignment — what failure mode it proves or disproves
4. Result — pass/fail with key output summary If any of the four are missing, the verification is incomplete.

Forbidden Moves

1. Treating HTTP 200 as full verification
2. Running broad suites before mapping the affected path
3. Mixing unrelated full-suite results into a focused conclusion
4. Using a single giant command when a smaller layered proof exists
5. Verifying frontend-only changes with backend-only evidence, or the reverse

Frontend Rule

For frontend-affecting work, prefer:

1. lint / type / route assertion
2. targeted browser test file or grep subset
3. concrete page entry proof
4. backend contract proof only if the UI depends on it

Backend Rule

For backend logic or orchestration work, prefer:

1. import / source assertion
2. targeted unit subset
3. targeted integration or API subset
4. health/debug/runtime proof
5. browser proof only if user-visible behavior changed

Project Index

Each project SHOULD keep a TEST_SCENARIOS.md (or equivalent) that maps:

• test file -> business path
• command -> scope
• when to run -> failure mode Without this index, test selection becomes improvisation.

§6 Test Agent Isolation

红线 G17 (见 RED_LINES_CATALOG.md) Full rules: see .codex/protocols/ directory

Grand Orchestrator Self-Check

• Sync Source: X:\workspace\AIStandard\codex_standard
• Check Command: pwsh X:\workspace\AIStandard\codex_standard\generators\sync-engine.ps1 -ProjectPath . -Status
• Sync Command: pwsh X:\workspace\AIStandard\codex_standard\generators\sync-engine.ps1 -ProjectPath . -Write -AutoApply
• Rule: Before starting any task, run -Status to check if sync is needed. If out-of-date, suggest user run sync.

AI 模型路由治理（workspace 统一 v1）

激活标志：本项目 .ai/PROJECT_ROUTING.yaml；权威规范：X:\workspace\AIStandard\codex_standard\universal\ai_model_routing.md（数据：同目录 routing\*.yaml）。本节为执行要点，完整矩阵/评分/降级链见权威规范。

每次任务第一个动作：输出 ROUTE 行，并追加一条 JSON 到 .codex/runs/routing/<YYYYMMDD>.ndjson（目录不存在则先创建）：

ROUTE: model=<别名> reasoning=<low|medium|high> risk=R<n> complexity=C<n> context=X<n> parallel=P<n> agents=<k> gate=<-|rev|MMPI>

角色别名（物理模型见 model-aliases.yaml，换代只改数据不改规则）：MODEL_EXPLORER（luna/low，搜索提取时间线）/ MODEL_ANALYST（terra/low，多文件阅读小改）/ MODEL_ENGINEER（sol/medium，开发重构）/ MODEL_REVIEWER（sol/high，审查与 R3 独立 gate）/ MODEL_ARCHITECT（astra，架构与 R4 最终 gate）。

执行铁律：

1. Risk Floor 优先于 Complexity；R4（本项目高危项见 overlay）最终结论/动作必须 MODEL_ARCHITECT（Astra）或人工，缺位挂起不降级。
2. ROUTE_MISMATCH 拦截：当前会话模型低于矩阵要求 → 只许取证（读码/搜索/整理），禁止宣布最终结论或执行最终动作，必须提示切换模型或人工裁决。
3. R3/R4 触发 MMPI（暂停 AI 生成 → 跨模型独立审查 → 人工解锁）；R2+ 结论必须附可回放证据（path:line / cmd#hash），证据不足不得宣布完成。
4. 升级纪律：同级 2 次无新证据 → 升一级并把 prompt 重构为 Task Packet；连续 2 次升级无进展 → 熔断停止转人工，禁止烧额度原样重试。
5. 大上下文先压缩（定位 → 精确 read → diff → packet）再给强模型；禁止全量日志/全量测试输出进上下文。
6. 用户 override 服从，但第 1/2/3 条安全规则不可被成本偏好覆盖。