trellis-skills

Create a Trellis MVP task tree from a full requirements document. Use when Codex receives a product brief, requirements document, PRD, zero-to-MVP request, large capability-building request, or partially implemented project and must do read-only analysis, assign stable REQ/AC IDs, build full requirement and MVP coverage matrices, lock project contracts, identify existing implementation evidence, use progressive batch and sub-agent planning to form a complete MVP parent/child task tree, draft PRD/design/implementation artifacts, and create Trellis tasks only after user confirmation; do not write application code.

从完整需求文档创建 Trellis MVP 任务树。用于 Codex 收到产品说明、需求文档、PRD、从 0 到 MVP 的新项目需求、大型能力建设请求，或已经手工实现了一部分功能的项目时，先做只读分析，分配稳定 REQ/AC 编号，生成需求追踪矩阵和项目契约，识别已有实现证据，通过渐进式批次规划和子代理协作形成完整 MVP 父子任务树、PRD、设计/实现产物计划，并在用户确认后创建 Trellis 任务；不编写应用代码。

Audit an existing MVP against a source requirements document and run a sustainable Trellis delivery loop. Use when Codex is asked to continue after an MVP, complete remaining requirements, perform a first full gap audit, maintain a Requirements Traceability Matrix, classify DONE/PARTIAL/MISSING/UNTESTED/UNCLEAR items, initialize or update .trellis/delivery-state.md, choose full/delta/early-exit audit scope, plan bounded gap-closing batches, require worktree + verifier gates, design automated test coverage, classify discovered bugs, or run final acceptance before delivery.

根据源需求文档审计已有 MVP，并执行 Trellis 完整交付 loop 的审计、状态维护、批次规划和最终验收。用于 Codex 被要求在 MVP 后继续、补齐剩余需求、执行 full/delta 差距审计、维护 Requirements Traceability Matrix、分类 DONE/PARTIAL/MISSING/UNTESTED/UNCLEAR 条目、初始化或更新 .trellis/delivery-state.md、规划有边界的补缺批次、要求 worktree + verifier 门、设计自动化测试覆盖、分类发现的 bug，或在交付前执行最终验收的场景。若由 trellis-delivery-controller-zh 调用，则遵循控制器已选定的 loop 模式和审计范围；若单独调用，则按本技能的 delivery-loop-policy 自行判定。

Trellis requirements-to-delivery entrypoint and workflow controller. Use when the user wants to invoke one Trellis skill that detects the current stage and routes work across requirements analysis, MVP task planning, post-MVP gap audit, delivery loop, TDD implementation, systematic debugging, two-stage review, and final acceptance. Coordinates trellis-zero-to-mvp, trellis-mvp-to-delivery, trellis-implement-tdd, trellis-debug-systematic, and trellis-review-twostage while enforcing safety gates.

Trellis 需求交付总入口和流程控制器。用于用户希望只调用一个 Trellis skill 自动判断当前阶段，并在需求分析、MVP 任务规划、MVP 后差距审计、交付 loop、TDD 实现、系统调试、双阶段评审和最终验收之间进行路由、推进和安全门控制的场景。适合基于 trellis-zero-to-mvp-zh、trellis-mvp-to-delivery-zh、trellis-implement-tdd-zh、trellis-debug-systematic-zh 和 trellis-review-twostage-zh 形成端到端交付闭环。

在 Trellis 子任务进入实现阶段时，用严格 TDD（红→绿→重构）逐条把验收标准落成代码。用于 Codex / Claude Code 等在 trellis-zero-to-mvp-zh 或 trellis-mvp-to-delivery-zh 生成主子任务后，按依赖顺序实现某个子任务，尤其当执行模型是 qwen3.6 35b 这类小参数本地模型时——把"理解并实现需求"转成"让这条失败测试变绿"的机械循环，每步都有客观信号。

在 Trellis 子任务自检全绿、标记完成之前，做两阶段代码评审：先机械核对规范符合（对照 PRD 验收标准/决策表/禁止事项/文件清单/挂载点），再做代码质量评审（对照 .trellis/spec/guides 的工程规范与设计纪律）。用于 Codex / Claude Code 等团队协作场景，尤其当实现由 qwen3.6 35b 小模型完成时——小模型评审自己的代码不可靠，Stage 2 应升级到强模型；critical 问题阻断完成。

Extract reusable engineering capabilities from external GitHub skill projects and produce safe Trellis absorption plans. Use when the user provides a GitHub repository URL, skill project URL, open-source agent skill collection, or asks to compare or update projects such as superpowers, gsd, CodeStable, loop-engineering, ponytail, or similar. Analyze README/SKILL.md/references/scripts/license and produce source analysis, capability extraction, absorption card, Trellis target mapping, risk judgment, update-scan delta report, and verification scenario recommendations. Do not copy third-party skill text directly; absorb capability patterns, workflow structure, and verifiable design only.

从外部 GitHub skill 项目中提取可复用工程能力，并按 Trellis skills 架构进行安全吸收规划。用于用户提供 GitHub 仓库地址、skill 项目地址、开源 agent skill 集合、或要求比较 superpowers、gsd、CodeStable、loop-engineering、ponytail 等项目更新时，分析 README/SKILL.md/references/scripts/license，生成 source analysis、capability extraction、absorption card、Trellis 落点、风险判断、update-scan delta 报告和验证场景建议。不要直接复制第三方 skill 原文；只吸收能力模式、流程结构和可验证设计。

When a test should be green but stays red, or a self-check command fails during Trellis subtask implementation/verification, use this non-deviable rigid script for systematic root-cause pinpointing and minimal fix. Used when Codex / Claude Code drives a small-parameter model like qwen3.6 35b—small models quickly spiral out of control when debugging freely; this skill uses "one change at a time, must re-run after change, no guessing" iron rules to lock it on a narrow path with defensive regression checks.

当 Trellis 子任务实现/验证过程中测试该绿不绿、或自检命令失败时，用一段不可偏离的刚性脚本做系统化根因定位与最小修复。用于 Codex / Claude Code 等驱动 qwen3.6 35b 这类小参数模型调试时——小模型一旦自由发挥就会到处乱改，本技能用"一次只改一处、改完必重跑、禁止猜"的铁律把它锁在窄路径上，并带防御性回归检查。

Implement a Trellis subtask using strict TDD (RED→GREEN→REFACTOR), turning acceptance criteria into code one by one. Used when Codex / Claude Code drives a small-parameter model like qwen3.6 35b—transforms "understand and implement requirements" into a mechanical "make this failing test green" loop with objective signals at every step.

Before marking a Trellis subtask complete after self-check all green, conduct two-stage code review: mechanically check spec compliance (against PRD acceptance criteria/decision table/forbidden/file manifest/mount points), then code quality review (against .trellis/spec/guides engineering standards and design discipline). Used in Codex / Claude Code team collaboration scenarios, especially when implementation is by a small-parameter model like qwen3.6 35b—small models reviewing their own code is unreliable, Stage 2 should escalate to strong model; critical issues block completion.