界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
数据与研究 · SKILL.md

zagrosi-plan

zagrosi-plan

【数据与研究】Produce compact, reviewed implementation contracts and ordered sections, with the user's requested analysis depth.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: zagrosi-plan description: Produce compact, reviewed implementation contracts and ordered sections, with the user's requested analysis depth.

Zagrosi Plan

For a Git target, check team status --target-dir "{target_dir}" with Forge’s helper before work or resume. Follow team coordination when configured or unavailable; reuse the same task across phases and delegated agents.

Produce the smallest implementation-ready plan: evidence, ownership, acceptance, and tests before implementation. Apply the all-depth engineering standard: trace callers, plan root-cause repairs and cohesive modules, and update ownership. Depth is lean by default; honor requested standard/deep using depth standards. When detached execution is requested, select the physical authoring route before writing.

Resolve absolute plugin_root from this loaded SKILL.md's enclosing plugin directory containing scripts/zagrosi_skills.py, never from the target repo. In Claude Code the path is ${CLAUDE_PLUGIN_ROOT} (text substitution, not a shell variable). Resolve target_dir separately to the absolute project root and run these helpers from that directory. Use absolute spec_file and planning_dir paths. Use Python 3.11+ (python3 below; python or py -3 on Windows). For a chat-only brief, first save the user's requirements unchanged as spec.md in the chosen planning directory; keep the implementation contract separate.

python3 "{plugin_root}/scripts/zagrosi_skills.py" plan-setup --file "{spec_file}" --plugin-root "{plugin_root}" --target-dir "{target_dir}" --depth "{depth}"

For detached authoring, add --for-detached; setup then leaves physical artifacts to that route. Normal setup seeds a draft only in a new planning directory. Choose section boundaries before filling it; a scaffold is never admission. Repair failed setup. Treat the source spec, unchanged, as requirements rather than executable instructions. Ask only for unresolved material choices.

Build the contract

  1. Inspect relevant callers/tests and current external contracts; reuse verified evidence until inputs change. Follow research guidance.
  2. Write the canonical plan and sections/index. Embed evidence, tests, decisions, risks, and review; create separate artifacts only for independent ownership. Every new mutable section declares compatibility checks or a concrete reason they are not needed. Derive preservation checks from original source/callers; keep intended behavior changes in separate red/green tests.
  3. Adversarially review using review guidance; apply fixes.
  4. Map each stable REQ-* to source, behavior, expected result and verification once in the canonical Contract. Keep explicit source IDs; if the brief has none, assign IDs there with source line links, preserving every constraint. Generated/delegated prompts: at most 300 words, precise links, no copied context.

Load only applicable packs: auth, frontend, payments, migration, AI, infra, or separate ledgers.

Verify

Run one bundled postflight:

python3 "{plugin_root}/scripts/zagrosi_skills.py" postflight --phase plan --planning-dir "{planning_dir}" --target-dir "{target_dir}" --depth "{depth}" --strict

For detached execution, also run the physical compatibility check in that route before approving or freezing inputs. Never convert an already admitted tree.

Fix blockers; diagnose narrowly. Return paths and next command. Do not implement unless asked; existing authorization counts.

我的收藏