本文へ移動
cccskills
無料GitHub で公開

vc-plan-discovery

Discover related plans for the current task: same feature folder full depth, other features active-only, general-plans active. Like vc-context-discovery but for plan artifacts.

インストール方法を見る

含まれるファイル(5)

  • SKILL.md3.8 KB
  • scripts/discover-plans.mjs5.9 KB
  • scripts/fixtures/validate-plan-discovery-output/fail.md807 B
  • scripts/fixtures/validate-plan-discovery-output/pass.md917 B
  • scripts/validate-plan-discovery-output.mjs2.2 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

vc-plan-discovery

Output style: Follow process/development-protocols/communication-standards.md — answer-first, plain language, no unexplained jargon, TL;DR on long responses.

Invocation

Primary method — run the auto-discovery script. It applies the Scope Rules below (same-feature full depth; other features active-only; general-plans active always) and extracts ONLY the leading YAML frontmatter block of each .md file (no whole-file reads):

node .claude/skills/vc-plan-discovery/scripts/discover-plans.mjs [--feature <name>] [--json]

It groups output into Active Plans / Backlog / Completed / Reports / References (each line: - [name]: description (path)) and prints a trailing Found N active, N backlog, N completed, N reports, N references line. It never throws on a missing root and exits 0 unless given a bad flag. Use --json for a machine-readable object. Prefer this over the manual scan below — it is deterministic and avoids loading huge files into context.

Per task-folder artefact colocation, the script scans one level into each {slug}_{date}/ task folder and surfaces co-located _PLAN_/_REPORT_/_REF_ artefacts; the sibling reports//references/ dirs are deprecated and only hold legacy artefacts.

The manual scan in Scope Rules below is the FALLBACK for when the script fails.

Purpose

Surface all plans relevant to the current task so agents have full plan context — what was tried, what is in progress, what is deferred, and what reports and references exist — before doing any phase work.

Scope Rules

  • Same feature folder (from task context or argument): read ALL of active/, backlog/, completed/, plus any legacy sibling reports/, references/ — surface every file with frontmatter
  • Other feature folders: read active/ only — surface plans whose frontmatter description or feature field matches the task domain
  • general-plans/active/: always scan
  • general-plans/completed/, legacy reports/, references/: scan only when same-feature folder is not identified

Per task-folder artefact colocation, expect every current artefact (plan, spec, reports, references) INSIDE its {slug}_{date}/ task folder — scan one level into each task folder. The sibling reports//references/ dirs are deprecated and only hold legacy artefacts.

Frontmatter Reading

Read frontmatter fields: name, description, type, feature, phase

Route by description field for relevance matching (same approach as vc-context-discovery).

Skip files without frontmatter or with incomplete frontmatter — log as "no frontmatter, skipped".

Output: grouped list by folder — Active Plans / Backlog / Completed / Reports / References — with name + description per file.

When To Invoke

  • First action alongside vc-context-discovery at the start of every loop step (research / validate / execute / update-process)
  • Any time an agent needs to know: what plans exist for this feature, what was tried before, what is deferred, what references exist

Output Format

### Active Plans
- [name]: description (path)

### Backlog
- [name]: description (path)

### Completed
- [name]: description (path)

### Reports
- [name]: description (path)

### References
- [name]: description (path)

Found N active, N backlog, N completed, N reports, N references

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

AI-optimized browser automation CLI with context-efficient snapshots. Use for long autonomous sessions, self-verifying workflows, video recording, and cloud browser testing (Browserbase).

日本語の概要は準備中です。原文の説明を表示しています。

withkynam/vibecode-pro-max-kit1,1502026年6月21日 更新

Evaluate 4 execution strategies (sequential, parallel-subagents, workflow, agent-team) for a phase or fan-out task. Outputs 7-signal score table, agent count math, cost guards, and strategy recommendation.

日本語の概要は準備中です。原文の説明を表示しています。

withkynam/vibecode-pro-max-kit1,1502026年6月21日 更新

Audit project context routing, shared-skill discoverability, and Claude/Codex wiring. Use when context docs or skill surfaces move, split, or drift.

日本語の概要は準備中です。原文の説明を表示しています。

withkynam/vibecode-pro-max-kit1,1502026年6月21日 更新

Audit active project plan files for staleness, completion, and routing truth. Use when cleaning up plans, reconciling active work, or archiving completed artifacts.

日本語の概要は準備中です。原文の説明を表示しています。

withkynam/vibecode-pro-max-kit1,1502026年6月21日 更新

Audit agent harness health: Claude/Codex agent parity, skill registry consistency, README.md sync, and protocol file wiring. Use when agents, skills, README.md, or development-protocol files move, split, or drift.

日本語の概要は準備中です。原文の説明を表示しています。

withkynam/vibecode-pro-max-kit1,1502026年6月21日 更新

Emit and validate the provisional goal block for Autopilot Mode. Owns the 9-field format and resume detection from a pasted goal block.

日本語の概要は準備中です。原文の説明を表示しています。

withkynam/vibecode-pro-max-kit1,1502026年6月21日 更新

withkynam のスキルをすべて見る

このスキルの問題を報告する