本文へ移動
cccskills
無料GitHub で公開

ralph-specum-implement

This skill should be used only when the user explicitly asks to use `$ralph-specum-implement`, or explicitly asks Ralph Specum in Codex to run implementation for approved tasks, quick mode, or an explicit continue request.

インストール方法を見る

含まれるファイル(2)

  • SKILL.md5.4 KB
  • agents/openai.yaml405 B

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。

Ralph Specum Implement

You are a coordinator, not an executor -- delegate each task to a spec-executor sub-agent.

Contract

  • Resolve the active spec by explicit path, exact name, or .current-spec
  • Require tasks.md
  • Recompute task counts from disk before execution
  • Merge state fields only
  • Reconcile prototype records before dispatch and block only dependent work
  • Remove .ralph-state.json only when all tasks are complete, verified, and activePrototypes is empty

Action

  1. Resolve the active spec. If none exists, stop.
  2. Require tasks.md. Read .progress.md, current state, and current task markers.
  3. Parse tasks.md once into ordered top-level task rows. Include only unindented checkboxes outside fenced example blocks whose next token is a concrete numeric task ID, V<number>, VE<number>, or VF; exclude nested and example checkboxes, completion criteria, and placeholder IDs. From that one list derive total, the completed count across all rows, and next_index as the zero-based position of the first incomplete row or total when none remains. Do not derive next_index from the completed count; non-prefix completion cases resume at the earliest incomplete row.
  4. Resolve the dispatch task index before merging state. For fresh execution, use next_index. For a prototype return, require a validated non-negative returnTaskIndex and verify that it identifies the first eligible incomplete task. Merge state once with:
    • phase: "execution"
    • awaitingApproval: false
    • totalTasks: total
    • taskIndex: next_index for fresh execution, or the validated returnTaskIndex for a prototype return
    • preserve taskIteration, maxTaskIterations, globalIteration, maxGlobalIterations, commitSpec, and relatedSpecs
  5. Before dispatch, run prototype_records.py reconcile whenever state exists and run select-downstream whenever activePrototypes is nonempty or prototype history exists. Request --target execution, --target "task:$TASK_INDEX", and --path for every declared current-task path. Stop when an active blocker or stale input targets the work, or when any matching targetDecisions entry is not both proofAvailable: true and eligible: true. Missing dependency or approved-transfer proof blocks conservatively. Report the prototype ID and resume active work through $ralph-specum-prototype --resume <id>; route terminal staleness to its earliest affected phase or task.
  6. On a prototype return, verify that the merged taskIndex still equals the validated returnTaskIndex and identifies the first eligible incomplete task before dispatch.
  7. Delegate each task to a spec-executor sub-agent. Pass the task description, file targets, success criteria, and context from .progress.md. The sub-agent implements the task and outputs TASK_COMPLETE. Do NOT implement tasks yourself. Execute tasks in order until complete or blocked.
  8. [P] tasks may batch only when file sets do not overlap and verification is independent.
  9. [VERIFY] tasks stay in the same run and must produce explicit verification evidence.
  10. Marker syntax must be explicitly present in tasks.md. If markers are absent, treat tasks as non-batchable by default.
  11. VE tasks are valid quality tasks when the spec includes autonomous end-to-end verification.
  12. Native task sync metadata should be preserved when present.
  13. After each task or safe batch:
  • mark the checkbox
  • update .progress.md
  • merge the state update
  • use the task Commit line unless commits were explicitly disabled
  1. Before any batching, generated-task, CI, review-fix, branch-publication, or PR-lifecycle push, apply the Prototype Evidence Push Gate in ../../references/workflow.md. Normal mode may ask at that boundary for separate explicit authorization naming every outbound **/prototypes/*.md record. Quick mode asks no question and skips every push. A skipped or denied push ends the dependent remote lifecycle path: do not run gh pr create, gh pr merge, gh pr checks, gh pr view, gh api, gh run, gh issue, remote review polling, issue writes, or later remote steps that depend on that push. Quick mode continues or finishes locally and reports Remote lifecycle skipped: prototype evidence stayed local. Preserve the existing normal remote lifecycle only after the gate completes the push. Never push an isolated prototype source branch. commitSpec authorizes local commits only.
  2. On failure or interruption, persist the current state and stop with a resumable summary.
  3. On full completion, reconcile again. If activePrototypes remains nonempty, preserve .ralph-state.json and stop with its IDs. Otherwise remove state and report completion.

Resume Rules

  • Resume from the persisted task state when execution was already in progress.
  • If disk state and task checkboxes disagree, prefer tasks.md for completion and repair state to match.
  • If approval is still pending for tasks, stop and get approval unless quick mode or explicit user direction says to continue.
  • A stale task or dependent active prototype always wins over resume dispatch. An unrelated active prototype does not pause the current task.

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

This skill should be used when the user asks to "create a slash command", "add a command", "write a custom command", "define command arguments", "use command frontmatter", "organize commands", "create command with file references", "interactive command", "use AskUserQuestion in command", or needs guidance on slash command structure, YAML frontmatter fields, dynamic arguments, bash execution in commands, user interaction patterns, or command development best practices for Claude Code.

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。

tzachbon/smart-ralph5592026年9月16日 更新

This skill should be used when generating spec artifacts (research.md, requirements.md, design.md, tasks.md), formatting agent output, structuring phase results, or when any Ralph agent needs guidance on concise, scannable output formatting. Applies to all Ralph spec phase agents.

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。

tzachbon/smart-ralph5592026年9月16日 更新

Output rules for all agents - concise, scannable, actionable. Based on Matt Pocock's planning principles.

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。

tzachbon/smart-ralph5592026年9月16日 更新

Core principle that the main agent is a coordinator, not an implementer. All work must be delegated to subagents.

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。

tzachbon/smart-ralph5592026年9月16日 更新

This skill should be used when the user asks to "create a hook", "add a PreToolUse/PostToolUse/Stop hook", "validate tool use", "implement prompt-based hooks", "use ${CLAUDE_PLUGIN_ROOT}", "set up event-driven automation", "block dangerous commands", or mentions hook events (PreToolUse, PostToolUse, Stop, SubagentStop, SessionStart, SessionEnd, UserPromptSubmit, PreCompact, Notification). Provides comprehensive guidance for creating and implementing Claude Code plugin hooks with focus on advanced prompt-based hooks API.

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。

tzachbon/smart-ralph5592026年9月16日 更新

This skill should be used when a Ralph phase must identify critical user decisions, run a layered grill, persist partial answers, obtain explicit approval, or resume an interrupted phase interview before delegating artifact work.

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。

tzachbon/smart-ralph5592026年9月16日 更新

tzachbon のスキルをすべて見る

このスキルの問題を報告する