本文へ移動
cccskills
無料GitHub で公開

handoff

Save provenance, decisions, phases, context policy, and next steps as kernel.handoff/v1. Triggers: handoff, save, pause, context, continue later.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md4.7 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。

<skill id="handoff"> <purpose> Compile the current task into an authoritative, resumable JSON manifest. The JSON file is CANONICAL. A markdown rendering may be generated for humans but is never a second source of truth.

Why manifest-first + bounded (EXP-L21): what a decision needs is flat (~50-70k tokens) regardless of session length; the accumulating transcript is almost pure decoration by late session. A handoff that says "read the conversation" recreates the waste. A handoff that pins exact selectors + phases lets the next session reconstruct ONLY the load-bearing state.

Reference: skills/context-mgmt/SKILL.md </purpose>

<on_start>

agentdb read-start

</on_start>

<phase id="1_extract" name="EXTRACT STATE"> Capture (these become manifest sections, not prose): - objective: goal, success_conditions (observable), non_goals - contract: governing _meta/contracts/ file if one exists, invariants, assumptions, decisions_accepted (+rationale), alternatives_rejected (+reason, REQUIRED by the schema even when empty) — every idea the user or you ruled out this session, so the next session never re-proposes it - workflow.phases: name each phase of the work; status inherited (done, with evidence) | required (still to do). Add invalidation_rules: what live-state change would flip an inherited phase back to required. Rules use typed `when.event` plus optional `when.path_glob`; never prose trigger strings. - open_threads, warnings (failed approaches — never re-explore silently) - tier (1/2/3) in identity </phase> <phase id="2_gather" name="GATHER EVIDENCE"> ```bash git branch --show-current && git rev-parse HEAD && git status --short git log --oneline -10 agentdb query "SELECT * FROM context WHERE type IN ('contract','checkpoint') ORDER BY ts DESC LIMIT 5" ``` Pin artifacts: for each file the next session must trust unchanged, `shasum -a 256 <path>` → provenance.artifacts[{path, sha256, purpose}]. </phase>

<ask_user> Use AskUserQuestion when: state extraction complete, before writing the manifest Ask: "Anything to add to the handoff? Blockers, decisions, or context I might have missed?" Options: looks complete, add context, skip handoff </ask_user>

<phase id="3_hygiene" name="GIT HYGIENE"> uncommitted: commit with "wip: checkpoint before handoff" or stash (document it) push: push current branch to remote — a local-only handoff is stranded work </phase> <phase id="4_compile" name="COMPILE THE MANIFEST"> Write `_meta/handoffs/{name}-{date}.json` conforming to schemas/kernel.handoff.v1.schema.json.

Context section — the attention budget is the design center:

  • policy.mode: advisory (default) | bounded (extra loads must be justified + ledgered) | sealed (forbidden globs hook-blocked; for controlled experiments)
  • required: SELECTORS, not whole files, wherever a section suffices: {path, heading: "## X"} · {path, lines: "A-B"} · {path, grep: "pat", context: N} · {git_diff: "revA..revB"} — each with a reason
  • forbidden: globs the resume must NOT touch (sealed mode enforces via guard-context hook)
  • budget: target_tokens (aim: boot + manifest + 2-12k selected history, per EXP-L21), max_tokens (hard stop)

Execution section: entry_phase, entrypoint (the literal first operation), stop_conditions, checkpoints (when to emit kernel.checkpoint/v1 during the resumed work).

Resume section: one-line continuation prompt: /kernel:ingest resume from _meta/handoffs/{file}.json

JSON strings are always quoted, so sha/commit fields cannot type-drift (the old YAML fails validation). </phase>

<phase id="5_validate" name="VALIDATE — MANDATORY"> ```bash "${CLAUDE_PLUGIN_ROOT:-.}/orchestration/manifest/kernel-manifest" validate _meta/handoffs/{file}.json ``` A handoff that does not validate DOES NOT EXIST. Fix and re-validate before reporting done. Optional human rendering: a .md next to it, first line `# RENDERED FROM {file}.json — NOT AUTHORITATIVE`. </phase> <delivery> 1. `kernel-manifest validate` passes (above) 2. Commit: "docs: handoff manifest for {name}" 3. Push to remote </delivery>

<on_complete>

agentdb write-end '{"skill":"handoff","saved_to":"_meta/handoffs/{file}.json","branch":"X","tier":N}'
# If non-local profile: post handoff summary to GitHub Discussions (Agent Logs category)
# Suggest /kernel:retrospective if learnings accumulated across sessions

</on_complete>

</skill>

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

Use AtomLane to compile and execute safe atomic parallel plans on macOS and native Windows Preview for worthwhile independent argv tasks, dependency DAGs, supported platform entrypoints, or Apple-silicon operators. Use at task start or an execution boundary when structured local work may contain two or more worthwhile units; skip plain answers, one quick command, and work whose effects cannot be safely bounded.

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。

hashgraph-online/awesome-codex-plugins1,2842026年10月11日 更新

add

無料

Register a deferred decision in the debt registry. Trigger by judgment, not a marker scan, whenever a future reader would ask "why this way?": an unmade decision, stub, loosened type, bypassed check, swallowed error, a default picked "for now", or a TODO/FIXME/HACK/XXX marker. Trigger immediately whenever you defer work, or when the user invokes $add. Over-register freely; the developer drops with "drop A", "drop A,C", or "drop all".

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。

hashgraph-online/awesome-codex-plugins1,2842026年10月11日 更新

ADK 框架适配层。为 LangChain / EINO / AutoGen / AgentScope / CrewAI 提供框架特定的 代码模板、惯用模式、API 映射和项目结构,供 agent-dev-workshop Phase 5 代码生成使用。 每个框架 reference 文件标注 verified_date 用于版本锁定。

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。

hashgraph-online/awesome-codex-plugins1,2842026年10月11日 更新

中文调试修复技能。用于报错、测试失败、页面异常、功能不符合预期、需要定位根因并做最小修复时。触发语包括"进入调试模式""帮我修问题""报错了""测试失败""页面坏了""找根因"。

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。

hashgraph-online/awesome-codex-plugins1,2842026年10月11日 更新

交互式 AI Agent 开发工作坊:通过 6 阶段深度协作对话,引导用户完成 Agent 需求分析、架构设计、 工具定义、Prompt 与编排设计、代码生成、验证迭代,产出可直接运行的 Agent 项目。 框架无关设计优先,支持 LangChain / EINO / AutoGen / AgentScope / CrewAI 等 ADK 框架。

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。

hashgraph-online/awesome-codex-plugins1,2842026年10月11日 更新

中文漂移审计技能。用于项目或学习过程变乱、上下文漂移、任务分叉、多个方案冲突、命名不一致、Codex 可能顺手改多了时。触发语包括"漂移检查""感觉跑偏了""项目变乱了""检查是否失控""分叉太多""上下文漂移"。

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。

hashgraph-online/awesome-codex-plugins1,2842026年10月11日 更新

hashgraph-online のスキルをすべて見る

このスキルの問題を報告する