diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
index 54f21c7..0511fc4 100644
--- a/.claude-plugin/marketplace.json
+++ b/.claude-plugin/marketplace.json
@@ -8,7 +8,7 @@
"name": "align",
"source": "./",
"description": "Ongoing personal alignment for project decisions",
- "version": "0.11.0"
+ "version": "0.12.0"
}
]
}
diff --git a/.claude-plugin/plugin.json b/.claude-plugin/plugin.json
index d3767b4..fc6111d 100644
--- a/.claude-plugin/plugin.json
+++ b/.claude-plugin/plugin.json
@@ -1,11 +1,11 @@
{
"name": "align",
- "version": "0.11.0",
+ "version": "0.12.0",
"description": "Keep project decisions aligned with personal criteria through an ongoing align skill.",
"author": {
"name": "hmmhmmhm"
},
- "homepage": "https://github.com/hmmhmmhm/aligns.md",
+ "homepage": "https://aligns.md",
"repository": "https://github.com/hmmhmmhm/aligns.md",
"license": "MIT",
"skills": "./skills/",
diff --git a/.codex-plugin/plugin.json b/.codex-plugin/plugin.json
index 4762cba..35e482a 100644
--- a/.codex-plugin/plugin.json
+++ b/.codex-plugin/plugin.json
@@ -1,11 +1,11 @@
{
"name": "aligns-md",
- "version": "0.11.0",
+ "version": "0.12.0",
"description": "Keep project decisions aligned with personal criteria through an ongoing align skill.",
"author": {
"name": "hmmhmmhm"
},
- "homepage": "https://github.com/hmmhmmhm/aligns.md",
+ "homepage": "https://aligns.md",
"repository": "https://github.com/hmmhmmhm/aligns.md",
"license": "MIT",
"skills": "./skills/",
diff --git a/AGENTS.md b/AGENTS.md
index 8c9c989..ed087d7 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -62,6 +62,12 @@ focused clarification when important judgment diverges, repeated corrections
suggest a missing criterion, the user is unsure of the next direction, or a loop
or goal exposes unresolved scope, completion, or stopping criteria. Compare the
request, configured objective, actual state, and results. Fix known errors directly.
+When open-ended work lacks a finish condition, the remaining scope is unclear,
+or repeated attempts bring no new evidence, use align to review the scope and
+suggest a bounded goal when useful. Reuse completed work and check overlapping
+assignments. Keep task lists in the plan and lasting criteria in ALIGNS.md.
+A suggestion does not create or start a goal or loop, change its limits, or
+replace existing authorization. Do not add goal setup to routine bounded tasks.
Before a long run with untested boundaries, offer hypothetical rehearsal without
executing scenarios or launching the run. A review does not authorize a restart,
budget increase, or override of a user stop. The skill acts during active turns;
diff --git a/ALIGNS.md b/ALIGNS.md
index 2e49b89..8d072ec 100644
--- a/ALIGNS.md
+++ b/ALIGNS.md
@@ -87,8 +87,8 @@ implying a full review.
## Project name
Use ALIGNS.md for the project and agreement file, and hmmhmmhm/aligns.md for the
-repository. Keep `align` as the skill name. The acquired aligns.md domain will
-be connected separately after DNS propagation. Preserve historical benchmark
+repository. Keep `align` as the skill name. Use https://aligns.md as the official website. The custom domain and SSL are
+active; keep public site links and metadata on this domain. Preserve historical benchmark
evidence and its hashes when updating current branding.
## Presentation and Licensing
@@ -214,6 +214,21 @@ execution error directly; ask only about unresolved judgment criteria.
The method must not imply that it can wake a stopped agent or that a review
alone authorizes restarting work or expanding its scope.
+## Goals, work scope and repetition
+
+Confirmed scope update: 2026-11-11 (Asia/Seoul). Partial confirmation only.
+
+The skill should suggest a bounded goal when open-ended work, unclear remaining
+scope or stalled repetition makes one useful. Review available work, dependencies,
+completed results and overlapping assignments, and propose completion evidence
+and relevant stopping boundaries. Keep ordinary bounded tasks moving without
+requiring goal setup. Distinguish productive iteration from repeated actions
+that bring no new evidence. Use existing paths 1 and 4 for this review.
+
+Keep task inventories and run settings in the plan or task system, with confirmed
+lasting criteria in ALIGNS.md. A proposal does not authorize creating or starting
+a goal or loop, changing budgets, scheduling work or restarting a stopped run.
+
## Verification and delivery
Confirmed from conversation and hypothetical rehearsal: 2026-10-18 16:58 (Asia/Seoul). Partial confirmation only.
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 20e5c28..d84b810 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,11 @@
# Changelog
+## 0.12.0
+
+- Suggest bounded goals when open-ended work needs a clear outcome and finish condition. Preserve the existing eight paths.
+- Review available scope, remaining effort, completed work and overlapping assignments before proposing more work.
+- Check stalled repetition against new evidence and agreed boundaries, while preserving useful iteration and existing execution permissions.
+
## 0.11.0
- Add a validated manual team-data export: reviewer policy, bound approvals, artifact hashes, context snapshots, withdrawals and shared task families.
diff --git a/README.md b/README.md
index c2dccf2..a6a83db 100644
--- a/README.md
+++ b/README.md
@@ -4,6 +4,8 @@
ALIGNS.md
+aligns.md
+
English · 한국어 · 日本語
@@ -59,39 +61,27 @@ Brief requests work too. The agent checks existing context, then asks only about
## Install align
-Paste this request into your agent. It sets up the maintained skill and an update
-policy, then connects it to your project.
+Paste this request into your agent:
```text
-Set up the maintained align skill from https://github.com/hmmhmmhm/aligns.md for this project. Read its current installation guide, check the existing installation and agent instructions, and use the supported installation route for my primary CLI. Ask which CLI I mainly use, and whether it is this one, only if that is unclear.
-
-Enable supported automatic method updates. If the client has no native auto-update option, use a targeted update at the start of each alignment session, preserving the installation scope. Keep a working installed version when offline. Explain which update mode is active; do not add a scheduler or update unrelated packages.
-
-Use the skill to establish or review ALIGNS.md and connect the primary instruction file to it. Preserve existing rules and distinct files; propose safe links for any other clients I use. Record the invocation and update policy in ALIGNS.md. Reuse clear answers, and suggest a focused alignment check after important judgment mismatches, before untested long autonomous work, or when I am unsure of the next direction. Method updates must never rewrite my personal criteria or count as agreement.
+Set up the maintained align skill from https://github.com/hmmhmmhm/aligns.md using its current installation guide. Check my existing setup, use the supported route for my primary CLI, connect ALIGNS.md to this project, and configure supported method updates. Preserve existing instructions and my confirmed criteria; ask only about unresolved choices.
```
-Or install the portable skill directly:
+Or install it directly:
```sh
npx skills add hmmhmmhm/aligns.md --skill align
```
-Then ask: **Use the align skill to set up ongoing alignment for this project.**
-
-Once the skill is available, ask “Help me align this project” whenever you need it. No command or exact phrase is required. If the purpose is open, choose a starting point by number: project criteria, a decision that missed expectations, history review, rehearsal before a long run, a recommendation, optional personal defaults across projects, adding the README badge when it is missing, or optional training-data collection and export. The agent outlines the steps and guides you through them. If you already say what you need, it skips the menu. It asks before saving when needed, then offers remaining topics and a finish choice. Completed or declined topics are left out. All remaining paths stay visible, including those not yet explored, with one recommended next step. Before offering finish alone, the agent accounts for all eight paths and explains any exclusions. It uses your client's choice interface when all choices fit, or a numbered list otherwise.
-
-A short return invitation appears at most once when a self-alignment session ends. It does not appear after ordinary project tasks, installation-only work, or changes to the skill itself. You can ask to omit it entirely.
+Then ask: **Let's do a self-alignment check for this project.** No command or exact phrase is required.
-Ask “Help me organize my preferences across projects” to set up a personal ALIGNS.md, optionally drawing on personal documents or notes you choose. No particular filename or format is required. Keep common preferences separate from workplace rules. Store the file locally, or connect a separate personal repository only if you want one. [Personal setup and storage](docs/usage.md#personal-defaults-across-projects).
+[Installation, updates and the eight starting points](docs/usage.md)
-Option **8** helps you collect decision examples for personal model training. Choose the sources, scope, storage and review checkpoints before collection starts. Review candidates, then request an SFT, eligible DPO, or evaluation export. Local storage outside the repository is the default; Git commits and uploads are separate choices. The skill does not train models or monitor every session in the background. [Training-data setup and export](docs/usage.md#personal-training-data). Teams can use a [manual submission and review workflow](docs/team-training-data.md) to prepare examples around shared criteria. The team handles training and evaluation. [MLOps and team adoption brief](docs/team-mlops-brief.md) includes the data contract, validation results and a pilot plan.
+## For teams
-Native plugin installation is available for Codex and Claude Code. Claude Code
-can update this marketplace automatically when enabled. Other routes can use a
-user-authorized update at the start of an alignment session. No background
-updater or universal automatic activation is assumed.
+Prepare reviewed decisions from code reviews and agent conversations as training data for your coding agent. Your team handles training and evaluation.
-[Installation and updates](docs/usage.md) · [Behavioral benchmark](BENCHMARK.md) · [Intervention pilot](EFFORT_BENCHMARK.md) · [Preference transfer](DISCOVERY_BENCHMARK.md) · [Preference retention](RETENTION_BENCHMARK.md) · [Validation](docs/validation.md) · [Distribution](docs/distribution.md)
+[MLOps and team adoption brief](docs/team-mlops-brief.md) · [Manual submission guide](docs/team-training-data.md)
## Add the badge to your project
diff --git a/VERSION b/VERSION
index d9df1bb..ac454c6 100644
--- a/VERSION
+++ b/VERSION
@@ -1 +1 @@
-0.11.0
+0.12.0
diff --git a/claude-hooks/session-start.sh b/claude-hooks/session-start.sh
index 0b884f4..db8270e 100644
--- a/claude-hooks/session-start.sh
+++ b/claude-hooks/session-start.sh
@@ -1,3 +1,3 @@
#!/bin/sh
# Static reminder only: no project reads, networking, writes, or self-update.
-printf '%s\n' '{"hookSpecificOutput":{"hookEventName":"SessionStart","additionalContext":"Read the project ALIGNS.md when present and apply its current criteria. Use the installed align skill (/align:align) for consequential judgment mismatches, repeated corrections that expose missing criteria, unclear direction, or unresolved criteria when an autonomous loop or goal is configured, stops, or is marked complete against expectations. Review before long runs with untested boundaries. A review does not authorize restarting, raising budgets, or overriding user stops. The skill cannot wake a stopped agent. Offer a brief review only when relevant. Fix clear errors directly, reuse existing answers, and do not repeat resolved or declined questions. Skill updates never change personal criteria or confirm agreement. If alignment is absent, suggest setup when relevant without blocking unrelated routine work."}}'
+printf '%s\n' '{"hookSpecificOutput":{"hookEventName":"SessionStart","additionalContext":"Read the project ALIGNS.md when present and apply its current criteria. Use the installed align skill (/align:align) for consequential judgment mismatches, repeated corrections that expose missing criteria, unclear direction, or unresolved criteria when an autonomous loop or goal is configured, stops, or is marked complete against expectations. When open-ended work lacks a finish condition or repetition brings no new evidence, use align to review scope and suggest a bounded goal when useful. Reuse completed work and check overlapping assignments. A suggestion does not create or start a goal or loop. Review before long runs with untested boundaries. A review does not authorize restarting, raising budgets, or overriding user stops. The skill cannot wake a stopped agent. Offer a brief review only when relevant. Fix clear errors directly, reuse existing answers, and do not repeat resolved or declined questions. Skill updates never change personal criteria or confirm agreement. If alignment is absent, suggest setup when relevant without blocking unrelated routine work."}}'
diff --git a/docs/distribution.md b/docs/distribution.md
index 17fb3eb..8473043 100644
--- a/docs/distribution.md
+++ b/docs/distribution.md
@@ -6,6 +6,7 @@ Do not describe a pending or unverified listing as published.
## Source
+- Website: https://aligns.md
- Repository: https://github.com/hmmhmmhm/aligns.md
- Skill: https://github.com/hmmhmmhm/aligns.md/tree/main/skills/align
- Name: `align`
@@ -14,24 +15,26 @@ Do not describe a pending or unverified listing as published.
## Service status
-Directory evidence was checked on 2026-10-12 under the former name `ALIGN.md`.
-The repository was renamed to `ALIGNS.md` on 2026-10-17. Third-party listings
-and their cached names have not been reverified since the rename. The URLs below
-preserve actual receipts; a new repository name does not prove a directory update.
+The repository was renamed to `ALIGNS.md` on 2026-10-17. Skills.sh, askill,
+Agent Skill Hub, and Skillstore were checked again on 2026-11-02 against source
+commit `43f2c30522085600883f4438ac2231ea076e87e2`, method version `0.11.0`.
+Other rows retain their earlier evidence and are not newly verified.
+The URLs below preserve actual receipts. A renamed listing or accepted refresh
+does not prove that its downloadable files have changed.
The maintained source now includes native Codex and Claude Code marketplaces.
These custom marketplaces are separate from official platform directory listings.
| Service | Route | Current evidence |
| --- | --- | --- |
| GitHub | Public source and release | Public source and the installable skill are available. Release packages include a skill ZIP and SHA-256 checksum. See the [latest release](https://github.com/hmmhmmhm/aligns.md/releases/latest). |
-| [skills.sh](https://skills.sh/docs/faq) | Installation telemetry and directory indexing | Public [skill page](https://skills.sh/hmmhmmhm/align.md/align) verified. One real remote installation succeeded for Codex, Claude Code, and Cursor in a test project. |
-| [askill](https://askill.sh/submit) | Submit GitHub URL | Submitted and indexed one skill. Public [listing](https://askill.sh/skills/704509) verified. |
+| [skills.sh](https://skills.sh/docs/faq) | Installation telemetry and directory indexing | The [current page](https://skills.sh/hmmhmmhm/aligns.md/align) works, and the old URL redirects to it. On 2026-11-02, its [download snapshot](https://skills.sh/api/download/hmmhmmhm/aligns.md/align) still contained method `0.6.1` and 12 files, compared with `0.11.0` and 25 tracked files in the source. No supported publisher refresh endpoint was found. The directory snapshot remains stale. This is separate from direct installation from GitHub. |
+| [askill](https://askill.sh/submit) | Submit GitHub URL | Reindexed the renamed source on 2026-11-02. The [current listing](https://askill.sh/skills/704733) shows method `0.11.0` and the eight alignment paths. The [legacy listing](https://askill.sh/skills/704509) was also submitted for reindexing, but its public page still showed the initial content at the follow-up check. Use the current listing. Supporting-file download completeness was not verified. |
| [OpenAgentSkill](https://www.openagentskill.com/submit) | Public submission queue | Submission accepted and marked public. [Listing](https://www.openagentskill.com/skills/hmmhmmhm-align-md-align) verified; static checks passed, not AI-reviewed. |
-| [Agent Skill Hub](https://doc.agentskillhub.dev/guide/publishing) | GitHub import API | Imported as `hmmhmmhm/align`, version `2026.10.12`. Public [record](https://agentskillhub.dev/api/v1/u/hmmhmmhm/skills/align) and search for `ALIGN.md` verified at the original check. Source commit: `86e1ecc`. |
+| [Agent Skill Hub](https://doc.agentskillhub.dev/guide/publishing) | Source refresh API | Refreshed the existing `hmmhmmhm/align` record on 2026-11-02 through the documented `refresh-from-source` endpoint. The [public record](https://agentskillhub.dev/api/v1/u/hmmhmmhm/skills/align) reports version `2026.11.02` and source commit `43f2c30`. Its SKILL.md matches the maintained source, and all 25 tracked skill paths are present in the manifest. Supporting-file contents were not individually compared. |
| [SkillsMP](https://skillsmp.com/about) | Public GitHub indexing | Public indexing documented; no self-service submission route found in the inspected site and documentation. A live search for the former `hmmhmmhm/align.md` returned no matches; listing is not yet verified. |
| [Skills Directory](https://skillsdir.dev/add) | GitHub issue submission | The site's linked repository and issue template both return 404. Submission currently unavailable. |
| [ClawHub](https://docs.openclaw.ai/clawhub) | Authenticated skill publication | CLI dry-run accepted `align-md` v0.1.0 with all three package files. The device authorization expired without login; a fresh owner authorization is required. Dry-run is not publication. |
-| [Skillstore](https://skillstore.io/submit) | Submit GitHub URL for audit and review | Submission processed successfully and [registration PR #3749](https://github.com/aiskillstore/marketplace/pull/3749) created. The service's automated report marks the skill safe, and all three PR checks passed. The PR remains open pending final review. [Submission status](https://skillstore.io/ko/submissions/3c6ca451-04aa-435e-95c3-5c51e1094fec). Publication is not yet confirmed. |
+| [Skillstore](https://skillstore.io/submit) | Submit GitHub URL for audit and review | The [existing listing](https://skillstore.io/skills/hmmhmmhm-align) is public, and [registration PR #3749](https://github.com/aiskillstore/marketplace/pull/3749) was merged on 2026-10-12. It still serves the initial snapshot. The [2026-11-02 refresh submission](https://skillstore.io/ko/submissions/220241ba-a49c-49ed-ae1b-f3eca9c223cb) failed before security review because the published source identity uses the former repository name. [Workflow evidence](https://github.com/aiskillstore/marketplace/actions/runs/41374403057). In [issue #4038](https://github.com/aiskillstore/marketplace/issues/4038#issuecomment-6576959897), the maintainer confirmed on 2026-11-02 that source migration is unsupported and closed the issue. The final instruction is to request removal of the old listing, wait for removal, then submit the renamed source. The 2026-11-05 attempt stopped at GitHub sign-in. No removal request or replacement submission was completed. Latest-version publication remains unconfirmed. |
| [Skills Directory](https://www.skillsdirectory.com/submit) | Authenticated submission | The submission route requires sign-in. The owner login request timed out with the sign-in page unchanged; no submission made. This is a different service from `skillsdir.dev`. |
## Platform directories
diff --git a/docs/team-mlops-brief.md b/docs/team-mlops-brief.md
index 1342d73..75c692a 100644
--- a/docs/team-mlops-brief.md
+++ b/docs/team-mlops-brief.md
@@ -2,7 +2,7 @@
A technical and adoption brief for engineering leaders, model teams and MLOps.
-[한국어](translated-briefs/ko.md) · [Manual submission guide](team-training-data.md) · [Data contract](../skills/align/references/team-training.md)
+**English** · [한국어](translated-briefs/ko.md) · [日本語](translated-briefs/ja.md) · [Manual submission guide](team-training-data.md) · [Data contract](../skills/align/references/team-training.md)
## The decision this supports
diff --git a/docs/translated-briefs/ja.md b/docs/translated-briefs/ja.md
new file mode 100644
index 0000000..aa859d3
--- /dev/null
+++ b/docs/translated-briefs/ja.md
@@ -0,0 +1,141 @@
+# ALIGNS.mdでチーム向けコーディングエージェントの学習データを準備する
+
+エンジニアリング責任者、モデル開発チーム、MLOpsチーム向けの導入検討資料です。
+
+[English](../team-mlops-brief.md) · [한국어](ko.md) · **日本語** · [手動提出ガイド](../team-training-data.md) · [データ仕様](../../skills/align/references/team-training.md)
+
+## 何を導入するのか
+
+ALIGNS.mdを使うと、コードレビューやエージェントとの会話でチームが下した判断を残し、承認された事例をチームの学習実験に渡せます。たとえば、パッチの変更範囲が広すぎる理由を説明したレビューには、マージ済みのコードだけでは分からない情報があります。当時の文脈、修正内容、受け入れた結果を一緒に残すことで、その判断を後の学習に使えます。
+
+最初は、一つのリポジトリや業務を対象にデータ収集を試すことを提案します。ALIGNS.mdが提供するのは、収集の手順、手動レビューの仕様、ローカルの書き出しツールです。データ、レビュアー、モデル、学習基盤はチームが管理します。スキルをインストールしても、オンラインの受付サービス、会話のバックグラウンド記録、モデル学習は始まりません。
+
+**現在は、MLOpsチームが主体となってデータ収集を試す段階に対応しています。** テキストと選択したツール呼び出しの出力は、バージョンを固定した一つのトークナイザーで読み込みとトークン境界の検査を通過しました。実際のトレーナーへの接続とコーディング能力の向上は、これから測定する必要があります。この資料では、実装の検査結果とモデルの性能を分けて報告します。
+
+## データとして残す価値があるもの
+
+| 出典 | すでに含まれる情報 | ALIGNS.mdで追加して残す情報 |
+| --- | --- | --- |
+| エージェントとの会話 | 依頼、提案された回答、その後の修正 | 判断時点で利用できた文脈、学習対象への明示的な承認、適用する基準 |
+| コードレビュー | パッチの議論と修正依頼 | その実装がチームに合う理由、受け入れたパッチと適用範囲 |
+| CIやローカル検査 | テストコマンド、出力、失敗 | 基準コミット、実行環境、レビュー対象のコードに結び付いた証拠 |
+| チームの現在の基準 | 合意した優先順位と制約 | 基準のバージョンと該当部分のスナップショット。古いルールを無意識に学習させないために使う |
+| 初回から受け入れられた結果 | 修正のやり取りが少ない成果物 | 明示的な承認。修正会話だけに偏ったデータを避ける |
+
+保守しやすい変更、APIの維持、適切なタイミングでの要件確認、実際に行った検査だけを報告する姿勢などが対象になります。実装を学習させるには、受け入れたパッチや行動を収集します。「変更範囲を絞ります」という回答は応答スタイルの例にはなりますが、その変更を実装できた証拠にはなりません。
+
+チームの基準は明示的な合意が必要です。頻出するコメントや一人のメンテナーの好みを、そのままチームの正解ラベルにはしません。個人の合意は別に扱います。
+
+## 出力ファイルと用途
+
+| ファイル | 内容 | 用途 | 注意点 |
+| --- | --- | --- | --- |
+| `sft-completion.jsonl` | プロンプトと、選択した一つのアシスタント出力 | テキストの判断、コードパッチ、選択したツール呼び出しのSFT | completion-only lossを設定し、実際のトレーナーのマスクを確認する |
+| `dpo.jsonl` | 同じ文脈のプロンプト、採用したテキスト、退けたテキスト | 選好最適化 | 人が承認した比較が必要。ツール呼び出しのDPOは対象外 |
+| `evaluation.jsonl` | 学習用から分離したタスク群と期待する回答 | チームの評価基準による採点や実行評価 | 期待する文章だけではコードの正しさを評価できない |
+| `manifest.json` | 出典、ハッシュ、レビュー、除外、件数、学習種別 | データの来歴確認と監査 | 非公開のメタデータ。トレーナーには入力しない |
+| 従来の `sft.jsonl` | 選択した回答を末尾に加えたテキスト会話 | 既存のテキスト処理系 | 過去のアシスタント発言も残る。通常は選択した出力を分離したファイルを使う |
+
+形式は[TRLのデータセット仕様](https://huggingface.co/docs/trl/dataset_formats)に対応しています。プロンプトと学習対象の出力を分けると、過去の誤答を文脈に残しつつ、選択した学習対象から外せます。学習側の設定も必要です。[TRLのcompletion-only loss](https://huggingface.co/docs/trl/sft_trainer#train-on-completion-only)を参照してください。
+
+ツール呼び出しのレコードには、関数スキーマ、観測できた呼び出しと結果、選択した一つのアシスタントの行動を残します。書き出しツールは呼び出しの対応関係と関数名を検証しますが、JSON Schemaの引数制約をすべて検証したり、エージェントの実行全体を再現したりはしません。実行全体の軌跡やモデル固有のrole・channel形式には、アダプターが必要です。
+
+## バージョン0.11.0で実装した検査
+
+[初期監査](../training-data-readiness.md)で、従来のテキスト出力に不足が見つかりました。今回の変更では再現できた問題に対処し、チームからの提出を扱う手順を加えました。提出された承認記録は確認すべき証拠であり、レビュアーの本人性を暗号学的に証明するものではありません。
+
+| 項目 | 初期状態 | 現在の実装 | チーム側で必要な対応 |
+| --- | --- | --- | --- |
+| チームの承認 | 手動の案内のみ | 許可レビュアーのポリシー、対象・ペアの明示的承認、レビュー参照、バージョンと日時の検査 | ポリシーの管理とチーム基盤での本人確認 |
+| ファイルの完全性 | 出典への参照 | 提出ファイルのバイト列ハッシュ、正規化したレコードのハッシュ、証拠ファイルと出力のハッシュ | ストレージの保護と原本の保持 |
+| 文脈 | 学習入力の外にしか存在しない場合があった | バージョン2ではプロンプトの抜粋と、確認済み入力スナップショットのダイジェストが必須 | 内容が十分かを確認し、判断後に判明した情報の混入を防ぐ |
+| 学習対象 | 過去のアシスタント回答もSFT会話に残った | プロンプトと学習対象の出力を分離 | トレーナーのラベルと文脈の切り詰めを確認 |
+| 矛盾する比較 | 逆向きのラベルも両方出力できた | 同じ文脈・基準で選好の有向サイクルを検出し拒否 | 意味上の矛盾や文脈をまたぐ矛盾をレビューで解決 |
+| コードの証拠 | 自由記述の検証情報 | 基準コミット、対象に含まれるパッチ、成功した検査の記録、ハッシュ付きの証拠を必須化 | 制限した環境で再実行し、テスト範囲を評価 |
+| ツール | テキスト専用の出力では拒否 | 関数スキーマ、対応する呼び出しと結果、選択した行動の出力 | ツールスキーマと選択したモデルの挙動を確認 |
+| データ分割 | ユーザー指定のタスク群とプロンプトの完全一致検査 | バッチをまたぐ共通のタスク群割り当て、完全一致の拒否、名前空間付きID | 割り当て前に関連タスクと意味上の重複を検出 |
+| 撤回 | 手動の案内 | ポリシーで名前空間付きレコードを除外 | 影響を受ける既存データセットと学習済みモデルを特定 |
+| 無効な提出 | 目視確認 | 書き込み前の検証、上書き禁止、読み取り専用の検証モード | 承認を捏造せずに不備を修正 |
+
+検証処理は、提出コードの実行、モデルのダウンロード、レコードのアップロード、学習開始を行いません。コードの証拠を調べる処理は、必要な証拠が申告どおりに存在し、変更されていないことを確認します。テストが実際に実行されたかを独立して証明するものではありません。
+
+## 再現可能な実装検証
+
+以下はメンテナーが合成データで実施した検査です。第三者認証やモデル品質のベンチマークではありません。テストデータに顧客の会話やチームの記録は含まれていません。
+
+| 検査 | 観測結果 | 確認できたこと |
+| --- | ---: | --- |
+| 既存の書き出しテスト | 13 / 13成功 | 個人向けテキスト出力の検査を維持 |
+| 新規のチームデータテスト | 23 / 23成功 | 承認済み経路と、改ざん・承認不足・文脈・ペア・コード証拠・呼び出し・分割に関する拒否動作 |
+| リポジトリのテスト全体 | 67 / 67成功 | 既存の検査と追加テスト |
+| Hugging Face Datasetsでの読み込み | SFT-completion 3行、DPO 2行 | このテストファイルをDatasets 5.1.0で読み込めた |
+| トークン境界 | 3 / 3成功 | 使用したトークナイザーで、判断・コード・ツール行動の学習対象境界を確認 |
+| ファインチューニングの効果 | 未測定 | 学習済みモデルや改善の主張はない |
+
+[個別テストの結果とソースのハッシュ](../../tests/audits/team-validation-2026-10-18.json)も公開しています。
+
+トークナイザーは `Qwen/Qwen2.5-0.5B-Instruct`、リビジョンは `7ae557604adf67be50417f59c2c2f167def9a775` です。Transformers 5.20.1、Tokenizers 0.23.3、Jinja2 3.1.6を使用し、モデルの重みとPyTorchは使っていません。三つの例の総トークン数は53、75、186、選択した出力は12、34、22トークンでした。検査時の上限は4,096トークンです。この短いテストデータでは、長い文脈や複数モデルとの互換性までは確認できません。
+
+[入力、出力、トークナイザーの検査結果](../../tests/audits/mlops-2026-10-18/)には、ハッシュとパッケージのバージョンを残しています。標準ライブラリによる検査と、任意の統合検査は次のように再現できます。
+
+```sh
+PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tests -p 'test_*.py'
+# In an isolated environment with the documented dependencies and local tokenizer:
+python3 tests/audits/mlops_smoke.py \
+ --tokenizer --output
+```
+
+## 過去のベンチマークで分かること
+
+既存の[確認済みの好みを保持する予備実験](../../RETENTION_BENCHMARK.md)は、基準を残す効果を調べる動機になります。ファインチューニングの実験ではありません。
+
+| 各条件で合成タスク12件を実施 | 一般的な指示、学習・保存は可能 | 確認済みの好みを初期入力し、alignを使用 |
+| --- | ---: | ---: |
+| 確認の回答なしで好みに合った件数 | 3 / 12 | 12 / 12 |
+| その後の確認に対する回答数 | 9 | 0 |
+| 最終的に好みに合った件数 | 12 / 12 | 12 / 12 |
+
+初期入力がある条件には、比較側が最初には持っていなかった情報に加え、スキルと参照手順も与えています。初期設定に必要な回答は、別途4回と仮定しました。ファイル名だけの効果や、同じ情報を持つAGENTS.mdに対する優位性を示すものではなく、人が費やした時間も測っていません。[情報量をそろえた比較](../../DISCOVERY_BENCHMARK.md)では同じ結果でした。
+
+これらの結果を踏まえ、対象を限定した作業で、確認済みの基準を保持すれば説明の繰り返しが減るかを検証できます。ファインチューニングの改善率、学習データの価値、投資収益には換算できません。
+
+## 試験導入の手順と担当
+
+| 段階 | 担当 | 成果物 | 完了条件 |
+| --- | --- | --- | --- |
+| 範囲の決定 | エンジニアリング責任者と参加者 | 一つの業務、合意した基準、許可する出典、非公開の保存先 | 参加者が収集内容と用途を理解している |
+| 記録 | 参加に同意した人とエージェント | 作業中または選択した履歴からの小さな候補レコード | 実際の文脈、対象、利用できる証拠を保持している |
+| 手動の提出確認 | 参加者 | 選択したバッチとファイル一覧 | 提出するファイルそのものが、指定用途について承認されている |
+| レビュー | 指定されたチームのレビュアー | 対象・ペアに結び付いた判断と共通のタスク群 | 未解決のラベルを学習データにしない |
+| 検証・書き出し | MLOps | バージョン付きデータとマニフェスト | 仕様検査に合格し、除外内容を確認している |
+| 学習基盤への接続 | MLOps | 固定したトークナイザーとトレーナー、collator検査、許可された試験学習 | マスクが正しく、切り詰めで対象を失わず、実行を再現できる |
+| 評価 | エンジニアリングのレビュアーとMLOps | 学習から分離したデータでの評価と回帰の報告 | 学習前に合意した基準を満たす |
+
+週次やスプリント単位の提出も選べますが、頻度はチームが決めます。意味のある判断時点で記録し、既存の承認を再利用して、不足しているラベルだけを確認します。最終提出は手動です。一人あたりのレコード数のノルマは設けません。
+
+## 件数だけでなく、使えるデータを測る
+
+| 指標 | 定義 | 判断に使う点 |
+| --- | --- | --- |
+| 承認率 | 承認された重複のない例 / レビューした候補 | 関連する事例を収集できているか |
+| レビュー効率 | 承認された重複のない例 / レビュアーの作業時間 | 手間が減ったのか、別工程に移っただけなのか |
+| タスクの網羅性 | 基準・業務・難易度別の異なるタスク群 | 追加の証拠が必要な領域 |
+| 証拠の充足率 | 再実行に必要な資料があるコード候補 / コード候補 | コーディング事例を検証できるか |
+| DPO適格率 | 承認された同一文脈のペア / レビューした比較 | 選好最適化に使えるデータがあるか |
+| 重複率・矛盾率 | 該当するレコード / レビューしたレコード。別々に報告 | 水増しやラベルの不一致がないか |
+| 評価の独立性 | 関連する学習タスクが知られていない評価用タスク群 | 汎化を評価できているか |
+| 学習後の修正負担 | 評価タスクごとのレビュアー介入数。コードの正しさと併記 | チームの実務が改善したか |
+
+しきい値は試験導入の費用と目的から決めます。一定件数が集まっただけでは、学習の準備ができたとは判断できません。生成した派生例は、把握している網羅性の不足を補うために使い、元のタスクとの関係を残し、別のラベルで区別します。言い換えを独立した証拠として数えないようにします。
+
+## 導入前にモデルを評価する
+
+同じベースモデルとALIGNS.mdを使う条件を、チューニング後のモデルに同じALIGNS.md、ツール、タスク予算を与える条件と比較します。派生例を作る前にタスク群を固定します。検証結果をチェックポイントの選択に使う場合は、最終テストセットを別に確保します。正しさ、不要な質問、範囲の判断、完了報告、レビュアーによる修正を確認し、一般的なコーディング課題や変更されたばかりの基準でも回帰を調べます。
+
+不確実性はタスク群の単位で報告します。修正回数が減ってもコードの品質が落ちたなら成功とはいえません。リポジトリの現在の事実や更新した基準は、引き続き実行時の文脈に入れる必要がある場合があります。ALIGNS.mdを変更しても、学習済みの挙動が自動更新されるわけではありません。
+
+## 導入の範囲
+
+収集・書き出しの実装は、このリポジトリのMITライセンスで利用できます。このライセンスは、参加者のコード、会話、第三者の資料を使う権利までは与えません。学習用途、保持期間、利用するサービスはチームが決めます。非公開の出典参照とマニフェストも、選択したレコードと同様に扱う必要があります。
+
+最初は範囲を限定したデータ収集を試し、その後にモデル固有の接続を確認する進め方を提案します。レビュアー認証の自動化、意味上の矛盾の解決、制限した環境でのコード再実行、オンライン受付、本番での効果測定は、現在の実装に含まれていません。導入の判断はチームが行い、ALIGNS.mdの合意内容を変えずに収集を停止できます。
diff --git a/docs/translated-briefs/ko.md b/docs/translated-briefs/ko.md
index 6444a97..777d6a5 100644
--- a/docs/translated-briefs/ko.md
+++ b/docs/translated-briefs/ko.md
@@ -2,7 +2,7 @@
엔지니어링 리더, 모델 개발팀, MLOps 팀을 위한 도입 검토 문서입니다.
-[English](../team-mlops-brief.md) · [수동 제출 가이드](../team-training-data.md) · [데이터 규격](../../skills/align/references/team-training.md)
+[English](../team-mlops-brief.md) · **한국어** · [日本語](ja.md) · [수동 제출 가이드](../team-training-data.md) · [데이터 규격](../../skills/align/references/team-training.md)
## 무엇을 도입하는가
diff --git a/docs/translated-readme/ja.md b/docs/translated-readme/ja.md
index 39fc8f7..38522e1 100644
--- a/docs/translated-readme/ja.md
+++ b/docs/translated-readme/ja.md
@@ -4,6 +4,8 @@
ALIGNS.md
+aligns.md
+
English · 한국어 · 日本語
@@ -61,35 +63,27 @@ ALIGNS.mdには、エージェントに目指してほしい目標と、何か
## alignをインストールする
-次の依頼文をエージェントに貼り付けてください。継続的にメンテナンスされるスキルをインストールし、更新方法を決めて、プロジェクトにつなぎます。
+次の依頼文をエージェントに貼り付けてください。
```text
-このプロジェクトに、https://github.com/hmmhmmhm/aligns.md のalignスキルをセットアップしてください。最新のインストールガイドを読み、既存のインストール状況とエージェントの指示を確認したうえで、私が主に使うCLIでサポートされている方法を選んでください。主に使うCLIが何か、今使っているCLIと同じかは、不明な場合だけ確認してください。
-
-サポートされている方法で、すり合わせの手順を自動更新するようにしてください。クライアントに自動更新機能がなければ、すり合わせのセッション開始時に、既存のインストール範囲を保ったまま、このスキルだけを更新してください。オフラインでは、動作するインストール済みバージョンを維持してください。どの更新方法が有効になっているかを説明し、スケジューラーの追加や無関係なパッケージの更新はしないでください。
-
-スキルを使ってALIGNS.mdを作成または見直し、主に使う指示ファイルから参照するようにしてください。既存のルールや個別に管理されているファイルを保ち、ほかに使っているクライアントには安全なリンク方法を提案してください。スキルの呼び出し方と更新方針をALIGNS.mdに記録してください。すでに明確に答えた内容は聞き直さないでください。大事な判断が意図とずれたとき、判断基準をまだ確かめていない長い自律作業の前、次の方向を決められないときには、必要な点に絞ったすり合わせを提案してください。手順を更新しただけで、私の判断基準を書き換えたり、私が合意したとみなしたりしてはいけません。
+https://github.com/hmmhmmhm/aligns.md の最新のインストールガイドに従って、alignスキルを設定してください。既存の設定を確認し、主に使うCLIでサポートされている方法でインストールして、このプロジェクトにALIGNS.mdを接続し、対応する更新方法を設定してください。既存の指示と確認済みの私の判断基準は保ち、まだ決まっていない点だけ質問してください。
```
-複数のツールで使えるスキルを直接インストールすることもできます。
+直接インストールすることもできます。
```sh
npx skills add hmmhmmhm/aligns.md --skill align
```
-インストール後に、こう依頼してください。**alignスキルを使って、このプロジェクトで継続的にすり合わせができるように設定してください。**
-
-スキルが使える環境なら、いつでも「このプロジェクトの判断基準をすり合わせたい」と話しかけてください。コマンドや決まった言い方は必要ありません。何から始めるか迷ったら、番号で選べます。プロジェクトの基準づくり、期待とずれた判断の見直し、履歴の確認、長い自律作業の前の判断練習、始め方の提案、複数のプロジェクトで使う個人の共通基準の設定、READMEへのバッジ追加、個人向けの学習データの収集・書き出しから選べます。7番のバッジ追加は、READMEにバッジがない場合だけ提案します。エージェントが手順を短く説明し、一緒に進めます。やりたいことをすでに伝えていれば、メニューは省略します。必要に応じて保存を確認し、一つ終わったら、完了済みの項目や断った項目を除いて次の候補と終了の選択肢を示します。まだ扱っていない項目も選択肢に残し、次に取り組むものを一つ勧めます。終了だけを提案する前に、八つの経路を確認し、候補から外した理由を伝えます。すべての候補が収まる場合はクライアントの選択画面を使い、収まらなければ番号付きの一覧を表示します。
-
-再びすり合わせを依頼できるという案内は、実際にすり合わせを行ったセッションの終了時に、一度だけ表示します。通常のプロジェクト作業、インストールだけの作業、スキル自体の修正では表示しません。案内を省くよう依頼することもできます。
+インストール後に、**「このプロジェクトの判断基準をすり合わせたい」**と依頼してください。コマンドや決まった言い方は必要ありません。
-「複数のプロジェクトで使う自分の好みを整理したい」と頼むと、個人用のALIGNS.mdを設定できます。手元の個人用ドキュメントやメモを参考に、共通の基準を整理できます。決まったファイル名や形式はありません。共通の好みと職場のルールは分けて扱います。ローカルファイルで管理でき、希望する場合だけ個人用の別リポジトリにも接続できます。[個人の基準の設定と保存先](../usage.md#personal-defaults-across-projects)(英語)。
+[インストール・更新と八つの始め方](../usage.md)(英語)
-**8番**では、個人向けのモデル学習に使う判断例を集められます。収集を始める前に、対象の履歴と範囲、保存先、確認のタイミングを決めます。候補を確認した後、SFT、条件を満たすDPO、評価用のデータとして書き出すよう依頼してください。既定の保存先はリポジトリ外のローカルフォルダです。Gitへのコミットとアップロードは別に選びます。スキルがモデルを学習させたり、すべてのセッションをバックグラウンドで収集したりすることはありません。[学習データの設定と書き出し](../usage.md#personal-training-data)(英語)。チームでは、[手動で提出・確認する手順](../team-training-data.md)(英語)に沿って、共通の基準に合う事例を準備できます。学習と結果の評価はチームが行います。 [MLOps・チーム向けの導入検討資料](../team-mlops-brief.md)(英語)に、データ仕様、検証結果、試験導入の進め方をまとめています。
+## チームで使う
-CodexとClaude Codeでは、各クライアントのプラグインとしてインストールできます。Claude Codeでは、このマーケットプレイスの自動更新を有効にすると、更新を受け取れます。ほかのインストール方法では、ユーザーの許可を得たうえで、すり合わせのセッション開始時に更新できます。バックグラウンド更新や、あらゆる環境での自動起動を前提にはしていません。
+コードレビューやエージェントとの会話で確認した判断例を、コーディングエージェントの学習データとして準備できます。学習と結果の評価はチームが行います。
-[インストールと更新](../usage.md) · [動作ベンチマーク](../../BENCHMARK.md) · [ユーザー介入の予備実験](../../EFFORT_BENCHMARK.md) · [好みの引き継ぎ](../../DISCOVERY_BENCHMARK.md) · [好みの保持](../../RETENTION_BENCHMARK.md) · [検証](../validation.md) · [配布](../distribution.md)
+[MLOps・チーム向けの導入検討資料](../translated-briefs/ja.md) · [手動提出ガイド](../team-training-data.md)(英語)
## プロジェクトにバッジを付ける
diff --git a/docs/translated-readme/ko.md b/docs/translated-readme/ko.md
index 31342c1..157b9a1 100644
--- a/docs/translated-readme/ko.md
+++ b/docs/translated-readme/ko.md
@@ -4,6 +4,8 @@
ALIGNS.md
+aligns.md
+
English · 한국어 · 日本語
@@ -61,35 +63,27 @@ ALIGNS.md에는 에이전트가 따라야 할 목표와, 선택이 필요할 때
## align 설치하기
-아래 요청문을 에이전트에 붙여 넣으세요. 지속적으로 관리되는 스킬을 설치하고 업데이트 방식을 정한 뒤, 프로젝트에 연결합니다.
+에이전트에게 다음 요청을 붙여 넣으세요.
```text
-이 프로젝트에 https://github.com/hmmhmmhm/aligns.md 의 align 스킬을 설치해 주세요. 최신 설치 가이드를 읽고, 기존 설치 상태와 에이전트 지시 파일을 확인한 뒤, 제가 주로 쓰는 CLI에서 지원하는 방법으로 설치해 주세요. 주로 쓰는 CLI가 무엇인지, 지금 쓰는 CLI와 같은지는 불분명할 때만 물어봐 주세요.
-
-지원되는 방식으로 정렬 방법의 자동 업데이트를 활성화해 주세요. 클라이언트에 자체 자동 업데이트 기능이 없다면, 정렬 세션을 시작할 때마다 기존 설치 범위를 유지하면서 이 스킬만 업데이트해 주세요. 오프라인에서는 정상 작동하는 설치 버전을 유지해 주세요. 어떤 업데이트 방식이 적용됐는지 알려주고, 스케줄러를 추가하거나 관련 없는 패키지를 업데이트하지 마세요.
-
-스킬로 ALIGNS.md를 작성하거나 검토하고, 주 지시 파일에서 이를 참조하게 해 주세요. 기존 규칙과 별도로 관리되는 파일은 보존하고, 제가 쓰는 다른 클라이언트에는 안전한 링크 방식을 제안해 주세요. ALIGNS.md에 스킬 호출법과 업데이트 방식을 기록해 주세요. 명확히 답한 내용은 다시 묻지 말고, 중요한 판단이 어긋났을 때, 아직 기준을 점검하지 않은 긴 자율 작업을 시작하기 전, 또는 제가 다음 방향을 정하지 못할 때 필요한 부분만 정렬하도록 제안해 주세요. 정렬 방법을 업데이트했다고 제 판단 기준을 바꾸거나 제가 동의한 것으로 간주해서는 안 됩니다.
+https://github.com/hmmhmmhm/aligns.md 의 최신 설치 가이드에 따라 align 스킬을 설정해 주세요. 기존 설정을 확인하고, 주로 쓰는 CLI에서 지원하는 방식으로 설치한 뒤 이 프로젝트에 ALIGNS.md를 연결하고 지원되는 방식으로 스킬을 업데이트하도록 설정해 주세요. 기존 지침과 확인된 제 기준은 보존하고, 아직 정해지지 않은 사항만 질문해 주세요.
```
-여러 도구에서 사용할 수 있는 스킬을 직접 설치해도 됩니다.
+직접 설치할 수도 있습니다.
```sh
npx skills add hmmhmmhm/aligns.md --skill align
```
-설치한 뒤 이렇게 요청하세요. **align 스킬로 이 프로젝트에서 지속적으로 정렬할 수 있도록 설정해 주세요.**
-
-스킬을 사용할 수 있는 환경이라면 언제든 “이 프로젝트 기준을 맞춰보자”처럼 말하면 됩니다. 명령어나 정해진 문구는 필요하지 않습니다. 무엇부터 할지 모르겠다면 번호로 고르세요. 프로젝트 기준 정리, 기대와 달랐던 판단 점검, 기록 검토, 긴 작업 전 판단 연습, 시작점 추천, 여러 프로젝트에 쓸 개인 공통 기준 설정, README 배지 추가, 개인화 학습용 데이터 축적·내보내기 중에서 선택할 수 있습니다. 7번 배지 추가는 README에 배지가 없을 때만 제안합니다. 에이전트가 진행 순서를 짧게 안내하고 함께 진행합니다. 원하는 작업을 이미 말했다면 메뉴는 건너뜁니다. 필요한 경우 저장을 확인하고, 한 항목이 끝나면 완료하거나 거절한 내용을 제외한 다음 선택지와 종료 선택지를 보여줍니다. 아직 다루지 않은 경로도 선택지에 남기고, 그중 다음에 하면 좋을 것을 추천합니다. 종료만 제안하기 전에는 8가지 경로를 모두 확인하고 제외한 이유를 알려줍니다. 선택지가 모두 들어가면 클라이언트의 선택형 질문 화면을 쓰고, 그렇지 않으면 번호 목록을 보여줍니다.
-
-다시 정렬할 수 있다는 안내는 실제 자가정렬 세션을 마칠 때 최대 한 번만 표시합니다. 일반 프로젝트 작업, 설치만 하는 작업, 스킬 자체를 수정하는 작업에서는 표시하지 않습니다. 안내를 생략해 달라고 요청해도 됩니다.
+설치 후 **“이 프로젝트 자가정렬 진행해봅시다”**라고 요청하세요. 명령어나 정확한 문구를 외울 필요는 없습니다.
-“여러 프로젝트에서 쓸 내 공통 취향을 정리해줘”라고 요청하면 개인용 ALIGNS.md를 설정할 수 있습니다. 기존 개인 문서나 메모를 참고해 공통 기준을 정리할 수 있습니다. 정해진 파일명이나 형식은 없으며, 공통 취향과 회사 업무 규칙은 구분합니다. 로컬 파일로 관리하거나, 원할 때만 별도 개인 저장소에 연결하세요. [개인 기준 설정과 저장 방식](../usage.md#personal-defaults-across-projects)(영어).
+[설치·업데이트 및 여덟 가지 시작점](../usage.md)(영어)
-**8번**에서는 개인 모델 학습에 쓸 판단 사례를 모을 수 있습니다. 수집 전에 출처, 범위, 저장 위치와 검토 시점을 정합니다. 후보를 검토한 뒤 SFT, 조건을 충족하는 DPO, 평가용 데이터로 내보내도록 요청하세요. 기본 저장 위치는 저장소 밖의 로컬 폴더이며, Git 커밋과 업로드는 따로 선택합니다. 스킬이 모델을 학습시키거나 모든 세션을 백그라운드에서 수집하지는 않습니다. [학습 데이터 설정과 내보내기](../usage.md#personal-training-data)(영어). 팀에서는 [수동 제출·검토 절차](../team-training-data.md)(영어)를 통해 공통 기준에 맞는 사례를 준비할 수 있습니다. 학습과 결과 검증은 팀이 진행합니다. [MLOps·팀 도입 검토 문서](../translated-briefs/ko.md)에서 데이터 규격, 검증 결과와 파일럿 계획을 확인할 수 있습니다.
+## 팀에서 활용하기
-Codex와 Claude Code에서는 자체 플러그인 방식으로 설치할 수 있습니다. Claude Code는 해당 마켓플레이스의 자동 업데이트를 켜면 업데이트를 받을 수 있습니다. 다른 설치 경로에서는 사용자가 허용한 경우 정렬 세션을 시작할 때 업데이트할 수 있습니다. 백그라운드 업데이트나 모든 환경에서의 자동 실행을 전제로 하지는 않습니다.
+코드 리뷰와 에이전트 대화에서 검토한 판단 사례를 코딩 에이전트 학습용 데이터로 준비할 수 있습니다. 학습과 결과 검증은 팀이 진행합니다.
-[설치 및 업데이트](../usage.md) · [동작 벤치마크](../../BENCHMARK.md) · [사용자 개입 예비 실험](../../EFFORT_BENCHMARK.md) · [선호 전이](../../DISCOVERY_BENCHMARK.md) · [선호 유지](../../RETENTION_BENCHMARK.md) · [검증](../validation.md) · [배포](../distribution.md)
+[MLOps·팀 도입 검토 문서](../translated-briefs/ko.md) · [수동 제출 가이드](../team-training-data.md)(영어)
## 프로젝트에 배지 붙이기
diff --git a/docs/usage.md b/docs/usage.md
index 7585c7d..6c67833 100644
--- a/docs/usage.md
+++ b/docs/usage.md
@@ -16,10 +16,10 @@ any time: “Help me align this project's criteria,” or “That result missed
meant; let's revisit the criteria.” No command or exact phrase is required. If
the purpose is open, the agent offers these choices in your language:
-1. Set or review project direction and personal criteria.
+1. Set or review project direction, scope, completion criteria, and personal criteria.
2. Review a decision that missed your expectations.
3. Find criteria in selected PR or conversation history.
-4. Rehearse decisions before a long autonomous run.
+4. Review execution boundaries and rehearse decisions before a long autonomous run.
5. Suggest a starting point from the current project.
6. Set up or review personal defaults across projects (optional).
7. Add the ALIGNS.md badge to the project README (only when absent).
@@ -44,6 +44,22 @@ confirm new preferences. Rehearsal discusses hypothetical decisions without
starting the autonomous run. Badge setup checks the README, adds the missing badge,
and verifies the result without asking preference questions.
+## Bound open-ended work
+
+When a request has no clear finish condition, the remaining scope is uncertain,
+or repeated attempts bring no new evidence, the agent can suggest a concrete
+goal. It checks available plans and work already completed or assigned, then
+proposes an outcome, scope and completion evidence. For larger work, it describes
+the likely effort and what remains uncertain. Paths 1 and 4 cover this review.
+
+A suggestion does not start a goal or loop. The agent reuses existing permission
+and asks only about choices that change the next step. Routine bounded tasks
+continue without goal setup. Productive iterations continue too, while stalled
+repetition prompts a changed approach, a narrower scope or a decision. A stalled
+run is still unfinished. Task lists and limits belong in the plan or task system,
+with confirmed lasting criteria in ALIGNS.md. See the
+[work scope and goal procedure](../skills/align/references/recurring.md#work-scope-and-goals).
+
## Continue from each answer
Choosing a client, source, or storage location completes one step. The agent uses
diff --git a/skills/align/SKILL.md b/skills/align/SKILL.md
index 3d37bf7..904e93d 100644
--- a/skills/align/SKILL.md
+++ b/skills/align/SKILL.md
@@ -1,9 +1,9 @@
---
name: align
-description: Establish and maintain project or optional cross-project personal alignment in ALIGNS.md, including review of selected personal documents, PR feedback, or conversations for scoped criteria. Use for natural-language self-alignment requests such as “자가정렬 진행해봅시다” or “Let’s do a self-alignment check,” or when goals or priorities are unclear, important decisions diverge from the user's expectations, repeated corrections expose a missing criterion, or a long autonomous run needs judgment calibration. Also use for requested collection or export of confirmed alignment decision examples for personal fine-tuning. Apply existing agreements without reopening routine decisions.
+description: Establish and maintain project or optional cross-project personal alignment in ALIGNS.md, including review of selected personal documents, PR feedback, or conversations for scoped criteria. Use for natural-language self-alignment requests such as “자가정렬 진행해봅시다” or “Let’s do a self-alignment check,” or when goals or priorities are unclear, important decisions diverge from the user's expectations, repeated corrections expose a missing criterion, a long autonomous run needs judgment calibration, open-ended work lacks a clear scope or finish condition, or repeated attempts stop producing new evidence. Also use for requested collection or export of confirmed alignment decision examples for personal fine-tuning. Apply existing agreements without reopening routine decisions.
license: MIT
metadata:
- version: "0.11.0"
+ version: "0.12.0"
---
# Align
@@ -53,10 +53,10 @@ For an open-purpose invocation, defer setup questions until a path is selected.
When the user invokes align without a clear purpose, show this short menu in
their language after reading the available project context. Keep numbers stable:
-1. Set or review the project's direction and personal criteria.
+1. Set or review the project's direction, scope, completion criteria, and personal criteria.
2. Review a recent decision that missed expectations.
3. Find candidate criteria in selected PR or conversation history.
-4. Rehearse judgment before a long autonomous run.
+4. Review execution boundaries and rehearse judgment before a long autonomous run.
5. Recommend a starting point from the current project context.
6. Set up or review personal ALIGNS.md defaults across projects (optional).
7. Add the ALIGNS.md badge to this project's README (only when absent).
@@ -206,6 +206,11 @@ Preserve other instructions and personal criteria.
| Long autonomous or repeated run | Check agreed boundaries and offer a short hypothetical rehearsal if consequential judgments are untested. |
| Loop or goal configured, stopped, or marked complete against expectations | Compare the request, configured objective, actual state, and results. Repair known errors; clarify only unresolved scope, completion, or stopping criteria. |
+When open-ended work needs a bounded goal, the remaining scope is unclear, or
+repetition stops producing useful evidence, follow
+[work scope and goals](references/recurring.md#work-scope-and-goals). Use paths 1
+and 4 without adding a menu step or restarting an interview.
+
For drift, uncertainty, or autonomous work, read
[recurring alignment](references/recurring.md). A suggestion is not a new
scheduled task or permission to run unattended.
diff --git a/skills/align/assets/instruction-block.md b/skills/align/assets/instruction-block.md
index 08d0266..7181bb2 100644
--- a/skills/align/assets/instruction-block.md
+++ b/skills/align/assets/instruction-block.md
@@ -32,6 +32,12 @@ focused clarification when important judgment diverges, repeated corrections
suggest a missing criterion, the user is unsure of the next direction, or a loop
or goal exposes unresolved scope, completion, or stopping criteria. Compare the
request, configured objective, actual state, and results. Fix known errors directly.
+When open-ended work lacks a finish condition, the remaining scope is unclear,
+or repeated attempts bring no new evidence, use align to review the scope and
+suggest a bounded goal when useful. Reuse completed work and check overlapping
+assignments. Keep task lists in the plan and lasting criteria in ALIGNS.md.
+A suggestion does not create or start a goal or loop, change its limits, or
+replace existing authorization. Do not add goal setup to routine bounded tasks.
Before a long run with untested boundaries, offer hypothetical rehearsal without
executing scenarios or launching the run. A review does not authorize a restart,
budget increase, or override of a user stop. The skill acts during active turns;
diff --git a/skills/align/references/recurring.md b/skills/align/references/recurring.md
index 211af64..eeb67c6 100644
--- a/skills/align/references/recurring.md
+++ b/skills/align/references/recurring.md
@@ -48,6 +48,68 @@ scope, supporting evidence and counterexamples. Apply already confirmed criteria
ask only about inferred or conflicting ones. A missed instruction still needs a
repair, not another preference interview.
+## Work scope and goals
+
+Use this review when an open-ended request has no clear finish condition, work
+spans several dependent steps, the remaining scope is uncertain, or repeated
+attempts no longer produce useful evidence. Apply existing agreements first.
+A routine bounded task needs neither a goal proposal nor another interview.
+
+Read the relevant request, project plans, available issue or run state, and work
+already completed or assigned to another agent. Describe the scope supported by
+those sources: delivered work, remaining outcomes, exclusions, dependencies and
+known blockers. Name coverage gaps; do not claim a complete inventory from a
+partial checkout or unavailable tracker. Broader conversation collection still
+uses the source and scope choice in [history review](history.md).
+Reuse verified results and coordinate overlapping assignments instead of doing
+completed work again. Recheck only when changes, failures or uncertainty justify it.
+
+Break substantial work into verifiable outcomes. Explain likely effort and the
+uncertainty that affects it, using observed throughput when available. Do not
+invent precise time estimates, progress percentages or completion evidence.
+Keep the task inventory and run settings in the current plan or supported task
+system. ALIGNS.md holds confirmed lasting criteria, not the task backlog.
+
+### Suggest a goal when it would help
+
+The agent may propose a concrete goal without waiting for the user to ask for one.
+Tie the suggestion to the actual gap, such as an unbounded request to keep
+improving a result. State the intended outcome, included and excluded work,
+completion evidence, and the checkpoint or stopping boundary that matters.
+Recommend one useful scope instead of inventing a roadmap. For example:
+"I can compare the remaining channels and prepare drafts for two suitable ones.
+That would complete this pass. Publishing would be a separate step."
+
+Reuse a clear objective and existing authorization. Ask only about unresolved
+choices that change the outcome or execution boundary. A goal proposal does not
+itself authorize creating or starting a goal, loop, automation or schedule.
+Use the client's supported controls only when their requirements and the user's
+authorization are satisfied. If goal tooling is unavailable, agree on a bounded
+plan without pretending a managed goal exists. Never infer a token, time or cost
+budget, raise an existing limit, or restart a stopped run from this suggestion.
+A declined proposal stays closed until relevant circumstances change.
+
+### Check whether repetition is making progress
+
+At meaningful checkpoints, compare each new attempt with the previous evidence
+and completion criteria. New feedback, a changed hypothesis, a code change, or
+a newly observed failure can justify another attempt. Repeating the same action
+without a changed input, new evidence or a testable reason to expect progress
+calls for a checkpoint. This includes polishing an already accepted result with
+no remaining target and checking unchanged sources with no revisit condition.
+
+When progress stalls, state what has already been tried and what remains
+unresolved. Reuse completed results, change the approach within authorized scope,
+or propose a smaller scope, missing prerequisite or user decision. Stop the
+unproductive repetition while continuing independent authorized work. Report
+unfinished outcomes accurately; a stall is not completion of the goal.
+Honor user-set limits and client stop controls. Suggest a task-specific attempt
+or effort limit when useful, not a universal retry count. Do not interrupt a
+justified recovery or demand approval for each ordinary iteration. A requested
+monitoring task may legitimately check for change at its agreed interval.
+This review operates during active turns and cannot impose a runtime limit on
+another process or wake a stopped agent.
+
## Before a long autonomous loop
Read existing permission, scope, budget, checkpoint and stopping agreements.