claude-alfred 0.2.0 → 0.3.1

package/README.ja.md

@@ -24,7 +24,7 @@ alfred は3つとも潰す。
 
 **適応する仕様。** 小さなバグ修正なら3ファイル。中規模なら5。大規模なら7。タスクに応じてスペックの深さを自動調整。bugfix専用テンプレート（再現手順、原因分析、修正戦略）も。
 
-**積み重なる記憶。** 「あの時こう決めた」「このバグ前にも見た」「Xは試したけどダメだった」— 全てが `.alfred/knowledge/` にMarkdownファイルとして保存される。Git フレンドリーで、人間が読めて、チームで共有できる。SQLite検索インデックスがセマンティック検索を提供。矛盾は自動検出。次に似た問題に当たったとき、聞く前に alfred が出してくる。
+**積み重なる記憶。** 「あの時こう決めた」「このバグ前にも見た」「Xは試したけどダメだった」— 全てが `.alfred/knowledge/` に構造化JSONとして保存される。3分類のみ: **decision**（一回きりの選択 + 理由 + 却下した代替案）、**pattern**（繰り返し実践 + 適用条件 + 期待結果）、**rule**（強制ルール + 優先度 + 根拠）。Git フレンドリーで、チームで共有できる。SQLite検索インデックスがセマンティック検索を提供。矛盾は自動検出。次に似た問題に当たったとき、聞く前に alfred が出してくる。
 
 **信頼性シグナル。** 仕様の全項目にgrounding level（`verified`/`reviewed`/`inferred`/`speculative`）が付く。どの要件が実証済みで、どれが推測なのか一目でわかる。typoも検出する。
 
@@ -36,7 +36,7 @@ alfred は3つとも潰す。
 
 **能動的スキル提案。** alfred は呼ばれるのを待たない。調査中か、設計中か、実装中か、バグ修正中かを検出し、適切なタイミングで適切なスキルを提案する。コードを探索し続けてる？「`/alfred:survey` 使ったら？」。調査結果が出た？「`ledger` に保存しよう」。タスクが溜まってきた？「`roster` でまとめよう」。
 
-**バイパスできない承認ゲート。** 実装前に仕様のレビューサイクルを回す。ブラウザダッシュボードで任意の行にコメントして、承認か差し戻し — GitHub の PR レビューと同じ体験を、仕様に対してやる。ステータスの手動書き換えでは突破できない。レビューファイルの存在まで検証する。
+**バイパスできない承認ゲート。** 実装前に仕様のレビューサイクルを回す。ブラウザダッシュボードで任意の行にコメントして、承認か差し戻し — GitHub の PR レビューと同じ体験を、仕様に対してやる。ステータスの手動書き換えでは突破できない。レビューファイルの存在まで検証する。**ハード強制**: PreToolUse フックが未承認 M/L/XL スペックへの Edit/Write を物理的にブロック。Stop フックがタスク完了・セルフレビュー・spec close まで Claude の停止を阻止。
 
 **張り付くプロジェクトコンテキスト。** Steering文書（プロダクトの目的、コード構造、技術スタック）がプロジェクトから自動生成され、すべての仕様に注入される。AIは常にアーキテクチャを理解している。
 
@@ -123,7 +123,7 @@ npm update -g claude-alfred        # CLI、hooks、MCP サーバー、ダッシ
 |--------|------|
 | `dossier` | 仕様のライフサイクル管理 — init（サイズ/タイプ指定可）, update, status, switch, complete, delete, history, rollback, review, validate |
 | `roster` | エピック管理 — タスクのグループ化、依存関係、進捗追跡 |
-| `ledger` | ナレッジ — search, save（Markdown+frontmatter）, promote, reflect, audit-conventions |
+| `ledger` | ナレッジ — search, save（構造化JSON: decision/pattern/rule）, promote（pattern→rule）, reflect, audit-conventions |
 
 ## Hook
 
@@ -131,10 +131,12 @@ npm update -g claude-alfred        # CLI、hooks、MCP サーバー、ダッシ
 
 | イベント | 動作 |
 |----------|------|
-| SessionStart | 仕様コンテキスト復元 + CLAUDE.md 取込 + プロジェクト成熟度に応じた注入量調整 + ナレッジ健康チェック提案 |
-| PreCompact | 決定抽出 → 構造化チャプターメモリ (Markdown) 保存 → エピック進捗同期 → 調査パターン検出・保存リマインド |
-| UserPromptSubmit | セマンティック検索 + ファイルコンテキストブースト + **スキルナッジ**（意図検出 → 適切なスキル提案）|
-| PostToolUse | Bash エラー検出 → 過去の修正を検索。コミット後: スペックドリフト検出。Read/Grep 5回以上: survey 提案 |
+| SessionStart | 仕様コンテキスト復元 + ナレッジ同期 + 1%ルール（スキル発動促進）+ 成熟度適応 |
+| PreCompact | スナップショット保存 + 決定抽出 + エピック進捗同期 + 調査パターン検出 |
+| UserPromptSubmit | セマンティック検索 + スキルナッジ + **spec承認ゲート**（未承認 M/L/XL に DIRECTIVE） |
+| PostToolUse | エラー検出 + Next Steps 自動チェック + ドリフト検出 + コミット時決定保存 |
+| **PreToolUse** | **ハード強制**: 未承認 M/L/XL スペックへの Edit/Write をブロック。`.alfred/` 編集と Read/Grep は常に許可 |
+| **Stop** | **品質ゲート**: Next Steps 全チェック + Wave セルフレビュー + dossier complete 完了まで停止を阻止 |
 
 ## ブラウザダッシュボード
 
@@ -147,7 +149,7 @@ alfred dashboard --url-only   # URLだけ出力
 | タブ | 表示内容 |
 |------|----------|
 | **Overview** | プロジェクトの健康状態 — タスク進捗とバリデーション結果、メモリ健康度（陳腐化数・矛盾数）、仕様の信頼度分布、エピック進捗、最近の意思決定 |
-| **Tasks** | タスクカード: 現在のフォーカス、次のアクション、エピック、バリデーションバッジ。仕様ファイルの閲覧。Review タブでインラインコメント |
+| **Tasks** | Active/Completed セクション分離。タスクをクリックで2カラム詳細ビュー: 左にメタデータ、右に折りたたみ可能なspecセクション（色分け付き）。Review タブでインラインコメント |
 | **Knowledge** | メモリ一覧（サブタイプ別タグ付き）。セマンティック検索（Voyage AI、300msデバウンス）。ローカルテキストフィルタ。メモリの有効/無効切り替え |
 | **Activity** | 操作タイムライン。イベントタイプ別フィルタ（init/complete/review）。エピックドリルダウン |
 
@@ -171,21 +173,32 @@ alfred dashboard --url-only   # URLだけ出力
 
 ## ナレッジアーキテクチャ
 
-ナレッジはYAMLフロントマター付きのMarkdownファイルとして保存される。真のソースはプロジェクトディレクトリにあり、バイナリDBではない。
+ナレッジは構造化JSONファイルとして保存される。3分類、曖昧さゼロ。真のソースはプロジェクトディレクトリにあり、バイナリDBではない。
 
 ```
 .alfred/knowledge/
-├── decisions/dec-auth-jwt.md    # 設計判断
-├── patterns/pat-error-handling.md  # 再利用パターン
-└── rules/rul-no-mock-db.md     # 強制ルール
+├── decisions/
+│   └── dec-auth-jwt.json        # 一回きりの選択 + 理由 + 却下した代替案
+├── patterns/
+│   └── pat-error-handling.json  # 繰り返し実践 + 適用条件 + 期待結果
+└── rules/
+    └── rule-no-mock-db.json     # 強制ルール + 優先度 + 根拠
 ```
 
+各タイプには厳密なスキーマ（[mneme](https://github.com/hir4ta/mneme) 互換）:
+- **Decision**: `title`, `decision`, `reasoning`, `alternatives[]`, `tags[]`, `status`
+- **Pattern**: `type` (good/bad/error-solution), `context`, `pattern`, `applicationConditions`, `expectedOutcomes`
+- **Rule**: `key`, `text`（命令形）, `category`, `priority` (p0/p1/p2), `rationale`, `sourceRef`
+
+全エントリはテンプレート化されたパラメータで保存（フリーテキスト不可）— セッション間のフォーマット揺れゼロ。
+
 - **Git フレンドリー**: ナレッジをコミットしてチームと共有、PRでレビュー
-- **人間が読める**: `cat` するだけで alfred が何を知っているか確認できる
+- **アトミック書き込み**: temp ファイル + rename でクラッシュ時の破損防止
 - **再構築可能**: SQLite検索インデックスはこれらのファイルから派生 — DB を削除しても次のセッションで再構築される
-- **サブタイプ減衰**: 仮定は30日で腐る。実証されたルールは120日持つ。メモリの種類ごとに半減期が違う。
+- **サブタイプ減衰**: パターンは90日で減衰。実証されたルールは120日持つ。種類ごとに半減期が違う。
+- **昇格**: パターンは検索ヒット15回以上でルールに自動昇格
 - **矛盾検出**: 2つのエントリが矛盾していたら（「JWT使え」vs「JWT避けろ」）、自動でフラグ。
-- **プロジェクト紐付け**: ナレッジは git リモートURLとブランチでタグ付けされ、プロジェクト横断検索が可能。
+- **多言語対応**: `ALFRED_LANG` で保存されるナレッジの言語を制御
 
 ## 適応的スペック
 
@@ -239,14 +252,16 @@ alfred dashboard --url-only   # URLだけ出力
   |
   v
 Hook（見えない）
-  |-- SessionStart     -> コンテキスト復元、プロジェクト成熟度に適応
-  |-- PreCompact       -> 決定をMarkdown保存、チャプターメモリ、エピック進捗
-  |-- UserPromptSubmit -> ベクトル検索 + FTS5 + ファイルブースト → メモリ注入
-  |-- PostToolUse      -> エラー検出 → 関連する過去の修正を提示
+  |-- SessionStart     -> コンテキスト復元、1%ルール、成熟度適応
+  |-- PreCompact       -> スナップショット保存、決定抽出、エピック進捗
+  |-- UserPromptSubmit -> ベクトル検索 + FTS5 + スキルナッジ + spec承認チェック
+  |-- PostToolUse      -> エラー検出、Next Steps自動チェック、ドリフト検出
+  |-- PreToolUse       -> 未承認 M/L/XL への Edit/Write をブロック
+  |-- Stop             -> 全タスク完了 + セルフレビュー + complete まで停止阻止
   |
   v
 ストレージ
-  |-- .alfred/knowledge/   -> Markdown+frontmatter（真のソース、Git フレンドリー）
+  |-- .alfred/knowledge/   -> JSON（decisions/, patterns/, rules/）— 真のソース
   |-- .alfred/specs/       -> 仕様ファイル + バージョン履歴 + レビュー
   |-- .alfred/epics/       -> エピック YAML + タスク依存関係
   |-- .alfred/steering/    -> プロジェクトコンテキスト（product, structure, tech）
@@ -266,6 +281,7 @@ Hook（見えない）
 | `.alfred/epics/` | 最初のエピック作成時 | `roster action=init` |
 | `.alfred/steering/` | `/alfred:init` 実行時 | プロジェクト初期化スキル |
 | `.alfred/templates/` | ユーザーが仕様・Steeringテンプレートをカスタマイズする時 | テンプレートオーバーライド用に手動作成 |
+| `.alfred/.state/` | 最初のセッションローカル状態保存時 | ナッジ抑制カウント、探索カウンター（gitignore対象） |
 | `.alfred/audit.jsonl` | 最初の仕様操作時またはコミット後ドリフト検出時 | `dossier init`、レビュー送信、PostToolUse ドリフト |
 
 ## トラブルシューティング

package/README.md

@@ -24,7 +24,7 @@ alfred fixes all three.
 
 **Specs that adapt.** Small bug? 3 files. Medium feature? 5. Large system? All 7. alfred auto-detects the right scope — or you can pick a bugfix template with surgical precision (reproduction steps, root cause, fix strategy).
 
-**Memory that compounds.** Every decision, every bug fix, every "we tried X and it didn't work" gets stored as Markdown files in `.alfred/knowledge/` — git-friendly, human-readable, team-shareable. A SQLite search index provides semantic search across all knowledge. Contradictions are detected automatically. Next time you hit a similar problem, alfred surfaces the relevant experience — before you even ask.
+**Memory that compounds.** Every decision, every bug fix, every "we tried X and it didn't work" gets stored as structured JSON files in `.alfred/knowledge/` — three types only: **decisions** (one-time choices with reasoning and rejected alternatives), **patterns** (repeatable practices with conditions and expected outcomes), and **rules** (enforceable standards with priority and rationale). Git-friendly, human-readable, team-shareable. A SQLite search index provides semantic search across all knowledge. Contradictions are detected automatically. Next time you hit a similar problem, alfred surfaces the relevant experience — before you even ask.
 
 **Reliability signals.** Every spec item gets a grounding level — `verified`, `reviewed`, `inferred`, or `speculative`. You can instantly see which requirements are battle-tested and which are guesswork. Typos in grounding values get caught, not silently ignored.
 
@@ -36,7 +36,7 @@ alfred fixes all three.
 
 **Proactive skill suggestions.** alfred doesn't wait to be asked. It detects what you're doing — researching, designing, implementing, fixing bugs — and suggests the right skill at the right time. Explored code for a while? "Try `/alfred:survey`." Got research findings? "Save them with `ledger`." Three tasks piling up? "Group them with `roster`."
 
-**Approval gates that can't be bypassed.** Specs go through a review cycle before implementation. Comment on any line in the browser dashboard, approve or request changes — like a GitHub PR review, but for your specs. The gate verifies both the review status *and* the existence of a signed review file, so manually editing the status won't get you past it.
+**Approval gates that can't be bypassed.** Specs go through a review cycle before implementation. Comment on any line in the browser dashboard, approve or request changes — like a GitHub PR review, but for your specs. The gate verifies both the review status *and* the existence of a signed review file, so manually editing the status won't get you past it. **Hard enforcement**: PreToolUse hooks physically block Edit/Write operations on unapproved M/L/XL specs. Stop hooks prevent Claude from finishing until all tasks are checked off, self-reviews are done, and the spec is closed.
 
 **Project context that sticks.** Steering documents (product purpose, code structure, tech stack) are auto-generated from your project and injected into every spec. Your AI always knows your architecture.
 
@@ -123,7 +123,7 @@ Run `alfred doctor` to verify both are in sync.
 |------|---------|
 | `dossier` | Spec lifecycle — init (with size/type), update, status, switch, complete, delete, history, rollback, review, validate |
 | `roster` | Epic management — group tasks with dependencies, track progress |
-| `ledger` | Knowledge — search, save (Markdown+frontmatter), promote, reflect, audit-conventions |
+| `ledger` | Knowledge — search, save (structured JSON: decision/pattern/rule), promote (pattern→rule), reflect, audit-conventions |
 
 ## Hooks
 
@@ -131,10 +131,12 @@ Run automatically. You don't touch these.
 
 | Event | What happens |
 |-------|-------------|
-| SessionStart | Restores spec context, ingests CLAUDE.md, adapts injection depth to project maturity, suggests `ledger reflect` when knowledge base needs attention |
-| PreCompact | Extracts decisions, saves structured chapter memory (Markdown), syncs epic progress, detects research patterns and reminds to save knowledge |
-| UserPromptSubmit | Semantic search + file context boost + **skill nudge** (detects intent → suggests the right skill) |
-| PostToolUse | Detects Bash errors + searches memory for similar past fixes. After commits: spec drift detection. After 5+ Read/Grep: suggests `/alfred:survey` |
+| SessionStart | Restores spec context, syncs knowledge index, adapts injection depth to project maturity, 1% rule skill activation |
+| PreCompact | Extracts decisions, saves chapter snapshots, syncs epic progress, detects research patterns |
+| UserPromptSubmit | Semantic search + file context boost + **skill nudge** + **spec approval gate** (blocks implement intent on unapproved M/L/XL specs) |
+| PostToolUse | Detects Bash errors + searches memory. After commits: spec drift detection + auto-save decisions. Edit/Write: auto-check Next Steps progress |
+| **PreToolUse** | **Hard enforcement**: blocks Edit/Write when active M/L/XL spec is not approved. `.alfred/` edits and Read/Grep always allowed |
+| **Stop** | **Quality gate**: blocks Claude from stopping until all Next Steps are checked, Wave self-reviews are done, and `dossier complete` is called |
 
 ## Browser dashboard
 
@@ -147,7 +149,7 @@ alfred dashboard --url-only   # print URL only
 | Tab | What you see |
 |-----|-------------|
 | **Overview** | Project health at a glance — task progress with validation badges, memory health (stale count, conflicts), confidence distribution across specs, epic progress, recent decisions |
-| **Tasks** | Enriched task cards: current focus, next action, epic context, validation status. Drill into spec files. Switch to Review tab for inline comments |
+| **Tasks** | Active/Completed sections. Drill into any task for a 2-column detail view: metadata on the left, collapsible spec sections (color-coded) on the right. Switch to Review tab for inline comments |
 | **Knowledge** | Browse all memories with sub-type tags. Semantic search (Voyage AI) with 300ms debounce. Local text filter. Toggle any memory on/off |
 | **Activity** | Timeline of all operations. Filter by event type (init/complete/review). Epic drill-down with task status badges |
 
@@ -171,21 +173,32 @@ Fuzzy matching catches typos: "authetication" still finds "authentication".
 
 ## Knowledge architecture
 
-Knowledge is stored as Markdown files with YAML frontmatter — the source of truth lives in your project directory, not a binary database.
+Knowledge is stored as structured JSON files — three types, no ambiguity. The source of truth lives in your project directory, not a binary database.
 
 ```
 .alfred/knowledge/
-├── decisions/dec-auth-jwt.md    # design decisions
-├── patterns/pat-error-handling.md  # reusable patterns
-└── rules/rul-no-mock-db.md     # enforced standards
+├── decisions/
+│   └── dec-auth-jwt.json        # one-time choices with reasoning + rejected alternatives
+├── patterns/
+│   └── pat-error-handling.json  # repeatable practices with conditions + expected outcomes
+└── rules/
+    └── rule-no-mock-db.json     # enforceable standards with priority + rationale
 ```
 
+Each type has a strict schema (inspired by [mneme](https://github.com/hir4ta/mneme)):
+- **Decisions**: `title`, `decision`, `reasoning`, `alternatives[]`, `tags[]`, `status`
+- **Patterns**: `type` (good/bad/error-solution), `context`, `pattern`, `applicationConditions`, `expectedOutcomes`
+- **Rules**: `key`, `text` (imperative), `category`, `priority` (p0/p1/p2), `rationale`, `sourceRef`
+
+All entries are saved via templated parameters (no freetext) — zero format drift across sessions.
+
 - **Git-friendly**: commit knowledge to share with your team, review in PRs
-- **Human-readable**: `cat` any file to see exactly what alfred knows
+- **Atomic writes**: temp file + rename prevents corruption on crash
 - **Rebuildable**: the SQLite search index is derived from these files — delete the DB, it rebuilds on next session
-- **Sub-type decay**: Assumptions decay in 30 days. Proven rules last 120 days. Each knowledge type has its own half-life.
+- **Sub-type decay**: Patterns decay in 90 days. Proven rules last 120 days. Each type has its own half-life.
+- **Promotion**: patterns auto-promote to rules at 15+ search hits
 - **Contradiction detection**: When two entries say opposite things ("use JWT" vs "avoid JWT"), alfred flags the conflict.
-- **Project-aware**: knowledge is tagged with git remote URL and branch, enabling cross-project search.
+- **Multilingual**: `ALFRED_LANG` controls the language of saved knowledge content
 
 ## Adaptive specs
 
@@ -239,14 +252,16 @@ You
   |
   v
 Hooks (invisible)
-  |-- SessionStart     -> restore context, adapt to project maturity
-  |-- PreCompact       -> save decisions as JSON, chapter memory, epic progress
-  |-- UserPromptSubmit -> vector search + FTS5 + file boost -> inject memories
-  |-- PostToolUse      -> detect errors -> surface related past fixes
+  |-- SessionStart     -> restore context, sync knowledge, 1% rule, adapt to project maturity
+  |-- PreCompact       -> save snapshots, extract decisions, epic progress
+  |-- UserPromptSubmit -> vector search + FTS5 + skill nudge + spec approval check
+  |-- PostToolUse      -> detect errors, auto-check Next Steps, drift detection
+  |-- PreToolUse       -> BLOCK Edit/Write on unapproved M/L/XL specs
+  |-- Stop             -> BLOCK stop until all tasks done + self-review + complete
   |
   v
 Storage
-  |-- .alfred/knowledge/   -> Markdown+frontmatter (source of truth, git-friendly)
+  |-- .alfred/knowledge/   -> JSON (decisions/, patterns/, rules/) — source of truth
   |-- .alfred/specs/       -> spec files + version history + reviews
   |-- .alfred/epics/       -> epic YAML + task dependencies
   |-- .alfred/steering/    -> project context (product, structure, tech)
@@ -266,6 +281,7 @@ Nothing is generated at install time. Files appear as you use alfred:
 | `.alfred/epics/` | First epic is created | `roster action=init` |
 | `.alfred/steering/` | Running `/alfred:init` | Project initialization skill |
 | `.alfred/templates/` | User customizes spec or steering templates | Manual creation for template override |
+| `.alfred/.state/` | First hook that needs session-local state | Nudge dismissals, exploration counter (gitignored) |
 | `.alfred/audit.jsonl` | First spec operation or commit drift detection | `dossier init`, `dossier delete`, review submission, PostToolUse drift |
 
 ## Troubleshooting

package/content/hooks/hooks.json

@@ -48,6 +48,31 @@
           }
         ]
       }
+    ],
+    "PreToolUse": [
+      {
+        "matcher": "Edit|Write",
+        "hooks": [
+          {
+            "command": "alfred hook PreToolUse",
+            "statusMessage": "alfred: checking spec approval...",
+            "timeout": 5,
+            "type": "command"
+          }
+        ]
+      }
+    ],
+    "Stop": [
+      {
+        "hooks": [
+          {
+            "command": "alfred hook Stop",
+            "statusMessage": "alfred: checking completion...",
+            "timeout": 5,
+            "type": "command"
+          }
+        ]
+      }
     ]
   }
 }

package/dist/{audit-DM4QaS4L.mjs → audit-BrTztz4z.mjs}

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { E as require_dist, S as rootDir, m as VALID_SLUG } from "./knowledge-DACIiaOh.mjs";
+import { E as require_dist, S as rootDir, m as VALID_SLUG } from "./knowledge-DItYbraO.mjs";
 import { appendFileSync, mkdirSync, readFileSync, readdirSync, renameSync, rmSync, statSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 //#region src/epic/index.ts
@@ -39,7 +39,7 @@ var EpicDir = class {
 	}
 	save(ep) {
 		const data = (0, import_dist.stringify)(ep);
-		const tmp = this.epicPath() + ".tmp";
+		const tmp = `${this.epicPath()}.tmp`;
 		writeFileSync(tmp, data);
 		renameSync(tmp, this.epicPath());
 	}
@@ -161,9 +161,7 @@ function listAllEpics(projectPath) {
 				total: tasks.length,
 				tasks
 			});
-		} catch {
-			continue;
-		}
+		} catch {}
 	}
 	return summaries;
 }
@@ -219,7 +217,7 @@ function writeActiveEpics(projectPath, state) {
 	mkdirSync(epicsDir(projectPath), { recursive: true });
 	const data = (0, import_dist.stringify)(state);
 	const path = epicActivePath(projectPath);
-	const tmp = path + ".tmp";
+	const tmp = `${path}.tmp`;
 	writeFileSync(tmp, data);
 	renameSync(tmp, path);
 }
@@ -233,7 +231,7 @@ function appendAudit(projectPath, entry) {
 			...entry,
 			timestamp: (/* @__PURE__ */ new Date()).toISOString()
 		});
-		appendFileSync(join(dir, "audit.jsonl"), line + "\n");
+		appendFileSync(join(dir, "audit.jsonl"), `${line}\n`);
 	} catch {}
 }
 //#endregion

package/dist/cli.mjs

@@ -369,9 +369,9 @@ const main = defineCommand({
 		serve: defineCommand({
 			meta: { description: "Start MCP server (stdio)" },
 			async run() {
-				const { Store } = await import("./store-wuybYYaO.mjs");
-				const { Embedder } = await import("./embedder-qPO74Ar-.mjs");
-				const { serveMCP } = await import("./server-HMIKVikr.mjs");
+				const { Store } = await import("./store-CuN3_dWu.mjs");
+				const { Embedder } = await import("./embedder-BRG2Xpr4.mjs");
+				const { serveMCP } = await import("./server-23A-5yhs.mjs");
 				const store = Store.openDefault();
 				let emb = null;
 				try {
@@ -397,9 +397,9 @@ const main = defineCommand({
 				}
 			},
 			async run({ args }) {
-				const { Store } = await import("./store-wuybYYaO.mjs");
-				const { Embedder } = await import("./embedder-qPO74Ar-.mjs");
-				const { startDashboard } = await import("./server-BVBSt-GA.mjs");
+				const { Store } = await import("./store-CuN3_dWu.mjs");
+				const { Embedder } = await import("./embedder-BRG2Xpr4.mjs");
+				const { startDashboard } = await import("./server-Dfefx0Gm.mjs");
 				const projectPath = process.cwd();
 				const store = Store.openDefault();
 				let emb = null;
@@ -422,7 +422,7 @@ const main = defineCommand({
 				description: "Event name"
 			} },
 			async run({ args }) {
-				const { runHook } = await import("./dispatcher-DNBIBPJS.mjs").then((n) => n.t);
+				const { runHook } = await import("./dispatcher-xZAhoK7h.mjs").then((n) => n.t);
 				await runHook(args.event);
 			}
 		}),
@@ -456,8 +456,8 @@ const main = defineCommand({
 				}
 				const dbPath = join(home, ".claude-alfred", "alfred.db");
 				check(existsSync(dbPath), `DB: ${dbPath}`, "not found — run: alfred (any command) to create");
-				check(!!process.env["VOYAGE_API_KEY"], "VOYAGE_API_KEY set", "not set — semantic search disabled, FTS5 fallback active");
-				const lang = process.env["ALFRED_LANG"];
+				check(!!process.env.VOYAGE_API_KEY, "VOYAGE_API_KEY set", "not set — semantic search disabled, FTS5 fallback active");
+				const lang = process.env.ALFRED_LANG;
 				check(true, `ALFRED_LANG: ${lang || "(not set, default: en)"}`);
 				const rulesDir = join(home, ".claude", "rules");
 				try {
@@ -498,9 +498,7 @@ async function resolveVersion() {
 		for (const rel of ["..", "../.."]) try {
 			const pkg = JSON.parse(readFileSync(join(thisDir, rel, "package.json"), "utf-8"));
 			if (pkg.version) return pkg.version;
-		} catch {
-			continue;
-		}
+		} catch {}
 	} catch {}
 	return "dev";
 }

package/dist/directives-iFMV2Mmx.mjs

@@ -0,0 +1,57 @@
+#!/usr/bin/env node
+import { n as emitAdditionalContext } from "./dispatcher-xZAhoK7h.mjs";
+//#region src/hooks/directives.ts
+const LEVEL_ORDER = {
+	DIRECTIVE: 0,
+	WARNING: 1,
+	CONTEXT: 2
+};
+const MAX_DIRECTIVES = 3;
+const MAX_DIRECTIVE_BLOCK_CHARS = 500;
+const SPIRIT_VS_LETTER = "Adapting or shortcutting these steps violates the rule, even if you believe the spirit is preserved.";
+/**
+* Build a single additionalContext string from directive items.
+* Order: DIRECTIVE → WARNING → CONTEXT.
+* Max 3 DIRECTIVE items (NFR-5). Excess DIRECTIVEs downgraded to WARNING.
+*/
+function buildDirectiveOutput(items) {
+	if (items.length === 0) return "";
+	let directiveCount = 0;
+	return items.map((item) => {
+		if (item.level === "DIRECTIVE") {
+			directiveCount++;
+			if (directiveCount > MAX_DIRECTIVES) return {
+				level: "WARNING",
+				message: item.message
+			};
+		}
+		return item;
+	}).slice().sort((a, b) => LEVEL_ORDER[a.level] - LEVEL_ORDER[b.level]).map((item) => {
+		let block = `[${item.level}] ${item.message}`;
+		if (item.level === "DIRECTIVE") {
+			const rats = (item.rationalizations ?? []).slice();
+			const suffix = item.spiritVsLetter ? `\n${SPIRIT_VS_LETTER}` : "";
+			while (rats.length > 0) {
+				if (`${block}\n${rats.map((r) => `- ${r}`).join("\n")}${suffix}`.length <= MAX_DIRECTIVE_BLOCK_CHARS) break;
+				rats.pop();
+			}
+			if (rats.length > 0) block += `\n${rats.map((r) => `- ${r}`).join("\n")}`;
+			block += suffix;
+			if (block.length > MAX_DIRECTIVE_BLOCK_CHARS) {
+				const messageOnly = `[${item.level}] ${item.message}`;
+				if (messageOnly.length > MAX_DIRECTIVE_BLOCK_CHARS) block = `${messageOnly.slice(0, MAX_DIRECTIVE_BLOCK_CHARS - 3)}...`;
+				else block = messageOnly;
+			}
+		}
+		return block;
+	}).join("\n");
+}
+/**
+* Emit directives via single emitAdditionalContext call (NFR-4).
+*/
+function emitDirectives(eventName, items) {
+	const output = buildDirectiveOutput(items);
+	if (output) emitAdditionalContext(eventName, output);
+}
+//#endregion
+export { emitDirectives as t };

package/dist/{dispatcher-DNBIBPJS.mjs → dispatcher-xZAhoK7h.mjs}

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { O as __exportAll } from "./knowledge-DACIiaOh.mjs";
+import { O as __exportAll } from "./knowledge-DItYbraO.mjs";
 import { resolve } from "node:path";
 //#region src/hooks/dispatcher.ts
 var dispatcher_exports = /* @__PURE__ */ __exportAll({
@@ -19,7 +19,7 @@ function emitAdditionalContext(eventName, context) {
 		additionalContext: context
 	} };
 	try {
-		process.stdout.write(JSON.stringify(out) + "\n");
+		process.stdout.write(`${JSON.stringify(out)}\n`);
 	} catch {}
 }
 function extractSection(content, heading) {
@@ -27,7 +27,7 @@ function extractSection(content, heading) {
 	const result = [];
 	let inSection = false;
 	for (const line of lines) {
-		if (line === heading || line.startsWith(heading + " ")) {
+		if (line === heading || line.startsWith(`${heading} `)) {
 			inSection = true;
 			continue;
 		}
@@ -51,13 +51,15 @@ async function runHook(event) {
 	} catch {
 		return;
 	}
-	if (ev.stop_hook_active) return;
+	if (ev.stop_hook_active && event !== "Stop" && event !== "PreToolUse") return;
 	if (ev.cwd) ev.cwd = resolve(ev.cwd);
 	const timeout = {
 		SessionStart: 4500,
 		PreCompact: 9e3,
 		UserPromptSubmit: 9e3,
-		PostToolUse: 5e3
+		PostToolUse: 5e3,
+		PreToolUse: 4500,
+		Stop: 4500
 	}[event] ?? 5e3;
 	const controller = new AbortController();
 	const timer = setTimeout(() => controller.abort(), timeout);
@@ -75,26 +77,40 @@ async function runHook(event) {
 			case "PostToolUse":
 				await handlePostToolUse(ev, controller.signal);
 				break;
+			case "PreToolUse":
+				await handlePreToolUse(ev, controller.signal);
+				break;
+			case "Stop":
+				await handleStop(ev, controller.signal);
+				break;
 		}
 	} finally {
 		clearTimeout(timer);
 	}
 }
 async function handleSessionStart(ev, signal) {
-	const { sessionStart } = await import("./session-start-DCp2_K0Z.mjs");
+	const { sessionStart } = await import("./session-start-D177nOGO.mjs");
 	await sessionStart(ev, signal);
 }
 async function handlePreCompact(ev, signal) {
-	const { preCompact } = await import("./pre-compact-BtHyMSFR.mjs");
+	const { preCompact } = await import("./pre-compact-CyNPLjSK.mjs");
 	await preCompact(ev, signal);
 }
 async function handleUserPromptSubmit(ev, signal) {
-	const { userPromptSubmit } = await import("./user-prompt-BZ9VhKi6.mjs");
+	const { userPromptSubmit } = await import("./user-prompt-X18njRds.mjs");
 	await userPromptSubmit(ev, signal);
 }
 async function handlePostToolUse(ev, signal) {
-	const { postToolUse } = await import("./post-tool-DCMLZ8BU.mjs");
+	const { postToolUse } = await import("./post-tool-G2PAotAa.mjs");
 	await postToolUse(ev, signal);
 }
+async function handlePreToolUse(ev, _signal) {
+	const { preToolUse } = await import("./pre-tool-B_CYjJPV.mjs");
+	await preToolUse(ev);
+}
+async function handleStop(ev, _signal) {
+	const { stop } = await import("./stop-uQEMYJUS.mjs");
+	await stop(ev);
+}
 //#endregion
 export { notifyUser as i, emitAdditionalContext as n, extractSection as r, dispatcher_exports as t };

package/dist/{embedder-qPO74Ar-.mjs → embedder-BRG2Xpr4.mjs}

@@ -12,7 +12,7 @@ function envIntOr(key, fallback) {
 	const v = process.env[key];
 	if (v) {
 		const n = parseInt(v, 10);
-		if (!isNaN(n) && n > 0) return n;
+		if (!Number.isNaN(n) && n > 0) return n;
 	}
 	return fallback;
 }
@@ -82,7 +82,6 @@ async function retryVoyage(fn, signal) {
 				if (err.status === 400 && (isVoyageTransient(err.detail) || isVoyageTransientStructured(err.errResp))) continue;
 				throw err;
 			}
-			continue;
 		}
 	}
 	throw lastErr;
@@ -116,7 +115,7 @@ var VoyageClient = class {
 			method: "POST",
 			headers: {
 				"Content-Type": "application/json",
-				"Authorization": `Bearer ${this.apiKey}`
+				Authorization: `Bearer ${this.apiKey}`
 			},
 			body: JSON.stringify(payload),
 			signal: signal ?? AbortSignal.timeout(3e4)
@@ -157,7 +156,7 @@ var VoyageClient = class {
 			method: "POST",
 			headers: {
 				"Content-Type": "application/json",
-				"Authorization": `Bearer ${this.apiKey}`
+				Authorization: `Bearer ${this.apiKey}`
 			},
 			body: JSON.stringify(payload),
 			signal: signal ?? AbortSignal.timeout(3e4)
@@ -185,7 +184,7 @@ var Embedder = class Embedder {
 		this.client = client;
 	}
 	static create() {
-		const apiKey = process.env["VOYAGE_API_KEY"];
+		const apiKey = process.env.VOYAGE_API_KEY;
 		if (!apiKey) throw new Error("VOYAGE_API_KEY is required but not set (get a key at https://dash.voyageai.com/)");
 		return new Embedder(new VoyageClient(apiKey));
 	}

package/dist/{fts-WNIfvJ7g.mjs → fts-CDEAxhSV.mjs}

@@ -1,12 +1,12 @@
 #!/usr/bin/env node
-import { c as mapRow, r as getKnowledgeByIDs, u as searchKnowledgeKeyword } from "./knowledge-DACIiaOh.mjs";
-import { n as deserializeFloat32, t as cosineSimilarity } from "./vectors-DByIf3R2.mjs";
+import { c as mapRow, r as getKnowledgeByIDs, u as searchKnowledgeKeyword } from "./knowledge-DItYbraO.mjs";
+import { n as deserializeFloat32, t as cosineSimilarity } from "./vectors-C_Yj-4sR.mjs";
 //#region src/store/fts.ts
 function subTypeHalfLife(subType) {
 	switch (subType) {
 		case "assumption": return 30;
 		case "inference": return 45;
-		case "general": return 60;
+		case "snapshot": return 30;
 		case "pattern": return 90;
 		case "decision": return 90;
 		case "rule": return 120;

package/dist/{helpers-BR57P8BX.mjs → helpers-DVDJYnXA.mjs}

@@ -1,19 +1,19 @@
 #!/usr/bin/env node
-import { r as getKnowledgeByIDs, s as incrementHitCount, u as searchKnowledgeKeyword } from "./knowledge-DACIiaOh.mjs";
-import { r as vectorSearchKnowledge } from "./vectors-DByIf3R2.mjs";
-import { i as subTypeHalfLife, n as searchKnowledgeFTS, r as subTypeBoost } from "./fts-WNIfvJ7g.mjs";
+import { r as getKnowledgeByIDs, s as incrementHitCount, u as searchKnowledgeKeyword } from "./knowledge-DItYbraO.mjs";
+import { r as vectorSearchKnowledge } from "./vectors-C_Yj-4sR.mjs";
+import { i as subTypeHalfLife, n as searchKnowledgeFTS, r as subTypeBoost } from "./fts-CDEAxhSV.mjs";
 //#region src/mcp/helpers.ts
 const RECENCY_FLOOR = .5;
 function truncate(s, maxLen) {
 	const runes = [...s];
 	if (runes.length <= maxLen) return s;
-	return runes.slice(0, maxLen).join("") + "...";
+	return `${runes.slice(0, maxLen).join("")}...`;
 }
 function recencyFactor(createdAt, subType, now) {
 	const halfLife = subTypeHalfLife(subType);
 	if (halfLife <= 0) return 1;
 	const parsed = Date.parse(createdAt);
-	if (isNaN(parsed)) return 1;
+	if (Number.isNaN(parsed)) return 1;
 	const ageDays = (now.getTime() - parsed) / (1e3 * 60 * 60 * 24);
 	if (ageDays <= 0) return 1;
 	const factor = Math.exp(-Math.LN2 * ageDays / halfLife);
@@ -55,7 +55,7 @@ async function searchPipeline(store, emb, query, limit, overRetrieve, precompute
 			rawDocs = ordered;
 			res.searchMethod = "vector";
 			if (rawDocs.length > limit) try {
-				const contents = rawDocs.map((d) => d.title + "\n" + d.content);
+				const contents = rawDocs.map((d) => `${d.title}\n${d.content}`);
 				const reranked = await emb.rerank(query, contents, limit);
 				if (reranked.length > 0) {
 					const reorderedDocs = [];
@@ -84,6 +84,7 @@ async function searchPipeline(store, emb, query, limit, overRetrieve, precompute
 			}
 		}
 	}
+	rawDocs = rawDocs.filter((d) => d.subType !== "snapshot");
 	res.scoredDocs = applyRecencySignal(rawDocs, res.searchMethod, /* @__PURE__ */ new Date());
 	if (res.scoredDocs.length > limit) res.scoredDocs = res.scoredDocs.slice(0, limit);
 	return res;