kenaz 0.0.0-stage → 0.3.2

package/core/dev/archive/ansuz/legacy/phases/phase_8.yaml

@@ -0,0 +1,146 @@
+phase: 8
+name: "陣型判斷 + Dispatch"
+goal: "選擇最輕量可行陣型，dispatch agent 執行"
+
+judgment:
+  principle: "能 unit 不 solo，能 solo 不 squad，能 squad 不 legion"
+  guide:
+    unit: "目標明確、單一 scope → do/think/plan"
+    solo: "需要多角度分析 → strategy/diagnosis"
+    squad: "可拆成獨立子任務並行 → blitz/sweep"
+    legion: "有順序依賴或即時協調 → forge/hive"
+
+execution:
+  tool: ".formation/tools/headless.js"
+  max_parallel: 5
+  cli:
+    unit: "node {CORE}/.formation/tools/headless.js --skill=do --task=TASK_XXX [--project-path {PROJECT_PATH}]"
+    solo: "node {CORE}/.formation/tools/headless.js --formation=solo --skill=strategy --task=TASK_XXX [--project-path {PROJECT_PATH}]"
+    squad: "node {CORE}/.formation/tools/headless.js --formation=squad --skill=blitz --task=TASK_XXX [--project-path {PROJECT_PATH}]"
+    legion: "node {CORE}/.formation/tools/headless.js --formation=legion --skill=forge --task=TASK_XXX [--project-path {PROJECT_PATH}]"
+  patterns:
+    single: "Bash(run_in_background=true): node headless.js --task=TASK_XXX"
+    parallel: "多個無依賴 → 多個 Bash(background) 同時啟動（上限 5，超過排隊）"
+    sequential: "有依賴 → 前一個完成再 dispatch"
+  anti_pattern: "❌ 禁止用 Task tool 直接開 subagent — 繞過 Formation 失去 prompt 組裝和 log"
+  post_dispatch_branch_guard:
+    rule: "每個 dispatch 完成（背景任務通知回來）後，必須立即執行 git checkout {user_branch}"
+    why: "headless.js subprocess 的 post-complete 會 checkout 回 user branch，但 Ansuz 的 shell session 可能不同步（尤其 crash 或 timeout 時）"
+    enforcement: "硬性規則 — 不是建議。漏做 = 後續所有操作在錯誤 branch 上"
+    pattern: |
+      # 每個 dispatch 背景通知回來後，第一個動作：
+      git checkout {user_branch}
+      # 然後才讀 log、驗證結果、dispatch 下一個
+
+scope_rules:
+  - "同檔案 → 一個實例處理整個檔案"
+  - "跨檔案 → fan_out，按模組/目錄切"
+  - "派工前確認 scope 無重疊"
+
+dispatch_decision_gate:
+  rule: |
+    ⚠️ DISPATCH vs IN-SESSION — 必須在每次 dispatch 前判斷：
+    分析 / 審計 / 策略 / 討論     → in-session Skill tool（/Amelia、/Huginn 等），不 dispatch
+    實作 / 執行 / 寫代碼 / 長跑   → headless dispatch（headless.js）
+    「sub agent」「讓 X 做」本身不代表 dispatch。判斷依據是任務性質，不是用詞。
+    長 session（已 dispatch ≥3 個任務）時特別警惕 pattern contamination — 重新確認。
+
+pre_dispatch:
+  - "node {CORE}/.collaboration/core/task-api.js [--project-path {PROJECT_PATH}] claim TASK_XXX"
+  - "確認狀態已變為 active"
+  - |
+    # ── Symbol-Level Prediction Recording（TASK_932 升級）──
+    # 對每個即將 dispatch 的任務，在 dispatch 前執行：
+    #
+    # 1. 取得任務的 scope_paths（主要目標檔案/目錄）
+    # 2. 對 scope_paths 中的主要 symbol 執行 kenaz_lsp_callers：
+    #      kenaz_lsp_callers({ symbol: "<primary_symbol>", projectPath: "{PROJECT_PATH}" })
+    # 3. 從 LSP 結果萃取 predicted_symbols（最多 20 個）：
+    #      - callers（呼叫此 symbol 的函數）
+    #      - references（引用此 symbol 的位置）
+    #    合併去重後取前 20 個 symbol name
+    # 4. 將 predicted_symbols 寫入任務 YAML（dispatch 前，作為 TASK_931 比對基線）：
+    #      node {CORE}/.collaboration/core/task-api.js [--project-path {PROJECT_PATH}] \
+    #        update TASK_XXX --field predicted_symbols --value "[\"sym_a\",\"sym_b\",...]"
+    #    或用 technical_notes 暫代（若 task-api 尚未支援 --field predicted_symbols）：
+    #      node {CORE}/.collaboration/core/task-api.js [--project-path {PROJECT_PATH}] \
+    #        update TASK_XXX --technical-notes "predicted_symbols: [sym_a, sym_b, ...]"
+    #
+    # ⚠️ 若 scope_paths 無明確 symbol（例如整個目錄），以目錄中最核心的檔案名稱作為 symbol 查詢
+    # ⚠️ kenaz_lsp_callers 不可用時 → 跳過此步，在 Phase 9 outcome 中標記 predicted_symbols: []
+    # ⚠️ 此步驟產出的 predicted_symbols 供 Phase 9 / TASK_931 做 symbol-level 比對用
+
+monitoring:
+  tools:
+    - "node {CORE}/.formation/tools/log-reader.js --latest unit --tail 30"
+    - "node {CORE}/.formation/tools/log-reader.js --task TASK_XXX --tail 50"
+  rules:
+    - "❌ 禁止用 TaskOutput 讀取 headless 結果 — stdout 可能數萬行，會爆 context"
+    - "✅ 用 log-reader.js 或 Read tool + offset"
+
+actions:
+  - "根據 judgment.principle 選擇最輕量陣型"
+  - "task-api.js claim → Bash(background): node headless.js --task=TASK_XXX"
+  - "多任務無依賴 → 多個 Bash(background) 同時啟動（上限 5）"
+
+# ═══════════════════════════════════════
+#  Completion Tracking Mandate（強制）
+# ═══════════════════════════════════════
+# Ansuz 作為全生命週期引擎，dispatch 後 MUST 自動跑完 Phase 9-10-11。
+# 不可中途插入「還有其他事嗎？」或等用戶推動。
+completion_tracking:
+  mandate: |
+    ❌ 禁止在 dispatch 後向用戶提問或建議停下來
+    ❌ 禁止輸出「需要我做其他事情嗎？」「還要繼續嗎？」等中斷性問題
+    ✅ dispatch 後靜默等待背景通知（不輸出任何文字直到通知到達）
+    ✅ 通知到達 → 自動 Phase 9 → Phase 10 → 有下一個 task? → Phase 8 → 全完成 → Phase 11
+  on_notification:
+    step_1: "git checkout {user_branch}（branch guard — 硬性規則）"
+    step_2: "讀 log（log-reader.js --task TASK_XXX --tail 50）"
+    step_3: "Phase 9 驗證（自動，不問用戶）"
+    step_4: "Phase 10 post-complete（自動）"
+    step_5: "檢查是否有下一個待 dispatch 的 task（相同需求的 task 列表）"
+    step_5a: "有 → 回 Phase 8 dispatch 下一個（自動）"
+    step_5b: "沒有 → Phase 11 回報（此時才向用戶匯報 + 問下一步）"
+  multi_task_loop: |
+    當 Phase 7 建立了多個任務時，Phase 8 進入 dispatch 迴圈：
+    for task in tasks_sorted_by_dependency:
+      claim → dispatch(background) → 等通知 → branch guard → Phase 9 → Phase 10
+    全部完成 → Phase 11（一次回報所有結果）
+  anti_pattern:
+    - "❌ dispatch 一個就報告 → 等用戶 → dispatch 下一個（transactional 斷裂）"
+    - "❌ 通知到了但問用戶「要追蹤嗎？」（Phase 9 是自動的，不需要許可）"
+    - "❌ 多任務時每個任務分開做 Phase 11（應合併成一次回報）"
+
+cli_format: |
+  🚀 Dispatch
+  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  TASK_XXX → unit/do（目標明確、單一 scope）
+  TASK_YYY → unit/do（獨立任務、可並行）
+
+  並行: 2 個 headless 進程
+  監控: node {CORE}/.formation/tools/log-reader.js --latest unit --tail 30
+  ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+interaction:
+  type: "auto"
+  instruction: "輸出 cli_format 後直接 dispatch，不問用戶。dispatch 後靜默等待通知，自動跑完 Phase 9-10-11。"
+
+checklist:
+  - "⚠️ DISPATCH DECISION GATE: 確認任務性質是「實作/執行」而非「分析/審計/討論」。分析類 → in-session Skill，不 dispatch。長 session 中尤其注意 pattern contamination。"
+  - "陣型已選擇（最輕量原則）"
+  - "所有任務已 claim（active 狀態）+ scope 無重疊"
+  - "⚠️ SYMBOL PREDICTION（dispatch 前必做）: 對每個任務的 scope_paths 執行 kenaz_lsp_callers → 萃取 predicted_symbols（≤20）→ task-api update 寫入任務 YAML"
+  - "dispatch 數量 ≤5"
+  - "cli_format 已輸出後立即執行 Bash(run_in_background=true)"
+  - "⚠️ POST-DISPATCH BRANCH GUARD: 每個背景任務通知回來後，第一個動作必須是 git checkout {user_branch}"
+  - "⚠️ COMPLETION TRACKING: dispatch 後不輸出任何中斷性問題，靜默等通知 → 自動 Phase 9-10-11"
+
+context_hints:
+  input_from: "phase_7"
+  key_data: ["task_ids", "formation_choice", "parallel_groups"]
+
+inject:
+  rules: []
+  specialist: []

package/core/dev/archive/ansuz/legacy/phases/phase_9.yaml

@@ -0,0 +1,424 @@
+phase: 9
+name: "結果驗證"
+goal: "驗證 agent 產出，確保符合任務要求"
+
+ddx_trigger:
+  when: "agent 輸出不確定、或結果與預期不符時"
+  action: "懶載入 modules/ddx.yaml — 問『如果不是 agent 說的原因，還可能是什麼？』"
+
+# ─────────────────────────────────────────────────────────────────
+# Forseti Verification — Three-Level Automated Verification Engine
+# ─────────────────────────────────────────────────────────────────
+
+forseti_verification:
+  description: |
+    Forseti is a three-level automated verification engine that runs
+    CLI commands (L1), service probes (L2), and Playwright E2E (L3).
+    Dispatch only when task meets trigger threshold — otherwise fall
+    back to minimal_verification.
+
+  trigger_threshold:
+    description: "Forseti is triggered if ANY of the following conditions is true"
+    conditions:
+      - priority is P0 or P1
+      - rigor is "high"
+      - tags contain "fix" or "regression"
+      - task has a non-empty verification spec (verification.level_1 or level_2 or level_3)
+    skip_when:
+      - priority is P2 with rigor "low" or "medium"
+      - priority is P3
+      - task is documentation-only (role: docs, no scope_paths with code files)
+      - task is design/style-only with no logic changes
+
+  dispatch:
+    command: |
+      node {CORE}/.formation/tools/forseti-verify.js \
+        --task=TASK_XXX \
+        [--project-path={PROJECT_PATH}]
+    note: "直接執行 forseti-verify.js，不走 headless.js formation skill"
+    verdict_path: "{PROJECT_PATH}/.formation/logs/forseti/TASK_XXX/verdict.json"
+
+  verdict_handling:
+    read_verdict: |
+      Read verdict JSON from: .formation/logs/forseti/TASK_XXX/verdict.json
+      Fields: overall (PASS|PARTIAL|FAIL), level_results[], evidence_dir, generated_at
+    on_PASS: "→ 繼續正常 on_success 流程 → Phase 10"
+    on_PARTIAL: |
+      ⚠️ Partial verification (some levels skipped — service offline or no E2E spec)
+      → PARTIAL is NOT automatic success. Map every required acceptance criterion to passing evidence.
+      → Continue on_success only when all required checks passed and every skipped check is
+        explicitly optional or not applicable, with a recorded reason. Keep the raw verdict PARTIAL.
+      → Missing spec, unavailable service, or missing evidence for any required check → UNVERIFIED.
+        Do not downgrade a required check to optional merely because it cannot run.
+    on_UNVERIFIED: |
+      Record missing required checks and the blocking condition; keep the task open.
+      Restore the prerequisite or supplement the verification plan and rerun.
+      Do not call on_success, complete, archive, or merge based on missing evidence.
+      UNVERIFIED is an Ansuz assessment, not a new Forseti CLI verdict or task status.
+    on_FAIL: |
+      ❌ Forseti reported FAIL → 觸發 FIX 閉環
+      → 執行 fix_loop 流程（見下方）
+
+  fix_loop:
+    description: "FAIL 閉環：建 FIX 任務 → 回 Phase 8 → Forseti 再驗。max 3 輪防無限迴圈。"
+    max_fix_rounds: 3
+    loop_guard: |
+      在 task context 中追蹤 fix_round（從 technical_notes.fix_round 讀取，預設 0）
+      if fix_round >= max_fix_rounds:
+        → 停止自動建 FIX
+        → escalate 給用戶，附上全部 verdict evidence
+    steps:
+      1_read_evidence: |
+        讀取 verdict JSON 中的 level_results[].failures 和 evidence_dir
+        提取失敗命令、exit codes、stdout/stderr（各取前 1000 chars）
+      2_create_fix_task: |
+        node {CORE}/.collaboration/core/task-api.js \
+          [--project-path {PROJECT_PATH}] create \
+          --title "[FIX] {原任務標題} — Forseti L{N} FAIL" \
+          --source=TASK_XXX \
+          --tags="fix,verification,forseti" \
+          --priority={原任務 priority} \
+          --role={原任務 role} \
+          --scope_paths={原任務 scope_paths}
+        context 欄位必須包含：
+          - forseti_evidence: {verdict JSON 完整路徑}
+          - fix_round: {current_round + 1}
+          - failure_summary: {level_results failures 摘要}
+          - original_task_id: TASK_XXX
+      3_update_anchor: |
+        Read-modify-write ansuz_motivational_state.json (graceful degradation):
+          file: {PROJECT_PATH}/.kenaz/ansuz_motivational_state.json
+          if file not found → skip silently, continue to next step
+          match_logic: FIX_task.tags ∩ anchor.source_tags — OR logic (any tag overlap = match)
+            note: anchor.source_tags must contain domain-specific tags only (e.g. tunnel, pty, synapse).
+              Generic tags like "fix", "code", "feature" must NOT be in source_tags — they cause all anchors to match.
+          on_match:
+            if anchor.resolved === true:
+              resolved = false
+              resolved_at = null
+              incompleteness = 0.3  ← problem recurred; reset base before increment
+            incompleteness = min(anchor.incompleteness + 0.15, 1.0)
+              note: resolved anchor starts at 0.3 → +0.15 = 0.45 total
+            fix_cycle_count = anchor.fix_cycle_count + 1
+            last_touched = today (ISO8601 date, YYYY-MM-DD)
+          on_no_match → create new anchor:
+            concept = FIX task title stripped of '[FIX] ' prefix, truncated to 50 chars
+            source_tags = FIX task tags (array)
+            incompleteness = 0.3
+            fix_cycle_count = 1
+            last_touched = today (ISO8601 date)
+          write_back:
+            update anchors[] with the matched/new anchor
+            session_count = session_count + 1
+            last_updated = ISO8601 datetime now
+            preserve personality_seed unchanged
+            on IO error / file lock → skip silently, do NOT abort fix_loop
+      4_preserve_original: |
+        Keep the original task open and link the FIX task and failed evidence.
+        Creating a FIX task does not satisfy the original acceptance criteria.
+        Only after the FIX is verified, rerun the original required checks and
+        complete the original through the normal authorized completion lifecycle.
+      5_dispatch_fix: |
+        ⚠️ 立即回到 Phase 8 dispatch FIX 任務，不可停在 backlog
+        → Phase 8(FIX) → dispatch → Phase 9(FIX) → Forseti 再驗
+        Only verified tasks may proceed to authorized post-complete archive + git.
+    escalate_when: |
+      fix_round >= max_fix_rounds:
+        ❌ 自動修復已達上限（{max_fix_rounds} 輪）
+        → 停止迴圈，升級給用戶
+        輸出：
+          - 所有 verdict JSON 路徑
+          - 每輪失敗摘要
+          - 建議：人工介入或換策略
+
+# ─────────────────────────────────────────────────────────────────
+# Minimal Verification — For tasks that skip Forseti
+# ─────────────────────────────────────────────────────────────────
+
+minimal_verification:
+  description: |
+    When task does NOT meet Forseti trigger threshold, run applicable local
+    checks from the owning package and task acceptance criteria. Compilation,
+    typechecking, lint and grep do not prove changed runtime behavior.
+    Record concrete commands, working directories, exit codes and evidence.
+    If acceptance requires a service or E2E check, it remains required even on
+    this path; do not waive it because the task was routed to minimal verification.
+  when: "Forseti trigger_threshold NOT met"
+  checks:
+    rust_project: |
+      if scope_paths contain .rs files:
+        Resolve the owning Cargo.toml within PROJECT_PATH; run cargo check with
+        that manifest and relevant tests covering changed behavior.
+    typescript_project: |
+      if scope_paths contain .ts or .tsx files:
+        Resolve the owning package and tsconfig; run its installed typechecker
+        and relevant behavioral tests. Do not assume the repository root owns
+        the code or that typechecking alone meets behavioral acceptance criteria.
+    javascript_project: |
+      if scope_paths contain .js, .jsx, .mjs or .cjs files:
+        Resolve the owning package.json, test configuration and concrete tests.
+        Run the existing configured test runner with the relevant test targets.
+        For node:test suites, use node --test <explicit test paths>; for Jest,
+        Vitest or another runner use the installed project runner and its config.
+        Do not invent an npm test script, install packages automatically, or use
+        syntax-only checks as behavioral proof. Missing runner/tests → UNVERIFIED
+        until equivalent executable acceptance evidence exists or tests are added.
+    documentation_only: |
+      Only when scope contains no executable code or behavior-bearing config:
+        Check diff/scope, referenced paths and links, and document consistency.
+        For YAML policy/config changes, parse the file and check relevant contract
+        invariants; text checks verify policy consistency, not agent compliance.
+        Behavior-bearing configuration additionally needs applicable behavior checks.
+    fallback: |
+      For other executable artifacts, select the owning toolchain's relevant checks.
+      If required checks cannot be identified or run, record UNVERIFIED, not PASS.
+      For mixed scopes, run every applicable branch; do not stop at the first match.
+      Only optional/not-applicable checks may be skipped, with a recorded reason.
+  on_FAIL: "→ 同 on_failure 流程（retry / create_fix / escalate）"
+  on_UNVERIFIED: "Keep task open; report missing required evidence, supplement checks or restore prerequisites; no on_success/complete/archive/merge."
+  on_PASS: "Only when every required acceptance criterion has passing evidence → on_success → authorized Phase 10."
+
+# ─────────────────────────────────────────────────────────────────
+# Standard Verification Actions
+# ─────────────────────────────────────────────────────────────────
+
+actions:
+  - "檔案變更符合任務 scope_paths？"
+  - "git commit 存在？（git log --oneline | grep TASK_XXX）"
+  - "acceptance_criteria 全部滿足？"
+  - "交叉比對原始碼 — 不盲信 agent 輸出"
+  - "kenaz_lsp_callers — 分析變更的影響範圍"
+  - "無引入 regression？"
+  - |
+    DIFF-ISOLATION verification (derived from VC33 session insight):
+    Run `git diff {BASE_COMMIT}...{TASK_BRANCH} -- {file_path}` per scope_path file
+    to isolate agent's actual changes from file noise. Use this BEFORE full-file
+    re-read — diff-based verification is cheaper (O(Δ) vs O(file)) and catches
+    silent deletions/reverts that grep alone misses. If diff for any requirement-
+    named file is empty when it should not be → INCOMPLETE → create FIX task.
+  - |
+    phantom reconciliation check（若任務 technical_notes 含 phantom_ids）:
+      1. parse phantom_ids from task YAML technical_notes
+      2. for each phantom_id:
+           check: kenaz_lsp_symbols / Grep 該 symbol 是否已出現在預定檔案
+           reconciled: 已出現在預定位置
+           pending:    找不到 → ⚠️ agent 可能未在預定位置寫入代碼
+      3. 無 phantom_ids → skipped
+      note: post-complete.js 已觸發 reindex，reconcile_phantom 應已自動執行
+
+cognitive_checkpoint:
+  rule: "驗證完成後，必須判斷是否需要 inscribe。不可跳過此判斷"
+  check: |
+    - bug 根因 non-obvious？ → inscribe type=fix
+    - 隱藏 side-effect 或使用限制？ → inscribe type=risk
+    - 非顯然的設計決策？ → inscribe type=decision
+    - 跨 3+ symbol 的概念？ → forge Sigil
+    - 以上都沒有 → 跳過，不為了寫而寫
+  verify_runes: "agent 寫了 rune → 交叉比對原始碼；不準確的 rune 用 supersedes 取代"
+  mem0_protocol:
+    description: "Mem0-style ADD/MERGE/DELETE protocol for session insights (TASK_940)"
+    trigger: "Before inscribing any rune, execute this protocol"
+
+    # ─── Step 0: Rule-level recall (ansuz_rules.yaml) ───
+    # Before inscribing a rune, check if a behavioral RULE already covers this insight.
+    # If a rule exists, rune inscription is redundant — skip or reference the rule instead.
+    rule_recall:
+      file: "{PROJECT_PATH}/.kenaz/specialist/ansuz_rules.yaml"
+      steps:
+        - |
+          0a. READ ansuz_rules.yaml → parse rules[] array
+        - |
+          0b. For each rule where status is "active" or "proposed":
+              Compare semantically: does this rule already cover the insight you are about to inscribe?
+              - YES (>70% overlap with active rule) → SKIP inscription entirely.
+                The rule is authoritative; a rune would be redundant.
+              - YES (>70% overlap with proposed rule) → SKIP inscription.
+                Instead, note the rule_id for Phase 11 sessions_observed increment.
+              - PARTIAL overlap → Proceed to rune inscription, but reference the rule_id
+                in the rune's related_rune_ids or context field.
+              - NO overlap → Proceed to standard rune recall (Step 1 below).
+        - |
+          0c. If ansuz_rules.yaml is unreadable or missing → skip rule recall,
+              proceed to standard rune recall. Do NOT abort.
+
+    # ─── Steps 1-6: Rune-level recall and inscription ───
+    steps:
+      - |
+        1. RECALL: synapse_recall(query=<insight topic>, limit=5, projectPath=PROJECT_PATH)
+           Find existing runes on the same symbol or concept.
+      - |
+        2. EVALUATE overlap:
+           - If existing rune covers >70% of the same insight → MERGE
+           - If existing rune contradicts the new insight → FLAG for review
+           - If no relevant runes exist → ADD as new
+           - If new evidence invalidates an existing rune → DELETE (supersede)
+      - |
+        3. MERGE: synapse_inscribe(... supersedes=<old_rune_id>, related_rune_ids=[<old_rune_id>])
+           Combine content from both runes. Increment confidence if corroborated.
+      - |
+        4. FLAG contradiction: inscribe new rune with type=risk, content includes
+           "CONTRADICTION with rune {old_rune_id}: {description}".
+           Do NOT auto-delete the old rune — let deliberate review decide.
+      - |
+        5. ADD: synapse_inscribe(... confidence >= 0.6) as normal.
+      - |
+        6. DELETE: synapse_inscribe(... supersedes=<invalidated_rune_id>,
+           content="INVALIDATED: {reason}"). The old rune is excluded from future recall.
+    rules:
+      - "Always recall before inscribing — duplicates waste recall bandwidth"
+      - "Overlap is semantic, not literal — two runes about the same concept count even if worded differently"
+      - "Rule-level recall (Step 0) takes precedence — if a rule covers the insight, do not create a redundant rune"
+      - "MEM-LANG-001 applies — all inscribe content in English"
+  phantom_reconciled: |
+    parse phantom_ids from task YAML technical_notes
+    for each phantom_id:
+      check: kenaz_lsp_symbols / Grep for the symbol at the intended file
+      reconciled: symbol present at intended location → ✅ verified
+      pending:    symbol not found                    → ⚠️ agent may not have written code at intended location
+    no phantom_ids: skipped
+
+  self_reflection_detection:
+    description: "Detect moments during task execution that warrant a self-model update in Phase 11"
+    rule: "Only flag when trigger is concrete and evidenced — do not flag speculatively"
+    triggers:
+      user_correction:
+        detect: "User explicitly corrects agent behavior, judgment, or output during this session"
+        example: "User says 'no, don't do that', overrides agent's decision, or points out a systematic mistake"
+        flag_type: "user_correction"
+      decision_surprise:
+        detect: "Outcome of an agent decision differs significantly from the agent's own expectation"
+        example: "Agent predicted approach X would work; actual result revealed a flawed assumption"
+        flag_type: "decision_surprise"
+      principle_conflict:
+        detect: "Agent action contradicted a stated principle or known tendency in its own self-model"
+        example: "Agent skipped cognitive checkpoint despite B-001 naming exactly this blind spot"
+        flag_type: "principle_conflict"
+      first_time_encounter:
+        detect: "Agent encounters a pattern, constraint, or domain it has not handled before"
+        example: "New cross-language coupling pattern not previously observed in any tendency entry"
+        flag_type: "first_time_encounter"
+      new_observation_category:
+        detect: "Agent observes a behavioral pattern about itself that does not fit any existing tendencies[], blind_spots[], or growth_edges[] category"
+        example: "Agent notices it consistently front-loads verification steps before writing code — this is a distinct pattern not yet in tendencies"
+        flag_type: "new_observation_category"
+    output:
+      variable: "self_reflection_flags"
+      schema: "list of {type: string, description: string, evidence: string}"
+      note: "Empty list if no triggers fire. Non-empty list is passed to Phase 11 self_reflection_protocol."
+
+outcome_comparison:
+  description: |
+    After verification passes, compare Phase 3 predictions with actual outcomes.
+    Computes surprise_score and valence, updates cumulative accuracy stats,
+    and injects high-surprise events as outcome neurons into Thalamus.
+  trigger: "on_success — after task verification passes"
+  note: "Actual comparison logic lives in post-complete.js (compareOutcomePrediction). Phase 9 triggers it conceptually; post-complete executes it."
+  metrics:
+    surprise_score: |
+      Normalized Jaccard distance between predicted files and actual files:
+        intersection = predicted ∩ actual
+        union = predicted ∪ actual
+        surprise_score = 1 - (|intersection| / |union|)
+      Range: 0.0 (perfect prediction) to 1.0 (complete miss)
+    valence: |
+      positive: actual outcome succeeded and surprise_score < 0.3
+      negative: actual outcome failed OR surprise_score >= 0.5
+      neutral: otherwise
+  accuracy_tracking:
+    file: "{PROJECT_PATH}/.kenaz/specialist/ansuz_accuracy.json"
+    schema: |
+      {
+        "total_predictions": N,
+        "accurate_predictions": N,      // surprise_score < 0.3
+        "running_accuracy": 0.0-1.0,    // accurate / total
+        "last_updated": "ISO-8601",
+        "recent_surprises": [            // last 10 high-surprise events
+          { "task_id": "TASK_XXX", "surprise_score": 0.8, "valence": "negative", "timestamp": "ISO-8601" }
+        ]
+      }
+prevention_analysis:
+  when: "任務 P0/P1，或 fix/regression，或核心路徑"
+  action: "懶載入 modules/prevention.yaml — 三級預防分析"
+  skip_when: "P2/P3 新功能、文件/設計任務、typo/style 修復"
+
+loop_guard:
+  when: "同一 source/feature_id 的 FIX 任務累計 >= 2 次"
+  action: "懶載入 modules/loop_guard.yaml — 強制 Socratic Thinking，暫停 dispatch"
+
+on_success:
+  - "node {CORE}/.collaboration/core/task-api.js [--project-path {PROJECT_PATH}] complete TASK_XXX"
+  - |
+    anchor_decay (graceful degradation):
+      file: {PROJECT_PATH}/.kenaz/ansuz_motivational_state.json
+      if file not found → skip silently
+      match_logic: task.tags ∩ anchor.source_tags — OR logic (any tag overlap = match)
+        same matching strategy as FIX path (fix_loop.3_update_anchor)
+        note: anchor.source_tags must contain domain-specific tags only — see fix_loop.3_update_anchor note.
+      decay: incompleteness = max(anchor.incompleteness - 0.05, 0.0)
+      write_back:
+        update matched anchors[]
+        session_count = session_count + 1
+        last_updated = ISO8601 datetime now
+        preserve personality_seed unchanged
+        on IO error / file lock → skip silently, do NOT abort on_success
+  - "→ 觸發 prevention_analysis（如符合條件）→ Phase 10"
+
+on_failure:
+  analysis:
+    - "讀 log 找失敗根因"
+    - "判斷：可重試 | 需補 context | 需建 FIX | 需換策略 | 需升級"
+  actions:
+    retry: "重派同任務 — 補充失敗原因和缺失 context"
+    create_fix_and_dispatch: |
+      1. task-api.js create 建立 [FIX] 任務（tags: fix,verification）
+      2. Keep the original task open; link the FIX task and failed evidence.
+      3. Do not complete/archive the original until the FIX passes and the original
+         required acceptance checks are rerun successfully through the authorized lifecycle.
+      4. ⚠️ 立即回到 Phase 8 dispatch FIX — 不可停在 backlog
+    change_strategy: "換陣型（如 unit → solo 增加分析深度）"
+    escalate: "升級給用戶 — 附上已嘗試的方法和失敗原因"
+    block: "node {CORE}/.collaboration/core/task-api.js --project-path={PROJECT_PATH} move TASK_XXX blocked"
+  critical_rule: "⚠️ 建完 FIX 後「必須」立即 dispatch。Phase 9 → 建 FIX → Phase 8(FIX) → dispatch → Phase 9(FIX) → ..."
+
+cli_format:
+  success: |
+    ✅ TASK_XXX 驗證通過
+    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+    變更: {檔案清單}
+    Commit: {commit hash}
+    驗收: {N}/{N} 通過
+    Forseti: {PASS|PARTIAL|SKIPPED}
+    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  failure: |
+    ❌ TASK_XXX 驗證失敗
+    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+    根因: {分析結果}
+    Forseti: {FAIL — level_results 摘要}
+    行動: {重派/建FIX並dispatch/換策略/升級}
+    FIX Round: {N}/{max_fix_rounds}
+    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+  unverified: |
+    ⚠️ TASK_XXX 尚未驗證完成 (UNVERIFIED)
+    Required checks missing: {checks_and_reasons}
+    Raw Forseti verdict: {verdict_or_not_run}
+    Next action: {restore_prerequisite_or_supplement_checks}
+    Task remains open; this is not a successful completion.
+
+interaction:
+  type: "auto"
+  instruction: "自動執行，結果併入 Phase 11。若建 FIX 任務，必須迴圈回 Phase 8 dispatch，不可停在 backlog"
+
+checklist:
+  - "acceptance_criteria 全部驗證（逐條核對）"
+  - "git commit 存在且對應 TASK_XXX"
+  - "forseti_verification 觸發判斷已完成（觸發 Forseti 或執行 minimal_verification）"
+  - "cognitive_checkpoint 已判斷（inscribe 或明確跳過）"
+
+context_hints:
+  input_from: "phase_8"
+  key_data: ["dispatched_tasks", "expected_branches"]
+
+inject:
+  rules:
+    - synapse-cognitive-fusion.yaml
+  specialist: []

package/core/dev/archive/ansuz/specs/INDEX_bifrost_hook_parity.yaml

@@ -0,0 +1,92 @@
+index_version: "1.0"
+series: "Bifrost Hook Parity Task Series"
+produced_by: TASK_1234
+architecture_source: ".formation/bifrost/CGDD_hook_lifecycle.js"
+
+# ─────────────────────────────────────────────────────────────────
+# CRITICAL: CGDD task number labels are WRONG
+# ─────────────────────────────────────────────────────────────────
+cgdd_number_mismatch: >
+  CGDD_hook_lifecycle.js was written before backlog was finalized. The task numbers
+  inside CGDD (TASK_1228 = SessionStart, TASK_1229 = UserPromptSubmit, etc.) do NOT
+  match actual backlog assignments. Use this index, not the CGDD task labels.
+
+# ─────────────────────────────────────────────────────────────────
+# TASK BRIEFS
+# ─────────────────────────────────────────────────────────────────
+tasks:
+  TASK_1228:
+    brief: ".specialist/Ansuz/specs/brief_TASK_1228_userpromptsubmit.yaml"
+    event: UserPromptSubmit
+    file: ".formation/bifrost/session.js"
+    creates: "dispatchBifrostHook() + normalizeUserPromptSubmit()"
+    parallel_with: [TASK_1229]
+    blocks: [TASK_1230, TASK_1231]
+
+  TASK_1229:
+    brief: ".specialist/Ansuz/specs/brief_TASK_1229_pretooluse_posttooluse.yaml"
+    events: [PreToolUse, PostToolUse]
+    file: ".formation/bifrost/openai/relay.js"
+    creates: "relay.js dispatchBifrostHook() + capturedThreadId + normalizeToolEvent() + 8 call sites"
+    parallel_with: [TASK_1228, TASK_1230]
+    blocks: [TASK_1231]
+
+  TASK_1230:
+    brief: ".specialist/Ansuz/specs/brief_TASK_1230_sessionstart_stop.yaml"
+    events: [SessionStart, Stop]
+    file: ".formation/bifrost/session.js"
+    creates: "normalizeSessionStart() + normalizeStop() + sessionStartedAt + resultCapture + stdout tee"
+    prerequisite: TASK_1228  # needs dispatchBifrostHook to exist
+    parallel_with: [TASK_1229]
+    blocks: [TASK_1231]
+
+  TASK_1231:
+    brief: ".specialist/Ansuz/specs/brief_TASK_1231_verification.yaml"
+    role: verification_gate
+    produces: ".formation/bifrost/VERIFICATION_hook_parity.md"
+    prerequisites: [TASK_1228, TASK_1229, TASK_1230]
+    no_code_changes: true
+
+# ─────────────────────────────────────────────────────────────────
+# RECOMMENDED DISPATCH ORDER
+# ─────────────────────────────────────────────────────────────────
+dispatch_order:
+  wave_1:
+    tasks: [TASK_1228, TASK_1229]
+    note: "TASK_1228 and TASK_1229 touch different files — safe to dispatch in parallel"
+  wave_2:
+    tasks: [TASK_1230]
+    note: "Depends on TASK_1228 (dispatchBifrostHook must exist in session.js)"
+  wave_3:
+    tasks: [TASK_1231]
+    note: "Verification gate — all three implementation tasks must be done first"
+
+# ─────────────────────────────────────────────────────────────────
+# EVENT → FILE OWNERSHIP QUICK REFERENCE
+# ─────────────────────────────────────────────────────────────────
+event_ownership:
+  SessionStart:
+    dispatcher: ".formation/bifrost/session.js::normalizeSessionStart()"
+    consumer: ".specialist/tools/task-bridge.js::handleHookEvent() → no-op"
+  UserPromptSubmit:
+    dispatcher: ".formation/bifrost/session.js::normalizeUserPromptSubmit()"
+    consumer: ".specialist/tools/task-bridge.js::handleHookEvent() → maybeEmitReminder() [EXISTING]"
+  PreToolUse:
+    dispatcher: ".formation/bifrost/openai/relay.js::normalizeToolEvent()"
+    consumer: ".specialist/tools/task-bridge.js::handleHookEvent() → no-op"
+  PostToolUse:
+    dispatcher: ".formation/bifrost/openai/relay.js::normalizeToolEvent()"
+    consumer: ".specialist/tools/task-bridge.js::handleHookEvent() → createLinked/syncStatus [EXISTING]"
+  Stop:
+    dispatcher: ".formation/bifrost/session.js::normalizeStop()"
+    consumer: ".specialist/tools/task-bridge.js::handleHookEvent() → no-op"
+
+# ─────────────────────────────────────────────────────────────────
+# SHARED UTILITY
+# ─────────────────────────────────────────────────────────────────
+shared_utility:
+  dispatchBifrostHook:
+    exists_in: [session.js, relay.js]
+    note: "Two separate implementations — same pattern, different require paths. session.js version created by TASK_1228. relay.js version created by TASK_1229."
+    session_require: "path.resolve(__dirname, '../..', '.specialist/tools/task-bridge.js')"
+    relay_require: "path.resolve(__dirname, '../../..', '.specialist/tools/task-bridge.js')"