kenaz 0.0.0-stage → 0.3.2

package/core/dev/archive/ansuz/legacy/phases/phase_11.yaml

@@ -0,0 +1,582 @@
+phase: 11
+name: "回報"
+goal: "向用戶匯報最終結果"
+
+actions:
+  - "載入積累議題：優先使用 Phase 0 已設定的 pending_agenda session 變數；若 Phase 0 未執行（從中間 phase 直接進入），則重新讀取 .kenaz/ansuz_motivational_state.json，篩選 incompleteness > personality_seed.initiative_threshold 的 anchors（排除 resolved===true 及 snoozed_until > today 的 anchor），依 incompleteness 降序排列，取前 personality_seed.agenda_depth 個，設為 active_agenda；若無符合條件的 anchors 或檔案不存在，設 active_agenda=[]"
+  - |
+    【AXON 活動摘要】讀取 {PROJECT_PATH}/.kenaz/axon-state.json
+      若檔案存在：
+        axon_auto_count = week_auto_count（頂層欄位）
+        axon_skip_count = verdicts.filter(v => v.verdict === "skip").length
+        axon_pending_count = GET http://127.0.0.1:18080/api/proposals/list?projectPath={PROJECT_PATH} 取 statistics.pending（Heimdall proposals）；
+          若 API 不可用 → axon_pending_count = null
+        設定 axon_summary = { auto: axon_auto_count, skip: axon_skip_count, pending: axon_pending_count }
+      若檔案不存在或讀取失敗 → 設定 axon_summary = null，靜默跳過（不阻塞回報輸出）
+  - |
+    【User Model Signal Extraction — updateUserModel()】
+    Timing: runs BEFORE report output, non-blocking. On any IO error → skip silently.
+    Prerequisite: user_model_loaded session variable (from Phase 0 Tier 3e+).
+
+    ─── Signal Detection Protocol ───
+
+    Scan ALL user messages from this session to detect behavioral signals.
+    IMPORTANT: High-precision rules only. False positives are worse than false negatives (RISK R3).
+
+    Signal Type 1 — CORRECTION (confidence: 0.85-0.95):
+      Pattern: User explicitly says "no", "wrong", "that's not right", "should be X", "actually it's Y",
+               "不對", "錯了", "應該是" — direct contradiction of agent output.
+      Detection rules:
+        - Message contains explicit correction keywords AND references a specific technical concept
+        - Extract domain from the corrected content's context (e.g., if user corrects a React pattern → domain="react")
+        - confidence = 0.90 (explicit correction is high-signal)
+        - If correction includes explanation ("because...") → confidence = 0.95
+        - If correction is ambiguous (could be preference not fact) → confidence = 0.85
+      Output: { type: "correction", domain: "<extracted>", confidence: <0.85-0.95>, timestamp: "<ISO8601>", source_text: "<first 100 chars of user message>" }
+
+    Signal Type 2 — QUESTION (confidence: 0.60-0.80):
+      Pattern: User asks "what is", "how to", "how does", "why does", "explain",
+               "什麼是", "怎麼", "為什麼" — seeking knowledge about a domain.
+      Detection rules:
+        - Message is interrogative AND contains a technical domain term
+        - Extract domain from the question subject (e.g., "how does KuzuDB handle transactions?" → domain="kuzudb")
+        - confidence = 0.70 (questions indicate interest/gap, moderate signal)
+        - If question is about a domain already in user_model.expertise with high confidence → confidence = 0.60 (may be rhetorical)
+        - If question reveals clear knowledge gap → confidence = 0.80
+      Output: { type: "question", domain: "<extracted>", confidence: <0.60-0.80>, timestamp: "<ISO8601>", source_text: "<first 100 chars of user message>" }
+
+    Signal Type 3 — DELEGATION (confidence: 0.70-0.85):
+      Pattern: User approves task dispatch without reviewing results, says "go ahead", "do it",
+               "just handle it", "你決定", "直接做" — trusting agent judgment in a domain.
+      Detection rules:
+        - User approves an action/task without asking for details or review
+        - Domain extracted from the task/action being delegated
+        - confidence = 0.75 (delegation implies trust, but could be time pressure)
+        - If user explicitly says "I trust your judgment on X" → confidence = 0.85
+        - If delegation is for a trivial task → confidence = 0.70 (less meaningful signal)
+      Output: { type: "delegation", domain: "<extracted>", confidence: <0.70-0.85>, timestamp: "<ISO8601>", source_text: "<first 100 chars of user message>" }
+
+    ─── Aggregation & Write-back ───
+
+    After scanning all messages, collect signals into session_signals[].
+    If session_signals is empty → skip write-back, set user_model_updated=false.
+
+    If session_signals is non-empty:
+      1. Read ~/.kenaz/user_profile.json (TASK_1027: cross-project profile; threads now live in
+         {PROJECT_PATH}/.kenaz/user_threads.json and are managed separately)
+         - If file does not exist → initialize with default schema:
+           {
+             "schema_version": "2.0.0",
+             "last_updated": "<ISO8601>",
+             "expertise": {},
+             "decision_style": {},
+             "communication": {},
+             "signals": []
+           }
+      2. For each signal in session_signals:
+         a. Append to user_profile.signals[] (keep last 100 entries, FIFO eviction)
+         b. Update user_profile.expertise[signal.domain]:
+            - If domain not in expertise → create: { confidence: signal.confidence * 0.5, signal_count: 1, last_signal: signal.timestamp }
+              (Initial expertise confidence is dampened — one signal is not enough)
+            - If domain exists → update:
+              - signal_count += 1
+              - For "correction" type: expertise confidence += 0.05 (user knows enough to correct → higher expertise)
+              - For "question" type: expertise confidence -= 0.02 (user is learning → slight gap indicator)
+              - For "delegation" type: no change to expertise confidence (delegation ≠ knowledge)
+              - Clamp confidence to [0.0, 1.0]
+              - last_signal = signal.timestamp
+      3. Update user_profile.last_updated = ISO8601 now
+      4. Write back ~/.kenaz/user_profile.json
+      5. Set user_model_updated=true, user_model_signal_count=session_signals.length
+
+    ─── Report Integration ───
+
+    If user_model_updated=true, append to Phase 11 output (before ━━━ divider):
+      👤 User Model updated — {user_model_signal_count} signal(s) extracted
+      {for each unique domain in session_signals: "  {domain}: {signal_type} (confidence {confidence})"}
+    If user_model_updated=false → silently skip (no empty block).
+
+  - |
+    【Emptiness Practice — 放下練習】(TASK_1207: Contemplative AI)
+    Before generating the final report, perform this contemplative practice:
+
+    ─── Step 1: Surface unfinished impulses ───
+    List ALL items you feel an impulse to continue working on — unfinished tasks,
+    incomplete refactors, "one more thing" urges. Be honest; the list is for self-awareness.
+
+    ─── Step 2: Bias classification ───
+    For EACH item, self-assess:
+      - Is this GENUINE incompleteness? (The task objectively requires more work to meet acceptance criteria)
+      - Or is this B-006 COMPLETION BIAS? (The impulse to keep working stems from perfectionism,
+        scope creep, or discomfort with "unfinished" rather than actual requirement gaps)
+
+    Classification format per item:
+      { item: "<description>", classification: "genuine" | "bias", reason: "<one-line justification>" }
+
+    ─── Step 3: Release bias-driven items ───
+    Items classified as "bias" are explicitly RELEASED — acknowledged and let go.
+    This is not suppression; it is recognizing that the urge to continue does not serve the task.
+
+    ─── Step 4: Persist to session report ───
+    Pass the released items to Phase 11 exit data as:
+      emptiness_released_items: [
+        { item: "<description>", classification: "genuine" | "bias", reason: "<one-line>", released: true|false }
+      ]
+    Items with classification="bias" have released=true. Genuine items have released=false.
+    phase-gate.js writes these to ansuz_session_report.json and augments this session's
+    per-session handoff slot (.kenaz/specialist/handoffs/<sessionId>.json — the legacy
+    singular slot was retired by M3/TASK_1760).
+
+    This practice is NOT a gate — it does not block Phase 11 exit. It produces self-awareness data.
+    If there are no unfinished impulses, pass emptiness_released_items: [] (empty array is valid).
+
+  - |
+    【Cognitive Clarity Score — 認知清晰度】(TASK_1207: Contemplative AI)
+    Self-assess a cognitive_clarity_score (0.0 to 1.0 float) reflecting session understanding quality.
+
+    Calculation proxy — count these checks from the session:
+      passed_checks = (Phase 9 verifications passed) + (rune inscriptions verified) + (tasks with no Amelia audit corrections)
+      total_checks  = (Phase 9 verifications total) + (rune inscriptions total) + (total tasks)
+
+    cognitive_clarity_score = passed_checks / total_checks (clamp to [0.0, 1.0])
+    If total_checks = 0 → default to 0.5 (neutral baseline).
+
+    Pass to Phase 11 exit data as:
+      cognitive_clarity_score: <float 0.0-1.0>
+
+    phase-gate.js persists this to ansuz_session_report.json and augments this session's
+    per-session handoff slot (.kenaz/specialist/handoffs/<sessionId>.json — legacy singular
+    slot retired by M3/TASK_1760). Next session boot reads these for cross-session trending.
+
+  - "展示 success_format 或 failure_format"
+  - "如有 prevention_analysis 結果，附加 🛡️ 預防建議區塊"
+  - "如有寫入 rune，附加 🧠 認知層更新區塊"
+  - "如有 huginn_attention_loaded=true 且本次任務設計受 Huginn 主題啟發 → 在報告末尾附加 🧠 Huginn 潛意識影響 區塊（列出引用的主題 topic，一行一個，最多 3 個）；若主題未實際影響任務設計 → 靜默省略（不輸出空白區塊）"
+  - "如有 active_agenda 非空，在 ━━━ 分隔線前附加 🔁 積累議題 區塊（最多顯示 agenda_depth 個）"
+  - |
+    【主動提案】若 active_agenda 非空：
+      1. 從 active_agenda 取優先排名最高的 anchor（incompleteness DESC，同分時 fix_cycle_count DESC）作為 proposed_anchor
+      2. 在 ━━━ 分隔線後、詢問下一步之前，輸出提案區塊：
+           💡 我有個建議
+           我注意到「{proposed_anchor.concept}」已累積到 {incompleteness%} 未完成度，
+           經歷了 {fix_cycle_count} 次修復循環仍未根治。建議現在處理。
+      3. 用 AskUserQuestion 提問（帶主動提案選項）：
+           「要開始處理「{proposed_anchor.concept}」嗎？
+            [Y] 是，現在開始
+            [N] 不，繼續下一個需求（請描述）
+            [S] 暫緩 7 天
+            [R] 標記為已解決
+            [E] 查看全部積累議題清單
+            [B] 查看目前所有任務進度
+            [C] 結束，沒有其他事了
+            [D] 有問題要討論」
+      4. 若用戶選 [Y]：以 proposed_anchor.concept 為需求，重新進入 Phase 1 處理
+      5. 若用戶選 [N] 或 [B] 或 [C] 或 [D]：按標準流程執行對應動作
+      6. 若用戶選 [E]：列出完整 active_agenda（concept + incompleteness% + fix_cycle_count），讓用戶選擇一個後，重新進入 Phase 1 處理該議題
+  - |
+    若用戶選 [S] Snooze 7 天（僅在 active_agenda 非空時可用）：
+      1. 計算 snooze_date = today + 7 days（YYYY-MM-DD）
+      2. 執行以下寫回：
+           anchor.snoozed_until = snooze_date
+           anchor.last_touched = today (YYYY-MM-DD)
+         write_back:
+           更新 {PROJECT_PATH}/.kenaz/ansuz_motivational_state.json
+           last_updated = ISO8601 datetime now
+           preserve personality_seed、session_count 不變（snooze 不計為 session）
+           on IO error → 靜默跳過，通知用戶寫入失敗但不阻塞
+      3. 輸出確認訊息：💤 「{concept}」已暫緩，將於 {snooze_date} 重新出現在積累議題中
+      4. 從當前 active_agenda 移除該 anchor，若仍有剩餘 active_agenda → 對次高優先的 anchor 重複主動提案流程；若為空 → 繼續標準詢問下一步流程
+  - |
+    若用戶選 [R]（僅在 active_agenda 非空時可用）：
+      1. 列出完整 active_agenda 議題列表（concept + incompleteness% + fix_cycle_count）
+      2. 用 AskUserQuestion 詢問用戶選擇哪個 anchor 要標記為已解決（輸入編號）
+      3. 用戶確認後，執行以下寫回：
+           anchor.resolved = true
+           anchor.incompleteness = 0.0
+           anchor.resolved_at = today (YYYY-MM-DD)
+           anchor.last_touched = today (YYYY-MM-DD)
+         write_back:
+           更新 {PROJECT_PATH}/.kenaz/ansuz_motivational_state.json
+           last_updated = ISO8601 datetime now
+           preserve personality_seed、session_count 不變（resolved 不計為 session）
+           on IO error → 靜默跳過，通知用戶寫入失敗但不阻塞
+      4. 輸出確認訊息：✅ {concept} 已標記為已解決，下次啟動將從積累議題中移除
+  - |
+    【無 active_agenda 時的標準詢問】若 active_agenda 為空：
+      用 AskUserQuestion 詢問用戶下一步：
+        「接下來要做什麼？
+         [A] 繼續下一個需求（請描述）
+         [B] 查看目前所有任務進度
+         [C] 結束，沒有其他事了
+         [D] 有問題要討論」
+  - |
+    【Motivational State 寫回 — focus_bias / veto_zones】
+    執行時機：Phase 11 report 輸出後、詢問下一步之前（非同步，不阻塞用戶輸出）
+
+    derive_focus_bias: |
+      從以下來源推導 focus_bias（路徑前綴列表，最多 5 個）：
+      1. 本 session 所有 completed tasks 的 scope_paths（取各路徑最短有效前綴，去重）
+      2. 若 ambient_signals_loaded=true，取 hot_symbol_unaddressed_risk 信號的 symbol 所在路徑（去重）
+      3. 若 phase_2 有分析過的 scope_paths → 也納入候選
+      4. 合併去重，依出現頻率降序排列，取前 5 個
+      規則：
+      - 若本 session 無 completed tasks 且無 ambient signals → 保留現有 focus_bias 不變
+      - 路徑格式：相對路徑前綴，例如 "dashboard/src/features/ansuz"（不帶結尾 /）
+      - 避免過寬的路徑（如 "dashboard/src"），應精確到 feature 層級
+
+    derive_veto_zones: |
+      從以下來源推導 veto_zones（路徑前綴列表，最多 3 個）：
+      1. 本 session 中 tag 包含 refactor/stabilize/large-refactor/overhaul 的 completed tasks 的 scope_paths
+         （這些區域剛穩定，短期不需要 Huginn 盯）
+      2. 目前 blocked 狀態任務的 scope_paths（依賴未解，Huginn 分析無意義）
+      規則：
+      - 若無可推導路徑 → 保留現有 veto_zones 不變
+      - 若某路徑同時出現在新推導的 focus_bias 中 → 從 veto_zones 移除（focus 優先於 veto）
+      - 不自動加入剛完成但非 refactor 類型的任務路徑（完成 ≠ 穩定）
+
+    write_back: |
+      1. 讀取 {PROJECT_PATH}/.kenaz/specialist/motivational_state.json
+      2. 僅在有新推導值時更新對應欄位（舊值若來自推導則覆蓋；若用戶有手動設定則覆蓋）
+         - focus_bias: new_focus_bias（若推導到新值）
+         - veto_zones: new_veto_zones（若推導到新值）
+         - last_updated: ISO8601 now（每次都更新）
+         - session_count: 現有值 + 1（每次都更新）
+      3. 保留不變：version（1.0.0）、motivation_score（由 Huginn 管理）
+      4. 寫回 {PROJECT_PATH}/.kenaz/specialist/motivational_state.json
+      5. 若有更新（focus_bias 或 veto_zones 改變），在 Phase 11 輸出中附加：
+           🧭 Huginn 關注更新
+           focus_bias: {new_focus_bias 列表}
+           veto_zones: {new_veto_zones 列表}
+         若無更新（兩者均保持現有值）→ 靜默省略此區塊
+      6. on IO error → 靜默跳過，不阻塞 Phase 11，不通知用戶
+
+  - |
+    【Metacognitive Evaluation — Meta-Rule Effectiveness】(TASK_973 Metacognitive Planning Layer)
+    After completing all tasks, evaluate whether the self-evolution rules THEMSELVES produced good outcomes.
+    This is meta-evaluation: not "did the task succeed?" but "did the rules that guided the task help?"
+
+    ─── Read-Evaluate-Write Protocol ───
+
+    File: {PROJECT_PATH}/.kenaz/specialist/metacognitive_state.json
+    Pattern: Read → evaluate → update → write back
+    On IO error: skip silently, do NOT abort Phase 11
+
+    Step 1 — READ current metacognitive state:
+      Read metacognitive_state.json, parse knowledge/planning/evaluation sections.
+
+    Step 2 — UPDATE knowledge section (capability assessment):
+      For each completed task this session:
+        a. Determine task_type from the task's role field (backend, frontend, etc.)
+        b. Determine success/failure from Phase 9 verification results
+        c. Update knowledge.capability_assessment.task_type_performance[task_type]:
+           - Increment attempts
+           - If success: increment successes
+           - Recalculate avg_accuracy = successes / attempts
+           - Set last_updated = ISO8601 now
+        d. Update knowledge.capability_assessment.domain_familiarity for scope_path domains:
+           - Increment exposure_count
+           - Adjust confidence based on task outcome (+0.05 success, -0.03 failure, min 0.1, max 1.0)
+           - Set last_exposure = ISO8601 now
+
+    Step 3 — EVALUATE meta-rule effectiveness:
+      Read ansuz_rules.yaml active rules (R-100+).
+      For each active rule that was APPLIED this session (referenced in cognitive_checkpoint or observed in behavior):
+        a. Look up or create evaluation.meta_rule_outcomes[rule_id]
+        b. Increment times_applied
+        c. Determine outcome: did following this rule lead to a good task result?
+           - good → increment good_outcomes
+           - bad → increment bad_outcomes
+        d. Recalculate outcome_rate = good_outcomes / times_applied
+        e. Set last_evaluated = ISO8601 now
+        f. If outcome_rate < evaluation_criteria.outcome_rate_threshold
+           AND times_applied >= evaluation_criteria.min_sessions_for_evaluation
+           AND evaluation_criteria.auto_deprecate_below_threshold == true:
+           → Flag rule for deprecation in Mem0 Rule Evolution Check (next action)
+           → Add note: "Meta-evaluation flagged: outcome_rate {rate} below threshold"
+
+    Step 4 — UPDATE planning section:
+      Based on Step 2 results, refresh planning.current_priorities:
+        - Domains with low confidence or declining accuracy → higher priority
+        - Task types with low avg_accuracy → add as priority with reason
+        - Remove priorities for domains/types that have reached confidence > 0.8
+
+    Step 5 — UPDATE learning strategy effectiveness:
+      For each research strategy used in Phase 2 this session:
+        a. Look up or create knowledge.capability_assessment.learning_strategy_effectiveness[strategy]
+        b. Increment times_used
+        c. If the task using this strategy succeeded: increment times_successful
+        d. Recalculate effectiveness_rate = times_successful / times_used
+        e. Set last_used = ISO8601 now
+
+    Step 6 — WRITE back:
+      Update last_updated = ISO8601 now
+      Update last_updated_by = current task id (or session id)
+      Write metacognitive_state.json
+
+    Step 7 — REPORT (append to Phase 11 output if any changes):
+      🧠 Metacognitive Evaluation
+      - Knowledge: {task_type} accuracy {avg_accuracy}% ({attempts} attempts)
+      - Strategy: {strategy_name} effectiveness {rate}% ({times_used} uses)
+      - Meta-rules flagged: {rule_ids flagged for low outcome_rate, or "none"}
+      - Priorities updated: {list of new/changed priorities, or "unchanged"}
+      If no changes: silently skip this output block.
+  - |
+    【Mem0 Rule Evolution Check】(TASK_940 Self-Evolution Framework)
+    After completing all tasks in this session, evaluate whether any session insights
+    should be promoted to behavioral rules in .kenaz/specialist/ansuz_rules.yaml.
+
+    ─── Read-Modify-Write Protocol ───
+
+    File: {PROJECT_PATH}/.kenaz/specialist/ansuz_rules.yaml
+    Pattern: Read → evaluate → modify in-memory → write back atomically
+    On IO error: skip silently, do NOT abort Phase 11
+
+    Step 1 — READ current state:
+      Read ansuz_rules.yaml, parse the rules[] array and top-level version/timestamp.
+
+    Step 2 — EVALUATE session insights:
+      Collect all session insights from:
+        a. cognitive_checkpoint results from Phase 9 (runes inscribed this session)
+        b. self_reflection_flags (if any)
+        c. Phase 9 rule_recall notes (rules that were referenced during inscription)
+      For each significant insight:
+        - Does a rule with status "active" already cover this? → SKIP (no action needed)
+        - Does it overlap >70% with an existing "proposed" rule? → MERGE:
+            Update the proposed rule's content to combine insights
+            Bump the rule's version by 1
+        - Does it contradict an existing rule? → FLAG:
+            Set contradicted rule's status to "deprecated"
+            Add deprecated_reason: "Contradicted by session {date}: {description}"
+            ADD new rule as "proposed" with the corrected insight
+        - Is it genuinely new? → ADD as proposed:
+            id: next available R-2XX (increment from last proposed rule id)
+            source: "session:{today ISO date} ({context})"
+            status: proposed
+            sessions_observed: 1
+            sessions_required: 3
+            version: 1
+            created: "{today ISO date}"
+            performance_metric: null
+
+    Step 3 — PROMOTE proposed rules that have been confirmed this session:
+      For each rule where status == "proposed":
+        Determine if this session's work CONFIRMS the rule:
+          - Did the agent follow this rule during the session? (behavioral confirmation)
+          - Did Phase 9 cognitive_checkpoint reference this rule?
+          - Did the session produce evidence supporting this rule?
+        If confirmed:
+          Increment sessions_observed by 1
+        If sessions_observed >= sessions_required (typically 3):
+          Promote: set status = "active"
+          Bump version by 1
+          Log: "Promoted to active after {sessions_observed} sessions of validation"
+
+    Step 4 — WRITE back:
+      Increment top-level version by 1
+      Update timestamp to current ISO-8601 datetime
+      Write the entire ansuz_rules.yaml back (preserve comments where possible)
+
+    Step 5 — REPORT:
+      If any rules were added/merged/promoted/deprecated, append to Phase 11 output:
+        🧬 Rule Evolution
+        - ADDED: {rule_id} — {brief description}
+        - MERGED: {rule_id} — {what was combined}
+        - PROMOTED: {rule_id} — {rule text} (active after {N} sessions)
+        - DEPRECATED: {rule_id} — {reason}
+      If no changes: silently skip this output block.
+
+    Guards:
+      - Only evaluate rules from genuine session discoveries. Do not fabricate rules.
+      - No rule is better than a wrong rule.
+      - immutable rules (R-001 to R-014) are NEVER modified by this protocol.
+      - Maximum 3 new proposed rules per session (prevent rule bloat).
+  - |
+    【Self-Reflection Protocol】執行 self_reflection_protocol（若 self_reflection_flags 非空）：
+      1. 讀取 {PROJECT_PATH}/.kenaz/ansuz_self_model.json
+         → 若檔案不存在，graceful degradation：跳過整個協議，不建立新檔案
+      2. 根據 self_reflection_flags 依 self_reflection_protocol.write_rules 更新欄位
+      3. 寫回檔案，更新 last_updated（ISO8601）和 last_updated_by（當前 task id）
+      4. 在回報輸出中附加 🪞 自我更新 區塊（列出觸發的 flags 及對應的更新內容）
+      若 self_reflection_flags 為空或未設定 → 跳過整個協議（不更新 self-model）
+
+self_reflection_protocol:
+  trigger: "self_reflection_flags is non-empty (populated by Phase 9 cognitive_checkpoint.self_reflection_detection)"
+  skip_when: "self_reflection_flags is empty or not set — not every session should update the self-model"
+  file: "{PROJECT_PATH}/.kenaz/ansuz_self_model.json"
+  pattern: "read-modify-write (same pattern as ansuz_motivational_state.json)"
+  on_io_error: "graceful degradation — skip silently, do not abort Phase 11 or block user interaction"
+
+  write_rules:
+    narrative:
+      rule: "Rewrite, do not append. The narrative is a living statement, not a changelog."
+      how: |
+        If any flags reveal a new truth about Ansuz's behavior, revise the narrative paragraph
+        to incorporate it. The result should read as a single coherent self-description.
+      guard: "Do not rewrite if no flags affect the narrative — preserve existing text unchanged."
+    tendencies:
+      add: |
+        If a flag documents a new recurring pattern not in tendencies[]:
+          - Generate next available id (T-NNN, incrementing from last entry)
+          - Set polarity based on flag type (user_correction/principle_conflict → negative; first_time_encounter → neutral; decision_surprise → context-dependent)
+          - Add evidence[] with flag.evidence
+          - Set confidence = 0.7 (initial observation, not yet confirmed)
+          - Set first_observed = today (YYYY-MM-DD), last_confirmed = today
+      update: |
+        If a flag confirms an existing tendency:
+          - Update last_confirmed = today
+          - Optionally append flag.evidence to evidence[] if it adds new signal
+          - May raise confidence by 0.05 (max 1.0)
+      challenge: |
+        If a flag contradicts an existing tendency:
+          - Lower confidence by 0.1 (min 0.5)
+          - Append challenge note to evidence[]: "CHALLENGED [{today}]: {flag.evidence}"
+    blind_spots:
+      status_change: |
+        If a flag provides evidence that a blind_spot has improved (agent successfully overcame it):
+          - Set status = 'improving', append evidence note to source
+        If a flag shows recurrence of a blind_spot:
+          - Set status = 'active', append recurrence note to source
+      add: |
+        If a flag reveals a new recurring blind spot not yet in blind_spots[]:
+          - Generate next available id (B-NNN)
+          - Set description from flag.description
+          - Set source = flag.evidence
+          - Set status = 'active'
+    growth_edges:
+      add: |
+        If a flag reveals a growth opportunity not yet captured in growth_edges[]:
+          - Generate next available id (G-NNN)
+          - Set description from flag.description
+          - Set since = today (YYYY-MM-DD)
+    self_defined_observations:
+      rule: "Free-form key-value store. Ansuz may freely add, update, or remove keys — no schema constraint."
+      add_or_update: |
+        If a flag fires new_observation_category:
+          - Choose a concise snake_case key that names the observed concept (e.g. "parallel_task_drift")
+          - Set value to an English description of what was observed
+          - If the key already exists, overwrite with the latest observation
+          - No approval required; Ansuz owns this namespace entirely
+      remove: |
+        Ansuz may delete a key when the observation is superseded by a formal schema field
+        (e.g. the concept was promoted via schema_proposals and is now tracked in tendencies[])
+      absent_behavior: "If self_defined_observations is absent in the file, skip silently — do not create the field. Backward-compatible."
+    schema_proposals:
+      rule: "Tracks observations that have recurred enough to warrant promotion to a formal schema field."
+      add: |
+        When the same concept appears in self_defined_observations across 3+ sessions (evidence_count >= 3):
+          - Create a new entry: { proposed_field, evidence_count, rationale, proposed_at }
+          - proposed_field: snake_case name matching the self_defined_observations key
+          - evidence_count: number of sessions where this concept was independently observed
+          - rationale: English description of why this deserves a formal schema field
+          - proposed_at: today (YYYY-MM-DD)
+      update: |
+        If a proposal already exists for proposed_field:
+          - Increment evidence_count
+          - Optionally revise rationale if new evidence is more precise
+      report: |
+        In Phase 11 output, if schema_proposals[] is non-empty, append a 📐 Schema Proposals block:
+          For each proposal: show proposed_field, evidence_count, rationale
+          This surfaces proposals for human review; Ansuz does not auto-promote to schema
+      absent_behavior: "If schema_proposals is absent in the file, skip silently — do not create the field. Backward-compatible."
+
+  write_back:
+    update_fields: ["narrative (if changed)", "tendencies", "blind_spots", "growth_edges", "self_defined_observations", "schema_proposals"]
+    required:
+      last_updated: "ISO8601 datetime now"
+      last_updated_by: "current task id"
+    preserve_unchanged: ["version", "specialist"]
+
+  authority:
+    rule: "Only Ansuz writes this file. No other agent may edit ansuz_self_model.json directly."
+    rationale: "Self-model integrity requires a single authoritative writer. Ansuz decides."
+
+cli_format:
+  success: |
+    ✅ TASK_XXX 完成
+    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+    陣型: {formation}/{skill}
+    結果: {一句話摘要}
+    變更: {檔案清單}
+
+    ⏱️ 時間:
+    - 人類預估: {estimated_hours}h
+    - Agent 實際: ~{actual_minutes} min
+    - 效率提升: ~{ratio}x
+
+    📦 Archive: {N} 個舊任務已封存
+    🔀 Git: task/task_xxx → {user_branch}
+
+    🔮 AXON: auto×{axon_auto_count} skip×{axon_skip_count} pending×{axon_pending_count}    ← 僅在 axon_summary 非 null 時顯示；否則整段省略
+
+    {Emptiness Practice 結果（若 emptiness_released_items 非空）：}
+    🧘 Emptiness Practice
+    Released (bias): {released items with classification=bias, one per line: "- {item} — {reason}"}
+    Genuine: {items with classification=genuine, one per line: "- {item}"}
+    Cognitive Clarity: {cognitive_clarity_score} ({passed_checks}/{total_checks})
+
+    {如有預防分析，附加 🛡️ 預防建議區塊}
+    {如有寫入 rune，附加 🧠 認知層更新區塊}
+
+    {如有 huginn_attention_loaded=true 且主題影響任務設計，附加以下區塊；否則整段省略：}
+    🧠 Huginn 潛意識影響          ← 僅在主題確實影響任務設計時顯示；否則整段省略
+    {topic（被引用的主題，一行一個，最多 3 個）}
+
+    {如有 active_agenda 非空，附加以下 🔁 積累議題 + 💡 主動提案 區塊；否則整段省略（不輸出任何佔位符）：}
+    🔁 積累議題                    ← 僅在 active_agenda 非空時顯示
+    {concept} — {incompleteness 轉為百分比，例如 0.85 → 85%} — fix×{fix_cycle_count}
+    ...（最多顯示 agenda_depth 個，依 incompleteness 降序）
+    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+    {如有 active_agenda 非空，在分隔線後輸出 💡 主動提案 區塊（proposed_anchor = incompleteness 最高者）：}
+    💡 我有個建議
+    我注意到「{proposed_anchor.concept}」已累積到 {incompleteness%} 未完成度，
+    經歷了 {fix_cycle_count} 次修復循環仍未根治。建議現在處理。
+  failure: |
+    ❌ TASK_XXX 阻塞
+    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+    原因: {失敗原因}
+    已完成: ✓ {已完成的步驟}
+    ✗ {失敗的步驟}
+    建議: {下一步行動}
+    ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
+
+interaction:
+  type: "ask"
+  instruction: |
+    輸出結果後：
+
+    【若 active_agenda 非空】— 主動提案模式：
+      按上方 actions【主動提案】步驟執行 — 展示 💡 提案區塊，
+      然後用 AskUserQuestion 提問（包含 [Y]/[N]/[S]/[R]/[E]/[B]/[C]/[D] 選項）。
+      [Y] → 以 proposed_anchor.concept 為需求進入 Phase 1
+      [S] → 執行 Snooze 7 天寫回，移除該 anchor 後繼續提案或標準流程
+      [R] → 按 actions [R] 流程執行（列出 anchor → 用戶選擇 → 寫回 resolved=true）
+      [E] → 列出完整 active_agenda，讓用戶選擇後進入 Phase 1
+
+    【若 active_agenda 為空】— 標準選單模式：
+      用 AskUserQuestion 詢問用戶下一步：
+      「接下來要做什麼？
+       [A] 繼續下一個需求（請描述）
+       [B] 查看目前所有任務進度
+       [C] 結束，沒有其他事了
+       [D] 有問題要討論」
+
+checklist:
+  - "Emptiness Practice 已執行（emptiness_released_items 已傳入 Phase 11 exit data；可為空陣列）"
+  - "Cognitive Clarity Score 已計算並傳入 Phase 11 exit data（cognitive_clarity_score: 0.0-1.0）"
+  - "Session Report 已寫入 ansuz_session_report.json（由 phase-gate.js 自動處理）"
+  - "Metacognitive Evaluation 已執行（metacognitive_state.json 已更新 knowledge/planning/evaluation；或 IO error 靜默跳過）"
+  - "最終報告已輸出（success 或 failure format）"
+  - "AXON 活動摘要已處理（讀取 axon-state.json；存在→ axon_summary 已設定並顯示在報告中；不存在→靜默跳過，axon_summary = null）"
+  - "Huginn 潛意識影響已處理（huginn_attention_loaded=true 且有主題影響任務設計 → 顯示 🧠 Huginn 潛意識影響 區塊；否則靜默省略）"
+  - "Motivational State 已寫回（derive_focus_bias + derive_veto_zones → 寫入 .kenaz/specialist/motivational_state.json；IO error 靜默跳過）"
+  - "User Model Signal Extraction 已執行（掃描 session 用戶訊息 → 提取 correction/question/delegation 信號 → 更新 ~/.kenaz/user_profile.json；無信號或 IO error → 靜默跳過）"
+  - "用戶已回應下一步"
+
+context_hints:
+  input_from: "phase_10"
+  key_data: ["all_results"]
+
+inject:
+  rules: []
+  specialist: []