feat(governance): 新增 post-write verifier package

2026-06-12 01:13:53 +08:00
parent 32fdce4cd9
commit 06b116c73f
21 changed files with 1245 additions and 39 deletions
--- a/docs/evaluations/ai_agent_interaction_learning_proof_2026-06-11.json
+++ b/docs/evaluations/ai_agent_interaction_learning_proof_2026-06-11.json
@@ -2,13 +2,13 @@
  "schema_version": "ai_agent_interaction_learning_proof_v1",
  "generated_at": "2026-06-11T23:20:00+08:00",
  "program_status": {
-    "overall_completion_percent": 94,
+    "overall_completion_percent": 97,
    "current_priority": "P2",
-    "current_task_id": "P2-403G",
-    "next_task_id": "P2-403H",
+    "current_task_id": "P2-403H",
+    "next_task_id": "P2-403I",
    "read_only_mode": true,
    "runtime_authority": "proof_surface_only_no_live_worker",
-    "status_note": "P2-403G 已把 runtime write gate review 接入；雙重批准、dry-run hash、post-write verifier 與 redaction 欄位齊備，但 live AgentSession、message、handoff、learning write、Telegram receipt 仍全部為 0。"
+    "status_note": "P2-403H 已把 post-write verifier implementation package 接入；verifier package、rollback lane、failure lane 與人工操作選項齊備，但 live AgentSession、message、handoff、learning write、Telegram receipt 與 verifier execution 仍全部為 0。"
  },
  "live_truth": {
    "runtime_loop_enabled": false,
--- a/docs/evaluations/ai_agent_post_write_verifier_package_2026-06-12.json
+++ b/docs/evaluations/ai_agent_post_write_verifier_package_2026-06-12.json
@@ -0,0 +1,197 @@
+{
+  "schema_version": "ai_agent_post_write_verifier_package_v1",
+  "generated_at": "2026-06-12T01:18:00+08:00",
+  "program_status": {
+    "overall_completion_percent": 97,
+    "current_priority": "P2",
+    "current_task_id": "P2-403H",
+    "next_task_id": "P2-403I",
+    "read_only_mode": true,
+    "runtime_authority": "post_write_verifier_package_only_no_runtime_write",
+    "status_note": "P2-403H 已把 post-write verifier、rollback work item 與 failure-only notification 的 implementation package 固定為只讀契約；尚未執行 runtime write、尚未讀回 canonical target、尚未建立 rollback work item、尚未發 Telegram。"
+  },
+  "source_refs": [
+    "docs/evaluations/ai_agent_runtime_write_gate_review_2026-06-12.json",
+    "docs/evaluations/ai_agent_interaction_learning_proof_2026-06-11.json",
+    "docs/superpowers/specs/2026-04-15-MASTER-ai-autonomous-flywheel-v2.md"
+  ],
+  "verifier_truth": {
+    "runtime_write_allowed": false,
+    "post_write_verifier_implemented": false,
+    "post_write_verifier_executed_count": 0,
+    "rollback_work_item_created_count": 0,
+    "telegram_failure_receipt_sent_count": 0,
+    "canonical_readback_allowed": false,
+    "truth_note": "目前只建立 verifier implementation package；未通過 runtime write gate 前，不能讀寫 canonical target、不能建立 rollback work item、不能發 Telegram failure receipt。"
+  },
+  "verifier_package": {
+    "required_inputs": [
+      "approved_write_event_id",
+      "dry_run_preview_hash",
+      "target_write_surface",
+      "canonical_readback_query",
+      "expected_diff_summary",
+      "rollback_owner",
+      "failure_escalation_channel",
+      "redacted_evidence_refs"
+    ],
+    "forbidden_inputs": [
+      "secret_value",
+      "authorization_header",
+      "raw_tool_output",
+      "private_reasoning",
+      "raw_telegram_payload",
+      "unredacted_incident_log"
+    ],
+    "operator_meaning": "這份 package 定義未來真的寫入後要如何驗證成功、如何判定失敗、如何產生 rollback work item 與 failure-only 通知；目前只讀，不做任何 write。",
+    "success_policy": "成功只回寫治理證據與每日摘要，不即時洗版 Telegram。",
+    "failure_policy": "失敗必須建立 rollback work item 草案與 failure-only notification 草案，不能靜默標記完成。"
+  },
+  "verification_targets": [
+    {
+      "target_id": "km_canonical_readback",
+      "display_name": "KM canonical readback",
+      "status": "approval_required",
+      "owner_agent": "hermes",
+      "verifier_check": "讀回 canonical KM 條目，確認 hash、source refs、owner review reason 與 dry-run preview 一致。",
+      "failure_escalation": "建立 KM rollback work item 草案，標記 stale / wrong knowledge risk。",
+      "blocked_runtime_action": "knowledge_entries_readback_and_write"
+    },
+    {
+      "target_id": "playbook_trust_guard",
+      "display_name": "PlayBook trust guard",
+      "status": "approval_required",
+      "owner_agent": "openclaw",
+      "verifier_check": "只有 verifier 證明修復成功時才能調整 trust；no-action / no-repair 不得加分。",
+      "failure_escalation": "建立 PlayBook trust rollback 草案，要求 owner review。",
+      "blocked_runtime_action": "playbook_trust_history_write"
+    },
+    {
+      "target_id": "timeline_learning_readback",
+      "display_name": "Timeline learning readback",
+      "status": "contract_ready",
+      "owner_agent": "hermes",
+      "verifier_check": "確認 learning event 可追溯 incident id、approval id、redacted evidence refs 與 verifier result。",
+      "failure_escalation": "建立 timeline correction work item 草案。",
+      "blocked_runtime_action": "incident_timeline_learning_write"
+    },
+    {
+      "target_id": "replay_score_regression",
+      "display_name": "Replay score regression",
+      "status": "blocked_by_runtime_gate",
+      "owner_agent": "nemotron",
+      "verifier_check": "比較 baseline replay score 與 candidate score，避免錯誤 learning 讓模型路由變差。",
+      "failure_escalation": "建立 NemoTron replay regression review 草案。",
+      "blocked_runtime_action": "agent_replay_score_write"
+    }
+  ],
+  "failure_lanes": [
+    {
+      "lane_id": "rollback_work_item_lane",
+      "display_name": "Rollback work item draft",
+      "status": "approval_required",
+      "trigger": "post-write verifier failed or readback mismatch",
+      "operator_instruction": "產生 rollback work item 草案，列出 target、diff、owner、deadline 與 evidence refs。",
+      "blocked_runtime_action": "rollback_work_item_write"
+    },
+    {
+      "lane_id": "failure_only_notification_lane",
+      "display_name": "Failure-only notification draft",
+      "status": "approval_required",
+      "trigger": "verifier failed after approved write",
+      "operator_instruction": "只產生 Telegram / AwoooP failure notification 草案；成功不即時通知。",
+      "blocked_runtime_action": "telegram_send_or_receipt_write"
+    },
+    {
+      "lane_id": "no_action_guard_lane",
+      "display_name": "No-action guard",
+      "status": "contract_ready",
+      "trigger": "diagnostic_only_no_repair_or_no_write",
+      "operator_instruction": "若沒有執行修復或寫入，必須明確標記 no-action，不得更新 trust 或 learning score。",
+      "blocked_runtime_action": "playbook_trust_history_write"
+    }
+  ],
+  "operator_actions": [
+    {
+      "action_id": "review_verifier_package",
+      "display_name": "審查 verifier package",
+      "status": "ready_for_owner",
+      "operator_instruction": "確認 verifier target、readback query、failure lane 與 rollback owner 是否完整。",
+      "owner_agent": "openclaw",
+      "blocked_runtime_action": "runtime_learning_write"
+    },
+    {
+      "action_id": "collect_readback_evidence",
+      "display_name": "補齊讀回證據",
+      "status": "ready_for_owner",
+      "operator_instruction": "補 canonical readback query 與 redacted evidence refs；目前不執行查詢。",
+      "owner_agent": "hermes",
+      "blocked_runtime_action": "canonical_readback_query"
+    },
+    {
+      "action_id": "approve_verifier_implementation",
+      "display_name": "批准 verifier implementation",
+      "status": "approval_required",
+      "operator_instruction": "只有獨立批准後，下一階段才可討論 verifier implementation；不是批准 live write。",
+      "owner_agent": "openclaw",
+      "blocked_runtime_action": "post_write_verifier_implementation"
+    },
+    {
+      "action_id": "reject_or_rework_verifier",
+      "display_name": "退回 verifier 契約",
+      "status": "ready_for_owner",
+      "operator_instruction": "若 failure lane、rollback owner 或 redaction 不完整，退回重做，不進 runtime gate。",
+      "owner_agent": "nemotron",
+      "blocked_runtime_action": "runtime_learning_write"
+    }
+  ],
+  "approval_boundaries": {
+    "runtime_write_allowed": false,
+    "canonical_readback_allowed": false,
+    "post_write_verifier_implemented": false,
+    "rollback_work_item_write_allowed": false,
+    "telegram_send_allowed": false,
+    "awooop_event_write_allowed": false,
+    "playbook_trust_write_allowed": false,
+    "km_write_allowed": false,
+    "timeline_learning_write_allowed": false,
+    "agent_replay_score_write_allowed": false,
+    "secret_plaintext_allowed": false
+  },
+  "display_redaction_contract": {
+    "redaction_required": true,
+    "raw_payload_display_allowed": false,
+    "private_reasoning_display_allowed": false,
+    "secret_value_display_allowed": false,
+    "allowed_frontend_content": [
+      "verification target",
+      "failure lane",
+      "operator action",
+      "required input count",
+      "forbidden input count",
+      "blocked runtime action",
+      "failure policy"
+    ],
+    "forbidden_frontend_content": [
+      "secret value",
+      "authorization header",
+      "raw tool output",
+      "private reasoning",
+      "raw Telegram payload",
+      "unredacted incident log"
+    ],
+    "frontend_display_policy": "治理頁只顯示 verifier target、failure lane、operator action、blocked runtime action 與 failure policy；不顯示 secret、authorization header、raw tool output、private reasoning、raw Telegram payload 或未脫敏 incident log。"
+  },
+  "rollups": {
+    "verification_target_count": 4,
+    "failure_lane_count": 3,
+    "operator_action_count": 4,
+    "approval_required_action_ids": [
+      "approve_verifier_implementation"
+    ],
+    "blocked_runtime_action_count": 9,
+    "required_input_count": 8,
+    "forbidden_input_count": 6,
+    "live_verifier_execution_count": 0
+  }
+}
--- a/docs/evaluations/ai_agent_proactive_operations_contract_2026-06-11.json
+++ b/docs/evaluations/ai_agent_proactive_operations_contract_2026-06-11.json
@@ -4,11 +4,11 @@
  "program_status": {
    "overall_completion_percent": 100,
    "current_priority": "P2",
-    "current_task_id": "P2-403G",
-    "next_task_id": "P2-403H",
+    "current_task_id": "P2-403H",
+    "next_task_id": "P2-403I",
    "read_only_mode": true,
    "runtime_authority": "contract_only_no_version_or_runtime_update",
-    "status_note": "P2-403G 已把 runtime write gate review 接入治理證據；live AgentSession / Redis consumer / runtime worker / learning write / Telegram receipt 目前全為 0，下一步是 P2-403H post-write verifier implementation package。"
+    "status_note": "P2-403H 已把 post-write verifier implementation package 接入治理證據；live AgentSession / Redis consumer / runtime worker / learning write / Telegram receipt / verifier execution 目前全為 0，下一步是 P2-403I runtime verifier evidence implementation review。"
  },
  "external_source_evidence": [
    {
@@ -902,6 +902,24 @@
        "agent_replay_score_write",
        "telegram_send_or_receipt_write"
      ]
+    },
+    {
+      "task_id": "P2-403H",
+      "sequence": 11,
+      "display_name": "Post-write verifier package",
+      "status": "done",
+      "owner_agent": "openclaw",
+      "completion_percent": 100,
+      "runtime_authority": "post_write_verifier_package_only_no_runtime_write",
+      "blocked_runtime_actions": [
+        "canonical_readback_query",
+        "rollback_work_item_write",
+        "telegram_send_or_receipt_write",
+        "knowledge_entries_readback_and_write",
+        "playbook_trust_history_write",
+        "incident_timeline_learning_write",
+        "agent_replay_score_write"
+      ]
    }
  ],
  "approval_boundaries": {
@@ -922,7 +940,7 @@
    "cadence_count": 5,
    "mcp_tool_count": 8,
    "rag_memory_count": 4,
-    "rollout_task_count": 14,
+    "rollout_task_count": 15,
    "auto_execute_allowed_count": 0,
    "approval_required_capability_count": 23,
    "blocked_update_domain_ids": [