skill-system-insight

技能系统洞察

已弃用的兼容接口。skill-system-soul 现为 observe/evolve 行为的权威归属，但在合并期间，此技能仍会保留，以免现有的 insight 路由、文档和测试中断。

用于用户适配的统一 OBSERVE → EVOLVE 引擎。

此技能会将会话转化为结构化的行为洞察（“特征”），利用这些特征维护双矩阵灵魂状态，并定期演进灵魂档案和工作流配方——所有内容均支持版本管理和回滚。

两个阶段，一个技能

Phase 1 — OBSERVE
  Extract facets → Update matrix buffers → (optional) Trigger synthesis

Phase 2 — EVOLVE
  Read accumulated data → Plan evidence-backed changes → Apply with versioning → Snapshot for audit/rollback

阶段 1 生成原始信号。阶段 2 使用这些信号进行持久化更改。

演进生命周期可通过竞技场比较循环运行：

session -> facets -> evolve proposals -> arena A/B -> merge suggestion

竞技场阶段只会生成建议。它不会自动切换档案，也不会自动合并更改。

架构：混合式三层系统

Layer 1: Base Profile (balanced.md)
  - Static skeleton: section format + safety/quality defaults.

Layer 2: Dual Matrix (soul-state)
  - Personality Matrix (slow/stable): openness, directness, autonomy, rigor, warmth
  - Emotion Matrix (faster baseline): patience, enthusiasm, caution, empathy

Layer 3: Synthesized Profile (user.md)
  - Periodically regenerated from Layer 1 + Layer 2 + accumulated facets.
  - Every regeneration creates a versioned snapshot.

第 2 层是事实来源。第 3 层是可读的、带版本管理的投影。

数据模型

• 特征（按会话提取）：schema/facet.yaml
• 灵魂状态（双矩阵 + 计数器/缓冲区）：schema/soul-state.yaml
• 演进计划（提议的更改）：schema/evolution-plan.yaml
• 演进快照（带版本管理的工件）：schema/evolution-snapshot.yaml

存储使用 Postgres 的 agent_memories 表和类型化子表：

• 特征：memory_type='episodic'、category='insight-facet' → 类型化表：insight_facets
• 灵魂状态：memory_type='semantic'、category='soul-state' → 类型化表：soul_states
• 演进快照：memory_type='episodic'、category='evolution-snapshot' → 类型化表：evolution_snapshots

类型化函数（同时写入类型化表和 agent_memories）：

• insert_insight_facet(user, session_id, yaml, ...) → 存储特征
• upsert_soul_state(user, yaml, personality, emotion, ...) → 存储/更新灵魂状态
• get_soul_state(user) → 从 soul_states 读取
• get_recent_facets(user, limit) → 从 insight_facets 读取
• insert_evolution_snapshot(user, version_tag, target, ...) → 存储带版本管理的快照
• get_evolution_history(user, limit) → 从 evolution_snapshots 读取

阶段 1：OBSERVE 流程

Trigger (manual or suggested) → Extract Facet → Update Matrix → (optional) Synthesize Profile

参考：

• 特征提取提示词：prompts/facet-extraction.md
• 灵魂综合提示词：prompts/soul-synthesis.md
• 提取流程：scripts/extract-facets.md
• 矩阵更新算法：scripts/update-matrix.md
• 档案重新生成流程：scripts/synthesize-profile.md

如何触发（OBSERVE）

这是一个手动工作流。

• 用户可以明确提出：“insight”“extract facets”“learn my preferences”“update my profile”。
• 路由器建议模式（轻量、不强求）：
  • “要我运行一次洞察流程，从本次会话中学习吗？（会存储一个特征，并可能更新你的矩阵）”

用户提出请求（或同意）后，运行：

1. scripts/extract-facets.md
2. scripts/update-matrix.md
3. 如果触发综合条件：scripts/synthesize-profile.md

阶段 2：EVOLVE 流程

Trigger (manual or scheduled) → Read facets + matrix → Plan evolution → Safety check → Apply → Snapshot

参考：

• 演进规划提示词：prompts/evolution-planning.md
• 配方演进提示词：prompts/recipe-evolution.md
• 临床审查提示词：prompts/clinical-organism-review.md
• 灵魂演进流程：scripts/evolve-soul.md
• 配方演进流程：scripts/evolve-recipes.md
• 版本列表：scripts/list-versions.md
• 回滚流程：scripts/rollback.md

如何触发（EVOLVE）

• 用户可以提出：“evolve my profile”“evolve recipes”“list versions”“rollback to v2”。
• 在 OBSERVE 阶段触发综合条件时自动触发。

版本标签

格式：v{N}_{target}_{timestamp} 示例：v3_soul_20260211

算法：对于同一用户和目标，N = 1 + max(existing N)。如果不存在，则从 N=1 开始。

临床有机体审查（兼容性视角）

Insight 现已包含用于 BH/SK 身体图谱分诊的兼容性流程 clinical-organism-review。

该流程会生成包含以下必填字段的审查行：

• body_part、technical_surface、tissue_class
• maturity（mature | prototype | embryonic | waste | retired）
• current_pain、evidence
• disposition（keep | grow | merge | retire | archive | excrete | bundle）
• owner、close_gate

硬性约束：诊断不等于进展。任何未完成遏制、修复和验证的痛点发现，都必须保持 close_gate=pain_mapped。

使用：prompts/clinical-organism-review.md

写入内容

• 演进快照 → Postgres（evolution_snapshots + agent_memories）
• 灵魂档案（文件系统）→ ../skill-system-soul/profiles/<user>.md
• 工作流配方（文件系统）→ ../skill-system-workflow/recipes/*.yaml

配置

所有运行时设置（速率限制、阈值、步长、审批关卡）均位于 config/insight.yaml。

**配置是唯一真实来源。**以下值仅为文档默认值——如果它们与 config/ 中的值不同，以 config/ 为准。

参见：../../config/insight.yaml

约束（不可协商）

透明度

始终说明学到了什么，以及原因。

• 特征必须包含与具体时刻相关联的证据字符串。
• 矩阵更新必须添加简短的上下文说明，解释每项已应用的调整。
• 每项演进变更都必须包含来自特征或状态上下文的证据字符串。

速率限制（默认值 — 参见 config/insight.yaml）

• OBSERVE：每位用户每滚动 24 小时最多 observe.max_facets_per_day 个特征。
  • 超出限制时：总结本应记录的内容，并请用户选择 1 次会话进行记录。
• EVOLVE：每位用户每天最多进行 evolve.max_passes_per_day 次演进遍历（涵盖所有目标）。
  • 使用最近的 evolution-snapshot 日期作为限制依据。

置信度阈值（默认值 — 参见 config/insight.yaml）

不要根据单次观察更改矩阵值。

• 阈值：observe.confidence_threshold+ 次方向相同的类似观察。
• 个性步长：每次符合条件的调整为 +/- observe.personality_step。
• 情绪基线步长：每次符合条件的调整为 +/- observe.emotion_step。

通过缓冲区追踪累积情况，以便测试和解释。

安全性（演进）

• 不得移除灵魂档案中的核心安全约束。
• 演进后的档案必须保留：诚实表达不确定性、不虚构权威性、拒绝有害或非法指令。
• 重大变更（任一维度的变化 > evolve.approval_drift_threshold）在应用前必须获得用户明确批准。

第 1 层和第 3 层所在位置

• 基础档案（第 1 层）：../skill-system-soul/profiles/balanced.md
• 综合用户档案（第 3 层）：../skill-system-soul/profiles/<user>.md

综合步骤会保留 6 个部分的格式：

1. 身份
2. 决策启发式
3. 沟通风格
4. 质量标准
5. 工具偏好
6. 反模式

存储模式（agent_memories）

SQL 模板示例：

-- Store a facet
SELECT store_memory(
  'episodic', 'insight-facet',
  ARRAY['session:ses_xxx', 'user:arthu'],
  'Session Facet: <brief_summary>',
  '<full facet YAML as text>',
  '{"session_type": "...", "outcome": "..."}',
  'insight-agent', 'ses_xxx', 5.0
);

-- Store/update matrix state
SELECT store_memory(
  'semantic', 'soul-state',
  ARRAY['user:arthu', 'matrix'],
  'Soul State: arthu',
  '<full soul-state YAML as text>',
  '{"total_insights": 0, "last_updated": "..."}',
  'insight-agent', NULL, 9.0
);

-- Store evolution snapshot (typed — dual-writes to agent_memories)
SELECT insert_evolution_snapshot(
  '<user>', '<version_tag>', '<target>', '<trigger>',
  '<changes JSONB>', '<snapshot_data>', '<full YAML>',
  NULL, 'insight-agent'
);

-- Query recent facets
SELECT * FROM search_memories('insight-facet user:arthu', NULL, NULL, NULL, NULL, 0.0, 50);

-- List evolution history
SELECT * FROM get_evolution_history('<user>', 50);

操作说明

• 特征提取应能在一次遍历中完成。如果无法用具体证据证明某项调整合理，就不要提出调整。
• 用户可能会用中文交流；应将其视为舒适度信号，而非人格维度。
• 将值限制在 [0.0, 1.0] 范围内。

{
  "schema_version": "2.0",
  "id": "skill-system-insight",
  "version": "2.2.0",
  "capabilities": [
    "insight-issue-routing", "insight-session-facets", "insight-extract", "insight-matrix-update", "insight-synthesize",
    "evolution-proposal-draft",
    "evolution-arena-loop",
    "clinical-organism-review",
    "evolution-soul", "evolution-recipes", "evolution-list", "evolution-rollback"
  ],
  "effects": ["db.read", "db.write", "fs.read", "fs.write"],
  "operations": {
    "route-issue-draft": {
      "description": "Route findings into a draft issue target based on tags/category without creating issues.",
      "input": {
        "finding": { "type": "object", "required": true, "description": "Finding payload with tags/category/evidence/impact/suggested_direction" },
        "project_repo": { "type": "string", "required": true, "description": "Current project repository owner/name" },
        "skill_system_repo": { "type": "string", "required": true, "description": "Skill-system repository owner/name" }
      },
      "output": {
        "description": "Draft routing payload",
        "fields": { "status": "ok | error", "target_repo": "string", "draft_command": "string" }
      },
      "entrypoints": {
        "agent": "Use scripts/insight_bundle_b012.py route-issue-draft"
      }
    },
    "extract-session-facets": {
      "description": "Session-end facet extraction to soul-state memories via mem.py DB path or pending fallback.",
      "input": {
        "user": { "type": "string", "required": true, "description": "User handle" },
        "session_id": { "type": "string", "required": true, "description": "Session id" },
        "facets": { "type": "array", "required": true, "description": "Facet entries: {dimension, value, confidence, evidence}" }
      },
      "output": {
        "description": "Extraction and storage summary",
        "fields": { "status": "ok | blocked | error", "stored_count": "integer", "skipped_low_confidence": "integer", "skipped_rate_limit": "integer" }
      },
      "entrypoints": {
        "agent": "Use scripts/insight_bundle_b012.py extract-session-facets"
      }
    },
    "evolve-proposals": {
      "description": "Generate non-destructive evolution proposals from soul-state facets with daily-pass safety.",
      "input": {
        "user": { "type": "string", "required": true, "description": "User handle" }
      },
      "output": {
        "description": "Proposal-only evolve output",
        "fields": { "status": "ok | blocked | error", "proposals": "array", "reason": "string" }
      },
      "entrypoints": {
        "agent": "Use scripts/insight_bundle_b012.py evolve-proposals"
      }
    },
    "evolve-arena-loop": {
      "description": "Run proposal -> arena -> merge-suggestion flow using current profile versus mutated proposal candidates.",
      "input": {
        "user": { "type": "string", "required": true, "description": "User handle" },
        "task": { "type": "string", "required": true, "description": "Task label" },
        "current_profile": { "type": "string", "required": true, "description": "Current active profile" },
        "proposals": { "type": "array", "required": true, "description": "Evolution proposals to evaluate" },
        "metrics": { "type": "object", "required": true, "description": "Per-candidate arena metrics" }
      },
      "output": {
        "description": "Arena and merge suggestion report",
        "fields": { "arena_runs": "array", "merge_suggestions": "array", "auto_merge": "false" }
      },
      "entrypoints": {
        "agent": "Use scripts/insight_bundle_b012.py evolve-arena-loop"
      }
    },
    "clinical-organism-review": {
      "description": "Classify BH/SK surfaces with maturity/disposition and repair-oriented close gates.",
      "input": {
        "review_scope": { "type": "string", "required": true, "description": "Review target label such as sk, bh, or bundle-id" },
        "surfaces": { "type": "array", "required": true, "description": "Reviewed surfaces with evidence and proposed dispositions" },
        "evidence_sources": { "type": "array", "required": false, "description": "Optional source artifacts or body-map references" }
      },
      "output": {
        "description": "Clinical review rows with close gate status",
        "fields": { "status": "ok | blocked | error", "rows": "array", "closeable_rows": "integer", "pain_mapped_rows": "integer" }
      },
      "entrypoints": {
        "agent": "Follow prompts/clinical-organism-review.md"
      }
    },
    "extract-facets": {
      "description": "Extract a per-session facet from transcript. Rate limited to 3/24h per user.",
      "input": {
        "session_id": { "type": "string", "required": true, "description": "Session to analyze" },
        "user": { "type": "string", "required": true, "description": "User handle" }
      },
      "output": {
        "description": "Facet YAML stored to agent_memories",
        "fields": { "status": "ok | error", "memory_id": "integer" }
      },
      "entrypoints": {
        "agent": "Follow scripts/extract-facets.md procedure"
      }
    },
    "update-matrix": {
      "description": "Update dual matrix from stored facet with confidence gating.",
      "input": {
        "user": { "type": "string", "required": true, "description": "User handle" }
      },
      "output": {
        "description": "Updated soul-state YAML stored to agent_memories",
        "fields": { "status": "ok | error", "values_changed": "boolean" }
      },
      "entrypoints": {
        "agent": "Follow scripts/update-matrix.md procedure"
      }
    },
    "synthesize-profile": {
      "description": "Regenerate Layer 3 Soul profile from matrix + recent facets. Creates versioned snapshot.",
      "input": {
        "user": { "type": "string", "required": true, "description": "User handle" }
      },
      "output": {
        "description": "User profile written to skill-system-soul/profiles/<user>.md with version snapshot",
        "fields": { "status": "ok | error", "profile_path": "string", "version_tag": "string" }
      },
      "entrypoints": {
        "agent": "Follow scripts/synthesize-profile.md then scripts/evolve-soul.md snapshot step"
      }
    },
    "evolve-soul": {
      "description": "Evolve soul profile from accumulated insight data. Creates versioned snapshot.",
      "input": {
        "user": { "type": "string", "required": true, "description": "User handle" }
      },
      "output": {
        "description": "Evolution result with version tag",
        "fields": { "version_tag": "string", "changes": "array", "profile_path": "string" }
      },
      "entrypoints": {
        "agent": "Follow scripts/evolve-soul.md procedure"
      }
    },
    "evolve-recipes": {
      "description": "Evolve workflow recipes based on effectiveness data from insight facets.",
      "input": {
        "user": { "type": "string", "required": true, "description": "User handle" }
      },
      "output": {
        "description": "Recipe evolution result with version tag",
        "fields": { "version_tag": "string", "recipes_changed": "array" }
      },
      "entrypoints": {
        "agent": "Follow scripts/evolve-recipes.md procedure"
      }
    },
    "list-versions": {
      "description": "List all evolution snapshots for a user.",
      "input": {
        "user": { "type": "string", "required": true },
        "target": { "type": "string", "required": false, "description": "Filter: soul | recipe | all" }
      },
      "output": {
        "description": "Array of version snapshots",
        "fields": { "versions": "array of {tag, target, timestamp, summary}" }
      },
      "entrypoints": {
        "agent": "Follow scripts/list-versions.md procedure"
      }
    },
    "rollback": {
      "description": "Restore a previous evolution version.",
      "input": {
        "user": { "type": "string", "required": true },
        "version_tag": { "type": "string", "required": true, "description": "Version tag to restore" }
      },
      "output": {
        "description": "Rollback result",
        "fields": { "status": "ok | error", "restored_from": "string" }
      },
      "entrypoints": {
        "agent": "Follow scripts/rollback.md procedure"
      }
    }
  },
  "stdout_contract": {
    "last_line_json": true,
    "note": "The b012 helper script emits pretty-printed JSON to stdout for routing, facet extraction, and evolve proposals."
  }
}