Reflect

Reflect Skill

Multi-pass adversarial extraction of personal insights from conversations.

Protocol reference: references/reflection-protocol.md — schemas, lens prompts, adversary rules, and storage conventions.

When to Use

• User types /reflect explicitly
• Asks to "reflect on this conversation", "复盘", "what did I learn"

Runtime

Step 0: Load Context

1. user_dna.json — Read state/user_dna.json. If missing: "没有现有的自我模型做对比，建议先运行 value-discovery。"
2. reflections.jsonl — Read state/reflections.jsonl. Parse each line as JSON. Run integrity checks. Report: "reflections.jsonl: [N] 条, [M] 条损坏已跳过".
3. protocol — Read references/reflection-protocol.md for lens prompts and schemas.
4. RAL records — Read state/records.jsonl. Filter for processed_at: null. Include as additional signal sources alongside the conversation transcript. If a record has value_tags, pass them as "user self-tagged" signals — higher confidence.
5. Pending experiments — Check most recent reflection for action_experiments with status: "active". Determine age:

Step 0.5: Preprocessing (long conversations)

• Short (<40 turns or <5k words): pass full transcript to all Lens agents
• Long (40+ turns or 5k+ words): extract 5-8 most signal-rich excerpts first

For long conversations, extract:

Moments with emotional weight, decisions, trade-offs, unprompted initiations, flow states. For each excerpt, include a one-line label and the verbatim exchange.

Lens agents receive: condensed signal map (primary) + full transcript (reference).

Step 1: Pass 1 — Parallel 3-Lens Extraction

Announce: "正在通过三个视角分析这次对话..."

Spawn three subagents in parallel using the Agent tool. Each receives the full conversation transcript (or condensed map from Step 0.5) + current user_dna.json + lens-specific prompt from references/reflection-protocol.md.

All three follow the same workflow: Segment → Focus → Extract.

1. Value Lens (label: "reflect:value-lens")

   • Extract what the user is pursuing — direction, attraction, energy
   • Schema: segments, focus_segments, candidate_values, attraction_signals, emotional_spikes, summary

2. Ability Lens (label: "reflect:ability-lens")

   • Extract demonstrated and emerging capabilities
   • Schema: segments, focus_segments, demonstrated_abilities, emerging_edges, new_connections, summary

3. Pattern Lens (label: "reflect:pattern-lens")

   • Identify recurring patterns, cross-domain connections, energy signature, abstraction layers
   • Schema: segments, focus_segments, identified_patterns, abstraction_layers, cross_domain_connections, energy_signature, recurring_dilemmas, decision_heuristics, summary

Step 1.5: Validation Gate

1. Parse JSON — each agent output must be valid JSON
2. Check required fields — segments (non-empty), focus_segments (non-empty), summary (non-empty)
3. Sanity check — do segment labels match actual conversation topics?
4. Classify each lens: passed | degraded | failed

Degraded mode:

Step 2: Pass 2 — Adversary Agent

Spawn adversary using the Agent tool (label: "reflect:adversary").

Receives all three lens outputs. Uses the adversary prompt from references/reflection-protocol.md. Three roles:

1. Truth calibration — verify claims have evidence
2. Meaning expansion — alternative framings for each finding
3. Action concretization — generate testable "If [trigger], then [action]" experiments

Step 3: Synthesize & Present

Present findings conversationally:

"这次对话中我注意到——"

情绪层面: [emotional highlights] 能力层面: [demonstrated abilities + emerging edges] 模式层面: [patterns + abstraction layers] 能量地图: [energizing vs. draining] 信号质量: [score] / 1.0

Present proposed user_dna.json diffs:

"基于以上信号，我建议对你的自我模型做以下调整："

• 价值观: [key] [from → to] — 证据: [evidence]
• 信念: [add/modify/remove]
• 决策准则: [context]: [rule]
• 偏好: [field]: [from → to]

Present action experiments:

"以下是基于本轮发现的行动实验：" | 洞察 | 行动规则 | 验证方式 |

Confirm: "请逐条确认——接受、拒绝、还是修改？"

Step 4: Persist

1. Apply accepted diffs to user_dna.json — merge changes, keep existing fields
2. Mark records processed — set processed_at + linked_reflection_id
3. Save experiments — selected: status: "active", unselected: status: "skipped"
4. Write reflection event to state/reflections.jsonl — full schema from protocol
5. Confirm: "已保存。复盘 ID: [id]。状态更新: user_dna.json 已更新 [N] 项 / reflections.jsonl 累计 [N] 条。"

Step 5: Auto-Suggest /distill

Calculate cumulative impact of unprocessed reflections. If impact >= threshold:

"你的复盘记录中累计影响分数为 [score]，建议运行 /distill 进行一次阶段性合成。现在跑还是稍后？"

Cold Start

First /reflect (no reflections.jsonl or empty):

• Run FULL 3-agent protocol. Don't simplify.
• Pattern Lens: "First reflection — cross-referencing user_dna.json only."
• Adversary: "No historical data available — calibrate against user_dna.json and cross-lens corroboration only."
• Output: "这是你的第一次复盘——历史模式会随着更多复盘数据而浮现。"

Edge Cases

Key Files