session-analysis（会话分析）

Conversation Analysis

Signal extraction → pattern detection → behavioral insights.

<when_to_use>

• User requests conversation analysis
• Identifying frustration, success, or workflow patterns
• Extracting user preferences and requirements
• Understanding task evolution and iterations

NOT for: real-time monitoring, content generation, single message analysis

</when_to_use>

<signal_taxonomy>

Confidence levels:

• High (0.8–1.0): Explicit keywords match taxonomy, no ambiguity, strong context
• Medium (0.5–0.79): Implicit signal, partial context, minor ambiguity
• Low (0.2–0.49): Ambiguous language, weak context, borderline classification

</signal_taxonomy>

Load the maintain-tasks skill for stage tracking. Stages advance only, never regress.

Task format:

- Parse Input { scope description }
- Extract Signals { from N messages }
- Detect Patterns { category focus }
- Synthesize Report { output format }

Edge cases:

• Small scope (<5 messages): Skip Extract Signals, jump to Synthesize
• Re-analysis: Resume at Detect Patterns
• Narrow focus (single signal type): Skip Detect Patterns

Workflow:

• Start: Create Parse Input in_progress
• Transition: Mark current completed, add next in_progress
• After delivery: Mark Synthesize Report completed

1. Define Scope

- Message range (all, recent N, date range) - Actors (user only, agent only, both) - Exclusions (system messages, tool outputs, code blocks) - Mark Parse Input completed, create Extract Signals in_progress

1. Extract Signals

- Scan messages for signal keywords - Match against taxonomy - Assign confidence (high/medium/low) - Record: type, subtype, message_id, timestamp, quote, context - Mark Extract Signals completed, create Detect Patterns in_progress

1. Detect Patterns

- Group signals by type/subtype - Find clusters (3+ related signals) - Identify evolution (signal changes over time) - Track repetition (recurring themes) - Spot correlations (tool chains, workflows) - Mark Detect Patterns completed, create Synthesize Report in_progress

1. Output

- Generate JSON with signals, patterns, summary - Include confidence, recommendations, action items - Append △ Caveats if gaps exist - Mark Synthesize Report completed

<pattern_detection>

Behavioral patterns from signal clusters:

Temporal patterns:

• Escalation: Increasing frustration/stronger requirements
• De-escalation: Frustration → success transition
• Cyclical: Same issue recurs across sessions

</pattern_detection>

<output_format>

JSON structure:

{
  "analysis": {
    "scope": {
      "message_count": N,
      "date_range": "YYYY-MM-DD to YYYY-MM-DD",
      "actors": ["user", "agent"]
    },
    "signals": [
      {
        "type": "success|frustration|workflow|request",
        "subtype": "specific_subtype",
        "message_id": "msg_123",
        "timestamp": "ISO8601",
        "quote": "exact text",
        "confidence": "high|medium|low",
        "context": "brief explanation"
      }
    ],
    "patterns": [
      {
        "pattern_type": "repetition|evolution|preference|tool_chain",
        "category": "success|frustration|workflow|request",
        "description": "pattern summary",
        "occurrences": N,
        "confidence": "strong|moderate|weak",
        "first_seen": "ISO8601",
        "last_seen": "ISO8601",
        "recommendation": "actionable next step"
      }
    ],
    "summary": {
      "total_signals": N,
      "by_type": { "success": N, "frustration": N, ... },
      "key_insights": ["insight 1", "insight 2"],
      "action_items": ["item 1", "item 2"]
    }
  }
}

</output_format>

ALWAYS:

• Create Parse Input at session start
• Update todos at stage transitions
• Include confidence levels for all signals
• Support patterns with 2+ signals minimum
• Mark Synthesize Report completed after delivery
• Apply recency weighting (recent overrides old)

NEVER:

• Skip stage transitions
• Extract low-confidence signals without marking them
• Claim patterns from single occurrences
• Regress stages
• Deliver without marking final stage complete
• Over-interpret neutral language

• signal-patterns.md — extended taxonomy, edge cases
• extraction-techniques.md — regex, heuristics
• sample-analysis.md — complete walkthrough