cta-usage-pattern

CTA Usage Pattern — Session Pattern Analysis

Analyze one or more sessions with the MCP classifier and convert the result into concrete harness guidance.

Workflow

Step 1: Freshness

Execute mcp__token-analyzer__sync_db when the user asks for "latest" or when the conversation likely depends on newly-created sessions. For a direct historical session_id lookup, sync_db is optional because classify_session_pattern reads JSONL directly.

Step 2: Pick Sessions

If the user provides a session_id, use it directly with mcp__token-analyzer__classify_session_pattern.

If the user does not provide a session_id:

1. Execute mcp__token-analyzer__analyze_global
2. Select up to 3 candidate sessions from top_sessions
3. Execute mcp__token-analyzer__classify_session_pattern for each selected session
4. Summarize the pattern mix, highest-severity result first

Step 3: Map Signals to Advice

Use the local skill reference file references/harness-signals-to-advice.md as the SSOT mapping.

Required output elements:

• detected pattern
• severity
• short signal summary (cache_hit_rate, subagent_count, repeated_edit_peak, turn_count, duration_minutes, topic_shift_count)
• 2-4 concrete workflow adjustments

Step 4: Optional Sparkline

If the user asks for trend context, execute mcp__token-analyzer__trend_report and render a short Unicode sparkline using the returned token totals. Keep it inline, for example: 14d token trend: ▁▂▃▅▄▆█

Reporting Template

## CTA 使用模式分析 — a1b2c3d4

- Pattern: `correction_spiral`
- Severity: `alert`
- Signals: cache_hit_rate 18.0%, repeated_edit_peak 8, output_token_ratio 61.0%, turn_count 42

### 建議
1. 把大檔案切成更小的編輯單元，避免同一檔案反覆來回修補。
2. 明確要求 diff-only 回覆，降低 output token 膨脹。
3. 如果需求已改變，先開新 session 或先 checkpoint，再繼續編輯。

Rules

1. Use 繁體中文 for prose; keep English for metric names and pattern IDs.
2. Quote exact numeric signals from MCP output; do not invent percentages or counts.
3. When severity is info, keep the tone observational instead of warning-heavy.
4. When classifying multiple sessions, order by severity first, then by cost if available.
5. If the MCP tool returns AMBIGUOUS_SESSION_ID, ask the user for a longer ID rather than guessing.