session-digest

name	session-digest
description	Summarize sessions across a multi-day range by paging session_summary in chunks, output bilingual, and write back stable soul traits.
description_zh	通过分段调用 session_summary 汇总多天会话，输出中英双语摘要，并在稳定特征出现时写回 soul。
version	1.0.0
author	Pulse Coder Team

Session Digest Skill

Summarize sessions over a long date range without blowing context by paging session_summary in windows and aggregating the results.

When to Use

• You need summaries for many days (e.g., 14/30/90 days)
• Single tool output is too large to fit in context
• You want a high-level digest with minimal raw excerpts

Inputs (JSON)

{
  "days": 30,
  "chunkDays": 7,
  "scope": "owner",
  "includeUserMessages": true,
  "includeAssistantMessages": true,
  "maxMessagesPerSession": 120,
  "focus": "optional: emphasize PRs, configs, incidents, or product decisions",
  "format": "themes"
}

• days (default 7): total days to summarize (UTC, counting back from today).
• chunkDays (default 7, max 30): size of each window.
• scope: owner (cross-channel) or platform (current channel).
• maxMessagesPerSession: reduce to avoid overlong excerpts.
• format: themes or timeline.

Required Execution Flow

1. Compute totalDays from input (default 7).
2. Compute chunkDays (default 7, clamp 1..30).
3. For each window index i = 0..:
   • offsetDays = i * chunkDays
   • windowDays = min(chunkDays, totalDays - offsetDays)
   • Call session_summary with:

     {
       "days": windowDays,
       "offsetDays": offsetDays,
       "scope": "owner",
       "includeUserMessages": true,
       "includeAssistantMessages": true,
       "maxMessagesPerSession": 120
     }
     

   • Summarize this window to 3-6 bullets, no raw excerpt dumps.
4. Aggregate across windows:
   • Deduplicate repeated items.
   • Highlight: PRs/MRs, code changes, config changes, incidents, decisions.
5. Produce bilingual output (EN + ZH) with the format below.
6. Soul writeback (no read):
   • Do NOT call memory_recall.
   • Infer 1-2 stable traits only when the same tendency appears in >= 2 sessions or windows.
   • Record with memory_record using kind: "soul" and a short, neutral sentence (no sensitive or user-facing text).
   • Do not mention the soul writeback in the visible output.

Output Format

• Start with the date range (UTC).
• Provide English then Chinese sections.
• Keep each language concise (target <= 1200 chars; avoid long excerpts).

Example:

Date Range (UTC): 2025-03-01 to 2025-03-07

EN
Overall:
- ...
- ...
Per-window:
- 2025-03-07: ...
- 2025-03-06: ...
Notable:
- ...

ZH
总体:
- ...
- ...
分段:
- 2025-03-07: ...
- 2025-03-06: ...
要点:
- ...

Output Rules

• Never paste the raw session_summary.sessions[].excerpt list.
• Prefer concise phrasing and avoid low-signal chat.
• If a window has no sessions, say so explicitly.
• Keep total output under ~300 lines.

Example Invocation

/skills session-digest {"days": 30, "chunkDays": 7, "scope": "owner"}