Analyze the current session context and output a structured summary in a fixed format.
Workflow
- Scan session history — Review the complete conversation, identifying all user inputs and assistant actions one by one
- Extract key information — Organize core topics, decisions, file changes, executed commands, and important notes
- Read output template — Refer to the fixed format in
references/summary-template.mdFixed format in - Generate short title — Generate a kebab-case short title (2-5 words) based on the session's core content, in the same language as the session's primary language
- Fill in summary and write to file — Generate the complete summary according to the template, use the Write tool to save it to the current working directory, file name format:
session-summary-YYYY-MM-DD-短标题.md
Fill-in instructions
Basic information
- Start time / End time: Record the session's start and end times, format YYYY-MM-DD HH:MM
- Interaction turns: Count the number of user messages and assistant replies in the conversation
- Token consumption: Estimate Input / Output / Cache Read / Cache Write / Total by dimension; when the quantities are large, use k (thousand) / m (million) abbreviations, e.g. ~120k, ~1.2m
- Primary language: Identify the primary language used in the session
Session overview
- Core topics: Sort by importance, one sentence each
- Decisions and conclusions: List only explicitly reached decisions; do not include pending items
- File changes: List only file operations that actually occurred in this session
- Code and command summary: List key Bash commands and code modifications
- Important notes: Record key context that affects the next session
Session details
- Output turn by turn, including every user input and the Agent's actions
- User input: Preserve verbatim in a markdown code block, without any modification or abbreviation
- Assistant reply: Compress into a summary of actions and replies, placed in a markdown code block