chat-compactor（聊天压缩器）

Chat Compactor

Generate structured summaries optimized for AI agent continuity across sessions.

Why This Exists

Human-written summaries and ad-hoc AI summaries lose critical context:

• Decision rationale gets lost (why X, not Y)
• Dead ends get forgotten (agent re-tries failed approaches)
• Implicit knowledge isn't captured (file locations, naming conventions, gotchas)
• State is unclear (what's done, what's pending, what's blocked)

This skill produces agent-optimized handoff documents that prime the next session.

Output Format

Generate a markdown file with this structure:

# Session: [Brief Title]
Date: [YYYY-MM-DD]
Duration: ~[X] messages

## Context Snapshot
[1-2 sentences: What project/task, what state it's in right now]

## What Was Accomplished
- [Concrete outcome 1]
- [Concrete outcome 2]

## Key Decisions & Rationale
| Decision | Why | Alternatives Rejected |
|----------|-----|----------------------|
| [Choice] | [Reason] | [What didn't work and why] |

## Current State
- **Working**: [files/features that are functional]
- **Broken/Blocked**: [what's not working and why]
- **Modified files**: [list with brief note on changes]

## Dead Ends (Don't Retry)
- ❌ [Approach that failed] — [why it failed]

## Next Steps (Prioritized)
1. [ ] [Most important next action]
2. [ ] [Second priority]

## Environment & Gotchas
- [Any setup notes, versions, quirks discovered]

## Key Code/Commands Reference
[Only if there are non-obvious commands or snippets the next session needs]

Workflow

1. Scan conversation for: decisions, outcomes, failures, file changes, blockers
2. Identify the "handoff moment" — what would a fresh agent need to continue?
3. Generate structured summary using format above
4. Save to file: session-[topic]-[date].md in project root or /home/claude/sessions/

Compaction Triggers

Invoke this skill when:

• User says: "compact", "wrap up", "save session", "summarize for next time"
• Context window exceeds ~50% capacity and task is ongoing
• Before major context switches
• End of debugging/implementation session

Quality Criteria

Good compactions are:

• Scannable: Next agent gets orientation in <30 seconds
• Actionable: Clear next steps, not vague summaries
• Defensive: Dead ends documented to prevent re-exploration
• Minimal: No fluff, every line earns its tokens

Anti-Patterns

Avoid:

• Narrative prose ("First we tried X, then Y, then Z...")
• Redundant context (don't repeat what's in code comments)
• Vague summaries ("Made good progress on the feature")
• Missing failure documentation (most valuable part!)

Example Trigger & Response

User: "Let's wrap up, compact this session"

Agent:

1. Reviews conversation for key decisions, outcomes, failures
2. Generates structured markdown per format above
3. Saves to session-[topic]-[date].md
4. Confirms: "Session compacted to session-auth-refactor-2025-01-06.md — ready for next time."