Agent-Skills.md

Agent Skills: Context Compactor

Automatic context summarization for long-running sessions. Use when context is approaching limits, summarizing completed work, preserving critical information, or managing token budgets.

UncategorizedID: adaptationio/skrillz/context-compactor

Repository

adaptationio/skrillz
adaptationio
61

Install this agent skill to your local

pnpm dlx add-skill https://github.com/adaptationio/Skrillz/tree/HEAD/.claude/skills/context-compactor

Skill Files

Browse the full folder contents for context-compactor.

Download Skill

Loading file tree…

.claude/skills/context-compactor/SKILL.md

Skill Metadata

Name
context-compactor
Description
Automatic context summarization for long-running sessions. Use when context is approaching limits, summarizing completed work, preserving critical information, or managing token budgets.

Context Compactor

Automatically summarizes and compacts context when approaching token limits while preserving critical information.

Quick Start

Check if Compaction Needed

from scripts.compactor import ContextCompactor

compactor = ContextCompactor()
if compactor.should_compact(context, limit=100000):
    compacted = compactor.compact(context)

Compact Context

compacted = compactor.compact(context)
# Preserves: architectural decisions, bugs, current state
# Discards: stale tool outputs, redundant messages

Preservation Strategy

┌─────────────────────────────────────────────────────────────┐
│                  PRESERVATION PRIORITY                       │
├─────────────────────────────────────────────────────────────┤
│                                                              │
│  MUST PRESERVE (Never remove)                               │
│  ├─ Architectural decisions                                 │
│  ├─ Unresolved bugs and blockers                           │
│  ├─ Current feature state                                   │
│  ├─ Recent file changes (last 5)                           │
│  └─ Error patterns and solutions                           │
│                                                              │
│  CAN SUMMARIZE (Compress to summary)                        │
│  ├─ Completed features (list of IDs)                       │
│  ├─ Resolved errors (brief mention)                        │
│  └─ Historical decisions (key points)                      │
│                                                              │
│  CAN DISCARD (Remove entirely)                              │
│  ├─ Redundant tool outputs                                 │
│  ├─ Stale search results                                   │
│  ├─ Superseded messages                                    │
│  └─ Verbose logging                                         │
│                                                              │
└─────────────────────────────────────────────────────────────┘

Compaction Process

  1. Score Importance: Assign score to each message/content
  2. Identify Critical: Mark must-preserve content
  3. Summarize Middle: Compress can-summarize content
  4. Discard Stale: Remove can-discard content
  5. Validate: Ensure compacted context is coherent

Token Budgets

| Context Type | Limit | Action | |--------------|-------|--------| | < 50% | Normal | No action | | 50-75% | Warning | Prepare for compaction | | 75-90% | Compact | Trigger compaction | | > 90% | Critical | Aggressive compaction |

Integration Points

  • autonomous-session-manager: Triggers compaction check
  • memory-manager: Stores compacted summaries
  • handoff-coordinator: Uses compacted state for handoffs

References

  • references/COMPACTION-STRATEGY.md - Detailed strategy
  • references/TOKEN-BUDGETS.md - Budget management

Scripts

  • scripts/compactor.py - Core ContextCompactor
  • scripts/importance_scorer.py - Message scoring
  • scripts/summarizer.py - Content summarization
  • scripts/token_estimator.py - Token counting