Context Monitor Skill

Real-time tracking of context window usage with intelligent preservation strategies.

Philosophy

"Context is precious. Spend it wisely, preserve what matters."

Claude Code operates within a context window. This skill helps manage that finite resource by tracking usage and ensuring critical content survives compaction.

Core Functions

•Track - Monitor context window utilization
•Reserve - Maintain headroom for essential operations
•Prioritize - Decide what to preserve vs. compact
•Alert - Warn when approaching limits

Context Zones

code

┌─────────────────────────────────────────────┐
│  0-70%: GREEN ZONE                          │
│  Normal operation, no constraints           │
├─────────────────────────────────────────────┤
│  70-85%: YELLOW ZONE                        │
│  Consider consolidation, avoid large reads  │
├─────────────────────────────────────────────┤
│  85-95%: ORANGE ZONE                        │
│  Active preservation, trigger scribe        │
├─────────────────────────────────────────────┤
│  95-100%: RED ZONE                          │
│  Compaction imminent, critical only         │
└─────────────────────────────────────────────┘

Reserved Headroom

Always maintain headroom for:

Purpose	Reserved	Priority
Scribe extraction	15K tokens	Critical
Session handoff	10K tokens	Critical
Emergency response	5K tokens	Critical
Agent communication	5K tokens	High
Tool responses	10K tokens	High

Total Reserved: ~45K tokens

Zone-Based Behavior

Green Zone (0-70%)

Normal operation:

•Read files freely
•Full agent spawning
•No consolidation pressure

Yellow Zone (70-85%)

Caution mode:

•Prefer targeted reads over full files
•Consider consolidating recent work
•Start identifying compaction candidates

yaml

yellow_zone_actions:
  - "Summarize recent tool outputs"
  - "Identify low-value context"
  - "Prepare preservation priorities"

Orange Zone (85-95%)

Active preservation:

•Trigger auto-scribe for uncaptured value
•Compress verbose outputs
•Avoid non-essential reads

yaml

orange_zone_actions:
  - "Run silent extraction"
  - "Capture key decisions"
  - "Prepare session handoff"
  - "Alert user of context pressure"

Red Zone (95-100%)

Compaction imminent:

•Complete active captures immediately
•Generate session continuity summary
•Preserve only critical state

yaml

red_zone_actions:
  - "Force scribe completion"
  - "Generate handoff document"
  - "Preserve working state"
  - "Prepare for restart"

Preservation Priorities

When compaction is needed, preserve in order:

Priority 1: Critical (Always Preserve)

•Current task and requirements
•Uncaptured decisions
•Active conflict resolutions
•Working file states

Priority 2: High (Preserve if Space)

•Recent tool outputs
•Key conversation turns
•Pattern discoveries
•Error resolutions

Priority 3: Medium (Summarize)

•File contents (keep paths, summarize content)
•Verbose command outputs
•Research results

Priority 4: Low (Discard)

•Duplicate information
•Superseded states
•Debugging artifacts
•Exploratory dead ends

Integration Points

With Auto-Scribe

When context hits 85%:

yaml

trigger_scribe:
  action: silent_extraction
  priority: immediate
  reserve_tokens: 15000
  message: "Context pressure - capturing valuable content"

With Session Continuity

When context hits 90%:

yaml

prepare_handoff:
  action: generate_summary
  include:
    - current_task
    - decisions_made
    - files_modified
    - next_steps
  output: "session-handoff.md"

With Long-Running Tasks

During autonomous work:

yaml

checkpoint_trigger:
  at_context_level: 80
  action: save_checkpoint
  include:
    - progress_summary
    - remaining_work
    - known_issues

User Notifications

Context Status Updates

When requested or at key thresholds:

code

## Context Status

Current: 72% (Yellow Zone)
Headroom: 28% (~56K tokens)

Reserved for:
- Scribe: 15K ✓
- Handoff: 10K ✓
- Emergency: 5K ✓

Recommendation: Consider consolidating recent work

Compaction Warnings

At 85%:

code

⚠️ Context at 85%

Triggering knowledge capture...
- 3 decisions captured
- 2 patterns logged
- Session handoff prepared

Safe to continue, but consider:
- Targeted file reads
- Completing current task
- Starting fresh session for new work

Anti-Patterns

Don't Do This

yaml

anti_patterns:
  - "Reading entire codebases at once"
  - "Keeping verbose error logs in context"
  - "Repeating full file contents on each edit"
  - "Ignoring context warnings"

Do This Instead

yaml

best_practices:
  - "Read specific functions, not entire files"
  - "Summarize error patterns, don't log verbatim"
  - "Reference files by path, read only changed sections"
  - "Heed context warnings, preserve proactively"

Estimation Heuristics

Rough token estimates:

Content Type	Tokens/Line	Example
Code	~10	100 lines ≈ 1K tokens
Prose	~20	50 lines ≈ 1K tokens
YAML/JSON	~8	125 lines ≈ 1K tokens
Tool output	~15	Varies widely

Configuration

yaml

context_monitor:
  thresholds:
    yellow: 70
    orange: 85
    red: 95

  reservations:
    scribe: 15000
    handoff: 10000
    emergency: 5000
    agent_comms: 5000
    tool_responses: 10000

  auto_actions:
    at_orange: trigger_scribe
    at_red: force_handoff

  notifications:
    show_status: on_request
    warn_at: yellow
    alert_at: orange

Benefits

•Prevents data loss - Critical content preserved before compaction
•Enables long sessions - Intelligent management extends useful context
•Supports handoffs - Sessions can continue across compactions
•Reduces waste - Focus context on high-value content
•Improves reliability - No surprise context exhaustion