context-management（上下文管理）

Context Management Skill

Purpose

This skill enables proactive management of Claude Code's context window through intelligent token monitoring, context extraction, and selective rehydration. Instead of reactive recovery after compaction, this skill helps users preserve essential context before hitting limits and restore it efficiently when needed.

Version 3.0 Enhancements:

• Predictive Budget Monitoring: Estimate when capacity thresholds will be reached
• Context Health Indicators: Visual indicators for statusline integration
• Priority-Based Retention: Keep requirements and decisions, archive verbose logs
• Burn Rate Tracking: Monitor token consumption velocity for early warnings

When to Use This Skill

• Token monitoring: Check current usage and get recommendations
• Approaching limits: Create snapshots at 70-85% usage
• After compaction: Restore essential context without full conversation
• Long sessions: Preserve key decisions and state proactively
• Complex tasks: Keep requirements and progress accessible
• Context switching: Save state when pausing work
• Team handoffs: Package context for others to continue
• Predictive planning: Get early warnings before capacity is reached
• Session health: Monitor context health for sustained productivity

Quick Start

Check Token Status

User: Check my current token usage

I'll use the context_manager tool to check status:

from context_manager import check_context_status

status = check_context_status(current_tokens=<current_count>)
# Returns: ContextStatus with usage percentage and recommendations

Create a Snapshot

User: Create a context snapshot named "auth-implementation"

I'll use the context_manager tool to create a snapshot:

from context_manager import create_context_snapshot

snapshot = create_context_snapshot(
    conversation_data=<conversation_history>,
    name="auth-implementation"
)
# Returns: ContextSnapshot with snapshot_id, file_path, and token_count

Restore Context

User: Restore context from snapshot <snapshot_id> at essential level

I'll use the context_manager tool to rehydrate:

from context_manager import rehydrate_from_snapshot

context = rehydrate_from_snapshot(
    snapshot_id="20251116_143522",
    level="essential"  # or "standard" or "comprehensive"
)
# Returns: Formatted context text ready to process

List Snapshots

User: List my context snapshots

I'll use the context_manager tool to list snapshots:

from context_manager import list_context_snapshots

snapshots = list_context_snapshots()
# Returns: List of snapshot metadata dicts

Detail Levels

When rehydrating context, choose the appropriate detail level:

• Essential (smallest): Requirements + current state only (~250 tokens)
• Standard (balanced): + key decisions + open items (~800 tokens)
• Comprehensive (complete): + full decisions + tools used + metadata (~1,250 tokens)

Start with essential and upgrade if more context is needed.

Actions

Action: status

Check current token usage and get recommendations.

Usage:

from context_manager import check_context_status

status = check_context_status(current_tokens=750000)
print(f"Usage: {status.percentage}%")
print(f"Status: {status.threshold_status}")
print(f"Recommendation: {status.recommendation}")

Returns:

• ContextStatus object with usage details
• threshold_status: 'ok', 'consider', 'recommended', or 'urgent'
• recommendation: Human-readable action suggestion

Action: snapshot

Create intelligent context snapshot.

Usage:

from context_manager import create_context_snapshot

snapshot = create_context_snapshot(
    conversation_data=messages,
    name="feature-name"  # Optional
)
print(f"Snapshot ID: {snapshot.snapshot_id}")
print(f"Token count: {snapshot.token_count}")
print(f"Saved to: {snapshot.file_path}")

Returns:

• ContextSnapshot object with metadata
• Snapshot saved to ~/.amplihack/.claude/runtime/context-snapshots/

Action: rehydrate

Restore context from snapshot at specified detail level.

Usage:

from context_manager import rehydrate_from_snapshot

context = rehydrate_from_snapshot(
    snapshot_id="20251116_143522",
    level="standard"  # essential, standard, or comprehensive
)
print(context)  # Display restored context

Returns:

• Formatted markdown text with restored context
• Ready to process and continue work

Action: list

List all available context snapshots.

Usage:

from context_manager import list_context_snapshots

snapshots = list_context_snapshots()
for snapshot in snapshots:
    print(f"{snapshot['id']}: {snapshot['name']} ({snapshot['size']})")

Returns:

• List of snapshot metadata dicts
• Includes: id, name, timestamp, size, token_count

Proactive Features (v3.0)

Predictive Budget Monitoring

Instead of just checking current usage, predict when thresholds will be reached:

# The system tracks token burn rate over time
# When checking status, you get predictive insights

status = check_context_status(current_tokens=500000)

# Status includes predictions (when automation is running):
# - Estimated tool uses until 70% threshold
# - Time estimate based on current burn rate
# - Early warning before you hit capacity

# Example output interpretation:
# "At current rate, you'll hit 70% in ~15 tool uses"
# "Consider creating a checkpoint before your next major operation"

How Prediction Works:

The automation tracks:

1. Token count at each check interval
2. Number of tool uses between checks
3. Average tokens consumed per tool use
4. Time elapsed between checks

From this data, it estimates:

• Tools remaining until threshold
• Approximate time until threshold
• Whether current task will complete before limit

Context Health Indicators

Visual indicators for session health, suitable for statusline integration:

Statusline Integration Example:

# In your statusline script, check context health:
# The automation state file contains health status

# Example statusline addition:
if [ -f ".claude/runtime/context-automation-state.json" ]; then
    LAST_PCT=$(jq -r '.last_percentage // 0' .claude/runtime/context-automation-state.json)
    if [ "$LAST_PCT" -lt 30 ]; then
        echo "[CTX:OK]"
    elif [ "$LAST_PCT" -lt 50 ]; then
        echo "[CTX:WATCH]"
    elif [ "$LAST_PCT" -lt 70 ]; then
        echo "[CTX:WARN]"
    else
        echo "[CTX:CRITICAL]"
    fi
fi

Priority-Based Context Retention

When creating snapshots, the system prioritizes content by importance:

High Priority (Always Retained):

• Original user requirements (first user message)
• Key architectural decisions
• Current implementation state
• Open items and blockers

Medium Priority (Retained in Standard+):

• Tool usage history
• Decision rationales
• Questions and clarifications

Low Priority (Only in Comprehensive):

• Verbose output logs
• Intermediate steps
• Debugging information

Usage Pattern:

# Create snapshot with priority awareness
snapshot = create_context_snapshot(
    conversation_data=messages,
    name='feature-checkpoint'
)

# Essential level (~200 tokens): Only high priority content
# Standard level (~800 tokens): High + medium priority
# Comprehensive level (~1250 tokens): Everything

# Start minimal, upgrade as needed:
context = rehydrate_from_snapshot(snapshot_id, level='essential')

Burn Rate Tracking

Monitor how fast you're consuming context:

# The automation tracks consumption velocity
# Adaptive checking frequency based on burn rate:

# Low burn rate (< 1K tokens/tool): Check every 50 tools
# Medium burn rate (1-5K tokens/tool): Check every 10 tools
# High burn rate (> 5K tokens/tool): Check every 3 tools
# Critical zone (70%+): Check every tool

# This means:
# - Normal development: Minimal overhead (checks rarely)
# - Large file operations: Increased monitoring
# - Approaching limits: Continuous monitoring

Burn Rate Thresholds:

Auto-Summarization Triggers

The system automatically creates snapshots before limits are hit:

# Automatic snapshot triggers (already implemented):
# - 30% usage: First checkpoint created
# - 40% usage: Second checkpoint created
# - 50% usage: Third checkpoint created (for 1M models)

# For smaller context windows (< 800K):
# - 55% usage: First checkpoint
# - 70% usage: Second checkpoint
# - 85% usage: Urgent checkpoint

# After compaction detected (30% token drop):
# - Automatically rehydrates from most recent snapshot
# - Uses smart level selection based on previous usage

Proactive Usage Workflow

Step 1: Monitor Token Usage

Periodically check status during long sessions:

status = check_context_status(current_tokens=current)

if status.threshold_status == 'consider':
    # Usage at 70%+ - consider creating snapshot
    print("Consider creating a snapshot soon")
elif status.threshold_status == 'recommended':
    # Usage at 85%+ - snapshot recommended
    create_context_snapshot(messages, name='current-work')
elif status.threshold_status == 'urgent':
    # Usage at 95%+ - create snapshot immediately
    create_context_snapshot(messages, name='urgent-backup')

Step 2: Create Snapshot at Threshold

When 70-85% threshold reached, create a named snapshot:

snapshot = create_context_snapshot(
    conversation_data=messages,
    name='descriptive-name'
)
# Save snapshot ID for later rehydration

Step 3: Continue Working

After snapshot creation:

• Continue conversation naturally
• Let Claude Code compact if needed
• Use /transcripts for full history if desired
• PreCompact hook saves everything automatically

Step 4: Rehydrate After Compaction

After compaction, restore essential context:

# Start minimal
context = rehydrate_from_snapshot(
    snapshot_id='20251116_143522',
    level='essential'
)

# If more context needed, upgrade to standard
context = rehydrate_from_snapshot(
    snapshot_id='20251116_143522',
    level='standard'
)

# For complete context, use comprehensive
context = rehydrate_from_snapshot(
    snapshot_id='20251116_143522',
    level='comprehensive'
)

Integration with Existing Systems

vs. PreCompact Hook

PreCompact Hook (automatic safety net):

• Triggered by Claude Code before compaction
• Saves complete conversation transcript
• Automatic, no user action needed
• Full conversation export to markdown

Context Skill (proactive optimization):

• Triggered by user when monitoring indicates
• Saves intelligent context extraction
• User-initiated, deliberate choice
• Essential context only, not full dump

Relationship: Complementary, not competing. Hook = safety net, Skill = optimization.

vs. /transcripts Command

/transcripts (reactive restoration):

• Restores full conversation after compaction
• Complete history, all messages
• Used when you need everything back
• Reactive recovery tool

Context Skill (proactive preservation):

• Preserves essential context before compaction
• Selective rehydration, not full history
• Used when you want efficient context
• Proactive optimization tool

Relationship: Transcripts for full recovery, skill for efficient management.

Storage Locations

• Snapshots: ~/.amplihack/.claude/runtime/context-snapshots/ (JSON)
• Transcripts: ~/.amplihack/.claude/runtime/logs/<session_id>/CONVERSATION_TRANSCRIPT.md
• No conflicts: Different directories, different purposes

Automatic Management

Context management runs automatically via the post_tool_use hook:

• Monitors token usage every Nth tool use (adaptive frequency)
• Creates snapshots at thresholds (30%, 40%, 50% for 1M models)
• Detects compaction (token drop > 30%)
• Auto-rehydrates after compaction at appropriate level

This happens transparently without user intervention.

Implementation

All context management functionality is provided by:

• Tool: ~/.amplihack/.claude/tools/amplihack/context_manager.py
• Hook Integration: ~/.amplihack/.claude/tools/amplihack/context_automation_hook.py
• Hook System: ~/.amplihack/.claude/tools/amplihack/hooks/tool_registry.py

See tool documentation for complete API reference and implementation details.

Common Patterns

Pattern 1: Preventive Snapshotting

Check before long operation and create snapshot if needed:

status = check_context_status(current_tokens=current)
if status.threshold_status in ['recommended', 'urgent']:
    create_context_snapshot(messages, name='before-refactoring')

Pattern 2: Context Switching

Save state when pausing work on one feature to start another:

# Pausing work on Feature A
create_context_snapshot(messages, name='feature-a-paused')

# [... work on Feature B ...]

# Resume Feature A later
context = rehydrate_from_snapshot('feature-a-snapshot-id', level='standard')

Pattern 3: Team Handoff

Create comprehensive snapshot for teammate:

snapshot = create_context_snapshot(
    messages,
    name='handoff-to-alice-api-work'
)
# Share snapshot ID with teammate
# Alice can rehydrate and continue work

Philosophy Alignment

Ruthless Simplicity

• Four single-purpose components in one tool
• On-demand invocation, no background processes
• Standard library only, no external dependencies
• Clear public API with convenience functions

Single Responsibility

• ContextManager coordinates all operations
• Token monitoring, extraction, rehydration in one place
• No duplicate code or scattered logic

Zero-BS Implementation

• No stubs or placeholders
• All functions work completely
• Real token estimation, not fake
• Actual file operations, not simulated

Trust in Emergence

• User decides when to snapshot, not automatic (unless via hook)
• User chooses detail level, not system
• Proactive choice empowers the user

Tips for Effective Context Management

1. Monitor regularly: Check status at natural breakpoints
2. Snapshot strategically: At 70-85% or before long operations
3. Start minimal: Use essential level first, upgrade if needed
4. Name descriptively: Use clear snapshot names for later reference
5. List periodically: Review and clean old snapshots
6. Combine tools: Use with /transcripts for full recovery option
7. Trust emergence: Don't over-snapshot, let context flow naturally

Resources

• Tool: ~/.amplihack/.claude/tools/amplihack/context_manager.py
• Hook: ~/.amplihack/.claude/tools/amplihack/context_automation_hook.py
• Philosophy: ~/.amplihack/.claude/context/PHILOSOPHY.md
• Patterns: ~/.amplihack/.claude/context/PATTERNS.md

Remember

This skill provides proactive context management through a clean, reusable tool. The tool can be called from skills, commands, and hooks. It complements existing tools (PreCompact hook, /transcripts) rather than replacing them. Use it to maintain clean, efficient context throughout long sessions.

Key Takeaway: Business logic lives in context_manager.py, this skill just tells you how to use it.