tao12345666333/context-management

Context Management Skill

Overview

This skill teaches you how to proactively manage your conversation context in AMCP to avoid LLM API errors caused by exceeding context window limits. Context management is critical when:

• Working on long coding sessions with many files
• Processing large tool outputs (e.g., grep results, file reads)
• Running multi-step debugging sessions
• Reviewing or refactoring large codebases

Understanding Context Windows

Different LLM models have different context window sizes:

| Model Family | Context Window | |--------------|----------------| | GPT-4 Turbo / GPT-4o | 128,000 tokens | | GPT-4.1 | 1,000,000 tokens | | Claude 3.5 Sonnet | 200,000 tokens | | DeepSeek V3 | 64,000 tokens | | Gemini 2.0 Flash | 1,000,000 tokens | | Qwen 2.5 | 128,000 tokens |

AMCP automatically detects most model context windows. For unknown models, it uses 32,000 tokens as the default.

Key Metrics

When managing context, track:

• Current tokens: Estimated size of current conversation history
• Threshold tokens: When compaction should trigger (default: 70% of context)
• Target tokens: What to aim for after compaction (default: 30% of context)
• Safety margin: Reserved for response generation (default: 10%)

How AMCP Compacts Context

AMCP's SmartCompactor automatically compresses context when it exceeds the threshold:

# In agent.py (line 500-505):
compactor = SmartCompactor(client, model)
if compactor.should_compact(history_to_add):
    history_to_add, _ = compactor.compact(history_to_add)

This happens automatically during conversation! You don't need to trigger it manually.

Compaction Strategies

AMCP supports four strategies (configurable via CompactionConfig):

1. SUMMARY (default): Uses LLM to create intelligent summary of old messages

   • Best for: Long sessions where earlier context is important
   • Preserves: Errors, working solutions, current task state, file paths

2. TRUNCATE: Simple removal of old messages, keeping first and last few

   • Best for: Fast compaction when context is less important
   • Fastest option

3. SLIDING_WINDOW: Keeps only the most recent messages that fit in target

   • Best for: Sessions where only recent context matters
   • Very efficient

4. HYBRID: Combines sliding window with summary of removed content

   • Best for: Balance between summary quality and speed

Best Practices for Context Management

1. Read Files Selectively

Instead of reading entire large files:

# BAD: Reads entire file
read_file(path="src/large_module.py")

# GOOD: Read specific sections
read_file(path="src/large_module.py", mode="indentation", offset=100, limit=50)

# GOOD: Read specific line ranges
read_file(path="src/large_module.py", mode="slice", ranges=["1-50", "200-250"])

Use read_file in indentation mode - it intelligently captures code blocks around your target, providing context without excessive content.

2. Use Grep with Context Limits

# BAD: Returns all matches with full context
grep(pattern="function", paths=["src/"])

# GOOD: Limited context
grep(pattern="function", paths=["src/"], context=2)

3. Iterate Over Small Batches

When processing multiple files:

# Instead of processing all files at once:
for file in files:
    # Process one file at a time
    result = process_file(file)
    # Save intermediate results

4. Clear Conversation When Starting New Tasks

After completing a complex task, consider suggesting:

"会话历史较长。如果开始新的无关任务，建议清除历史或创建新会话以减少上下文。"

5. Be Strategic with Tool Calls

• Avoid redundant tool calls (e.g., reading the same file multiple times)
• Use grep to find what you need before reading files
• Process results incrementally rather than all at once

Monitoring Context Usage

You can check context usage programmatically:

from amcp import SmartCompactor

# Create compactor
compactor = SmartCompactor(client, model="gpt-4-turbo")

# Get detailed usage info
usage = compactor.get_token_usage(messages)
print(f"Current: {usage['current_tokens']:,} tokens")
print(f"Usage: {usage['usage_ratio']:.1%} of context")
print(f"Headroom: {usage['headroom_tokens']:,} tokens")
print(f"Should compact: {usage['should_compact']}")

Configuration

Context compaction is configured via CompactionConfig:

from amcp import SmartCompactor, CompactionConfig, CompactionStrategy

config = CompactionConfig(
    strategy=CompactionStrategy.SUMMARY,
    threshold_ratio=0.7,  # Compact at 70% usage
    target_ratio=0.3,     # Aim for 30% after compaction
    preserve_last=6,      # Keep last 6 user/assistant messages
    preserve_tool_results=True,  # Preserve recent tool results
    max_tool_results=10,  # Max tool results to preserve
    min_tokens_to_compact=5000,  # Don't compact tiny contexts
    safety_margin=0.1,    # 10% margin for responses
)

You can configure compaction in ~/.config/amcp/config.toml:

[chat]
model = "deepseek-chat"

[compaction]
strategy = "summary"  # summary, truncate, sliding_window, hybrid
threshold_ratio = 0.7
target_ratio = 0.3

Automatic Summary Structure

When using SUMMARY or HYBRID strategies, the summary follows this structure:

<current_task>
What we're working on now - be specific about files and goals
</current_task>

<completed>
- Task 1: Brief outcome + key changes made
- Task 2: Brief outcome + key changes made
</completed>

<code_state>
Key files and their current state - signatures + key logic only
Include file paths that were modified
</code_state>

<important>
Any crucial context: errors, decisions made, constraints, blockers
</important>

Token Estimation

AMCP provides accurate token estimation:

from amcp import estimate_tokens

tokens = estimate_tokens(messages)

• Uses tiktoken library when available (recommended)
• Falls back to character-based estimation (4 chars ≈ 1 token)
• Accounts for message role overhead

Common Issues and Solutions

Issue 1: "Context length exceeded" API Error

Symptom: LLM API returns error about context window being exceeded.

Causes:

• Large tool outputs (e.g., grep or find results)
• Reading many large files
• Long multi-step task conversations

Solutions:

1. Let AMCP's auto-compaction handle it (it will compact automatically)
2. Reduce tool output size (use grep --limit, read_file with ranges)
3. Process files in batches, not all at once
4. Suggest clearing conversation history if starting new task

Issue 2: Context Losing Important Information

Symptom: After compaction, agent forgets critical details.

Causes:

• SUMMARY strategy missed important context
• Old critical messages were removed

Solutions:

1. Use preserve_last to keep more recent messages
2. Consider HYBRID strategy (keeps sliding window + summary)
3. Manually restate critical context in your prompt

Issue 3: Summary Losing Code Details

Symptom: Agent forgets specific code changes after compaction.

Solutions:

1. Use preserve_tool_results=True (default) to keep recent tool outputs
2. Increase max_tool_results (default: 10)
3. Review summary and add missing details

Progressive Disclosure

AMCP uses progressive disclosure to manage skill instructions:

# In skills.py (line 297):
skills_summary = skill_manager.build_skills_summary()

This means:

• You only get a compact summary of all skills initially
• Full skill content is available when needed
• Reduces initial context overhead

Memory System vs Context

AMCP separates:

1. Conversation Context: The messages sent to the LLM (limited by context window)
2. Persistent Memory: Long-term storage in .amcp/memory/ (unlimited)

Use the memory system to store important information that should persist long-term:

memory(action="write", content="# Project Notes\n- Uses PostgreSQL database")

This information is NOT in the conversation context - it's stored separately and retrieved when relevant.

Practical Workflow

For complex tasks with large context:

1. Start: Use grep or find to locate relevant files

   grep(pattern="class User", paths=["src/"])
   

2. Explore: Read specific sections using indentation mode

   read_file(path="src/models/user.py", mode="indentation", offset=1)
   

3. Process: Work incrementally, one file at a time

4. Monitor: If context gets large, compaction happens automatically

5. Persist: Save important findings to memory if needed for future sessions

   memory(action="append", content="Found authentication bug in src/auth.py:45")
   

Event Monitoring

AMCP emits events when compaction occurs:

from amcp import get_event_bus, EventType

@get_event_bus().on(EventType.CONTEXT_COMPACTED)
async def on_compaction(event):
    data = event.data
    print(f"Context compacted: {data['original_tokens']} -> {data['compacted_tokens']}")

Key Takeaways

1. Auto-compaction happens automatically - you don't need to trigger it
2. Read files selectively - use indentation mode or ranges
3. Monitor context usage - be aware of large tool outputs
4. Persist important info to memory - for long-term retention
5. Process incrementally - handle large tasks in steps
6. Configure as needed - adjust CompactionConfig for your use case

When to Suggest Clearing Context

If you notice:

• Conversation is very long (>50 messages)
• You're starting a completely different task
• Context from earlier sessions is no longer relevant

Then suggest:

"会话历史较长。如果开始新的无关任务，建议清除历史或创建新会话以减少上下文。"

This is proactive context management!