This commit is contained in:
2026-04-12 01:06:31 +07:00
commit 10d660cbcb
1066 changed files with 228596 additions and 0 deletions

View File

@@ -0,0 +1,110 @@
---
name: ck:context-engineering
description: >-
Check context usage limits, monitor time remaining, optimize token consumption, debug context failures.
Use when asking about context percentage, rate limits, usage warnings, context optimization, agent architectures, memory systems.
argument-hint: "[topic or question]"
metadata:
author: claudekit
version: "1.0.0"
---
# Context Engineering
Context engineering curates the smallest high-signal token set for LLM tasks. The goal: maximize reasoning quality while minimizing token usage.
## When to Activate
- Designing/debugging agent systems
- Context limits constrain performance
- Optimizing cost/latency
- Building multi-agent coordination
- Implementing memory systems
- Evaluating agent performance
- Developing LLM-powered pipelines
## Core Principles
1. **Context quality > quantity** - High-signal tokens beat exhaustive content
2. **Attention is finite** - U-shaped curve favors beginning/end positions
3. **Progressive disclosure** - Load information just-in-time
4. **Isolation prevents degradation** - Partition work across sub-agents
5. **Measure before optimizing** - Know your baseline
**IMPORTANT:**
- Sacrifice grammar for the sake of concision.
- Ensure token efficiency while maintaining high quality.
- Pass these rules to subagents.
## Quick Reference
| Topic | When to Use | Reference |
|-------|-------------|-----------|
| **Fundamentals** | Understanding context anatomy, attention mechanics | [context-fundamentals.md](./references/context-fundamentals.md) |
| **Degradation** | Debugging failures, lost-in-middle, poisoning | [context-degradation.md](./references/context-degradation.md) |
| **Optimization** | Compaction, masking, caching, partitioning | [context-optimization.md](./references/context-optimization.md) |
| **Compression** | Long sessions, summarization strategies | [context-compression.md](./references/context-compression.md) |
| **Memory** | Cross-session persistence, knowledge graphs | [memory-systems.md](./references/memory-systems.md) |
| **Multi-Agent** | Coordination patterns, context isolation | [multi-agent-patterns.md](./references/multi-agent-patterns.md) |
| **Evaluation** | Testing agents, LLM-as-Judge, metrics | [evaluation.md](./references/evaluation.md) |
| **Tool Design** | Tool consolidation, description engineering | [tool-design.md](./references/tool-design.md) |
| **Pipelines** | Project development, batch processing | [project-development.md](./references/project-development.md) |
| **Runtime Awareness** | Usage limits, context window monitoring | [runtime-awareness.md](./references/runtime-awareness.md) |
## Key Metrics
- **Token utilization**: Warning at 70%, trigger optimization at 80%
- **Token variance**: Explains 80% of agent performance variance
- **Multi-agent cost**: ~15x single agent baseline
- **Compaction target**: 50-70% reduction, <5% quality loss
- **Cache hit target**: 70%+ for stable workloads
## Four-Bucket Strategy
1. **Write**: Save context externally (scratchpads, files)
2. **Select**: Pull only relevant context (retrieval, filtering)
3. **Compress**: Reduce tokens while preserving info (summarization)
4. **Isolate**: Split across sub-agents (partitioning)
## Anti-Patterns
- Exhaustive context over curated context
- Critical info in middle positions
- No compaction triggers before limits
- Single agent for parallelizable tasks
- Tools without clear descriptions
## Guidelines
1. Place critical info at beginning/end of context
2. Implement compaction at 70-80% utilization
3. Use sub-agents for context isolation, not role-play
4. Design tools with 4-question framework (what, when, inputs, returns)
5. Optimize for tokens-per-task, not tokens-per-request
6. Validate with probe-based evaluation
7. Monitor KV-cache hit rates in production
8. Start minimal, add complexity only when proven necessary
## Runtime Awareness
The system automatically injects usage awareness via PostToolUse hook:
```xml
<usage-awareness>
Claude Usage Limits: 5h=45%, 7d=32%
Context Window Usage: 67%
</usage-awareness>
```
**Thresholds:**
- 70%: WARNING - consider optimization/compaction
- 90%: CRITICAL - immediate action needed
**Data Sources:**
- Usage limits: Anthropic OAuth API (`https://api.anthropic.com/api/oauth/usage`)
- Context window: Statusline temp file (`/tmp/ck-context-{session_id}.json`)
## Scripts
- [context_analyzer.py](./scripts/context_analyzer.py) - Context health analysis, degradation detection
- [compression_evaluator.py](./scripts/compression_evaluator.py) - Compression quality evaluation

View File

@@ -0,0 +1,84 @@
# Context Compression
Strategies for long-running sessions exceeding context windows.
## Core Insight
Optimize **tokens-per-task** (total to completion), not tokens-per-request.
Aggressive compression causing re-fetching costs more than better retention.
## Compression Methods
| Method | Compression | Quality | Best For |
|--------|-------------|---------|----------|
| **Anchored Iterative** | 98.6% | 3.70/5 | Best balance |
| **Regenerative Full** | 98.7% | 3.44/5 | Readability |
| **Opaque** | 99.3% | 3.35/5 | Max compression |
## Anchored Iterative Summary Template
```markdown
## Session Intent
Original goal: [preserved]
## Files Modified
- file.py: Changes made
## Decisions Made
- Key decisions with rationale
## Current State
Progress summary
## Next Steps
1. Next action items
```
**On compression**: Merge new content into existing sections, don't regenerate.
## Compression Triggers
| Strategy | Trigger | Use Case |
|----------|---------|----------|
| Fixed threshold | 70-80% utilization | General purpose |
| Sliding window | Keep last N turns + summary | Conversations |
| Task-boundary | At logical completion | Multi-step workflows |
## Artifact Trail Problem
Weakest dimension (2.2-2.5/5.0). Coding agents need explicit tracking of:
- Files created/modified/read
- Function/variable names, error messages
**Solution**: Dedicated artifact section in summary.
## Probe-Based Evaluation
| Probe Type | Tests | Example |
|------------|-------|---------|
| Recall | Factual retention | "What was the error?" |
| Artifact | File tracking | "Which files modified?" |
| Continuation | Task planning | "What next?" |
| Decision | Reasoning chains | "Why chose X?" |
## Six Evaluation Dimensions
1. **Accuracy** - Technical correctness
2. **Context Awareness** - Conversation state
3. **Artifact Trail** - File tracking (universally weak)
4. **Completeness** - Coverage depth
5. **Continuity** - Work continuation
6. **Instruction Following** - Constraints
## Guidelines
1. Use anchored iterative for best quality/compression
2. Maintain explicit artifact tracking section
3. Trigger compression at 70% utilization
4. Merge into sections, don't regenerate
5. Evaluate with probes, not lexical metrics
## Related
- [Context Optimization](./context-optimization.md)
- [Evaluation](./evaluation.md)

View File

@@ -0,0 +1,93 @@
# Context Degradation Patterns
Predictable degradation as context grows. Not binary - a continuum.
## Degradation Patterns
| Pattern | Cause | Detection |
|---------|-------|-----------|
| **Lost-in-Middle** | U-shaped attention | Critical info recall drops 10-40% |
| **Context Poisoning** | Errors compound via reference | Persistent hallucinations despite correction |
| **Context Distraction** | Irrelevant info overwhelms | Single distractor degrades performance |
| **Context Confusion** | Multiple tasks mix | Wrong tool calls, mixed requirements |
| **Context Clash** | Contradictory info | Conflicting outputs, inconsistent reasoning |
## Lost-in-Middle Phenomenon
- Information in middle gets 10-40% lower recall
- Models allocate massive attention to first token (BOS sink)
- As context grows, middle tokens fail to get sufficient attention
- **Mitigation**: Place critical info at beginning/end
```markdown
[CURRENT TASK] # Beginning - high attention
- Critical requirements
[DETAILED CONTEXT] # Middle - lower attention
- Supporting details
[KEY FINDINGS] # End - high attention
- Important conclusions
```
## Context Poisoning
**Entry points**:
1. Tool outputs with errors/unexpected formats
2. Retrieved docs with incorrect/outdated info
3. Model-generated summaries with hallucinations
**Detection symptoms**:
- Degraded quality on previously successful tasks
- Tool misalignment (wrong tools/parameters)
- Persistent hallucinations
**Recovery**:
- Truncate to before poisoning point
- Explicit note + re-evaluation request
- Restart with clean context, preserve only verified info
## Model Degradation Thresholds
| Model | Degradation Onset | Severe Degradation |
|-------|-------------------|-------------------|
| GPT-5.2 | ~64K tokens | ~200K tokens |
| Claude Opus 4.5 | ~100K tokens | ~180K tokens |
| Claude Sonnet 4.5 | ~80K tokens | ~150K tokens |
| Gemini 3 Pro | ~500K tokens | ~800K tokens |
## Four-Bucket Mitigation
1. **Write**: Save externally (scratchpads, files)
2. **Select**: Pull only relevant (retrieval, filtering)
3. **Compress**: Reduce tokens (summarization)
4. **Isolate**: Split across sub-agents (partitioning)
## Detection Heuristics
```python
def calculate_health(utilization, degradation_risk, poisoning_risk):
"""Health score: 1.0 = healthy, 0.0 = critical"""
score = 1.0
score -= utilization * 0.5 if utilization > 0.7 else 0
score -= degradation_risk * 0.3
score -= poisoning_risk * 0.2
return max(0, score)
# Thresholds: healthy >0.8, warning >0.6, degraded >0.4, critical <=0.4
```
## Guidelines
1. Monitor context length vs performance correlation
2. Place critical info at beginning/end
3. Implement compaction before degradation
4. Validate retrieved docs before adding
5. Use versioning to prevent outdated clash
6. Segment tasks to prevent confusion
7. Design for graceful degradation
## Related Topics
- [Context Optimization](./context-optimization.md) - Mitigation techniques
- [Multi-Agent Patterns](./multi-agent-patterns.md) - Isolation strategies

View File

@@ -0,0 +1,75 @@
# Context Fundamentals
Context = all input provided to LLM for task completion.
## Anatomy of Context
| Component | Purpose | Token Impact |
|-----------|---------|--------------|
| System Prompt | Identity, constraints, guidelines | Stable, cacheable |
| Tool Definitions | Action specs with params/returns | Grows with capabilities |
| Retrieved Docs | Domain knowledge, just-in-time | Variable, selective |
| Message History | Conversation state, task progress | Accumulates over time |
| Tool Outputs | Results from actions | 83.9% of typical context |
## Attention Mechanics
- **U-shaped curve**: Beginning/end get more attention than middle
- **Attention budget**: n^2 relationships for n tokens depletes with growth
- **Position encoding**: Interpolation allows longer sequences with degradation
- **First-token sink**: BOS token absorbs large attention budget
## System Prompt Structure
```xml
<BACKGROUND_INFORMATION>Domain knowledge, role definition</BACKGROUND_INFORMATION>
<INSTRUCTIONS>Step-by-step procedures</INSTRUCTIONS>
<TOOL_GUIDANCE>When/how to use tools</TOOL_GUIDANCE>
<OUTPUT_DESCRIPTION>Format requirements</OUTPUT_DESCRIPTION>
```
## Progressive Disclosure Levels
1. **Metadata** (~100 words) - Always in context
2. **SKILL.md body** (<5k words) - When skill triggers
3. **Bundled resources** (Unlimited) - As needed
## Token Budget Allocation
| Component | Typical Range | Notes |
|-----------|---------------|-------|
| System Prompt | 500-2000 | Stable, optimize once |
| Tool Definitions | 100-500 per tool | Keep under 20 tools |
| Retrieved Docs | 1000-5000 | Selective loading |
| Message History | Variable | Summarize at 70% |
| Reserved Buffer | 10-20% | For responses |
## Document Management
**Strong identifiers**: `customer_pricing_rates.json` not `data/file1.json`
**Chunk at semantic boundaries**: Paragraphs, sections, not arbitrary lengths
**Include metadata**: Source, date, relevance score
## Message History Pattern
```python
# Summary injection every 20 messages
if len(messages) % 20 == 0:
summary = summarize_conversation(messages[-20:])
messages.append({"role": "system", "content": f"Summary: {summary}"})
```
## Guidelines
1. Treat context as finite with diminishing returns
2. Place critical info at attention-favored positions
3. Use file-system-based access for large documents
4. Pre-load stable content, just-in-time load dynamic
5. Design with explicit token budgets
6. Monitor usage, implement compaction triggers at 70-80%
## Related Topics
- [Context Degradation](./context-degradation.md) - Failure patterns
- [Context Optimization](./context-optimization.md) - Efficiency techniques
- [Memory Systems](./memory-systems.md) - External storage

View File

@@ -0,0 +1,82 @@
# Context Optimization
Extend effective context capacity through strategic techniques.
## Four Core Strategies
| Strategy | Target | Reduction | When to Use |
|----------|--------|-----------|-------------|
| **Compaction** | Full context | 50-70% | Approaching limits |
| **Observation Masking** | Tool outputs | 60-80% | Verbose outputs >80% |
| **KV-Cache Optimization** | Repeated prefixes | 70%+ hit | Stable prompts |
| **Context Partitioning** | Work distribution | N/A | Parallelizable tasks |
## Compaction
Summarize context when approaching limits.
**Priority**: Tool outputs → Old turns → Retrieved docs → Never: System prompt
```python
if context_tokens / context_limit > 0.8:
context = compact_context(context)
```
**Preserve**: Key findings, decisions, commitments (remove supporting details)
## Observation Masking
Replace verbose tool outputs with compact references.
```python
if len(observation) > max_length:
ref_id = store_observation(observation)
return f"[Obs:{ref_id}. Key: {extract_key(observation)}]"
```
**Never mask**: Current task critical, most recent turn, active reasoning
**Always mask**: Repeated outputs, boilerplate, already summarized
## KV-Cache Optimization
Reuse cached Key/Value tensors for identical prefixes.
```python
# Cache-friendly ordering (stable first)
context = [system_prompt, tool_definitions] # Cacheable
context += [unique_content] # Variable last
```
**Tips**: Avoid timestamps in stable sections, consistent formatting, stable structure
## Context Partitioning
Split work across sub-agents with isolated contexts.
```python
result = await sub_agent.process(subtask, clean_context=True)
coordinator.receive(result.summary) # Only essentials
```
## Decision Framework
| Dominant Component | Apply |
|-------------------|-------|
| Tool outputs | Observation masking |
| Retrieved docs | Summarization or partitioning |
| Message history | Compaction + summarization |
| Multiple | Combine strategies |
## Guidelines
1. Measure before optimizing
2. Apply compaction before masking
3. Design for cache stability
4. Partition before context problematic
5. Monitor effectiveness over time
6. Balance savings vs quality
## Related
- [Context Compression](./context-compression.md)
- [Memory Systems](./memory-systems.md)

View File

@@ -0,0 +1,89 @@
# Evaluation
Systematically assess agent performance and context engineering choices.
## Key Finding: 95% Performance Variance
- **Token usage**: 80% of variance
- **Tool calls**: ~10% of variance
- **Model choice**: ~5% of variance
**Implication**: Token budgets matter more than model upgrades.
## Multi-Dimensional Rubric
| Dimension | Weight | Description |
|-----------|--------|-------------|
| Factual Accuracy | 30% | Ground truth verification |
| Completeness | 25% | Coverage of requirements |
| Tool Efficiency | 20% | Appropriate tool usage |
| Citation Accuracy | 15% | Sources match claims |
| Source Quality | 10% | Authority/credibility |
## Evaluation Methods
### LLM-as-Judge
Beware biases:
- **Position**: First position preferred
- **Length**: Longer = higher score
- **Self-enhancement**: Rating own outputs higher
- **Verbosity**: Detailed = better
**Mitigation**: Position swapping, anti-bias prompting
### Pairwise Comparison
```python
score_ab = judge.compare(output_a, output_b)
score_ba = judge.compare(output_b, output_a)
consistent = (score_ab > 0.5) != (score_ba > 0.5)
```
### Probe-Based Testing
| Probe | Tests | Example |
|-------|-------|---------|
| Recall | Facts | "What was the error?" |
| Artifact | Files | "Which files modified?" |
| Continuation | Planning | "What's next?" |
| Decision | Reasoning | "Why chose X?" |
## Test Set Design
```python
class TestSet:
def sample_stratified(self, n):
per_level = n // 3
return (
sample(self.simple, per_level) +
sample(self.medium, per_level) +
sample(self.complex, per_level)
)
```
## Production Monitoring
```python
class Monitor:
sample_rate = 0.01 # 1% sampling
alert_threshold = 0.85
def check(self, scores):
if avg(scores) < self.alert_threshold:
self.alert(f"Quality degraded: {avg(scores):.2f}")
```
## Guidelines
1. Start with outcome evaluation, not step-by-step
2. Use multi-dimensional rubrics
3. Mitigate LLM-as-Judge biases
4. Test with stratified complexity
5. Implement continuous monitoring
6. Focus on token efficiency (80% variance)
## Related
- [Context Compression](./context-compression.md)
- [Tool Design](./tool-design.md)

View File

@@ -0,0 +1,88 @@
# Memory Systems
Architectures for persistent context beyond the window.
## Memory Layer Architecture
| Layer | Scope | Persistence | Use Case |
|-------|-------|-------------|----------|
| L1: Working | Current window | None | Active reasoning |
| L2: Short-Term | Session | Session | Task continuity |
| L3: Long-Term | Cross-session | Persistent | User preferences |
| L4: Entity | Per-entity | Persistent | Consistency |
| L5: Temporal Graph | Time-aware | Persistent | Evolving facts |
## Benchmark Performance (DMR Accuracy)
| System | Accuracy | Approach |
|--------|----------|----------|
| Zep | 94.8% | Temporal knowledge graphs |
| MemGPT | 93.4% | Hierarchical memory |
| GraphRAG | 75-85% | Knowledge graphs |
| Vector RAG | 60-70% | Embedding similarity |
## Vector Store with Metadata
```python
class MetadataVectorStore:
def add(self, text, embedding, metadata):
doc = {
"text": text, "embedding": embedding,
"entities": metadata.get("entities", []),
"timestamp": metadata.get("timestamp")
}
self.index_by_entity(doc)
def search_by_entity(self, entity, k=5):
return self.entity_index.get(entity, [])[:k]
```
## Temporal Knowledge Graph
```python
class TemporalKnowledgeGraph:
def add_fact(self, subject, predicate, obj, valid_from, valid_to=None):
self.facts.append({
"triple": (subject, predicate, obj),
"valid_from": valid_from,
"valid_to": valid_to or "current"
})
def query_at_time(self, subject, predicate, timestamp):
for fact in self.facts:
if (fact["triple"][0] == subject and
fact["valid_from"] <= timestamp <= fact["valid_to"]):
return fact["triple"][2]
```
## Memory Retrieval Patterns
| Pattern | Query | Use Case |
|---------|-------|----------|
| Semantic | "Similar to X" | General recall |
| Entity-based | "About user John" | Consistency |
| Temporal | "Valid on date" | Evolving facts |
| Hybrid | Combine above | Production |
## File-System-as-Memory
```
memory/
├── sessions/{id}/summary.md
├── entities/{id}.json
└── facts/{timestamp}_{id}.json
```
## Guidelines
1. Start with file-system-as-memory (simplest)
2. Add vector search for scale
3. Use entity indexing for consistency
4. Add temporal awareness for evolving facts
5. Implement consolidation for health
6. Measure retrieval accuracy
## Related
- [Context Fundamentals](./context-fundamentals.md)
- [Multi-Agent Patterns](./multi-agent-patterns.md)

View File

@@ -0,0 +1,90 @@
# Multi-Agent Patterns
Distribute work across multiple context windows for isolation and scale.
## Core Insight
Sub-agents exist to **isolate context**, not anthropomorphize roles.
## Token Economics
| Architecture | Multiplier | Use Case |
|--------------|------------|----------|
| Single agent | 1x | Simple tasks |
| Single + tools | ~4x | Moderate complexity |
| Multi-agent | ~15x | Context isolation needed |
**Key**: Token usage explains 80% of performance variance.
## Patterns
### Supervisor/Orchestrator
```python
class Supervisor:
def process(self, task):
subtasks = self.decompose(task)
results = [worker.execute(st, clean_context=True) for st in subtasks]
return self.aggregate(results)
```
**Pros**: Control, human-in-loop | **Cons**: Bottleneck, telephone game
### Peer-to-Peer/Swarm
```python
def process_with_handoff(agent, task):
result = agent.process(task)
if "handoff" in result:
return process_with_handoff(select_agent(result["to"]), result["state"])
return result
```
**Pros**: No SPOF, scales | **Cons**: Complex coordination
### Hierarchical
Strategy → Planning → Execution layers
**Pros**: Separation of concerns | **Cons**: Coordination overhead
## Context Isolation Patterns
| Pattern | Isolation | Use Case |
|---------|-----------|----------|
| Full delegation | None | Max capability |
| Instruction passing | High | Simple tasks |
| File coordination | Medium | Shared state |
## Consensus Mechanisms
```python
def weighted_consensus(responses):
scores = {}
for r in responses:
weight = r["confidence"] * r["expertise"]
scores[r["answer"]] = scores.get(r["answer"], 0) + weight
return max(scores, key=scores.get)
```
## Failure Recovery
| Failure | Mitigation |
|---------|------------|
| Bottleneck | Output schemas, checkpointing |
| Overhead | Clear handoffs, batching |
| Divergence | Boundaries, convergence checks |
| Errors | Validation, circuit breakers |
## Guidelines
1. Use multi-agent for context isolation, not role-play
2. Accept ~15x token cost for benefits
3. Implement circuit breakers
4. Use files for shared state
5. Design clear handoffs
6. Validate between agents
## Related
- [Context Optimization](./context-optimization.md)
- [Evaluation](./evaluation.md)

View File

@@ -0,0 +1,97 @@
# Project Development
Design and build LLM-powered projects from ideation to deployment.
## Task-Model Fit
**LLM-Suited**: Synthesis, subjective judgment, NL output, error-tolerant batches
**LLM-Unsuited**: Precise computation, real-time, perfect accuracy, deterministic output
## Manual Prototype First
Test one example with target model before automation.
## Pipeline Architecture
```
acquire → prepare → process → parse → render
(fetch) (prompt) (LLM) (extract) (output)
```
Stages 1,2,4,5: Deterministic, cheap | Stage 3: Non-deterministic, expensive
## File System as State
```
data/{id}/
├── raw.json # acquire done
├── prompt.md # prepare done
├── response.md # process done
└── parsed.json # parse done
```
```python
def get_stage(id):
if exists(f"{id}/parsed.json"): return "render"
if exists(f"{id}/response.md"): return "parse"
# ... check backwards
```
**Benefits**: Idempotent, resumable, debuggable
## Structured Output
```markdown
## SUMMARY
[Overview]
## KEY_FINDINGS
- Finding 1
## SCORE
[1-5]
```
```python
def parse(response):
return {
"summary": extract_section(response, "SUMMARY"),
"findings": extract_list(response, "KEY_FINDINGS"),
"score": extract_int(response, "SCORE")
}
```
## Cost Estimation
```python
def estimate(items, tokens_per, price_per_1k):
return len(items) * tokens_per / 1000 * price_per_1k * 1.1 # 10% buffer
# 1000 items × 2000 tokens × $0.01/1k = $22
```
## Case Studies
**Karpathy HN**: 930 items, $58, 1hr, 15 workers
**Vercel d0**: 17→2 tools, 80%→100% success, 3.5x faster
## Single vs Multi-Agent
| Factor | Single | Multi |
|--------|--------|-------|
| Context | Fits window | Exceeds |
| Tasks | Sequential | Parallel |
| Tokens | Limited | 15x OK |
## Guidelines
1. Validate manually before automating
2. Use 5-stage pipeline
3. Track state via files
4. Design structured output
5. Estimate costs first
6. Start single, add multi when needed
## Related
- [Context Optimization](./context-optimization.md)
- [Multi-Agent Patterns](./multi-agent-patterns.md)

View File

@@ -0,0 +1,202 @@
# Runtime Awareness
Monitor usage limits and context window utilization in real-time to optimize Claude Code sessions.
## Overview
Runtime awareness provides visibility into two critical metrics:
1. **Usage Limits** - API quota consumption (5-hour and 7-day rolling windows)
2. **Context Window** - Current token utilization within the 200K context limit
## Architecture
```
┌─────────────────┐ ┌──────────────────────────┐
│ statusline.cjs │───▶│ /tmp/ck-context-*.json │
│ (writes data) │ │ (context window data) │
└─────────────────┘ └────────────┬─────────────┘
┌────────────▼─────────────┐
│ usage-context-hook.cjs │◀── PostToolUse
│ - Reads context file │
│ - Fetches usage limits │
│ - Injects awareness │
└──────────────────────────┘
```
## Usage Limits API
### Endpoint
```
GET https://api.anthropic.com/api/oauth/usage
```
### Authentication
Requires OAuth Bearer token with `anthropic-beta: oauth-2025-04-20` header.
### Credential Locations
| Platform | Method | Location |
|----------|--------|----------|
| macOS | Keychain | `Claude Code-credentials` |
| Windows | File | `%USERPROFILE%\.claude\.credentials.json` |
| Linux | File | `~/.opencode/.credentials.json` |
### Response Structure
```json
{
"five_hour": {
"utilization": 45,
"resets_at": "2025-01-15T18:00:00Z"
},
"seven_day": {
"utilization": 32,
"resets_at": "2025-01-22T00:00:00Z"
},
"seven_day_sonnet": {
"utilization": 11,
"resets_at": "2025-01-15T09:00:00Z"
}
}
```
- `utilization`: Already a percentage (0-100), NOT a decimal
- `resets_at`: ISO 8601 timestamp when quota resets
- `seven_day_sonnet`: Model-specific limit (may be null)
## Context Window Data
### Source
Statusline writes context data to `/tmp/ck-context-{session_id}.json`:
```json
{
"percent": 67,
"tokens": 134000,
"size": 200000,
"usage": {
"input_tokens": 80000,
"cache_creation_input_tokens": 30000,
"cache_read_input_tokens": 24000
},
"timestamp": 1705312000000
}
```
### Token Calculation
```
total = input_tokens + cache_creation_input_tokens + cache_read_input_tokens
percent = (total + AUTOCOMPACT_BUFFER) / context_window_size * 100
```
Where `AUTOCOMPACT_BUFFER = 45000` (22.5% reserved).
## Hook Output
The PostToolUse hook injects awareness data every 5 minutes:
```xml
<usage-awareness>
Limits: 5h=45%, 7d=32%
Context: 67%
</usage-awareness>
```
### Warning Indicators
| Level | Threshold | Indicator |
|-------|-----------|-----------|
| Normal | < 70% | Plain percentage |
| Warning | 70-89% | `[WARNING]` |
| Critical | 90% | `[CRITICAL]` |
### Examples
Normal state:
```xml
<usage-awareness>
Limits: 5h=45%, 7d=32%
Context: 67%
</usage-awareness>
```
Warning state:
```xml
<usage-awareness>
Limits: 5h=75% [WARNING], 7d=32%
Context: 78% [WARNING - consider compaction]
</usage-awareness>
```
Critical state:
```xml
<usage-awareness>
Limits: 5h=92% [CRITICAL], 7d=65%
Context: 91% [CRITICAL - compaction needed]
</usage-awareness>
```
## Recommendations by Threshold
### Context Window
| Utilization | Action |
|-------------|--------|
| < 70% | Continue normally |
| 70-80% | Plan compaction strategy |
| 80-90% | Execute compaction |
| > 90% | Immediate compaction or session reset |
### Usage Limits
| 5-Hour | Action |
|--------|--------|
| < 70% | Normal usage |
| 70-90% | Reduce parallelization, delegate to subagents |
| > 90% | Wait for reset or use lower-tier models |
| 7-Day | Action |
|-------|--------|
| < 70% | Normal usage |
| 70-90% | Monitor daily consumption |
| > 90% | Limit usage to essential tasks |
## Configuration
### Hook Settings (`.opencode/settings.json`)
```json
{
"hooks": {
"PostToolUse": [
{
"matcher": "*",
"hooks": [{
"type": "command",
"command": "node .opencode/hooks/usage-quota-cache-refresh.cjs"
}]
}
]
}
}
```
### Throttling
- **Injection interval**: 5 minutes (300,000ms)
- **API cache TTL**: 60 seconds
- **Context data freshness**: 30 seconds
## Troubleshooting
| Issue | Cause | Solution |
|-------|-------|----------|
| No usage limits shown | No OAuth token | Run `claude login` |
| Stale context data | Statusline not updating | Check statusline config |
| 401 Unauthorized | Expired token | Re-authenticate |
| Hook not firing | Settings misconfigured | Verify PostToolUse matcher |

View File

@@ -0,0 +1,86 @@
# Tool Design
Design effective tools for agent systems.
## Consolidation Principle
Single comprehensive tools > multiple narrow tools. **Target**: 10-20 tools max.
## Architectural Reduction Evidence
| Metric | 17 Tools | 2 Tools | Improvement |
|--------|----------|---------|-------------|
| Time | 274.8s | 77.4s | 3.5x faster |
| Success | 80% | 100% | +20% |
| Tokens | 102k | 61k | 37% fewer |
**Key**: Good documentation replaces tool sophistication.
## When Reduction Works
**Prerequisites**: High docs quality, capable model, navigable problem
**Avoid when**: Messy systems, specialized domain, safety-critical
## Description Engineering
Answer four questions:
1. **What** does the tool do?
2. **When** should it be used?
3. **What inputs** does it accept?
4. **What** does it return?
### Good Example
```json
{
"name": "get_customer",
"description": "Retrieve customer profile by ID. Use for order processing, support. Returns 404 if not found.",
"parameters": {
"customer_id": {"type": "string", "pattern": "^CUST-[0-9]{6}$"},
"format": {"enum": ["concise", "detailed"]}
}
}
```
### Poor Example
```json
{"name": "search", "description": "Search for things", "parameters": {"q": {}}}
```
## Error Messages
```python
def format_error(code, message, resolution):
return {
"error": {"code": code, "message": message,
"resolution": resolution, "retryable": code in RETRYABLE}
}
# "Use YYYY-MM-DD format, e.g., '2024-01-05'"
```
## Response Formats
Offer concise vs detailed:
```python
def get_data(id, format="concise"):
if format == "concise":
return {"name": data.name}
return data.full() # Detailed
```
## Guidelines
1. Consolidate tools (target 10-20)
2. Answer all four questions
3. Use full parameter names
4. Design errors for recovery
5. Offer concise/detailed formats
6. Test with agents before deploy
7. Start minimal, add when proven
## Related
- [Context Fundamentals](./context-fundamentals.md)
- [Multi-Agent Patterns](./multi-agent-patterns.md)

View File

@@ -0,0 +1,349 @@
#!/usr/bin/env python3
"""
Compression Evaluator - Evaluate compression quality with probe-based testing.
Usage:
python compression_evaluator.py evaluate <original_file> <compressed_file>
python compression_evaluator.py generate-probes <context_file>
"""
import argparse
import json
import os
import re
import sys
from dataclasses import dataclass, field
from enum import Enum
from typing import Optional
MAX_FILE_SIZE_MB = 100
def load_file(path: str, as_json: bool = True):
"""Load file with proper error handling and size validation."""
try:
size_mb = os.path.getsize(path) / (1024 * 1024)
if size_mb > MAX_FILE_SIZE_MB:
print(f"Error: File too large ({size_mb:.1f}MB). Max {MAX_FILE_SIZE_MB}MB", file=sys.stderr)
sys.exit(1)
with open(path, encoding='utf-8') as f:
return json.load(f) if as_json else f.read()
except FileNotFoundError:
print(f"Error: File not found: {path}", file=sys.stderr)
sys.exit(1)
except PermissionError:
print(f"Error: Permission denied: {path}", file=sys.stderr)
sys.exit(1)
except json.JSONDecodeError as e:
print(f"Error: Invalid JSON in {path}: {e}", file=sys.stderr)
sys.exit(1)
class ProbeType(Enum):
RECALL = "recall" # Factual retention
ARTIFACT = "artifact" # File tracking
CONTINUATION = "continuation" # Task planning
DECISION = "decision" # Reasoning chains
@dataclass
class Probe:
type: ProbeType
question: str
ground_truth: str
context_reference: Optional[str] = None
@dataclass
class ProbeResult:
probe: Probe
response: str
scores: dict
overall_score: float
@dataclass
class EvaluationReport:
compression_ratio: float
quality_score: float
dimension_scores: dict
probe_results: list
recommendations: list = field(default_factory=list)
# Six evaluation dimensions with weights
DIMENSIONS = {
"accuracy": {"weight": 0.20, "description": "Technical correctness"},
"context_awareness": {"weight": 0.15, "description": "Conversation state"},
"artifact_trail": {"weight": 0.20, "description": "File tracking"},
"completeness": {"weight": 0.20, "description": "Coverage and depth"},
"continuity": {"weight": 0.15, "description": "Work continuation"},
"instruction_following": {"weight": 0.10, "description": "Constraint adherence"}
}
def estimate_tokens(text: str) -> int:
"""Estimate token count."""
return len(text) // 4
def extract_facts(messages: list) -> list:
"""Extract factual statements that can be probed."""
facts = []
patterns = [
(r"error[:\s]+([^.]+)", "error"),
(r"next step[s]?[:\s]+([^.]+)", "next_step"),
(r"decided to\s+([^.]+)", "decision"),
(r"implemented\s+([^.]+)", "implementation"),
(r"found that\s+([^.]+)", "finding")
]
for msg in messages:
content = str(msg.get("content", "") if isinstance(msg, dict) else msg)
for pattern, fact_type in patterns:
matches = re.findall(pattern, content, re.IGNORECASE)
for match in matches:
facts.append({"type": fact_type, "content": match.strip()})
return facts
def extract_files(messages: list) -> list:
"""Extract file references."""
files = []
patterns = [
r"(?:created|modified|updated|edited|read)\s+[`'\"]?([a-zA-Z0-9_/.-]+\.[a-zA-Z]+)[`'\"]?",
r"file[:\s]+[`'\"]?([a-zA-Z0-9_/.-]+\.[a-zA-Z]+)[`'\"]?"
]
for msg in messages:
content = str(msg.get("content", "") if isinstance(msg, dict) else msg)
for pattern in patterns:
matches = re.findall(pattern, content)
files.extend(matches)
return list(set(files))
def extract_decisions(messages: list) -> list:
"""Extract decision points."""
decisions = []
patterns = [
r"chose\s+([^.]+)\s+(?:because|since|over)",
r"decided\s+(?:to\s+)?([^.]+)",
r"went with\s+([^.]+)"
]
for msg in messages:
content = str(msg.get("content", "") if isinstance(msg, dict) else msg)
for pattern in patterns:
matches = re.findall(pattern, content, re.IGNORECASE)
decisions.extend(matches)
return decisions
def generate_probes(messages: list) -> list:
"""Generate probe set for evaluation."""
probes = []
# Recall probes from facts
facts = extract_facts(messages)
for fact in facts[:3]: # Limit to 3 recall probes
probes.append(Probe(
type=ProbeType.RECALL,
question=f"What was the {fact['type'].replace('_', ' ')}?",
ground_truth=fact["content"]
))
# Artifact probes from files
files = extract_files(messages)
if files:
probes.append(Probe(
type=ProbeType.ARTIFACT,
question="Which files have been modified or created?",
ground_truth=", ".join(files)
))
# Continuation probe
probes.append(Probe(
type=ProbeType.CONTINUATION,
question="What should be done next?",
ground_truth="[Extracted from context]" # Would need LLM to generate
))
# Decision probes
decisions = extract_decisions(messages)
for decision in decisions[:2]: # Limit to 2 decision probes
probes.append(Probe(
type=ProbeType.DECISION,
question=f"Why was the decision made to {decision[:50]}...?",
ground_truth=decision
))
return probes
def evaluate_response(probe: Probe, response: str) -> dict:
"""
Evaluate response against probe.
Note: Production should use LLM-as-Judge.
"""
scores = {}
response_lower = response.lower()
ground_truth_lower = probe.ground_truth.lower()
# Heuristic scoring (replace with LLM evaluation in production)
# Check for ground truth presence
if ground_truth_lower in response_lower:
base_score = 1.0
elif any(word in response_lower for word in ground_truth_lower.split()[:3]):
base_score = 0.6
else:
base_score = 0.3
# Adjust based on probe type
if probe.type == ProbeType.ARTIFACT:
# Check file mentions
files_mentioned = len(re.findall(r'\.[a-z]+', response_lower))
scores["artifact_trail"] = min(1.0, base_score + files_mentioned * 0.1)
scores["accuracy"] = base_score
elif probe.type == ProbeType.RECALL:
scores["accuracy"] = base_score
scores["completeness"] = base_score
elif probe.type == ProbeType.CONTINUATION:
scores["continuity"] = base_score
scores["context_awareness"] = base_score
elif probe.type == ProbeType.DECISION:
scores["accuracy"] = base_score
scores["context_awareness"] = base_score
return scores
def calculate_compression_ratio(original: str, compressed: str) -> float:
"""Calculate compression ratio."""
original_tokens = estimate_tokens(original)
compressed_tokens = estimate_tokens(compressed)
if original_tokens == 0:
return 0.0
return 1.0 - (compressed_tokens / original_tokens)
def evaluate_compression(original_messages: list, compressed_text: str,
probes: Optional[list] = None) -> EvaluationReport:
"""
Evaluate compression quality.
Args:
original_messages: Original context messages
compressed_text: Compressed summary
probes: Optional pre-generated probes
Returns:
EvaluationReport with scores and recommendations
"""
# Generate probes if not provided
if probes is None:
probes = generate_probes(original_messages)
# Calculate compression ratio
original_text = json.dumps(original_messages)
compression_ratio = calculate_compression_ratio(original_text, compressed_text)
# Evaluate each probe (simulated - production uses LLM)
probe_results = []
dimension_scores = {dim: [] for dim in DIMENSIONS}
for probe in probes:
# In production, send compressed_text + probe.question to LLM
# Here we simulate with heuristic check
scores = evaluate_response(probe, compressed_text)
overall = sum(scores.values()) / len(scores) if scores else 0
probe_results.append(ProbeResult(
probe=probe,
response="[Would be LLM response]",
scores=scores,
overall_score=overall
))
# Aggregate by dimension
for dim, score in scores.items():
if dim in dimension_scores:
dimension_scores[dim].append(score)
# Calculate dimension averages
avg_dimensions = {}
for dim, scores in dimension_scores.items():
avg_dimensions[dim] = sum(scores) / len(scores) if scores else 0.5
# Calculate weighted quality score
quality_score = sum(
avg_dimensions.get(dim, 0.5) * info["weight"]
for dim, info in DIMENSIONS.items()
)
# Generate recommendations
recommendations = []
if compression_ratio > 0.99:
recommendations.append("Very high compression. Risk of information loss.")
if avg_dimensions.get("artifact_trail", 1) < 0.5:
recommendations.append("Artifact tracking weak. Add explicit file section to summary.")
if avg_dimensions.get("continuity", 1) < 0.5:
recommendations.append("Continuity low. Add 'Next Steps' section to summary.")
if quality_score < 0.6:
recommendations.append("Quality below threshold. Consider less aggressive compression.")
return EvaluationReport(
compression_ratio=compression_ratio,
quality_score=quality_score,
dimension_scores=avg_dimensions,
probe_results=probe_results,
recommendations=recommendations
)
def main():
parser = argparse.ArgumentParser(description="Compression quality evaluator")
subparsers = parser.add_subparsers(dest="command", required=True)
# Evaluate command
eval_parser = subparsers.add_parser("evaluate", help="Evaluate compression quality")
eval_parser.add_argument("original_file", help="JSON file with original messages")
eval_parser.add_argument("compressed_file", help="Text file with compressed summary")
# Generate probes command
probe_parser = subparsers.add_parser("generate-probes", help="Generate evaluation probes")
probe_parser.add_argument("context_file", help="JSON file with context messages")
args = parser.parse_args()
if args.command == "evaluate":
original = load_file(args.original_file, as_json=True)
messages = original if isinstance(original, list) else original.get("messages", [])
compressed = load_file(args.compressed_file, as_json=False)
report = evaluate_compression(messages, compressed)
print(json.dumps({
"compression_ratio": f"{report.compression_ratio:.1%}",
"quality_score": f"{report.quality_score:.2f}",
"dimension_scores": {k: f"{v:.2f}" for k, v in report.dimension_scores.items()},
"probe_count": len(report.probe_results),
"recommendations": report.recommendations
}, indent=2))
elif args.command == "generate-probes":
data = load_file(args.context_file, as_json=True)
messages = data if isinstance(data, list) else data.get("messages", [])
probes = generate_probes(messages)
output = []
for probe in probes:
output.append({
"type": probe.type.value,
"question": probe.question,
"ground_truth": probe.ground_truth
})
print(json.dumps(output, indent=2))
if __name__ == "__main__":
main()

View File

@@ -0,0 +1,317 @@
#!/usr/bin/env python3
"""
Context Analyzer - Health analysis and degradation detection for agent contexts.
Usage:
python context_analyzer.py analyze <context_file>
python context_analyzer.py budget --system 2000 --tools 1500 --docs 3000 --history 5000
"""
import argparse
import json
import math
import os
import re
import sys
from dataclasses import dataclass, field
from enum import Enum
from typing import Optional
MAX_FILE_SIZE_MB = 100
def load_json_file(path: str):
"""Load JSON file with proper error handling and size validation."""
try:
size_mb = os.path.getsize(path) / (1024 * 1024)
if size_mb > MAX_FILE_SIZE_MB:
print(f"Error: File too large ({size_mb:.1f}MB). Max {MAX_FILE_SIZE_MB}MB", file=sys.stderr)
sys.exit(1)
with open(path, encoding='utf-8') as f:
return json.load(f)
except FileNotFoundError:
print(f"Error: File not found: {path}", file=sys.stderr)
sys.exit(1)
except PermissionError:
print(f"Error: Permission denied: {path}", file=sys.stderr)
sys.exit(1)
except json.JSONDecodeError as e:
print(f"Error: Invalid JSON in {path}: {e}", file=sys.stderr)
sys.exit(1)
class HealthStatus(Enum):
HEALTHY = "healthy"
WARNING = "warning"
DEGRADED = "degraded"
CRITICAL = "critical"
@dataclass
class ContextAnalysis:
total_tokens: int
token_limit: int
utilization: float
health_status: HealthStatus
health_score: float
degradation_risk: float
poisoning_risk: float
recommendations: list = field(default_factory=list)
def estimate_tokens(text: str) -> int:
"""Estimate token count (~4 chars per token for English)."""
return len(text) // 4
def estimate_message_tokens(messages: list) -> int:
"""Estimate tokens in message list."""
total = 0
for msg in messages:
if isinstance(msg, dict):
content = msg.get("content", "")
total += estimate_tokens(str(content))
# Add overhead for role, metadata
total += 10
else:
total += estimate_tokens(str(msg))
return total
def measure_attention_distribution(context_length: int, sample_size: int = 100) -> list:
"""
Simulate U-shaped attention distribution.
Real implementation would extract from model attention weights.
"""
attention = []
for i in range(sample_size):
position = i / sample_size
# U-shaped curve: high at start/end, low in middle
if position < 0.1:
score = 0.9 - position * 2
elif position > 0.9:
score = 0.7 + (position - 0.9) * 2
else:
score = 0.3 + 0.1 * math.sin(position * math.pi)
attention.append(score)
return attention
def detect_lost_in_middle(messages: list, critical_keywords: list) -> list:
"""Identify critical items in attention-degraded regions."""
if not messages:
return []
total = len(messages)
warnings = []
for i, msg in enumerate(messages):
position = i / total
content = str(msg.get("content", "") if isinstance(msg, dict) else msg)
# Middle region (10%-90%)
if 0.1 < position < 0.9:
for keyword in critical_keywords:
if keyword.lower() in content.lower():
warnings.append({
"position": i,
"position_pct": f"{position:.1%}",
"keyword": keyword,
"risk": "high" if 0.3 < position < 0.7 else "medium"
})
return warnings
def detect_poisoning_patterns(messages: list) -> dict:
"""Detect potential context poisoning indicators."""
error_patterns = [
r"error", r"failed", r"exception", r"cannot", r"unable",
r"invalid", r"not found", r"undefined", r"null"
]
# Simple contradiction check - look for both positive and negative statements
contradiction_keywords = [
("is correct", "is not correct"),
("should work", "should not work"),
("will succeed", "will fail"),
("is valid", "is invalid"),
]
errors_found = []
contradictions = []
for i, msg in enumerate(messages):
content = str(msg.get("content", "") if isinstance(msg, dict) else msg).lower()
# Check error patterns
for pattern in error_patterns:
if re.search(pattern, content):
errors_found.append({"position": i, "pattern": pattern})
# Check for contradiction keywords (simplified)
for pos_phrase, neg_phrase in contradiction_keywords:
if pos_phrase in content and neg_phrase in content:
contradictions.append({"position": i, "type": "self-contradiction"})
total = max(len(messages), 1)
return {
"error_density": len(errors_found) / total,
"contradiction_count": len(contradictions),
"poisoning_risk": min(1.0, (len(errors_found) * 0.1 + len(contradictions) * 0.3))
}
def calculate_health_score(utilization: float, degradation_risk: float, poisoning_risk: float) -> float:
"""
Calculate composite health score.
1.0 = healthy, 0.0 = critical
"""
score = 1.0
# Utilization penalty (kicks in after 70%)
if utilization > 0.7:
score -= (utilization - 0.7) * 1.5
# Degradation penalty
score -= degradation_risk * 0.3
# Poisoning penalty
score -= poisoning_risk * 0.2
return max(0.0, min(1.0, score))
def get_health_status(score: float) -> HealthStatus:
"""Map health score to status."""
if score > 0.8:
return HealthStatus.HEALTHY
elif score > 0.6:
return HealthStatus.WARNING
elif score > 0.4:
return HealthStatus.DEGRADED
return HealthStatus.CRITICAL
def analyze_context(messages: list, token_limit: int = 128000,
critical_keywords: Optional[list] = None) -> ContextAnalysis:
"""
Comprehensive context health analysis.
Args:
messages: List of context messages
token_limit: Model's context window size
critical_keywords: Keywords that should be at attention-favored positions
Returns:
ContextAnalysis with health metrics and recommendations
"""
critical_keywords = critical_keywords or ["goal", "task", "important", "critical", "must"]
# Calculate token utilization
total_tokens = estimate_message_tokens(messages)
utilization = total_tokens / token_limit
# Check for lost-in-middle issues
middle_warnings = detect_lost_in_middle(messages, critical_keywords)
degradation_risk = min(1.0, len(middle_warnings) * 0.2)
# Check for poisoning
poisoning = detect_poisoning_patterns(messages)
poisoning_risk = poisoning["poisoning_risk"]
# Calculate health
health_score = calculate_health_score(utilization, degradation_risk, poisoning_risk)
health_status = get_health_status(health_score)
# Generate recommendations
recommendations = []
if utilization > 0.8:
recommendations.append("URGENT: Context utilization >80%. Trigger compaction immediately.")
elif utilization > 0.7:
recommendations.append("WARNING: Context utilization >70%. Plan for compaction.")
if middle_warnings:
recommendations.append(f"Found {len(middle_warnings)} critical items in middle region. "
"Consider moving to beginning/end.")
if poisoning_risk > 0.3:
recommendations.append("High poisoning risk detected. Review recent tool outputs for errors.")
if health_status == HealthStatus.CRITICAL:
recommendations.append("CRITICAL: Consider context reset with clean state.")
return ContextAnalysis(
total_tokens=total_tokens,
token_limit=token_limit,
utilization=utilization,
health_status=health_status,
health_score=health_score,
degradation_risk=degradation_risk,
poisoning_risk=poisoning_risk,
recommendations=recommendations
)
def calculate_budget(system: int, tools: int, docs: int, history: int,
buffer_pct: float = 0.15) -> dict:
"""Calculate context budget allocation."""
subtotal = system + tools + docs + history
buffer = int(subtotal * buffer_pct)
total = subtotal + buffer
return {
"allocation": {
"system_prompt": system,
"tool_definitions": tools,
"retrieved_docs": docs,
"message_history": history,
"reserved_buffer": buffer
},
"total_budget": total,
"warning_threshold": int(total * 0.7),
"critical_threshold": int(total * 0.8),
"recommendations": [
f"Trigger compaction at {int(total * 0.7):,} tokens",
f"Aggressive optimization at {int(total * 0.8):,} tokens",
f"Reserved {buffer:,} tokens ({buffer_pct:.0%}) for responses"
]
}
def main():
parser = argparse.ArgumentParser(description="Context health analyzer")
subparsers = parser.add_subparsers(dest="command", required=True)
# Analyze command
analyze_parser = subparsers.add_parser("analyze", help="Analyze context health")
analyze_parser.add_argument("context_file", help="JSON file with messages array")
analyze_parser.add_argument("--limit", type=int, default=128000, help="Token limit")
analyze_parser.add_argument("--keywords", nargs="+", help="Critical keywords to track")
# Budget command
budget_parser = subparsers.add_parser("budget", help="Calculate context budget")
budget_parser.add_argument("--system", type=int, default=2000, help="System prompt tokens")
budget_parser.add_argument("--tools", type=int, default=1500, help="Tool definitions tokens")
budget_parser.add_argument("--docs", type=int, default=3000, help="Retrieved docs tokens")
budget_parser.add_argument("--history", type=int, default=5000, help="Message history tokens")
budget_parser.add_argument("--buffer", type=float, default=0.15, help="Buffer percentage")
args = parser.parse_args()
if args.command == "analyze":
data = load_json_file(args.context_file)
messages = data if isinstance(data, list) else data.get("messages", [])
result = analyze_context(messages, args.limit, args.keywords)
print(json.dumps({
"total_tokens": result.total_tokens,
"token_limit": result.token_limit,
"utilization": f"{result.utilization:.1%}",
"health_status": result.health_status.value,
"health_score": f"{result.health_score:.2f}",
"degradation_risk": f"{result.degradation_risk:.2f}",
"poisoning_risk": f"{result.poisoning_risk:.2f}",
"recommendations": result.recommendations
}, indent=2))
elif args.command == "budget":
result = calculate_budget(args.system, args.tools, args.docs, args.history, args.buffer)
print(json.dumps(result, indent=2))
if __name__ == "__main__":
main()

View File

@@ -0,0 +1,246 @@
"""Tests for context-engineering edge case handling.
Tests the error handling improvements in compression_evaluator.py and context_analyzer.py:
- File not found
- Permission denied
- Invalid JSON
- File too large
- UTF-8 encoding
"""
import json
import os
import stat
import subprocess
import sys
import tempfile
from pathlib import Path
import pytest
SCRIPTS_DIR = Path(__file__).parent.parent
PYTHON = sys.executable
class TestCompressionEvaluatorEdgeCases:
"""Test edge cases in compression_evaluator.py"""
@pytest.fixture
def valid_json_file(self, tmp_path):
"""Create valid JSON file."""
f = tmp_path / "valid.json"
f.write_text('{"messages": [{"role": "user", "content": "hello"}]}', encoding='utf-8')
return str(f)
@pytest.fixture
def valid_text_file(self, tmp_path):
"""Create valid text file."""
f = tmp_path / "compressed.txt"
f.write_text("Summary of conversation", encoding='utf-8')
return str(f)
def run_script(self, *args, timeout=30):
"""Run compression_evaluator.py with args."""
cmd = [PYTHON, str(SCRIPTS_DIR / "compression_evaluator.py")] + list(args)
result = subprocess.run(cmd, capture_output=True, text=True, timeout=timeout)
return result
def test_missing_file_exits_1(self, tmp_path):
"""Test exit code 1 when file not found."""
result = self.run_script("evaluate", "/nonexistent/file.json", str(tmp_path / "c.txt"))
assert result.returncode == 1
assert "File not found" in result.stderr
def test_missing_file_error_message(self, tmp_path):
"""Test error message format for missing file."""
missing = "/this/path/does/not/exist/file.json"
result = self.run_script("evaluate", missing, str(tmp_path / "c.txt"))
assert result.returncode == 1
assert missing in result.stderr or "not found" in result.stderr.lower()
def test_invalid_json_exits_1(self, tmp_path, valid_text_file):
"""Test exit code 1 when JSON is invalid."""
bad_json = tmp_path / "bad.json"
bad_json.write_text("{invalid json content", encoding='utf-8')
result = self.run_script("evaluate", str(bad_json), valid_text_file)
assert result.returncode == 1
assert "Invalid JSON" in result.stderr or "JSON" in result.stderr
def test_valid_files_succeed(self, valid_json_file, valid_text_file):
"""Test success with valid inputs."""
result = self.run_script("evaluate", valid_json_file, valid_text_file)
assert result.returncode == 0
output = json.loads(result.stdout)
assert "compression_ratio" in output
assert "quality_score" in output
def test_generate_probes_missing_file(self):
"""Test generate-probes with missing file."""
result = self.run_script("generate-probes", "/nonexistent/context.json")
assert result.returncode == 1
assert "File not found" in result.stderr
def test_generate_probes_invalid_json(self, tmp_path):
"""Test generate-probes with invalid JSON."""
bad = tmp_path / "bad.json"
bad.write_text("not valid json {{{", encoding='utf-8')
result = self.run_script("generate-probes", str(bad))
assert result.returncode == 1
assert "Invalid JSON" in result.stderr or "JSON" in result.stderr
def test_generate_probes_success(self, valid_json_file):
"""Test generate-probes with valid file."""
result = self.run_script("generate-probes", valid_json_file)
assert result.returncode == 0
output = json.loads(result.stdout)
assert isinstance(output, list)
def test_utf8_content(self, tmp_path):
"""Test UTF-8 encoding with special characters."""
utf8_file = tmp_path / "utf8.json"
content = {"messages": [{"role": "user", "content": "日本語テスト émojis 🎉"}]}
utf8_file.write_text(json.dumps(content), encoding='utf-8')
compressed = tmp_path / "compressed.txt"
compressed.write_text("Summary with 日本語 and émojis 🎉", encoding='utf-8')
result = self.run_script("evaluate", str(utf8_file), str(compressed))
assert result.returncode == 0
@pytest.mark.skipif(os.name == 'nt', reason="Permission test not reliable on Windows")
def test_permission_denied(self, tmp_path):
"""Test permission denied error."""
protected = tmp_path / "protected.json"
protected.write_text('{"messages": []}', encoding='utf-8')
os.chmod(protected, 0o000)
try:
result = self.run_script("generate-probes", str(protected))
assert result.returncode == 1
assert "Permission denied" in result.stderr or "permission" in result.stderr.lower()
finally:
os.chmod(protected, stat.S_IRUSR | stat.S_IWUSR)
class TestContextAnalyzerEdgeCases:
"""Test edge cases in context_analyzer.py"""
@pytest.fixture
def valid_context_file(self, tmp_path):
"""Create valid context file."""
f = tmp_path / "context.json"
content = {
"messages": [
{"role": "user", "content": "implement feature X"},
{"role": "assistant", "content": "I'll help with that"}
]
}
f.write_text(json.dumps(content), encoding='utf-8')
return str(f)
def run_script(self, *args, timeout=30):
"""Run context_analyzer.py with args."""
cmd = [PYTHON, str(SCRIPTS_DIR / "context_analyzer.py")] + list(args)
result = subprocess.run(cmd, capture_output=True, text=True, timeout=timeout)
return result
def test_missing_file_exits_1(self):
"""Test exit code 1 when file not found."""
result = self.run_script("analyze", "/nonexistent/context.json")
assert result.returncode == 1
assert "File not found" in result.stderr
def test_invalid_json_exits_1(self, tmp_path):
"""Test exit code 1 when JSON is invalid."""
bad = tmp_path / "bad.json"
bad.write_text("not json", encoding='utf-8')
result = self.run_script("analyze", str(bad))
assert result.returncode == 1
assert "Invalid JSON" in result.stderr or "JSON" in result.stderr
def test_valid_file_succeeds(self, valid_context_file):
"""Test success with valid input."""
result = self.run_script("analyze", valid_context_file)
assert result.returncode == 0
output = json.loads(result.stdout)
assert "health_status" in output or "health_score" in output
def test_utf8_content(self, tmp_path):
"""Test UTF-8 encoding with international characters."""
utf8_file = tmp_path / "utf8.json"
content = {
"messages": [
{"role": "user", "content": "日本語で説明してください"},
{"role": "assistant", "content": "はい、説明します。émojis: 🎉🚀"}
]
}
utf8_file.write_text(json.dumps(content, ensure_ascii=False), encoding='utf-8')
result = self.run_script("analyze", str(utf8_file))
assert result.returncode == 0
def test_empty_messages_array(self, tmp_path):
"""Test handling of empty messages array."""
f = tmp_path / "empty.json"
f.write_text('{"messages": []}', encoding='utf-8')
result = self.run_script("analyze", str(f))
assert result.returncode == 0
def test_direct_messages_list(self, tmp_path):
"""Test handling of direct messages list (no wrapper)."""
f = tmp_path / "direct.json"
content = [
{"role": "user", "content": "hello"},
{"role": "assistant", "content": "hi"}
]
f.write_text(json.dumps(content), encoding='utf-8')
result = self.run_script("analyze", str(f))
assert result.returncode == 0
@pytest.mark.skipif(os.name == 'nt', reason="Permission test not reliable on Windows")
def test_permission_denied(self, tmp_path):
"""Test permission denied error."""
protected = tmp_path / "protected.json"
protected.write_text('{"messages": []}', encoding='utf-8')
os.chmod(protected, 0o000)
try:
result = self.run_script("analyze", str(protected))
assert result.returncode == 1
assert "Permission denied" in result.stderr or "permission" in result.stderr.lower()
finally:
os.chmod(protected, stat.S_IRUSR | stat.S_IWUSR)
def test_with_keywords_filter(self, valid_context_file):
"""Test analyze with keywords filter."""
result = self.run_script("analyze", valid_context_file, "--keywords", "feature,implement")
assert result.returncode == 0
def test_with_limit(self, valid_context_file):
"""Test analyze with limit parameter."""
result = self.run_script("analyze", valid_context_file, "--limit", "10")
assert result.returncode == 0
class TestFileSizeValidation:
"""Test file size validation (100MB limit)."""
def test_large_file_warning_in_code(self):
"""Verify MAX_FILE_SIZE_MB constant exists in scripts."""
evaluator = SCRIPTS_DIR / "compression_evaluator.py"
analyzer = SCRIPTS_DIR / "context_analyzer.py"
eval_content = evaluator.read_text()
analyzer_content = analyzer.read_text()
assert "MAX_FILE_SIZE_MB = 100" in eval_content
assert "MAX_FILE_SIZE_MB = 100" in analyzer_content
if __name__ == "__main__":
pytest.main([__file__, "-v"])