README.md

---
license: apache-2.0
---
Dataset: 22 Real Claude Code Sessions

To validate Suffix Decoding's applicability in Agentic Coding scenarios, we collected 22 complete Claude Code session recordings.

### Dataset Overview

| Metric | Value |
|--------|-------|
| Collection date | December 2025 |
| Total sessions | 22 |
| Total conversation turns | 17,487 |
| Total runtime | 50 hours |
| Total input tokens | 6,996,619 |
| Total output tokens | 6,094,906 |

### Session Scale Distribution

| Statistic | Min | Max | Average |
|-----------|-----|-----|---------|
| Conversation turns | 273 | 1,992 | 795 |
| Session duration | 48 min | 505 min | 136 min |
| Input tokens | 108,737 | 757,931 | 318,028 |
| Output tokens | 100,256 | 682,035 | 277,041 |

### Project Type Coverage

These 22 sessions cover **15 different types** of software development tasks:

| Project Type | Call Count | Project Type | Call Count |
|--------------|------------|--------------|------------|
| Instant Messaging (im) | 1,191 | Cloud Storage (netdisk) | 518 |
| Game (mario) | 796 | Travel (travel) | 444 |
| Finance (stock) | 731 | Music (music) | 439 |
| Utility (calculator) | 645 | Game (snake) | 426 |
| Ticketing (ticket) | 559 | Reminder (reminder) | 402 |
| Memo (memo) | 534 | Social (inlove) | 381 |
| Fitness (fitness) | 522 | Video (video) | 350 |

### Agent Architecture

Claude Code employs a multi-agent collaboration architecture:

| Agent Type | Percentage | Responsibility |
|------------|------------|----------------|
| Main Agent | 73.6% | Primary control, task decomposition and execution |
| Explore Agent | 26.2% | Code exploration, file search |
| Plan Agent | 0.1% | Architecture design, implementation planning |

### Agentic Behavior Patterns

Analysis of key phrases in response text:

| Behavior Pattern | Occurrences | Description |
|------------------|-------------|-------------|
| "Let me..." | 3,729 | Proactive task execution |
| "Now let me..." | 2,817 | Step transitions |
| Test-related | 2,807 | Running tests, validating results |
| Create file/code | 2,795 | Generating new code |
| Update/modify code | 1,267 | Iterative improvements |
| Error fix related | 1,196 | Self-correction |


This dataset has been tested with Novita.ai's Suffix decoding via sglang. While originally structured using Anthropic's protocols, it has been converted to OpenAI's format to enhance open-source compatibility. The following commands can be used to test the dataset with custom endpoints.

```
python concurrent_session_test.py \
      --input e22_sessions_openai.json \
      --num-sessions 22 \
      --selection-mode first \
      --api-url http://127.0.0.1:8006/v1/chat/completions \
      --api-key YOUR_KEY \
      --model YOUR_MODEL \
      --provider "Novita-GLM4" \
      --skip-first-turns 40 \
      --warmup-turns 5 \
      --cooldown-turns 5 \
      --max-concurrent 22 \
      --min-concurrent 10 \
      --max-turns 40 \
      --output benchmark_results/e22_sessions_test.json \
      --generate-charts \
      --chart-format both \
      --min-output-tokens 16 \
      --show-content-threshold 100000
```