分布式追踪
Rnix 提供跨多智能体工作流的分布式因果追踪，支持根因分析和上下文内存分析。
Trace ID 与 Span
每个 Compose 工作流会生成一个唯一的 Trace ID，通过 IPC 在智能体间自动传播，形成跨进程的因果链。
Trace: abc-123
├── Span: analyzer (PID 5) [0ms - 3800ms] tokens: 1450
│ ├── Open /dev/llm/claude [13ms - 14ms]
│ ├── Write /dev/llm/claude [14ms - 5214ms] ← LLM call
│ └── Read /dev/fs/main.go [5216ms - 5218ms]
├── Span: doc-gen (PID 6) [3800ms - 8000ms] tokens: 1180
│ └── ...
└── Span: checker (PID 7) [8000ms - 10500ms] tokens: 890每个智能体记录 Span（起止时间、syscall 序列、token 消耗）。Span 通过父子关系形成追踪树。
查看追踪
bash
$ rnix trace <trace-id>
Trace abc-123: Code review workflow (10.5s, 3520 tokens)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Timeline:
[0.0s ─── 3.8s] analyzer (PID 5) ████████░░░░░░░
[3.8s ─── 8.0s] doc-gen (PID 6) ████████░░░
[8.0s ── 10.5s] checker (PID 7) █████
Token distribution:
analyzer: 1450 (41%) doc-gen: 1180 (34%) checker: 890 (25%)瓶颈分析
自动识别关键路径——即瓶颈节点：
bash
$ rnix trace blame <trace-id>
Root cause analysis for trace abc-123:
Slowest path: analyzer → doc-gen → checker (10.5s)
Bottleneck: analyzer (3.8s, 36% of total)
└── LLM call at step 2 (5.2s) ← Primary bottleneck
Highest token consumer: analyzer (1450 tokens, 41%)
Recommendations:
- Consider using a faster model for analyzer (e.g., haiku)
- analyzer's context could be trimmed (cold segments: 23%)上下文内存分析器
分析任意智能体的上下文使用情况，识别浪费并预测耗尽时间：
bash
$ rnix ctx-profile <pid>
Context analysis for PID 5 (code-analyst):
Total tokens: 4,567 / budget: 8,192 (56%)
Breakdown by category:
System prompt: 1,245 tokens (27%) ████████░░
Skill bodies: 890 tokens (19%) ██████░░░░
Dialog history: 980 tokens (21%) ███████░░░
Tool results: 1,452 tokens (32%) █████████░
Activity classification:
Active (current step): 2,100 tokens
Warm (recent): 1,200 tokens
Cold (unreferenced): 800 tokens ← optimization target
Leaked: 467 tokens ← unused but not freed
Growth prediction:
Current rate: ~450 tokens/step
Estimated exhaustion: step 8 of 10
⚠ Warning: may exceed budget before completion