pi-llm-rounds

Visualize pi's agent loop in the terminal: which LLM call is streaming, how many API calls one message triggered, and a separator line between rounds.

你：帮我看看这个目录
🔄 LLM 调用 #1（消息 #1 的 agent loop）          ← live, while streaming
（模型回复 + 工具调用）
──── LLM 第 1 轮 · 3.4s · ↑1.5k ↓300 tok · 2 工具 ────────────
（模型带着工具结果再回复）
──── LLM 第 2 轮 · 1.2s · ↑2.1k ↓120 tok · 0 工具 ────────────
==== 消息 #1 完成 · 共 2 次 LLM 调用 · 2 工具 · 6.1s ════════

pi's agent loop, made visible:

• One user message → one agent run (agent_start … agent_end)
• Each LLM API call inside the run → one turn (turn_start / turn_end)

Install

pi install npm:pi-llm-rounds
# or from git
pi install git:github.com/nilpo/pi-llm-rounds

What you get

• Live indicator — while streaming, the working loader shows 🔄 LLM 调用 #N; the status bar shows R1·T2 (run 1, turn 2), then ✓ 消息 #1: 2 次 LLM 调用.
• Turn separators — after every LLM call, a dim rule line in the transcript with duration, input/output tokens, and tool-call count. Expand tool output (Ctrl+X) to also see model name and cache-read tokens.
• Run summary — a ==== line after each run: total LLM calls, tools, wall time. Directly answers "how many API calls did this message trigger?"
• /llm-stats — session totals (runs, calls, tools, tokens), rebuilt from the session branch so it stays correct after resume or branch switches.

Markers are stored as custom session entries: they render in the transcript and persist across restarts, but are never sent to the LLM — zero context pollution.

Notes

• Automatic retries / post-compaction reruns start a new run and are counted honestly.
• Subagents (pi-subagents) run in their own sessions; their turns are marked in their own session files.

Uninstall

pi remove npm:pi-llm-rounds