pi-prompt-tracker

Track complete prompts sent to the LLM gateway — system prompt, tools, messages, responses, and usage — for debugging, caching analysis, and cost optimization.

Install

pi install git:github.com/<your-username>/pi-prompt-tracker

Or from a local path:

pi install /path/to/pi-prompt-tracker

Or try it once without installing:

pi -e /path/to/pi-prompt-tracker

What It Logs

For every LLM request, writes a structured Markdown log file containing:

| Data | Detail | |------|--------| | System prompt | Full system prompt text sent to the model | | Tool definitions | Every tool's name, description, and JSON schema | | All messages | User, assistant (text + tool calls + thinking), tool results | | Full provider payload | Raw JSON of the complete request sent to the API | | Response status | HTTP status code and response headers | | Assistant response | Final output with token usage and cost |

Log Files

Written to ~/.pi/agent/prompt-logs/ by default (configurable in source).

~/.pi/agent/prompt-logs/
├── prompt-log_2026-07-02--20-31-12_session-c0983527.md
├── prompt-log_2026-07-03--10-15-00_session-a1b2c3d4.md
└── ...

Each session gets one log file. Entries are appended as requests/responses happen.

Commands

| Command | Description | |---------|-------------| | /prompt-tracker-log | Open the latest log file in your editor | | /prompt-tracker-dir | Show the log directory path and file count | | /prompt-tracker-clear | Clear all logs (with confirmation) |

Use Cases

• Debug what's actually sent — verify system prompt, tools, and message ordering
• Caching analysis — see which parts of the prompt are static vs dynamic to optimize prompt caching (Anthropic/OpenAI)
• Cost tracking — monitor token usage and cost per request
• Prompt engineering — inspect the final payload to refine system prompts and tool descriptions
• Provider comparison — log the exact payload format for different providers

Configuration

Edit the constants at the top of extensions/prompt-tracker.ts:

const LOG_DIR = join(homedir(), ".pi", "agent", "prompt-logs"); // Change log directory
const SHOW_NOTIFICATIONS = true;                                   // Toggle startup notification

Example Snippet

## Request #3 — Jul 02, 2026, 20:35:12

**Model:** deepseek-v4-flash-free

### System Prompt
\`\`\`
You are an expert coding assistant operating inside pi...
\`\`\`

### Tool Definitions
### 1. read
- **Description:** Read the contents of a file...
- **Input schema:** {...}

### Messages (42 total)
### user
can we create pi extension...

### assistant
Let me read the docs...

### Full Provider Payload
\`\`\`json
{ "model": "...", "messages": [...], "tools": [...] }
\`\`\`

Disclaimer

This extension was generated by Pi Coding Agent through vibe coding. It logs every prompt sent to the LLM, which may include sensitive data. Use with caution and review the code before deploying in production.

Requirements

• Pi coding agent
• No external npm dependencies (uses only pi's built-in packages and Node.js built-ins)

License

MIT