pi-http-sniff

By Jaylen Douglas (jaylendouglas.com)

A Pi Coding Agent extension that sniffs and logs HTTP traffic to/from LLM providers during chat sessions, with configurable model filtering and output formatting.

Features

• Logs full request and response payloads for LLM API calls
• Filters by model — log all models or only specific ones
• Pretty-print or compact JSON output
• Session lifecycle tracking — start, shutdown, requests, and responses
• Queue-based request/response correlation per model for better latency tracking under concurrent requests
• Non-blocking async log writes that preserve event order
• Session summary includes token totals, cache totals, estimated cost, pretty mode, and model filter
• Persistent config stored in ~/.pi/pi-http-sniff.json
• UI notifications for config changes and errors

What it logs

Event	Description
session_start	Session initialization with metadata
session_shutdown	Session termination event
before_provider_request	Full request payload sent to the LLM provider
after_provider_response	Provider response metadata (status and headers) captured before stream consumption
message_end	Assistant message with actual token counts, timing, and costs

Installation

# Clone and build
git clone <repo-url>
cd pi-http-sniff
pnpm install
pnpm build

Usage

Command

Register a session command via httpsniff:

httpsniff [modelName|all] [pretty] | pretty|summary|stats|help

Argument	Description	Example
modelName	Specific model ID/name to filter, or all	httpsniff gpt-4o
pretty	When used with a model, enable pretty output; when used alone, toggle pretty mode	httpsniff pretty
summary / stats	Show current session token, cache, cost, pretty mode, and filter summary	httpsniff summary
help	Show the built-in usage text	httpsniff help

Examples

httpsniff all                      # Log all models, compact JSON
httpsniff gpt-4o                   # Log only gpt-4o, compact JSON
httpsniff pretty                   # Toggle pretty mode on/off
httpsniff all pretty               # Log all models, pretty-printed
httpsniff claude-sonnet-4-20250514 pretty  # Log specific model, pretty-printed
httpsniff summary                  # Show session summary
httpsniff stats                    # Show session summary
httpsniff help                     # Show usage text

Config file

Configuration is saved to ~/.pi/pi-http-sniff.json and managed via the httpsniff command:

{
  "modelFilter": "all",
  "prettyPrint": false
}

Log output

Logs are written to ~/.pi/logs/pi-http-sniff-{sessionId}.jsonl.

• Compact mode (prettyPrint: false) writes one JSON object per line (JSONL-compatible).
• Pretty mode (prettyPrint: true) writes human-readable multiline JSON blocks.
• after_provider_response events are always logged. The hook payload currently includes status and headers but no model identifier, so strict per-model filtering is not available for that specific event.
• When pretty is used on its own, it toggles the persisted prettyPrint setting and displays a warning that pretty logs increase file size.

Use cases

• Debugging provider issues (what's actually being sent/received)
• Understanding prompt construction and token usage
• Auditing API traffic
• Monitoring multi-model sessions
• Latency debugging — response_time_ms in enriched logs
• Cost tracking — actual input_tokens, output_tokens, and cost_usd per request

Session Summary

View a real-time summary of the current session's token usage and costs:

httpsniff summary

🔍 pi-http-sniff Session Summary
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  Requests:          5
  Input tokens:      12,450
  Output tokens:     3,210
  Total tokens:      15,660
  Cache read:        8,192
  Cache write:       512
  Estimated cost:    $0.003240

The summary is also available via the stats alias: httpsniff stats.

Log Enrichment

All logged events include a sniff_enriched field with additional metadata:

before_provider_request enrichment

Field	Description
request_time	Epoch milliseconds when the request was sent
request_time_iso	ISO 8601 timestamp of the request

message_end enrichment

Field	Description
model	Model ID (e.g. gpt-4o, claude-sonnet-4-20250514)
stop_reason	Why generation stopped (stop, length, toolUse, error, aborted)
response_time_ms	Time from request send to first token received
input_tokens	Actual input tokens consumed
output_tokens	Actual output tokens generated
total_tokens	Sum of input + output tokens
cache_read_tokens	Tokens read from prompt cache
cache_write_tokens	Tokens written to prompt cache
cost_usd	Estimated cost in USD for this request

Development

# Build
pnpm build

# Lint
pnpm lint

# Format
pnpm format

# Auto-fix lint issues
pnpm fix