Pi Retro

Generate a coaching retrospective report for your current Pi session. /retro analyzes your conversation, extracts actionable insights, and produces an HTML report that helps you collaborate better with AI agents.

How It Helps

• Learn from your session — See what went well, what could be improved, and what to avoid next time
• Better prompts — Get concrete examples of how to rephrase your prompts for clearer results
• Understand agent behavior — Learn how the agent interpreted your requests and where corrections were needed
• Actionable takeaways — Get specific, personalized advice you can apply to future sessions

Install

pi install npm:@petechu/pi-retro
/reload

Usage

Run /retro at any point during a session:

/retro

The extension waits for the agent to become idle, then sends a hidden retrospective request through the active Pi agent. The extension does not call provider APIs directly.

The Report

The report is saved as an HTML file and automatically opened in your browser:

pi-retro-2026-06-05T12-30-00.html

By default, the report is saved to the git repository root if available, otherwise to the current working directory.

What's Inside

Each report includes:

• Session metrics — turns, duration, tool calls, corrections detected
• Session summary — a concise overview of what happened
• Timeline — key moments and milestones
• What went well — effective approaches worth repeating
• What could be improved — areas where communication or approach could be clearer
• What not to do — patterns to avoid
• Better prompts — concrete examples with original/improved/why
• Agent behavior notes — insights into how the agent interpreted your requests
• Actionable takeaways — personalized advice for next time
• Detailed breakdowns — tool calls by name, files read/edited/written, verification commands

Design & Workflow

The extension has three layers: collection, analysis, and reporting.

Collection

When /retro runs, the extension walks the current session branch and extracts:

• Conversation transcript — user prompts, assistant responses, with truncation for long sessions
• Tool call metrics — which tools were called, how often, and which failed
• File operations — which files were read, edited, or written
• Verification commands — test, lint, and typecheck commands that ran
• Correction signals — moments where the user clarified or redirected the agent

Analysis

The extension sends a hidden prompt to the active Pi agent with:

• The condensed transcript (truncated at ~40k characters for long sessions)
• Structured metrics as JSON
• A system prompt that asks the agent to analyze the collaboration from the user's perspective

The agent is instructed to:

• Focus on how well it interpreted user intent
• Identify where user corrections were needed
• Provide specific, actionable feedback
• Generate concrete prompt rewrites with explanations
• Call retro_save_report exactly once as its final action

Reporting

When the agent calls retro_save_report, the extension:

1. Renders the analysis as a styled HTML report
2. Saves it to the git root or current working directory
3. Opens the report in your browser (or shows a notification if this fails)

The HTML report includes:

• Light/dark theme toggle with system preference detection
• Responsive layout that works on mobile and desktop
• Print-friendly styles for archiving
• All content stored locally — no external requests

Privacy

Reports may contain excerpts from your session, including:

• User prompts and assistant responses
• File paths and tool arguments
• Commands and their outputs
• Agent behavior observations

Review the report before sharing or committing. The extension saves reports locally and does not transmit them anywhere.

Notes

• The transcript is truncated at ~40,000 characters for long sessions — this preserves early turns and errors, but mid-session content may be summarized or omitted
• Correction detection uses heuristics — phrases like "actually," "I meant," "not what I wanted" — and may miss some corrections or flag false positives
• The /retro request itself is persisted in the session context; future versions may isolate this to avoid polluting the main conversation history