📝 What I Did — GitHub Copilot Impact Report

One command. See everything you built with Copilot — and the leverage you're getting from your seat.

🚀 All Analytics Hub Reports

Found this useful? ⭐ Star this repo to help others discover it!

Get Started ↓  ·  What You'll See ↓  ·  How It Works ↓  ·  Requirements ↓  ·  Privacy ↓

The fastest way is the VS Code extension — one click, no terminal. Prefer the command line? The Copilot CLI plugin and a plain clone work exactly the same way.

1. Open the Extensions view (Ctrl+Shift+X / Cmd+Shift+X), search What I Did, and click Install (or install the bundled what-i-did-copilot-*.vsix via Extensions: Install from VSIX…).
2. Click ✦ What I Did in the status bar, or run What I Did: Open Report Builder from the Command Palette.
3. Pick a time range, flip any toggles, and click Generate my report. The finished report opens right inside VS Code.

Option A — Copilot CLI plugin (requires the GitHub Copilot CLI)

In any Copilot CLI session:

/plugin install whatidid@awesome-copilot

Then just run:

whatidid                # defaults to a 7-day lookback

Option B — VS Code chat plugin (Awesome Copilot marketplace)

1. Open the Command Palette (Ctrl+Shift+P / Cmd+Shift+P) → Chat: Plugins
   (or type @agentPlugins in the Extensions search view)
2. Search whatidid → click Install
3. In Copilot Chat, just ask:

   "give me a 7-day Copilot report"  ·  "whatidid this month"  ·  "summarize what Copilot helped me with today"

💡 Power user tip: the plugin is a full Python checkout under
~/.copilot/installed-plugins/awesome-copilot/whatidid/ (Windows: %USERPROFILE%\.copilot\...).
cd there and run python whatidid.py --30D directly. Note this is the marketplace-pinned release — run copilot plugin update whatidid@awesome-copilot to refresh, or use Option C for true latest-main.

Option C — Clone the repo (maximum control, dev-mode)

Open a terminal (Windows Terminal, Terminal.app, or your favourite shell — including VS Code's built-in terminal), then:

gh auth login                                           # one-time: authenticate to GitHub (enables AI analysis)
git clone https://github.com/microsoft/What-I-Did-Copilot.git
cd What-I-Did-Copilot
python whatidid.py                                      # defaults to a 7-day lookback

Prerequisites: Python 3.10+, git, and GitHub CLI (gh). See the full requirements table below.

That's it. A report opens in your browser showing your last 7 days with Copilot.

📊 A note on the numbers. Credit and cost figures are estimates, calculated from the token counts in your local session logs and GitHub's published per-model rates. They give an accurate picture of the shape of your AI usage, but your actual GitHub bill can differ depending on your plan, included credit allowance, and billing details that aren't visible in local logs.

• Learning intent expansion — the classifier now catches genuine learning queries: "how do I X", "help me understand", "what's the best way to...", "tell me about", "pros and cons of...", "primer on...", plus dozens more phrasings. A hand-holding bypass routes "I don't understand how X works" to Learning instead of Course-correcting.
• New methodology doc — docs/credit-optimization.md synthesizes credit-burn pattern the tool detects with the exact Anthropic / OpenAI / GitHub source it implements, including short excerpts.
• AI Investment and Consumption Patterns — new sub-segments for Model Mix, Top Sessions and behavior pattern analysis that drive up consumption, based on guidance from OpenAI, Anthropic, and GitHub.

whatidid --14D                        # last 14 days
whatidid --30D                        # last 30 days
whatidid --date 2026-03-19            # specific date
whatidid --from 2026-03-01 --to 2026-03-31   # date range
whatidid --7D --email                 # send via Outlook
whatidid --7D --email you@company.com # send to a specific address
whatidid --refresh                    # force re-analysis

~/.copilot/session-state/<uuid>/events.jsonl          (Copilot CLI sessions)
<AppData>/Code/User/globalStorage/                    (VS Code Copilot Chat
   emptyWindowChatSessions/<uuid>.jsonl                sessions — same pipeline)
                │
                ▼
           harvest.py    → scan sessions, extract messages, tools, files, intents
                │
                ▼
           analyze.py    → AI categorization via GitHub Models API (gpt-4o-mini)
                │         → calibrated effort estimation with quantitative signals
                ▼
           report.py     → HTML report: story arc, donut charts, heatmaps, ROI
                │
                ▼
         whatidid.py     → opens report in browser; --email sends via Outlook COM

See docs/architecture.md for session file formats, token cost model, and leverage calculation details.

See docs/effort-estimation-methodology.md for the research basis, signal definitions, and calibration logic behind effort estimates — grounded in 13 peer-reviewed sources including Alaswad et al. 2026, Cambon et al. 2023 (Microsoft Research), Ziegler et al. 2024 (CACM), and the SPACE framework (Forsgren et al. 2021).

See docs/credit-optimization.md for the credit-burn pattern catalogue — every detector is tied to published guidance from Anthropic, OpenAI, or GitHub, with the per-model rates sourced from Copilot's own session metadata when available.

Your data stays on your machine. This tool is completely local-first:

• Reads only local files — session logs from ~/.copilot/session-state/ (Copilot CLI) and <AppData>/Code/User/globalStorage/emptyWindowChatSessions/ (VS Code Copilot Chat) that already exist on your machine
• No telemetry, no tracking, no cloud uploads — the tool never phones home
• AI analysis is optional — uses GitHub Models API (authenticated via your own gh CLI token) to semantically interpret sessions. Without API access, a local heuristic fallback produces estimates with zero external calls
• Email is optional — the --email flag sends the report via your own Outlook client. If you don't use it, the HTML file stays on disk
• No one has access to your report unless you share it — the output is a standalone HTML file saved to your local project directory

The tool processes the same session data that GitHub Copilot already stores locally. It adds nothing new to disk beyond the HTML report and a small analysis cache in cache/.

No pip install needed — the core report generator (harvest.py, analyze.py, report.py, whatidid.py) uses only the Python standard library + GitHub Models API.

This tool ships as a Copilot CLI agent. Anyone who clones the repo gets it automatically — run /agent in Copilot CLI and select whatidid, or just ask naturally:

"What did I build this week?"

See .github/agents/whatidid.agent.md for the agent definition.

MIT

_{Keywords: GitHub Copilot ROI, Copilot usage report, Copilot activity tracker, AI productivity metrics, token usage analysis, Copilot impact measurement, developer productivity, AI-assisted development analytics}