feat: initial import

[jm-cnaught]

Summary

Single-commit initial import of the AI coding agent adoption study. Repo is private; will be flipped to public after final review.

The previous internal working repo (CNaught-Inc/coding-agent-adoption) stays private for ongoing internal work — this is a clean snapshot built from the working tree of that repo's chore/public-release-prep branch, with no history bleed (no prior commits, no deleted files reachable via git log).

What's in this repo

Pipeline (9 scripts):

• github_ai_daily.py — daily commit-attribution fetcher (GitHub Search API).
• fetch_branch_activity.py, fetch_branch_creates.py, fetch_daily_totals.py — BigQuery fetchers for GH Archive signals.
• estimate_carbon.py — Jegham-framework carbon/energy estimation.
• dashboard.py — Plotly HTML dashboard.
• run_pipeline.py — orchestrator (daily/weekly modes, Slack alerting, git commit-back).
• anomaly_analysis.py — signature-drift scanner.
• build_carbon_workbook.py — Excel QA workbook.

Intervention analysis (10 modules):

• Primary estimators: event-window CAR, BEAST changepoint detection, first-difference ITS.
• Robustness: BSTS, synthetic DiD, compositional VAR, 3-param logistic fits.
• Worked example: Oct 20, 2025 Claude Code web launch.
• Vendored synthdid (PyPI build broken on Python 3.14).

Data (7 CSVs, ~2 MB):

• ~450 days of commit data across 32 tools.
• ~392 days of BigQuery branch/push signals.
• Bot-donor pool for SDID.
• Derived carbon estimates.

Documentation:

• README.md — three no-credential quickstarts + full credentialed setup.
• AGENTS.md — operating manual for AI coding agents (file map, autonomous walkthrough, walk-the-user-through-it mode, token-availability cheat sheet, guardrails, footguns).
• METHODOLOGY.md — full methodology, signal definitions, caveats.
• CONTRIBUTING.md, LICENSE (Apache 2.0), .env.example.
• references/ — 9 cited academic papers.

Tests: 95-test pytest suite, runs in ~4s, GitHub Actions CI configured.

Test plan

• python3 -m pytest tests/ -v — 95 passed
• python3 dashboard.py builds the dashboard from shipped CSVs (464 days, 30 tools, ~1.6 MB output)
• python3 run_intervention_analysis.py produces outputs/intervention/ with 12 artifacts; headline CAR = 60,067 commits, z = 5.59
• Secret scan against all tracked files — clean
• Dogfood test passed: a fresh subagent constrained to only the public docs successfully built the dashboard, ran the intervention analysis, and ran the full test suite

Before flipping to public

• Decide whether to leave the repo name as-is (coding-agent-emissions) or pick something else.
• Add GH_SEARCH_TOKEN, GCP_SA_KEY, optional SLACK_WEBHOOK_URL as repo secrets if scheduled CI should run after publication.
• Verify the .github/workflows/ schedules are in the state you want before public.
• Flip visibility: Settings → General → Danger Zone.

🤖 Generated with Claude Code