feat: integrate Entire (entire.io) to log AI workflows with consent-gated upload

[qyli00]

Summary

Integrates Entire into the CLI to capture AI agent sessions (prompts, responses, tool calls, files touched) as users solve problems, and — only with the user's consent — uploads them for research. Capture is local-only until consent; uploaded data is used solely for research and is de-identified/anonymized before use.

Logging is set up at fetch/download time and uploaded at submit time. Everything degrades gracefully: if the entire binary isn't installed, the core fetch/submit flow is never blocked — instead the user is pointed at aicodinggym configure (which offers to install it).

How it works

• configure — offers to install the Entire CLI, records the writable submission-repo URL, and accepts --upload-logs/--no-upload-logs to pre-set consent.

• swe fetch / cr fetch / mle download — install Entire's hooks so the session is captured locally as the user works. MLE workspaces are git init'd (they aren't repos by default).

• swe submit / cr submit / mle submit — consent-gated upload. The first submit (if not already configured) prompts once:

  AI Coding Gym can upload this AI coding session (prompts, responses, files changed) for research only. Data is de-identified/anonymized before use.

  The choice is saved to ~/.aicodinggym/config.json. Non-interactive sessions never upload without a recorded choice.

Log identification & target

• One repo, many branches: all three benchmarks log to the user's single writable repo (recorded from SWE fetch / configure), so each upload is identifiable by its branch even when one repo holds many problems.
• Logs land on aicodinggym-logs/<benchmark>/<problem_id> with an aicodinggym-meta.json file at the tip (problem id, benchmark, user, tool, timestamp), injected via git plumbing without touching the working tree or Entire's branch.
• CR's cloned PR repo is read-only, so CR/MLE logs target the user's own repo, never the review repo.

MLE code push

mle submit also pushes the user's solution code (notebooks/scripts/CSV; data/ excluded) to a branch named after the competition (e.g. spaceship-titanic), gated by the same log-upload consent. The prediction CSV still goes to the scoring API as before.

Files

• entire_logging.py (new) — thin best-effort wrapper around the entire binary: setup, ensure_git_repo, commit_workspace + push_branch (MLE), flush (CR), has_sessions, upload.
• config.py — persist entire_logging_consent and submission_repo_url; get_logging_consent/set_logging_consent.
• cli.py — wire logging into configure/fetch/download/submit; consent + upload helpers.
• README.md — document the feature, consent flow, unified repo, MLE code push, privacy.
• Version bump 0.5.1 → 0.6.0.

Testing

• All modules byte-compile; every command's --help exits 0.
• SWE: upload pushes to the per-problem branch with aicodinggym-meta.json injected and session data carried over.
• MLE: ensure_git_repo → commit_workspace → push_branch pushes solution files to the competition branch with data/ excluded.
• Config: consent + submission_repo_url round-trip through the field allowlist.

Notes for reviewers

• CR/MLE upload depends on submission_repo_url, which is auto-recorded after any SWE fetch, or from configure if the backend /api/configure response includes a repo_url (it currently returns only repo_name). Until then a CR/MLE-only user gets a graceful "no repo configured — pass --logs-remote" note instead of an upload.
• The consent prompt centers on the AI session (the privacy-sensitive part) but also governs the MLE code push (per design). Easy to broaden the wording if preferred.

🤖 Generated with Claude Code