[mountainowl]

Requested by Ajay Singh · project thread

What changed

• Before: the Action keeps its state in ${{ runner.temp }}/bubo, which starts empty on every run. Every run therefore created a new random install_id and new project IDs. One repository using the Action looked like one install per run, which inflated install counts and pushed retention towards zero.
• After: the Action restores var/state/install_id and var/state/analytics_projects.sqlite with actions/cache before Bubo starts. The cache is keyed by github.repository_id, so a repository's runs share one anonymous install.
• How:
  • The new step is pinned by commit (actions/cache v6.1.0), like the other actions.
  • The cached files hold only random IDs. Nothing about the repository is sent, and the cache key never leaves GitHub.
  • GitHub's cache scoping still applies: PR runs read caches from their own PR and from the base branch. A repository's PRs share one id once a run on the default branch has saved the cache.

Verification

• tests/test_action_install_id.py runs the real bubo init and the real identity code in two fresh roots. It carries over only the paths listed in the Action's cache step.
• On main there's no cache step, and two runs give two install ids. With this change both runs share one id. The test also checks that the cache step runs before Initialize Bubo workspace, is keyed by repository ID, and is pinned by SHA.
• The layout tests pass, and ruff check, ruff format --check and cz check pass.
• docs/pages/github-action.mdx lists the new step.

---

Generated by Claude Code