Skip to content

Roadmap

Work lands in small pull requests, one concern each, tracked in havelin-labs/chakshu.

Done (0.1)

  • Scaffold, Postgres schema, deterministic IDs, encrypted settings
  • React portal in the Havelin design system
  • GitHub App setup through the manifest flow
  • Sign in with GitHub, org roles and admins
  • Reviewer catalog synced from each org's reviewers repo, plus a template repo
  • Webhooks: mentions, review requests, automatic runs, reviewers-repo sync
  • Queue and worker on Postgres
  • Review engine on the Claude Agent SDK with read-only tools
  • Publisher: inline comments, suggestions, no repeats, one check run per reviewer
  • Portal pages for reviews, repositories and settings
  • Release workflow with signed multi-arch images

Done (0.2)

  • Webhooks stored in Postgres and handled by the worker, with retries and a failures list
  • Graceful worker shutdown and heartbeat-based recovery
  • Monthly spend cap
  • Later reviews look only at commits added since the last one
  • Chakshu resolves its own outdated threads
  • Insights: per-skill fixed and dismissed rates
  • Jira connector

Done (0.3)

  • Reviews run in GitHub Actions in the reviewed repository, with logs in its Actions tab
  • Repository setup from the portal: workflow pull request and Actions secrets
  • Missing app permissions listed in Settings
  • Call any reviewer by name
  • Create and edit reviewers in the portal, through pull requests
  • Chakshu answers replies to its comments
  • The chakshu team is created during setup
  • A Cloudflare tunnel profile for laptops
  • Analytics for owners (usage, outcomes and spend by repository, reviewer, skill and person) and My activity for everyone

Next

  1. Tag a release, so the workflow can pin a version of the action.
  2. Deploy to a server and publish these docs at chakshu.havelin.in/docs.

Planned

  • A People directory: a card for every member and a page per person (#89)
  • Analytics: teams, spend alerts and a weekly digest, token counts, review speed, and retention controls (#84 to #88)

Later

  • More connectors: Linear, Confluence, Sentry
  • Sandboxed scripts from skills, such as linters or schema checks
  • Other engines (OpenAI, Copilot SDK)
  • GitLab and Bitbucket