Skip to content

What Sentinel does

Sentinel watches a Sentry project. When a new or regressed issue fires, it posts an alert card to Microsoft Teams and then hands the issue to an unattended Claude Code session, which checks the repository out at the release the error came from and works out what went wrong. The findings come back as a reply in the same Teams thread. If a finding is confident enough, that same session goes on to write the fix and open a pull request.

A new Sentry issue arrives with a stack trace, a release, and nobody who has looked at it yet. Someone has to open it, find the commit, read the surrounding code, and decide whether it matters. That work is mechanical for the first several minutes and it is the reason issues sit unread.

Sentinel does those first several minutes automatically, and does them at the exact commit the error came from rather than at whatever the default branch happens to be today.

  1. Alert. Sentry webhooks the issue to a receiver, which verifies the signature, renders an Adaptive Card, and posts it to a Teams channel through its own bot identity.
  2. Gate and enqueue. Alerts that are warning level or above (or an N+1 performance issue), in an investigated environment, and carry a release that resolves to a commit SHA are queued with a debounce, so a burst of related issues becomes one batch rather than a stampede of sessions.
  3. Investigate. A scheduled sweep fires one unattended session per batch. It checks out the target repository at the release SHA, investigates each issue, and posts a findings document back, which the receiver renders as a threaded reply.
  4. Autofix. A finding whose confidence and fixability both clear their configured minimums, and whose intended behavior is already decided somewhere it names, earns a grant, returned in the reply to that findings document. The same session writes the fix and its test and sends the change back as a patch; the receiver opens the pull request as the GitHub App.

See How it works for the same pipeline with its gates and failure handling, or See it work for the artifacts it produces.

Nothing merges automatically. Every pull request waits for a human reviewer. The session that writes a fix has no way to write to GitHub, and the credential the receiver uses to open the pull request cannot reach past a single repository, two permissions, and an hour. It notably cannot push a workflow file, and the receiver refuses to carry one.

Both of the aggressive behaviours ship switched off. A new deployment starts in shadow mode, which runs the whole gate and records what it would have fired without firing anything, and the autofix gate ships enabled: false. Turning either on is a config change and a deploy, which is deliberate: it means nobody enables unattended pull requests by accident.

Sentinel is infrastructure, not a plugin. Running it means bringing five things you already own or are willing to create:

  • a Sentry organization, with permission to create an internal integration
  • a Microsoft 365 tenant with Azure, for the app registration and the bot
  • an AWS account, for the receiver
  • a GitHub App, only if you want autofix
  • Claude Code access, for the unattended sessions

What you need first walks that list before you start, and the security model covers what the unattended sessions can and cannot touch.