Clear stuck pull requests and failed GitHub builds from chat

The agent works with GitHub through its API: repositories, issues, pull requests with their comments and reviews, commits and GitHub Actions runs. On top of that it computes what the interface does not show at a glance: which PRs have been waiting for review the longest and who is worth a nudge, which commit and which step the build broke on, what was opened and closed in a repository over the day.

For an Actions run the status and the conclusion are different things: a workflow completion event means the run finished, not that it passed, so the agent looks at the outcome separately and does not count a failed build as green. Events follow the same logic: pull_request on its own says nothing, the meaning is in the action — opened, reopened, closed, updated with a new commit or a review requested.

Events arrive over a webhook signed with HMAC-SHA256: without a shared secret the delivery is rejected, and a repeat delivery of the same event is dropped by its delivery identifier. That is why there is no need to poll the API in circles. The agent is careful about writing: a comment on a PR, a new issue, a re-run of a failed workflow or closing a task are separate actions on your explicit request, not a side effect of the analysis.

The agent's visibility is bounded by the app installation: a repository outside it does not exist for the agent, and instead of an empty digest it says that access there was never granted. History is read page by page and one repository at a time rather than a whole organisation at once, and the request rate is kept within GitHub limits — so a large report takes longer to assemble than a quick question about one pull request. The agent has exactly the rights the installation has.