fixing-ci

Fixing CI

Overview

This skill drives an iterative CI fix session: investigate the failure (via the Buildkite investigation skill), apply a fix, verify it locally, push, watch the new build, and iterate until green or until you've hit the iteration cap and need to step back.

The scope is the fix loop only — investigation is delegated to buildkite:investigating-builds, which already covers bktide snapshot, log reading, and failure-pattern recognition.

When to Use

• You have a failing CI build and want to push fixes until it goes green
• You want to verify changes locally before each push (saves CI cycles)
• You're iterating on fixes across multiple builds and need to know when to step back

When NOT to Use

• First-time "why did this fail" investigation with no fix-and-push intent → use buildkite:investigating-builds directly
• Authoring or modifying pipeline YAML → use buildkite:developing-pipelines
• GitHub Actions, CircleCI, or non-Buildkite CI

Input Contract

The skill needs to know what to fix. Any one of these inputs is sufficient:

• build_url — A Buildkite build URL (e.g. https://buildkite.com/org/pipeline/builds/123)
• pr — A GitHub PR number (resolves to its latest failing build for that branch)
• branch — A git branch name (resolves to its open PR's latest failing build)

Resolving inputs

If the caller provides one of the above, use it directly. Otherwise:

1. From cwd: read the current git branch (git branch --show-current) and look for an open PR (gh pr view --json number,headRefName). Then find the branch's latest failing build:
   • Work out the pipeline slug. It usually matches the repo name; a monorepo may name it after the main repo.
   • List the branch's builds with MCP buildkite:list_builds (branch: <branch>, detail_level: "summary"), or npx bktide@latest builds --format json --branch <branch> <org>/<pipeline>. Without --branch, bktide builds lists every branch, so filter on branch yourself.
   • On the newest build, check in order: overall state, job summary (passed/failed/broken), annotations, then logs for the failed jobs.
2. If cwd doesn't resolve cleanly (no branch, no open PR, no failing build): the caller (slash-command wrapper or workflow agent) is responsible for either asking the user or failing with a clear message. This skill assumes resolution is done before its loop starts.