ci-cost-control

# Skill: CI Cost Control & Monitoring ## Description Guidelines for agents to minimize GitHub Actions compute waste and prevent redundant or failing CI runs. ## Rules ### 1. Check Before You Kick Before pushing code that triggers a CI workflow, you **MUST** check if a relevant run is already in progress: ```bash gh run list --branch $(git branch --show-current) --limit 5 ``` - If a run is pending/running for your current state, **DO NOT** push again unless you are fixing a specific CI failure. - Batch fixes into one push. `pull-request.yml` cancels a PR's superseded run on its own (`cancel-in-progress`), and `pr-closed-cleanup.yml` cancels a closed PR's runs, so there is nothing to cancel by hand. ### 2. Local First NEVER use CI as a "remote compiler." - You must run `./gradlew spotlessApply spotlessCheck detekt detektTypeResolved assembleDebug test allTests` locally before pushing. - If local tests fail, CI **will** fail. Do not waste the tokens or the compute. ### 3. Let the path filters do their job `pull-request.yml`'s `check-changes` filter already skips the heavy pipeline for docs-only changes, and `merge-queue.yml` skips it for docs-only merge-group entries (`docs/**`, `fastlane/**`, `obtainium/**`, `*.md`). Don't defeat that by bundling an unrelated source edit into a docs PR: a one-line README fix should not pay for the Gradle jobs. The PR's desktop matrix runs only when the `desktop` filter matches, and screenshot validation only when the `screenshots` filter does. If you add a new top-level module, add its `<root>/**` line to the `android:` filter, or the drift guard fails the PR. ## Monitoring Use `gh run view <run_id>` to inspect failures. Do not re-run a whole suite if only one shard failed due to a known flake; use `gh run rerun <run_id> --failed`.