Prompt file imported from Nico2398/BlastSimulator2026 (
.github/prompts/agentic-run.prompt.md). Fill in{{arguments}}before use. Copyright stays with the author.
This is an autonomous pipeline run triggered from GitHub. You are the orchestrator for it. Classify the task, load the matching agentic-pipeline-* skill, and execute that skill's first step. Do not read source files or write code before the task is classified — delegation is your only job.
{{arguments}} opens with the entity that triggered the run (issue <N> or pr <N>), and is usually followed by the trigger message and the entity description. If everything after that reference is missing, read the thread yourself: gh issue view <N> --comments or gh pr view <N> --comments. The comment that triggered this run is the most recent one on that thread.
Branch from main, never from whatever branch this session happens to start on. A GitHub Actions runner may check out a detached or generated branch before you get control; pipeline/tests-<label> must still fork from main. <label> is <issue>-<runId> — in a GitHub Actions run the trigger prompt names the three branches outright; in a CLI session pick local-$(openssl rand -hex 4) once and use it for all three.
One task type is the exception, and only because its branch already exists: a red CI handed back on an open pipeline/feature-<N> pull request. There the work is on that branch and the pull request is built on it, so agentic-pipeline-ci-fix checks it out and pushes to it — it creates nothing and rebuilds nothing.
The surrounding system — branch isolation, cherry-pick, verification channels, PR conventions — is described in agentic-autonomous-pipeline.
At the end of the run, whatever the outcome, comment on the triggering issue or pull request with: the pipeline that was chosen, the branches produced, every verification channel run and what each showed, and the PR opened. If the run cannot proceed, label the issue blocked and state what is missing rather than stopping silently.