Skip to main content

The CI/CD pipeline: push to green

A CI/CD pipeline is the automated path from a code push to a verified, deployable state. In Agentic Pipeline, that path is: webhook intake, an orchestrated run on a pre-warmed machine, your recipe's stages, and a required GitHub check — median push-to-green is about 21 seconds.

Legend: green = engine-owned hops · blue = GitHub-side · red = failure loop an agent closes over MCP.

The four stages of an Agentic Pipeline CI/CD pipeline​

  1. Webhook intake. GitHub delivers a push or PR event; the control plane verifies the HMAC signature and rejects invalid payloads.
  2. Orchestration. The orchestrator queues the run and picks a pre-warmed executor that already has your repo and dependencies.
  3. Run stages. The executor runs your recipe's stages — sequential or one-level parallel groups — each stage a named shell command with optional env, timeout, and secrets.
  4. Required check. When all stages pass, the control plane sets the RunsGreen check to green and auto-merge can proceed.

A pre-warmed executor already holds your repository checkout, installed dependencies, and cached Docker layers before the push lands. The control plane dispatches the run over the private network with no cold-start clone or install. That is why median push-to-green is about 21 seconds.

Speed is also a contract: CI and sandbox work stays under a six-minute ceiling; production deploys have a separate bounded eight-minute ceiling so healthy remote rollout work is not clipped after setup. A run that hits its applicable wall is killed and its check says so. Both bounds keep pipelines from quietly growing without globally weakening the fast PR contract. Below the ceiling sits a softer per-repo speed budget (120 seconds by default): a green run that exceeds it still passes, but its check summary says so and asks you to stabilize and accelerate — and a repo that stays over budget gets an automatic recipe-improvement proposal from the platform's advisor.

When a stage fails​

When a stage fails, the RunsGreen check goes red and merge is blocked until the run passes. A coding agent can close the loop over the MCP server: fetch a diagnosis of the failing stage, tail the log, re-run it, or test a fix in a sandbox before pushing again. You do not need to leave the agent session to inspect CI state on GitHub — and for the green path, ship and follow streams the whole chain from push to live in a single call.