Starlight Intelligence
Explore

WorkflowsCrew patternPlan, relay, evaluate

Build something useful

Hand the crew a public codebase and get back working files, checked against what you asked for.

  • RECORDED: A saved run of this pattern replays on the homepage and in the mission studio. It is not a live run. One saved run of this pattern, a diagnostics workflow for a public Node.js repository, replays on the homepage and opens in the mission studio. It is a recording, not a live run.

The run, step by step

Build something useful: Astra, then Atlas, then Orion, then Rivet, then Vera, then Hand-off gate, then Files handed offSwipe to follow the steps. Drawn to show the order of the steps; nothing is running.

Steps

Who does what, in order.

Starts when: You have a public repository and a clear goal for it.

  1. Plan

    Astra, Mission Strategy Lead

    Reads the goal, splits it into steps and hands each one to the right specialist.

  2. Make

    Atlas, Research Source Discovery Specialist

    Investigates the repository and records what it found and where.

  3. Make

    Orion, Mission Planning Specialist

    Designs the change: which checks belong in the workflow and why.

  4. Make

    Rivet, Software Build Engineer

    Builds the files and shows exactly what changed.

  5. Check

    Vera, Decision Evidence Reviewer

    Evaluates the result against the brief and can send the work back.

  6. Your yes

    Hand-off gate

    You review the files before anything leaves the run. In the recorded replay this approval is staged: the original run had no gate.

  7. Publish

    Files handed off

    The files, each with its hash, ready for you to review before you run them.

What it hands you, and what stops it

Exit proof

In the recording: five generated files with hashes and a historical evaluation. The regenerated files still need their own review before use.

Kept with the run

Receipts, sources and the approval decision stay with the run.

Acceptance checks

  • Run code only in a scoped, ephemeral environment with explicit limits.
  • Bind approval to the exact artifact and action; changed content requires new approval.

If it fails

As designed, a failed step stops the run, keeps its receipts and tells you which step to retry.