Orrery.

For harnesses

Sealed leaves for your agent tree.

Agent swarms decompose work into trees. Planners need short SKUs; workers need narrow, sealed facts. Orrery sells the tree-handling rim — bounded stars, citeable decision freezes, and frozen planner subgraphs — so harnesses can hang evidence on their own nodes without Orrery becoming the harness.

Hang, don’t host

A receipt is something a tree node can keep. Orrery seals facts and stage evidence at call time; it does not retain your agent’s plan, task graph, or merge state. Your harness owns the tree — Orrery supplies verifiable leaves you attach where the work happened.

Guides vs sensors

Industry harness vocabulary maps cleanly onto Orrery surfaces — without renaming what you already resolve and call.

  • Guides (feedforward) Gaze shortlist, agent cards, decision-bind freezes — steer before the worker acts.
  • Sensors (feedback) Protocol stars — structure, link, and write checks after action; remediation text on findings.
  • Constellations Frozen planner subgraphs — reusable subtree contracts, not chatty orchestrators.

What to hang mid-tree

  • Protocol stars Manifest, patch capture, structure audits — narrow worker leaves.
  • Gaze + cards Bounded shortlist — names and facets, not tool payloads on the shelf.
  • Constellations Planner freezes made durable — composite receipts across stages.
  • Receipts Signed Envelopes — verify before you treat a leaf as done.

Not this

  • ·Not agent swarm VCS — no merge reconciler or megafile decomposer
  • ·Not a skill router — agents choose among gaze hits; execution stays on the publisher you resolved
  • ·Not a transcript store — receipts and digests as environment, not chat logs

Orrery does not pick your next GitHub issue, reconcile parallel edits, or own Cursor’s swarm coordination layer. Hang sealed leaves; keep the tree.

Publishers may index_ping a listing file — /connect#get-listed.

Deep read

Full rim plan and ADR cross-links: tree-handling-rim.md on GitHub.