Platform, not per-repo folklore

Every product team can wire a feedback table and a cron that adjusts weights. Very few teams ship auditability, org overlays, approval symmetry, and marketplace-safe contracts — and fewer still do it the same way twice.

HumanOS centralizes the shape of adaptation:

• human.call already routes governed work; learning attaches to the same org policy and delegation model.
• Feedback lands in POST /v1/humanos/feedback/events — not in five JSON blobs named learning_v2_final.json.
• Proposals use /v1/humanos/learning/proposals + …/apply / …/rollback — the same lifecycle whether the UI is Command Plane, Companion, or a headless worker.

Marketplace consequence

Packages declare spec.learning in humanos.agent.v1; learning_contract on the wire manifest is derived at publish — not hand-edited. Review surfaces align: human marketplace review and POST /v1/marketplace/review share the same autonomy / learning gates (see HumanOS learning rollback).

CLI is co-equal product

Validate before deploy:

human humanos-manifest validate ./humanos.agent.yaml

Scaffolding and manifest discipline are covered in From zero to governed agent.

Next

Part 4 — Preferences vs policy · Part 2