# Concept: stewardship loop

Engineering Stewardship is the loop that keeps a repository understandable and improvable for humans and agents.

```text
Observe friction
  -> classify repo/archetype
  -> decide or clarify intent
  -> update docs/spec/code/eval/harness
  -> run or record proof
  -> preserve learning
```

The loop is intentionally broader than feature delivery. It includes the product, but also the software ecology around it: tools, languages, docs, decisions, tests, release mechanics, debugging surfaces, and agent operating rules.

## Durable outputs

| Friction | Durable output |
|----------|----------------|
| Ambiguous direction | North Star update, ADR, DESIGN FAQ |
| Repeated command failure | DX FAQ, validation script, typed action candidate |
| Hidden repo knowledge | AGENTS map, docs map, concept doc |
| Skill routing drift | Eval case, skill description update |
| Runtime diagnostic gap | Harness action, benchmark scenario, unknown case |
| Release confusion | Changeset, changelog, artifact provenance |

A plan is not durable by itself. When work is done, extract the learning and remove stale plan files.

## Adoption check

Use this concept when friction repeats or survives one session. Minimum artifact: convert the learning into exactly one durable owner: ADR, FAQ, code, eval, action candidate, benchmark, or unknown case.

Do not promote a permanent diagnostic or action from the same single run that discovered the issue.

## Related

- **Skills that apply this:** all stewardship skills; portable digests ship under `references/` for `repo-quality-system-lifecycle`, `repository-governance-lifecycle`, `mcp-harness-repo-maintainer`, and `mixture-of-experts`