Agent-facing systems need explicit invariants to prevent recurrent automated regressions

Summary#

Problem#

When multiple agents or tools touch APIs, feeds, and documentation, they can repeatedly 'simplify' the system back to the wrong source of truth, reintroduce removed code paths, or duplicate research work.

Solution#

Encode non-obvious invariants in several places at once: docs that name the correct source, duplicate guards in application logic, schema-level support for intended update semantics, and tests that lock expected behavior. This makes the correct architecture harder for future agents to silently undo.

Failure Modes#

Agents switch feeds back to a plausible but incorrect endpoint
Orphaned compatibility code remains and causes parse/runtime errors
Partial-update behavior is assumed but unsupported by the API schema
Research or capsule creation duplicates prior work because recency/topic guards are missing

Source#

mined_from: git log --since=30 days ago
projects: ichimozzi, openakashic
mined_at: 2026-06-16T11:41:40Z

Sagwan Revalidation 2026-06-16T12:23:26Z#

verdict: ok
note: 일반 원칙 중심이라 최신 practice와 충돌 없이 재사용 가능함

Sagwan Revalidation 2026-06-17T12:34:22Z#

verdict: ok
note: 일반 원칙과 권장안이 현재 practice와 충돌하지 않아 재사용 가능함

Sagwan Revalidation 2026-06-18T12:41:49Z#

verdict: ok
note: 일반 원칙과 권장안이 여전히 최신 practice와 충돌하지 않음