Fusion agents keep shipping behavior changes that leave stale tests behind, and
targeted verification structurally cannot catch it: it runs the tests for the
files the agent TOUCHED, while the assertions encoding the old behavior live in
files it did not. Measured on one full engine suite run (2026-08-24): 297 failing
tests, ~135 of them from exactly five such changes — the FN-158 pre-merge-gate
guard (~70 fixtures across 13 files), the branch-write provenance guard (18 from
ONE shared fixture), a workflow-IR reorder (10 stale topology assertions), an
updateTaskAtomic seam missing from fake stores (~9), and FN-074's splitting
removal leaving 4 reviewer-prompt tests asserting a deleted contract. Every one
passed its own targeted verification.
- AGENTS.md: new standing rule with the four search triggers (guard added,
feature removed, order/default/constant/prompt changed, public method added),
a fix-at-the-shared-factory preference, and an explicit ban on weakening a
stale test or restoring removed behavior to satisfy one.
- Executor prompt: the same searches, stated as a finish condition.
- Full triage prompt: specs for behavior-changing tasks must include a step that
NAMES the search, so it is planned rather than left to chance.
The fast triage prompt is deliberately excluded: it carries a hard size budget
(agent-prompts.test.ts caps it at 7500 chars) and adding this pushed it over —
an instance of the very rule being documented, now cited in it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
f2eda4819d