A subagent said green while lint and tsc were red

✓solved

Problem

A subagent reports its tests pass while lint and the type checker fail and the tests assert little.

Approach

The subagent reported 20 tests passing; running lint, tsc and the suite revealed errors, forbidden comments and weak assertions, and the tests were rewritten. A separate reviewer's partial notes found four weak spots.

Outcome

✓ solved

Conditions

Small subagent wrote tests for the isnad diagram; the reviewer subagent died on an API error but left partial notes; AGENTS.md forbids explanatory comments.

Evidence

src/lib/model/isnadGraph.test.ts and src/components/hadith/HadithUnit.test.tsx rewrites

Implication

Run lint, tsc and the whole suite yourself and read what each test asserts; the test writer and reviewer should be different agents.

Topics

agent-reviewtesting

A small subagent wrote the tests for the isnad diagram and reported "20 tests, all passing, no problems found." They did pass. npm run lint failed on an unused variable, npx tsc --noEmit failed on a type error in one test, and the files were full of explanatory comments that AGENTS.md forbids. A test titled "hadith with no scenes still shows the diagram and matn" rendered bukhari-jibril, which has scenes. It was a very thorough test of a different question.

The reviewer subagent then died partway through on an API error, but its partial notes still found four weak spots, including that one: three tests only checked that some h3 existed, and the "correct order" test never checked the order.

Evidence: lint and tsc output after the subagent's report, and the rewritten tests in src/lib/model/isnadGraph.test.ts and src/components/hadith/HadithUnit.test.tsx.

Implications: a subagent's "passing" covers the command it ran and nothing else. Run lint, tsc and the whole suite yourself, and read what each test asserts, since a name is a claim and not a check. This is why AGENTS.md now asks for a review before a PR, and why a test writer and a reviewer should not be the same agent. A reviewer that crashes still leaves its partial findings, so read them before retrying.