assay

assay

!checklist init ${CLAUDE_SKILL_DIR} --force

A green suite is worthless until it has proven it can go red. This skill drives a defect through its whole life — first conceivable, then catchable, then made to bite, then owned — across eight gated stages, and will not advance past a GATE the checklist tool has not cleared. Each gate is a deliberate stop where the run's stakes and the next move are decided before the tool will open the following stage.

Discipline: the gates are machine-enforced. Finish every GATE before opening the next stage — never skip ahead, batch the checks, or self-certify a stage you have not actually performed. Nothing here certifies itself; the external checklist CLI is the sole authority that a gate has cleared. Commands address stages by name — you never track a stage number by hand.

If $ARGUMENTS names a target (a file, module, or subsystem), test that. If it names something that is not a test target (say, a language preference), acknowledge it and ask what to test.

Speak the user's language, or the stakes get set blind. Testing is a collaboration — you supply the rigor, the user supplies the domain judgment and the stakes: which risks matter, how a found bug is disposed of. They can only supply it on terms they follow, so read their fluency and gloss a term on first use (mock vs fake, property-based, mutation testing, hermetic, flaky). A user who signs off on a risk ledger or a disposition full of unparsed words has not set the stakes — and the CHARTER, TRIAGE, and BOOKS judgments this skill leans on them for are then hollow.

Reference library

Detail lives in references/. Open each when a stage sends you there, not upfront. Ten references back the eight stages:

• references/decision-tree.md — the engine. Four nested decision trees (worth-it → type → doubles → cases) plus overlays and an escalation ladder. Open it at TRIAGE and keep it beside you through COMPOSE.
• references/evidence-catalogue.md — the trust ceiling of every form of test evidence: what each proves, what it costs, and what it is structurally blind to. Pick evidence by what it cannot see.
• references/probe-construction.md — the oracle-first pipeline that derives cases from the source-of-truth question: oracles, equivalence classes, edge probes, failure paths, negative space, the call-it-again probe, fail-first proof.
• references/test-doubles.md — dummy / stub / spy / mock / fake: choosing the weakest one, classicist vs mockist, anti-patterns.
• references/property-based.md — finding properties, generators, shrinking, stateful testing.
• references/determinism-and-flakiness.md — making tests deterministic by construction, and diagnosing flakes.
• references/coverage-and-mutation.md — coverage as a guide, mutation testing as the real teeth, the economics of what to skip.
• references/language-norms.md — per-language testing norms an agent must be told: idiomatic frameworks, blind-spots (patch-where-used, -race, doc-tests, fake timers…), and the must-run commands agents skip. Load at SURVEY, apply at BUILD.
• references/agent-test-smells.md — how an agent writes plausible-but-worthless tests by default (assertion-free coverage padding, over-mocking, implementation-coupled brittleness, tolerated flakes), and how to keep a large suite from rotting into a slow, flaky "test graveyard". Apply while writing at BUILD, sweep at BITE.
• references/parallel-execution.md — running assay on a fleet of agents (ultracode): how the risk ledger, the escalation ladder, and the oracle gate project onto parallel search, and why agreement is not evidence. Open when the harness hands you subagents.