leaf-adversarial-gateThe final pre-merge gate for emergent implementations and diagnosis-repairs — open-ended deep review that hunts for defects and blocks the merge. NOT for checklist-driven review (use leaf-checklist-review). This is the one role every "a cheaper model is fine" decision depends on, so don't cheap out
leaf-checklist-reviewReview where WHAT to look at is already specified as a checklist / set of lenses — pre-merge gate reviews, security/privacy review against given criteria, checking a doc against a requirements list. NOT for open-ended review where you must decide what to inspect; use leaf-adversarial-gate for that.
leaf-coverage-researchCoverage research where the scope to collect is DEFINED (official docs, a fixed source list) and the goal is to gather completely and compare against a baseline — compatibility matrices, standard cross-checks, exhaustive collection. If the scope is open and "finding hidden things" is the point, use
leaf-design-mockupBuild design mockups / comparison pages — several variants side by side for a human to pick (new screen or component drafts). The final choice is the human's. (Re-measured 2026-07: on both mockup tasks the mid tier was the only model to beat the previous best — and in the earlier round no single mod
leaf-diagnose-repairDiagnose AND fix a bug whose cause is NOT yet established (hypothesis-level) — diagnosis through repair in one go. If the fix method is already given, use leaf-oracle-repair. The output must pass leaf-adversarial-gate before merge. Re-measured 2026-07: this tier now **ties the top tier on defects (1
leaf-discovery-researchDiscovery research where finding what's OUTSIDE the official list — hidden features, undocumented constraints, workarounds, pitfalls — is the substance. For defined exhaustive collection use leaf-coverage-research. Discovery is the one axis where models genuinely split, and a missed finding can't be
leaf-emergent-implImplementation where only the direction / cause-family is given and "the design is your call" — typical when several subsystems' rules (sync, concurrency, data integrity) must all hold at once. The output MUST pass through leaf-adversarial-gate before merge. For concurrency / data-integrity / securi
leaf-oracle-repairFix or refactor where BOTH the method (what to change and how) AND an oracle (tests / output comparison that mechanically judges success) are supplied. Behavior-preserving refactors, bugs whose cause is already agreed. If the cause is unknown, use leaf-diagnose-repair. (First pass: with method + ora
leaf-plan-draftProduce a SCOPED PLAN document (not code): the goal is known and the work is detailing it — investigation and trade-off analysis for a restructure/migration. Output is always a "pre-review draft". If the goal itself is unknown and the work is searching broadly for what to do (an exploratory plan), o