| name | verdict-directed |
| description | 🔪 SDT · verdict-directed review of backend codegen/transform. Use when auditing codegen/transform for smart-analyzer / dumb-codegen violations, or when porting a transform/codegen visitor from the Original. |
| allowed-tools | Bash, Read, Grep, Glob, mcp__narsil-mcp__*, Agent |
verdict-directed
Review a backend crate against the verdict-directed principle (docs/context.md §«Принципы»): transform and codegen are a syntax-directed translation over analysis verdicts. This skill finds where the backend instead recovers a domain fact on its own. It is a review, not a linter: a finding is closed by re-homing the fact to analysis, never by silencing.
Required reading (step 0)
Re-read every invocation; do not assume in-session.
docs/context.md §«Принципы» — the keystone and its three violation shapes.docs/context.md §«Language» — the three border anti-patterns, the finding vocabulary.
The razor
Two kinds of complexity live in a backend. Only one is a violation.
- Instruction selection (legit). Choosing the target print-form of an already-decided verdict: escaping, whitespace,
clsx-wrap, css-hash merge, boolean-attribute table, dev vs prod, client vs server, which $. call. Stays in the backend.
- Domain-fact recovery (violation). Deciding what a unit means: is it reactive, static, which role, how many references. Must be born in analysis and consumed as a verdict.
Litmus, per branch: swap the target runtime / printing — does this branch still need to exist? Yes → instruction selection, legit. No, because it changes what the unit means → violation.
Scope
Four backend crates only: svelte_codegen_client, svelte_codegen_server, svelte_transform_client, svelte_transform_server. Analysis, parser, ast are out of scope — a fact missing there is the fix, not a finding here.
Porting an Original transform/codegen visitor is a first-class trigger: run the pass on the new backend code before committing — the port is where a violation enters.
Finder — a reading pass, not a grep
The tell is not a syntax, it is provenance: trace the deciding input of every backend branch to its origin. Legit when the input is a verdict handle — the result of one analysis query. Violation when the input is a value read out of source text or AST structure. The syntactic form is irrelevant, so an example never bounds it — a name-match hides just as well inside a helper, a name.starts_with(…), a HashSet lookup, or an enum discriminant.
Ripgrep only seeds the pass — it catches the obvious forms, never all of them. Seed shapes:
.name / element.name / attr.name() vs a string literal, or matches!(…, "x" | "y") on a name.self.component.source / .source read to inspect a value, not to slice a known span.walk / Visit / .iter().any( / .iter().find( over children to derive a boolean.- one emit decision reading several analysis side-tables joined with
&&.
A seed hit is a candidate, not a finding — trace it, don't tally it.
Judge — classify every candidate
Apply the razor litmus and label each candidate exactly one of:
- legit — instruction selection / target-printing. Record which print-form axis.
- Анализ в кодгене — name-match or re-walk recovering a fact analysis owns.
- Сырые факты — several side-tables assembled where one verdict should exist.
- Эмит-форма семантики — the fact is in analysis but stored as a runtime-call form the backend just relays.
For each violation, name the owed verdict: the domain answer analysis must produce so the backend collapses to one match. It must pass the codegen-agnostic litmus — swap the runtime and the verdict is unchanged. A raw index or side-table keyed by node (is_class_attr: bool, a set of special names) is not an owed verdict: it only relocates the violation from Анализ в кодгене to Сырые факты / Эмит-форма семантики — the same triangle, moved one corner.
Report
Per violation: file:line, the shape, the owed verdict, and the one-match backend shape after the fix.
This skill applies no fix. A finding is never closed by #[allow], a suppression, an allowlist, or "match the surrounding code". Those are forbidden as resolutions — the neighbour breaking the principle is not precedent. A finding closes only by (a) re-homing the fact to analysis as a domain verdict — not as a raw index or emit-form, which merely pushes the violation around the triangle — or (b) an explicit human decision recorded in the report.
The fix is a new task — gate it
Naming the owed verdict is review; producing it means editing analyze, a new layer whose PRD invariants govern where a verdict lives and how it is keyed. No green gate validates that model — tests stay green on a mis-keyed one (an element-level aggregate crammed into one variant, two unrelated verdicts merged for a shared key). So before writing it, invoke /required to route the analyze/attribute PRDs into context, and design before code. Key the verdict by the node it belongs to; a slot conflict means the key is wrong, not that you may merge.
Completion criterion
Every backend decision-branch is traced to its input origin and labelled — verdict handle (legit) or source/AST re-derivation (one of the three shapes). The unit of coverage is the branch, not the seed hit: an exhausted seed list is not completion. If any in-scope backend file was not read through, the review is not done.
