feat(spec): llm() accepts optional instructions

pkg/spec/src/action-tree.ts

@@ -63,7 +63,7 @@ export type GeneratorNode =
   | { kind: "weighted"; branches: ReadonlyArray<readonly [number, GeneratorNode]> }
   | { kind: "actions"; generate: () => ActionDescriptor[] }
   | { kind: "builtin"; verb: BuiltinVerb }
-  | { kind: "llm"; config: { model: string } };
+  | { kind: "llm"; config: { model: string; instructions?: string } };
 
 // Candidate is one host-enumerated target for a builtin verb. The host
 // resolves geometry (and a native selector) so no element handle crosses into

pkg/spec/src/actions.ts

@@ -38,9 +38,11 @@ export function actions(generator: () => Action[]): GeneratorNode {
 // random candidate, Go drives an OpenAI-compatible model that chooses which
 // candidate to act on from the screenshot + the candidate list. The returned
 // marker is inert on the JS picker (pick.ts walks it to null); Go reads its
-// config.model off globalThis.actions. API key comes from OPENROUTER_API_KEY
+// config off globalThis.actions. API key comes from OPENROUTER_API_KEY
 // (OpenRouter) or OPENAI_API_KEY (OpenAI); OpenRouter wins when both are set.
-export function llm(config: { model: string }): GeneratorNode {
+// Optional `instructions` are appended to the model's prompt as generic
+// bug-hunting guidance for this spec; the model still only ranks candidates.
+export function llm(config: { model: string; instructions?: string }): GeneratorNode {
   return { kind: "llm", config };
 }