@lingjingai/scriptctl 0.19.0 → 0.19.1

package/dist/domain/direct-core.d.ts

@@ -154,6 +154,18 @@ export declare function clipText(text: unknown, limit?: number): string;
 export declare function sceneExample(scene: Dict, limit?: number): string;
 export declare function buildMetadataContext(script: Dict): Dict;
 export declare function deterministicExtractMetadata(script: Dict): Dict;
+export interface MetadataMdSections {
+    worldviewRaw: string;
+    actors: Map<string, {
+        role: string;
+        description: string;
+    }>;
+    locations: Map<string, string>;
+    props: Map<string, string>;
+}
+export declare function parseMetadataMdSections(raw: string): MetadataMdSections;
+export declare function metadataMdSectionsEmpty(s: MetadataMdSections): boolean;
+export declare function overlayMetadataSections(parsed: MetadataMdSections, script: Dict): Dict;
 export declare function parseMarkdownMetadata(raw: string, script: Dict): Dict;
 export declare function applyMetadataToScript(script: Dict, metadata: Dict): void;
 interface AssetUsage {

package/dist/domain/direct-core.js

@@ -2544,7 +2544,10 @@ export function deterministicExtractMetadata(script) {
     return {
         confidence: "high",
         worldview: "现代",
-        worldview_raw: "mock provider generated metadata for CLI tests",
+        // Neutral default that mirrors the worldview enum — NOT a mock/test string,
+        // since this baseline is also the real provider path's fallback when the
+        // model omits the worldview section (see overlayMetadataSections).
+        worldview_raw: "现代",
         actors,
         locations: asList(script["locations"]).map((loc) => ({
             location_id: loc["location_id"],
@@ -2557,10 +2560,15 @@ export function deterministicExtractMetadata(script) {
     };
 }
 // Split a metadata markdown document (see METADATA_MD_SPEC) into its raw pieces.
-// Tolerant by design: unknown headings reset the section, non-bullet lines under
-// asset sections are ignored, and a missing field just yields "". The id-keyed
-// maps and the raw worldview line are validated/backfilled by the caller.
-function parseMetadataMdSections(raw) {
+// This is a DISTINCT grammar from parseMarkdownBatch (id-keyed `id | role | desc`
+// rows, not the name-keyed `名: 描述` asset registry), so it intentionally does
+// not share that parser. Strict on section identity: only an EXACT top-level
+// heading (世界观 / 人物 / 场景 / 道具, any number of leading `#`) opens a section,
+// so a stray sub-heading like "## 人物关系说明" can't hijack the bullets under it.
+// A worldview value written inline on the heading (`# 世界观：赛博朋克`) is still
+// captured. Pipe-split fields keep any extra pipes inside the description. Missing
+// fields yield "".
+export function parseMetadataMdSections(raw) {
     const actors = new Map();
     const locations = new Map();
     const props = new Map();
@@ -2572,14 +2580,20 @@ function parseMetadataMdSections(raw) {
             continue;
         if (line.startsWith("#")) {
             const heading = line.replace(/^#+/, "").trim();
-            if (heading.includes("世界观"))
-                section = "worldview";
-            else if (heading.includes("人物"))
+            const inlineWorldview = heading.match(/^世界观[：:]\s*(.+)$/);
+            if (heading === "人物")
                 section = "actors";
-            else if (heading.includes("场景"))
+            else if (heading === "场景")
                 section = "locations";
-            else if (heading.includes("道具"))
+            else if (heading === "道具")
                 section = "props";
+            else if (heading === "世界观")
+                section = "worldview";
+            else if (inlineWorldview) {
+                section = "worldview";
+                if (!worldviewRaw)
+                    worldviewRaw = inlineWorldview[1].trim();
+            }
             else
                 section = "";
             continue;
@@ -2591,46 +2605,61 @@ function parseMetadataMdSections(raw) {
         }
         if (section === "")
             continue;
-        const body = line.replace(/^[-*]\s*/, "").trim();
-        const parts = body.split("|").map((s) => s.trim());
+        // Keep any extra pipes inside the description (the model may ignore the
+        // no-pipe rule) by rejoining the trailing parts rather than dropping them.
+        const parts = line.replace(/^[-*]\s*/, "").split("|").map((s) => s.trim());
         const id = parts[0] ?? "";
         if (!id)
             continue;
         if (section === "actors")
-            actors.set(id, { role: parts[1] ?? "", description: parts[2] ?? "" });
+            actors.set(id, { role: parts[1] ?? "", description: parts.slice(2).join(" | ").trim() });
         else if (section === "locations")
-            locations.set(id, parts[1] ?? "");
+            locations.set(id, parts.slice(1).join(" | ").trim());
         else
-            props.set(id, parts[1] ?? "");
+            props.set(id, parts.slice(1).join(" | ").trim());
     }
     return { worldviewRaw, actors, locations, props };
 }
-// Parse a metadata markdown response into the same dict shape as
-// deterministicExtractMetadata. Strategy: start from the deterministic baseline
-// (guarantees every script asset is present with a valid role/description), then
-// overlay the model's worldview/role/description wherever it supplied a usable
-// value. So a truncated or partial response still produces a complete,
-// schema-valid result — missing assets simply keep their deterministic default.
-export function parseMarkdownMetadata(raw, script) {
+// True when nothing recognizable parsed — no worldview and no asset rows. The
+// provider boundary uses this to reject a non-conforming response (prose refusal,
+// wrong format) instead of silently degrading to a name-only baseline.
+export function metadataMdSectionsEmpty(s) {
+    return !s.worldviewRaw.trim() && s.actors.size === 0 && s.locations.size === 0 && s.props.size === 0;
+}
+// Resolve a raw worldview line to a valid enum value, or null if it can't be
+// pinned down. Exact match wins; otherwise a substring match is accepted ONLY
+// when exactly one enum value appears, so a line naming two (e.g. "古代历史…古风
+// 架空…") stays ambiguous instead of resolving by array order. Null → caller
+// keeps its default.
+function coerceWorldview(worldviewRaw) {
+    if (WORLDVIEW_VALUES.includes(worldviewRaw))
+        return worldviewRaw;
+    const hits = WORLDVIEW_VALUES.filter((v) => worldviewRaw.includes(v));
+    return hits.length === 1 ? hits[0] : null;
+}
+// Overlay parsed markdown sections onto the deterministic baseline (which already
+// guarantees every script asset is present with a valid role/description). The
+// model's worldview/role/description are applied wherever usable, so a partial
+// response still yields a complete, schema-valid result.
+export function overlayMetadataSections(parsed, script) {
     const base = deterministicExtractMetadata(script);
-    const parsed = parseMetadataMdSections(raw);
     const worldviewRaw = parsed.worldviewRaw.trim();
     if (worldviewRaw) {
-        const matched = WORLDVIEW_VALUES.includes(worldviewRaw)
-            ? worldviewRaw
-            : WORLDVIEW_VALUES.find((v) => worldviewRaw.includes(v));
-        if (matched)
-            base["worldview"] = matched;
         base["worldview_raw"] = worldviewRaw;
+        base["worldview"] = coerceWorldview(worldviewRaw) ?? base["worldview"];
     }
+    // If the model is assigning roles at all, don't let the deterministic
+    // "first actor = 主角" default survive for an actor it didn't mark as a lead —
+    // otherwise an omitted first-actor line would leave the script with two 主角.
+    const modelAssignsRoles = [...parsed.actors.values()].some((a) => ROLE_TYPE_VALUES.includes(a.role));
     for (const actor of base["actors"]) {
         const item = parsed.actors.get(strOf(actor["actor_id"]));
-        if (!item)
-            continue;
-        if (ROLE_TYPE_VALUES.includes(item.role))
-            actor["role_type"] = item.role;
-        if (item.description)
+        if (item?.description)
             actor["description"] = item.description;
+        if (item && ROLE_TYPE_VALUES.includes(item.role))
+            actor["role_type"] = item.role;
+        else if (modelAssignsRoles)
+            actor["role_type"] = "配角";
     }
     for (const loc of base["locations"]) {
         const desc = parsed.locations.get(strOf(loc["location_id"]));
@@ -2645,6 +2674,12 @@ export function parseMarkdownMetadata(raw, script) {
     base["confidence"] = "high";
     return base;
 }
+// Parse a metadata markdown response into the deterministicExtractMetadata dict
+// shape. Total by design (always returns a complete dict); the provider boundary
+// guards conformance via metadataMdSectionsEmpty before calling the overlay.
+export function parseMarkdownMetadata(raw, script) {
+    return overlayMetadataSections(parseMetadataMdSections(raw), script);
+}
 export function applyMetadataToScript(script, metadata) {
     script["worldview"] = metadata["worldview"];
     script["worldview_raw"] = strOf(metadata["worldview_raw"]).trim();