Add capability-probed structured outputs and a prototype/instance creation graph.

Probe API profiles for json_schema/tools, flatten worker artifacts onto a strict schema, and let users spawn or revise graph nodes instead of a one-shot DAG. Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-31 01:57:22 +08:00
parent 872f742d3e
commit bd5d88abfa
65 changed files with 7625 additions and 942 deletions
@@ -1,6 +1,7 @@
 import type { Blackboard } from "../blackboard/blackboard.js";
 import type { LlmProvider, StreamCallbacks, ChatMessage } from "../llm/client.js";
-import { supportsContentStream } from "../llm/stream-complete.js";
+import { completeStructured } from "../llm/structured-complete.js";
+import { resolveActiveProfile } from "../config/settings.js";
 import { loadWorkerSkillWithContext } from "../skills/loader.js";
 import { filterInputsForRolePerspective } from "../skills/worker-llm.js";
 import { collectUserInputTranscript } from "../intake/intake.js";
@@ -14,6 +15,10 @@ import {
  looksLikeProseNotSpec,
  parseWorkerSetYaml,
 } from "../skills/worker-set-parse.js";
+import {
+  selectJsonPayload,
+  tryParseJsonDoc,
+} from "../parse/json-doc.js";
 import { assembleWorkerContext } from "../skills/context-segments.js";
 import { isPlayLayerActive } from "../skills/play-turn.js";
 import { loadAppSettings } from "../config/settings.js";
@@ -21,6 +26,10 @@ import { resolveActivePreset } from "../preset/store.js";
 import { assemblePlayWorkerMessages } from "../preset/play-frame.js";
 import { worldInfoPackFromSegments } from "../preset/world-info-pack.js";
 import { PRESENT_TAG } from "../skills/present-packet.js";
+import {
+  coerceToFlatArtifact,
+  flatArtifactToFragmentDoc,
+} from "../skills/flat-artifact.js";
 import {
  gmChanceToolsEnabled,
  GM_CHANCE_HARNESS_INSTRUCTION,
@@ -36,7 +45,10 @@ import {
  isUsableContextFragment,
  looksLikeFragmentDoc,
 } from "../skills/context-fragment.js";
-import { isProgressPointerTag } from "../skills/creation-flow.js";
+import {
+  isProgressPointerTag,
+  looksLikeCreationFlowDoc,
+} from "../skills/creation-flow.js";
 import {
  CONTEXT_ORDER_TAG,
  parseContextOrder,
@@ -93,28 +105,19 @@ const WORKER_OUTPUT_INSTRUCTION = `

 ---

-## 运行时输出协议（必须遵守）
+## 运行时输出协议

-请输出 **单个 JSON 对象**（不要 markdown 代码块），字段：
-{
-  "outputs": { "<outputTag>": "<内容字符串>" },
-  "summary": "50字以内产物摘要",
-  "askUser": null 或 问题数组
-}
+输出 **一个 JSON 对象**（由程序按模型能力用 schema / tool / json_object 约束投递）。

-askUser 每项可为：
- 字符串："需要用户补充的问题"
- 或结构化：{ "id": "q1", "prompt": "问题", "options": [{ "id": "A", "label": "可编辑完整句选项" }], "allowOther": true }
+按本步 SKILL 规定的产物形状写。有正文的技能稿按该步 \`正文\` 对象写（柱下到条目/诊断三元组），不要把业务字段改成一段 \`sections[].text\`。
+「设计.创作流程」必须是带 \`steps\` 数组的 DAG，worker 集必须是带 \`workers\` 的规格。

- outputs 的 key 必须是要求的 outputTags
- **设计.worker集 / 设计.worker集.草稿**：value 必须是 JSON 对象文本（以 { 开头），禁止中文说明、元叙述、提问长文；禁止 YAML
- **context-fragment.v1**：必须放在 outputs["<本步 artifact tag>"]，禁止把片段当根对象。题目只写在产物「追问」（建议选项 + 示例）；顶层 askUser 必须为 **null**。程序会把「追问」挂到询问卡。禁止同一问再抄一份 askUser。禁止写入「创作.当前步骤」等进度指针
- **其它产物**：优先同时给 outputs + askUser；有产物时追问挂在产物下（用户可直接接受而不作答）
- **残稿也要交**：JSON 不完美、字段不全、截断，只要有可展示正文就必须写入 outputs，交给用户验收（可打回重生成，也可接着改）。不要因为渲染不出就把稿扔掉改成提问
- 仅当完全没有可展示正文、只有提问时，才留空 outputs、只填 askUser
- 能推断选项时 **必须**给 options（完整句、可改写）；不要只丢裸问题逼用户写长段
- 若 inputs 中 \`用户.需求\` / \`用户.博弈需求\`（或 book.brief）已有实质内容，禁止 askUser 要求用户重复提供其中已写明的情境、角色、规则等；仅对 genuinely 缺失且无法推断的要点提问
- summary 用于界面展示`;
+也兼容：
+1. 直接输出本步产物对象（创作流程 / context-fragment.v1 / worker 集等）
+2. \`{ "outputs": { "<本步 tag>": { … } }, "summary":"…", "askUser": null }\`（values 用对象）
+
+不要写入进度指针。追问用 \`questions\`（或旧 \`追问.题目\`）；\`questions: []\` 表示可验收。
+「开放问题」留给后续步。残稿也要交。summary 用于界面展示。`;

 function slotValueForTag(
  tag: string,
@@ -193,6 +196,13 @@ function assignOutput(
 }

 /** 模型常把片段当根对象，或写进「创作.当前步骤」 */
+function existingOutputIsProduct(text: string | undefined): boolean {
+  if (!text?.trim()) return false;
+  if (isUsableContextFragment(text)) return true;
+  const parsed = tryParseJsonDoc(text);
+  return looksLikeCreationFlowDoc(parsed) || looksLikeFragmentDoc(parsed);
+}
+
 function recoverFragmentOutputs(
  obj: Record<string, unknown>,
  outputs: Record<string, string>,
@@ -200,7 +210,7 @@ function recoverFragmentOutputs(
 ): void {
  const target = productOutputTags(outputTags)[0];
  if (!target) return;
-  if (outputs[target] && isUsableContextFragment(outputs[target])) return;
+  if (existingOutputIsProduct(outputs[target])) return;

  const outputsRaw =
    obj.outputs && typeof obj.outputs === "object" && !Array.isArray(obj.outputs)
@@ -212,7 +222,11 @@ function recoverFragmentOutputs(
      if (key === target) continue;
      const text = stringifyOutputValue(val);
      if (!text) continue;
-      if (isUsableContextFragment(text) || looksLikeFragmentDoc(val)) {
+      if (
+        isUsableContextFragment(text) ||
+        looksLikeFragmentDoc(val) ||
+        looksLikeCreationFlowDoc(val)
+      ) {
        outputs[target] = text;
        return;
      }
@@ -227,6 +241,13 @@ function recoverFragmentOutputs(
  if (looksLikeFragmentDoc(obj) || looksLikeFragmentDoc(rest)) {
    const source = looksLikeFragmentDoc(obj) ? obj : rest;
    outputs[target] = JSON.stringify(source);
+    return;
+  }
+  if (looksLikeCreationFlowDoc(obj) || looksLikeCreationFlowDoc(rest)) {
+    const source = looksLikeCreationFlowDoc(obj) ? obj : rest;
+    if (!outputs[target]?.trim()) {
+      outputs[target] = JSON.stringify(source);
+    }
  }
 }

@@ -283,14 +304,26 @@ const INCOMPLETE_FRAGMENT_ASK = normalizeQuestions([
  },
 ]);

+function lookupOutputValue(
+  raw: Record<string, unknown>,
+  tag: string,
+): unknown {
+  if (raw[tag] != null) return raw[tag];
+  const leaf = tag.includes(".") ? tag.slice(tag.lastIndexOf(".") + 1) : tag;
+  if (!leaf) return undefined;
+  for (const [key, val] of Object.entries(raw)) {
+    if (val == null) continue;
+    if (key === leaf || key.endsWith(`.${leaf}`)) return val;
+  }
+  return undefined;
+}
+
 function parseWorkerResponse(
  raw: string,
  outputTags: string[],
 ): WorkerRunResult {
-  let parsed: unknown;
-  try {
-    parsed = JSON.parse(raw);
-  } catch {
+  const parsed = tryParseJsonDoc(raw);
+  if (parsed == null || typeof parsed !== "object" || Array.isArray(parsed)) {
    const target = productOutputTags(outputTags)[0] ?? outputTags[0];
    if (target && looksLikeStructuredDraft(raw)) {
      return finalizeParsedOutputs(
@@ -322,17 +355,43 @@ function parseWorkerResponse(
    };
  }

-  if (!parsed || typeof parsed !== "object") {
-    throw new Error("Worker 返回了无效 JSON");
+  const obj = parsed as Record<string, unknown>;
+
+  // 创作流程 DAG：根对象就是产物，不能先收成扁形（brief 会被当成稿、steps 丢掉）
+  if (looksLikeCreationFlowDoc(obj)) {
+    const target = productOutputTags(outputTags)[0] ?? outputTags[0];
+    if (target) {
+      return finalizeParsedOutputs(
+        { [target]: JSON.stringify(obj) },
+        obj,
+        outputTags,
+      );
+    }
+  }
+
+  // 扁形产物 → 旧 fragment 外壳。已是 context-fragment 深树则原样收下（保留自评与嵌套正文）。
+  if (!looksLikeFragmentDoc(obj)) {
+    const asFlat = coerceToFlatArtifact(obj);
+    if (asFlat) {
+      const frag = flatArtifactToFragmentDoc(asFlat);
+      const target = productOutputTags(outputTags)[0] ?? outputTags[0];
+      if (target) {
+        return finalizeParsedOutputs(
+          { [target]: JSON.stringify(frag) },
+          { ...frag, summary: asFlat.summary || obj.summary },
+          outputTags,
+        );
+      }
+    }
  }

-  const obj = parsed as Record<string, unknown>;
  const outputsRaw = obj.outputs;
  const outputs: Record<string, string> = {};

-  if (outputsRaw && typeof outputsRaw === "object") {
+  if (outputsRaw && typeof outputsRaw === "object" && !Array.isArray(outputsRaw)) {
+    const bag = outputsRaw as Record<string, unknown>;
    for (const tag of productOutputTags(outputTags)) {
-      assignOutput(outputs, tag, (outputsRaw as Record<string, unknown>)[tag]);
+      assignOutput(outputs, tag, lookupOutputValue(bag, tag));
    }
  }
  recoverFragmentOutputs(obj, outputs, outputTags);
@@ -484,25 +543,6 @@ function extractQuestionsFromText(text: string): string[] {
  return [];
 }

-async function completeWorkerPreferStream(
-  llm: LlmProvider,
-  messages: Parameters<LlmProvider["complete"]>[0],
-  options: Parameters<LlmProvider["complete"]>[1],
-  stream?: WorkerStreamCallbacks,
-): Promise<Awaited<ReturnType<LlmProvider["complete"]>>> {
-  const callbacks: StreamCallbacks = {
-    onReasoningDelta: (delta) => stream?.onThinkingDelta?.(delta),
-    onContentDelta: (delta) => stream?.onOutputDelta?.(delta),
-  };
-  if (supportsContentStream(llm)) {
-    return llm.completeStream!(messages, options, callbacks);
-  }
-  const result = await llm.complete(messages, options);
-  if (result.reasoning) callbacks.onReasoningDelta?.(result.reasoning);
-  if (result.content) callbacks.onContentDelta?.(result.content);
-  return result;
-}
-
 export async function runWorkerSkill(params: WorkerRunParams): Promise<WorkerRunResult> {
  const workerId = resolveWorkerId(params.workerId);
  let worker: ParsedWorkerSkill;
@@ -519,9 +559,15 @@ export async function runWorkerSkill(params: WorkerRunParams): Promise<WorkerRun
      params.slots.creationAcceptedUnits ??
      params.blackboard.getContentByTag("创作.已验收单位");
    let acceptedStepNames: string[] = [];
+    let inheritExisting = false;
    try {
-      const { parseAcceptedSteps } = await import("../skills/creation-flow.js");
+      const { parseAcceptedSteps, isInheritExistingFlag } = await import(
+        "../skills/creation-flow.js"
+      );
      acceptedStepNames = parseAcceptedSteps(acceptedRaw);
+      inheritExisting = isInheritExistingFlag(
+        params.blackboard.getContentByTag("创作.继承修改"),
+      );
    } catch {
      acceptedStepNames = [];
    }
@@ -538,6 +584,7 @@ export async function runWorkerSkill(params: WorkerRunParams): Promise<WorkerRun
        acceptedStepNames,
        selectedRecipeRef,
        filledArtifactTags,
+        inheritExisting,
      },
    );
    worker = loaded.worker;
@@ -616,20 +663,36 @@ export async function runWorkerSkill(params: WorkerRunParams): Promise<WorkerRun
      stream: params.stream,
      caller: `worker:${worker.id}`,
    });
-    return parseWorkerResponse(harness.content, worker.outputTags);
+    return parseWorkerResponse(selectJsonPayload(harness.content), worker.outputTags);
  }

-  const result = await completeWorkerPreferStream(
-    params.llm,
-    messages,
-    {
-      responseFormat: "json_object",
-      caller: `worker:${worker.id}`,
-    },
-    params.stream,
-  );
+  const streamCallbacks: StreamCallbacks | undefined = params.stream
+    ? {
+        onReasoningDelta: (delta) => params.stream?.onThinkingDelta?.(delta),
+        onContentDelta: (delta) => params.stream?.onOutputDelta?.(delta),
+      }
+    : undefined;

-  return parseWorkerResponse(result.content, worker.outputTags);
+  const result = await completeStructured(params.llm, messages, {
+    schema: {
+      type: "object",
+      additionalProperties: true,
+    },
+    name: "submit_worker_result",
+    schemaIsLoose: true,
+    capabilities: resolveActiveProfile()?.capabilities,
+    caller: `worker:${worker.id}`,
+    stream: streamCallbacks,
+  });
+
+  const raw =
+    typeof result.parsed === "string"
+      ? result.parsed
+      : JSON.stringify(result.parsed);
+  return parseWorkerResponse(
+    selectJsonPayload(raw, result.reasoning) || raw,
+    worker.outputTags,
+  );
 }

 /** @internal 供单测 */