Initial TrueGrowth source import

openspec/changes/refactor-ai-json-response-parser/design.md

@@ -0,0 +1,18 @@
+## Context
+大模型 JSON 输出不是稳定的裸 JSON：可能被 `<think>`、代码块、自然语言说明或多个候选 JSON 包裹。业务模块需要根据自身 schema 选择正确候选，而不是各自用正则或贪婪匹配。
+
+## Goals
+- 提供无组件/服务依赖的纯工具，避免循环依赖。
+- 支持对象和数组协议，并允许 predicate 选择业务匹配候选。
+- 让业务模块保留现有校验、normalize、修复和降级策略。
+
+## Non-Goals
+- 不替换 localStorage、GitHub 同步、备份、HTTP API body 等普通 JSON 解析。
+- 不引入流式增量 JSON 解析。
+- 不统一各业务返回 schema。
+
+## Decisions
+- 工具放在 `packages/drawnix/src/utils/llm-json-extractor.ts`。
+- 候选查找顺序优先使用 fenced JSON、去除完整 `<think>` 后的内容、`</think>` 后内容、最后回退完整文本。
+- 使用括号平衡扫描，并识别 JSON 字符串内的括号和转义，避免字符串内容破坏平衡。
+- `extractJsonValue` 返回已解析值，`extractJsonSource` 返回原始 JSON 字符串，兼容需要二次修复/解析的旧逻辑。

openspec/changes/refactor-ai-json-response-parser/proposal.md

@@ -0,0 +1,18 @@
+# Change: Refactor AI JSON response parsing
+
+## Why
+部分大模型会在 JSON 前后输出 `<think>`、Markdown、解释文本或多个候选片段，现有各业务入口解析方式不一致，容易误取思考区伪 JSON 或因前缀文本报错。
+
+## What Changes
+- 新增纯工具统一提取大模型返回中的 JSON 对象/数组候选。
+- 只迁移 AI/LLM 返回 JSON 的解析入口，保留普通存储、同步、HTTP 响应 JSON 解析现状。
+- 将爆款视频、爆款 MV、漫画、音乐、长视频、PPT outline、Agent workflow、技能解析等业务入口改为共享候选提取，业务层继续负责 schema 校验、normalize 和降级策略。
+- 为共享工具和关键业务入口补充回归测试。
+
+## Impact
+- Affected specs: ai-json-response-parsing
+- Affected code:
+  - `packages/drawnix/src/utils/llm-json-extractor.ts`
+  - `packages/drawnix/src/components/*`
+  - `packages/drawnix/src/services/*`
+  - `packages/drawnix/src/mcp/tools/*`

openspec/changes/refactor-ai-json-response-parser/specs/ai-json-response-parsing/spec.md

@@ -0,0 +1,21 @@
+## ADDED Requirements
+
+### Requirement: AI JSON responses must use shared candidate extraction
+
+When parsing JSON returned by AI/LLM text generation, the system SHALL use a shared candidate extraction utility that supports object and array contracts, Markdown code fences, `<think>` blocks, surrounding prose, and multiple JSON candidates.
+
+#### Scenario: Thinking text contains a mismatched JSON candidate
+- **GIVEN** an AI response contains a `<think>` block with a JSON-like candidate that does not match the business schema
+- **AND** the final response body contains a valid JSON object that matches the schema predicate
+- **WHEN** the business parser extracts the AI JSON response
+- **THEN** it selects the matching final JSON object instead of the thinking candidate
+
+#### Scenario: JSON strings contain bracket characters
+- **GIVEN** an AI response contains JSON with string values that include `{`, `}`, `[`, or `]`
+- **WHEN** the shared extractor scans for balanced candidates
+- **THEN** bracket characters inside strings do not terminate or corrupt the candidate
+
+#### Scenario: Non-AI JSON parsing remains unchanged
+- **GIVEN** code parses stored settings, backups, sync payloads, or HTTP API bodies
+- **WHEN** those payloads are not AI/LLM free-text responses
+- **THEN** the system may continue using direct JSON parsing and does not route them through the AI JSON extractor

openspec/changes/refactor-ai-json-response-parser/tasks.md

@@ -0,0 +1,6 @@
+## 1. Implementation
+- [x] 1.1 Add `llm-json-extractor` pure utility and unit tests.
+- [x] 1.2 Migrate component-side AI JSON parsing entrances.
+- [x] 1.3 Migrate service/tool-side AI JSON parsing entrances while preserving business fallbacks.
+- [x] 1.4 Update focused regression tests.
+- [x] 1.5 Run OpenSpec validation, targeted Vitest, and drawnix typecheck.