v0.0.12

v0.0.12（2026-07-24）

🇬🇧 English

Opinion-layer fault tolerance (empty-result detection, single retry, degraded fusion for partial inputs), visual translator fallback second-role, declarative agent preconditions, pipeline stage visualization, unified progress reporting, and 闪电侠 priority parallel shortcut.

New (P1)

• Opinion-layer fault tolerance: empty-result detection, single retry, and degraded fusion for partial inputs. Prevents pipeline stalls when one of 3 parallel opinion agents fails or returns empty.
• Visual translator fallback second-role: when no screenshots, decomposes error logs, ASCII diagrams, long documents, and ambiguous user intent — distinct from tool-handler's raw I/O.
• Declarative agent preconditions: each agent declares activation conditions in precondition metadata, replacing hardcoded routing rules. Router reads preconditions instead of encoding conditional logic inline.
• Pipeline stage visualization: routing decisions output stage identifiers (工具层|意见层|融合层|实现层), mapping to user-facing pipeline phases.
• Unified progress reporting: both success and failure paths follow [Pipeline] mode=... stage=... status=... format with five statuses (idle|in_progress|complete|degraded|stuck). Agent names never exposed.
• 闪电侠 priority parallel shortcut: when the main pipeline is executing, 闪电侠 can be dispatched in parallel for independent simple subtasks (e.g., searching TODOs while architecture plans are being designed). Limited scope: same-dependency and opinion-fusion tasks excluded.
• 残差提取者 supports partial inputs (marks missing perspectives, adjusts consensus coverage denominator to N/M).
• 旗舰·融合 supports degraded fusion (2/3 inputs: skip missing perspective, fuse available ones; 1/3 input: pass-through with confidence penalty).
• 门童路由员 added result validation step, [Partial] N/3 inputs labeling, and precondition-based activation. README.md / README.zh.md updated with new sections.

Fix (P1)

• Pipeline no longer stalls (STUCK) when a single opinion agent fails — degraded mode keeps the pipeline moving.
• 融合·保底 now also covers opinion-layer partial failure cases.
• 视觉翻译官 no longer idles when there's no screenshot — provides value through content decomposition.

Model (P1)

• flag-fuse (旗舰·融合) model changed qwen3.7-max → kimi-k3 (2.8T params, 1M context)
• README model table + intro updated; OpenCode Go quota 2x from 7/24 (140 → 280 / 5h)

---

🇨🇳 中文

意见层容错（空结果检测、单次重试、部分输入降级融合）、视觉翻译官降级第二职责、声明式 Agent 预条件、流水线阶段可视化、统一进度报告格式、闪电侠优先并行快捷通道。

新增（P1）

• 意见层容错：空结果检测、单次重试、部分输入降级融合。当3条并行意见链中任意一条返回空或超时时，流水线不再卡死，而是以降级模式继续。
• 视觉翻译官降级第二职责：无截图时解构错误日志、ASCII 图、长文档和含糊用户输入，与工具人的原始 I/O 形成能力互补。
• 声明式 Agent 预条件：每个 Agent 在 YAML frontmatter 的 precondition 元数据中声明激活条件，替代硬编码路由规则。路由器读取 preconditions 而非编码条件激活逻辑。
• 流水线阶段可视化：路由决策输出阶段标识（工具层|意见层|融合层|实现层），对应用户可理解的流水线阶段。
• 统一进度报告格式：成功与失败路径均遵循 [Pipeline] mode=... stage=... status=... 格式。五种状态（idle|in_progress|complete|degraded|stuck）让用户无论流水线成功还是失败，都能以相同格式理解当前状态。不暴露内部 Agent 名。
• 闪电侠优先并行快捷通道：主流水线执行期间，闪电侠可并行处理独立的简单子任务（如架构方案设计中顺便搜索 TODO）。限定范围：与主流水线有依赖关系或参与意见融合的任务除外。
• 残差提取者支持部分输入（标记缺失视角，共识覆盖率分母调整为 N/M）。
• 旗舰·融合支持降级融合（2/3 输入：跳过缺失视角，融合可用方案；1/3 输入：直通并施加置信度惩罚）。
• 门童路由员新增结果校验步骤、[Partial] N/3 inputs 标注、基于 precondition 的条件激活。README.md / README.zh.md 新增章节。

修复（P1）

• 单条意见代理失败时流水线不再卡死（STUCK），降级模式保持流程继续。
• 融合·保底现在也覆盖意见层部分失败场景。
• 视觉翻译官在没有截图时不再闲置，而是通过内容解构持续提供价值。

模型（P1）

• flag-fuse（旗舰·融合）模型由 qwen3.7-max 改为 kimi-k3（2.8T 参数，1M 上下文）
• README 模型表与介绍同步更新；OpenCode Go 额度 7/24 起翻倍（140 → 280 / 5h）

---