Spec Harness v0.26.0

[0.26.0] - 2026-11-07

Added

• 最高原生推理请求：受控的新会话与并行 worker 统一请求各模型已公布的最高推理档位，并记录控制层和限制；当前宿主仅执行深度工作指令，不把启动参数或提示词冒充模型实际推理效果。
• Maximum native reasoning requests: Managed fresh sessions and parallel workers request each model's highest advertised reasoning level and record control limits; current-host sessions use deep-work instructions without presenting flags or prompts as proof of actual model reasoning.
• 自主问题闭环：可查证事实和范围内可逆选择绑定机器任务，填入结论并通过匹配验证后才能关闭，减少无必要的人工等待；权限、不可逆操作与真实偏好仍由用户决定。
• Evidence-backed autonomous inquiry: Discoverable facts and in-scope reversible choices bind to machine tasks and close only after a recorded conclusion passes matching verification, reducing unnecessary waiting while leaving permissions, irreversible actions, and genuine preferences to the user.

Changed

• BREAKING — 受控启动约束：自定义原始 worker 命令与无法确认最高档位的启动被拒绝，旧链恢复也执行新策略；不支持最高档时不会静默降档，非受控宿主保持如实标注。
• BREAKING — Managed launch policy: Raw custom worker commands and launches without a verified maximum level are rejected, including resumed chains; unsupported maximum levels never silently fall back, and unmanaged hosts retain honest control labels.
• 深度规划与证伪：规划和执行契约统一先调查上下游、比较最小方案、尝试证伪再验证，并复用已有假设与修正记录，避免增加第二套工作状态。
• Deep planning and falsification: Planning and execution investigate surrounding constraints, compare minimal options, seek counterevidence, and verify results using the existing assumption and correction records rather than a second workflow state.

Fixed

• 人工决策保护：新会话、并行 worker 和验证过程共用决策保护；执行中改写、关闭或重分类既有人工问题会留下异常并阻断续跑，强制终止后的未完成保护记录也不会被当作成功。
• Human-decision protection: Fresh sessions, parallel workers, and verification share decision guards; rewriting, closing, or reclassifying existing human questions during execution records failure and blocks continuation, and unfinished guards after forced termination never count as success.
• 结论与证据保鲜：自主问题的结论纳入运行证据指纹，修改问题、选择或理由后必须重验；旧版无分类问题继续保持人工阻断，避免借重新标记绕过授权。
• Conclusion freshness: Autonomous conclusions are fingerprinted into execution evidence, so changing a question, choice, or rationale requires re-verification; legacy unclassified questions keep blocking for human review instead of allowing relabeling to bypass authority.