diff --git a/mateclaw-server/src/main/resources/docs/en/managed-json-acceptance.md b/mateclaw-server/src/main/resources/docs/en/managed-json-acceptance.md index 9de8b1a7..6f39b611 100644 --- a/mateclaw-server/src/main/resources/docs/en/managed-json-acceptance.md +++ b/mateclaw-server/src/main/resources/docs/en/managed-json-acceptance.md @@ -50,3 +50,5 @@ V197 makes expiry authoritative in epoch seconds, independent of the JVM/JDBC ti The host clock, database credentials and service host remain trusted. Run arbitrary external code without service/database credentials and outside the service's storage permissions if it is not trusted; setting a working directory or scanning paths does not provide OS isolation. Use coordinated backups of the database for requirements, bodies, pointers and bindings; workspace-file backups alone cannot restore managed acceptance. V198 also stores absolute scheduler lease deadlines. Existing leases expire during upgrade and are recovered from their persisted checkpoints: safe work may receive a new attempt; uncertain side effects remain blocked for review. Expired owners cannot renew, checkpoint, settle or use managed JSON tools. Renewal checks both current lease records after acquiring the goal lock, so a delayed scheduler tick cannot reuse an old timestamp to revive its owner. New valid owners can continue under the existing requirements. + +Validation snapshot (2026-09-14): the full default backend test run passed 5,249 executed tests with 33 conditional skips; the frontend passed 386 tests. The managed contract also has real compiled ReAct/Plan graph tests for account and scheduled-owner execution. Their model choices and semantic verdicts are controlled fixtures, not online-model benchmarks. Specialized integration profiles and the proprietary Kingbase engine are outside that full-default-suite claim. diff --git a/mateclaw-server/src/main/resources/docs/zh/managed-json-acceptance.md b/mateclaw-server/src/main/resources/docs/zh/managed-json-acceptance.md index aaf7a62e..92b82848 100644 --- a/mateclaw-server/src/main/resources/docs/zh/managed-json-acceptance.md +++ b/mateclaw-server/src/main/resources/docs/zh/managed-json-acceptance.md @@ -50,3 +50,5 @@ V197 使用 epoch 秒作为有效期依据,不受 JVM/JDBC 时区变化影响 宿主时钟、数据库凭据和服务宿主仍属于可信基础。若外部任意代码不可信,应在没有服务/数据库凭据、没有服务存储权限的隔离环境执行;仅设置工作目录或扫描路径不是操作系统隔离。备份应协调保存数据库中的要求、正文、指针和绑定,仅备份工作区文件无法恢复受管验收。 V198 同样以绝对时间保存调度租约截止。升级时旧租约失效,按已有检查点恢复:安全工作可建立新 attempt,不确定副作用仍阻断并要求核实。过期 owner 不能续租、提交检查点、结算或使用受管 JSON 工具。续租在获得 Goal 锁后检查两侧当前租约,迟到调度 tick 不能利用旧时间戳复活 owner;新的有效 owner 可按现有要求继续。 + +验证快照(2026-09-14):默认后端完整测试集实际通过 5,249 项、条件跳过 33 项,前端通过 386 项。受管契约还覆盖真实编译的 ReAct/Plan 图及账户/调度 owner 四种组合;模型选择和语义评价是受控夹具,不是在线模型基准。专门集成 profile 与 Kingbase 专有引擎不包含在“默认完整测试集”结论中。