Skip to content

RLCR methodology feedback: stale stop-gate stall, pre-commit secret gate, and non-code deliverable mode (6 suggestions from a completed session) #222

Description

@ZziTaiLeo

Voluntary, fully sanitized methodology feedback from a completed RLCR session (auto-analysis phase). No project-specific information is included.

TL;DR (EN)

A documentation/investigation-type plan (zero source-code changes allowed) ran through the RLCR loop: the entire plan completed in round 0, plus one review-phase fix. Output quality was good (goal anchor prevented drift, deviations were honestly logged, adversarial review caught a real credential leak AND refuted false positives). Two structural gaps dominated the experience:

  1. Stop-gate stalled silently for ~23 hours on a stale background-task marker combined with a session-identifier mismatch, until a human manually removed the marker. The actual work took minutes-to-hours; the handoff state machine cost a day.
  2. A plaintext credential reached committed evidence artifacts and was only caught by post-hoc code review, forcing a local history rewrite (safe only because nothing had been pushed).

Plus four more suggestions: a deliverable-type switch for non-code plans; quarantining generated evidence before it enters history; adaptive iteration budget with a single-round fast path; explicitly distinguishing verification-review from adversarial-review.

Suggestions (detail, zh-CN)

1. 为「非代码交付物」增设专门模式,替换代码专属门禁

模式:面向代码变更的循环被用在纯文档/调查型交付上,收尾的「代码精简」、测试产出、代码引用核对等内建阶段沦为空转或需执行者临场替代。
改进:循环启动时增加交付物类型开关(代码/文档/调查/混合);非代码类型自动以「引用完整性/证据链完整性/事实核验/产物清单齐备度」等价门禁替换「测试通过/代码精简/源码 diff 检查」。

2. 给停止判定加入陈旧检测与心跳,消除单点空转(优先落地)

模式:停止判定被陈旧后台任务标记 + 会话标识不匹配同时卡住,静默空转近一天,只能人工删除标记恢复。
改进:(a) 阻塞性标记带时间戳与 TTL,超时判定陈旧并跳过或告警;(b) 会话标识不匹配应显式检测上报(或自动对账),而非静默等待;(c) 停止判定挂起超过阈值(如 30–60 分钟)时主动发出停滞告警。核心原则:循环绝不应无声卡在自己无法自愈、也不报警的状态里。

3. 隐私/凭据检查从「事后审查」前移为「提交前门禁」(优先落地)

模式:防泄露规则存在于手册,但只在捕获后靠人工审查发现;泄密内容已入本地历史,补救被迫历史重写。
改进:对循环生成且会入库的证据/产物,增设提交前密钥/凭据扫描门禁,未通过则阻止提交——把脱敏规则从「靠人自觉」变为「机器强制」。

4. 生成的证据产物在通过安全门禁前不进入主线历史

模式:证据直接提交进主线历史后,一处泄密迫使整段历史重写。
改进:证据默认先落暂存/隔离区,通过隐私与完整性门禁后再纳入正式历史;与建议 3 互补(一个防发生,一个降补救代价)。

5. 迭代预算自适应,识别「单轮即可完成」的计划走快车道

模式:循环配置了数十轮预算与「第 N 轮全量深审」检查点,而计划第 0 轮即整体完成——深审门槛永远到不了。
改进:探测计划是否可单轮完成,若是则在第一轮后立即触发延后的深审门禁;迭代上限按计划体量自适应。

6. 显式区分「核验型审查」与「批判型审查」

模式:轮间审查实际扮演核验/记账角色,真正的对抗性批判来自执行期内嵌证伪与审查阶段的泄密捕获;两种目的混在同一「审查」名义下,易使「审查通过」被误读为「已被对抗性挑战过」。
改进:方法论中显式区分两类审查,审查小结标注类别;把执行期对抗证伪制度化为标准步骤,确保批判型审查尽早作用于最终交付物。

Overall

产出质量层面循环表现优秀(目标锚防漂移、偏差诚实留痕、审查抓真且剪伪、轮数精简成比例);短板集中在流程贴合度与状态机健壮性。建议优先落地 #2(停止判定健壮性)与 #3(隐私前移门禁)。

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions