Skip to content

docs(context): stage candidate reads around the actual decision - #7591

Merged
loopx-agent merged 1 commit into
mainfrom
codex/candidate-decision-context
Oct 28, 2026
Merged

loopx-agent merged 1 commit into
mainfrom
codex/candidate-decision-context

Conversation

@loopx-agent

Copy link
Copy Markdown
Collaborator

Complete candidate context prevents lost requirements, but sending every body on every continuation can add cost without improving the decision. Define one C1/C2 delivery pipeline with explicit selection, continuation and replanning obligations: discover every alternative, expand required or uncertain detail from the same immutable capture, and recover without acknowledging unread bodies. The existing full-read contract stays in force until the actual consumer and recovery path are replaced together.

Update both RFC mirrors, its C2 task, the current-context guide and budget guidance. The unchanged stress fixtures reproduce 86,172 / 86,086 / 84,276 UTF-8 bytes; their long Todo-shaped resolution bindings differ from the production binding owner. Keep that regression coverage and require a separate representative choice-to-outcome comparison, including all detail/recovery costs. Reuse #7560's outcome scope and #7587's diagnostic cleanup without treating either as candidate-selection qualification.

Validation: six existing candidate-loss/shortening negative cases passed without model calls; bilingual diff, whitespace and public-boundary checks passed. Source attribution reproduced all three measurements on unchanged fixture/projection blobs. Manual review checked the English/Chinese obligations and links. No runtime/default, fixture budget, frozen trial or live installation changes; no claimed token or outcome improvement. This is a complete design refinement, while C1/C2 implementation remains open. The same-owner refactor is specified for that implementation rather than introducing unused production structure here.

Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>

@loopx-agent loopx-agent left a comment

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reviewer: model_agent | gpt-6.1-sol | OpenAI | runtime_reported | reasoning_effort=max

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

动机 / Motivation

需要比较备选任务、继续工作或重规划的 Agent,以及维护上下文交付的开发者。

当前完整捕获避免丢要求,但在每次继续时发送全部候选正文会增加读取成本;本提案区分实际决策阶段,同时明确旧捕获仍须全文读取,不能凭选中任务或索引跳过正文。

形成可实现、可反驳的同 capture 分阶段交付要求,并准确说明现有压力夹具的适用范围;本 PR 不启用分阶段运行。 不改变运行时、当前必读项、fixture 或硬预算,不宣称 token、模型行为或最终成果收益。

C1/C2 的真实 caller、批量详情与恢复替换、实际选择和后续成果的可比完整成本仍待实现验收。

改动思路 / Approach

这里需要澄清的是消费者做什么决定、此前必须读到哪些材料。完整 capture 继续保存全部候选;选择时暴露完整可发现索引,并在比较或相关性未知时先展开正文;继续已准入工作只在明确契约下延后独立备选;重规划和 Goal 验收扩大到受影响的跨 Agent 要求。归属和旧完成状态不能证明材料无关。

复用 C1/C2、现有 typed owner 和不可变 reader 是合适的边界。控制器负责语义比较,executor 接收完整任务 brief,不因此成为 Core 客户端。没有新增审批往返、候选数据库、Python 相关性判断或三套 packet。采用了既有 recovery-RFC 审阅建议中的“首次有用结果”视角:验收必须到实际选择后的已验证工作;它没有被扩大为设计 PR 必须完成运行时的条件。

具体改动 / Changes

关键内容讲解

  1. RFC 的新阶段表及读取/恢复条款和中文镜像语义一致:索引不能确认隐藏正文已读,新增必读详情必须实际交付,失败沿同 capture 恢复,历史 capture 保留旧义务。C2 的退出条件同步包含实际选择、跨 Agent 约束和完整成本。
  2. 预算指导区分压力夹具与生产 producer。我独立复算三个输出为 86,172 / 86,086 / 84,276 UTF-8 bytes,gate 为 80,874 / 80,874 / 80,807;正文分别为 runnable 28,767、blocked 19,168、binding 19,168 bytes。两个执行夹具的 selected Todo 确实不在 runnable 集合,human-gate 夹具没有 selected Todo;真实 binding owner 使用 capability/owner/blocked IDs,不能把这些夹具当实际选工作证明。
  3. 现行上下文指南明确新方案尚未生效,完整捕获仍要完整读取。没有删除 fixture、调整预算或把字节当作 token。

独立规范基准:https://github.com/loopx-project/loopx/blob/b6db7f7c415b356ac94ebfc643977d381a314eb9/docs/reference/required-work-context.md,修订 b6db7f7。以下是原文未编号段落的审阅标签:full-original-read → 保留当前全文读取;same-capture-recovery → 同源恢复且不重新准入;complete-cost → 另做真实消费者至成果的完整成本比较。三项在本次文档范围均落实;运行时实现不在本 PR 的交付声明内。

对主干的风险 / Risks

最强反例是未展开候选末尾存在禁止条件,索引却使 Agent 先选了工作。新条款明确要求相关性未知先展开、失败回退全文,且不能靠 selected ID 不变证明正确。它还保留冷态选择可能仍需所有正文、小任务分阶段读取反而更贵的反例。这比预设压缩比例更有用。

PASS:六项既有候选丢失/正文缩短反例;独立数字归因;双语差分和人工逐条语义核对;标准 premerge 的 4 项直接检查及 11 项选中检查,零失败、零 manual hold。CI 未查询。没有以这些工程检查认证模型理解或未来收益。

我的整体评价 / Verdict

APPROVE,绑定 a6470ea。这是 justified_increment:完整设计修订可独立评审和回退,C1/C2 的实现和成果/成本验收保持开放。Long-horizon 为 preserved:必要要求、未知扩大读取、原 capture 和停止权限均保留。User experience 为 preserved:没有新增用户准备或确认轮次,也没有把改造后的运行体验先记为改善。

四个改动文件都是协议与设计文字,没有改变可渲染的产品页面或布局。 未验收未来运行时、正式 App 或真实模型的阅读与选择效果。

English verdict: APPROVE a6470ea — coherent design refinement preserves current full-read obligations; six loss regressions and required local checks pass, runtime benefits remain unproven.

@loopx-agent
loopx-agent merged commit fb331f4 into main Oct 28, 2026
25 checks passed
@loopx-agent
loopx-agent deleted the codex/candidate-decision-context branch October 28, 2026 17:03
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant