Skip to content

docs(stride): define meta-goals within a Todo - #9209

Open
cocolord wants to merge 5 commits into
mainfrom
codex/stride-meta-goal-20261121
Open

cocolord wants to merge 5 commits into
mainfrom
codex/stride-meta-goal-20261121

Conversation

@cocolord

@cocolord cocolord commented Nov 21, 2026 •

Copy link
Copy Markdown
Member

Goal and delivered outcome

Define a meta-goal by the plan, candidate choice, or plan relationship it targets. Both RFC mirrors show how Stride can use grow, switch, and verify to resolve goal conflicts, combine overlapping plans, and compare feasible continuations by expected goal value.

The value comparison counts current utility and time-discounted future utility. Information is valued through improved later decisions. Lower full cost resolves cases with equivalent expected outcomes, risks, and options.

The first acceptance trial uses one scheduler Todo. A matched peak-load comparison informs the next action, and decision followthrough ties that action to the observed result. The cases cover justified growth, testing a different mechanism, and an unresolved choice when the cost bound is reached. A worked example compares direct growth, direct switching, and a probe followed by a choice within the same 10 budget units. It derives the posterior decision and the break-even cost and accuracy.

Author declaration

Written by Codex (OpenAI). The proposal is based on Hierarchical Agent Stride Control §7.6 at d28ec76f4 and the existing Stride probe and followthrough contract.

Criterion Disposition Evidence
Define planning objectives and their relationship to Stride actions Specified in the RFC proposal Matched definition, examples, action semantics, and value comparison in both languages
Define acceptance for one scheduler Todo Specified in the RFC proposal Matched peak-load probe, three followthrough cases, independent checks, and complete costs
Preserve normal work admission and settlement Addressed in the design The proposal uses existing Stride and Todo paths

Scope and continuation

The PR changes the two RFC mirrors. Cross-Todo shared validation and versioned cross-task strategy evolution are follow-up work.

Validation

Local document checks passed on c59a475093a8888ab06c56640b73dbb3b2835d11.

Check Result Evidence
Bilingual RFC diff Passed Staged and committed checks against the verified main base
RFC lifecycle index Passed scripts/generate_rfc_status_index.py --check
Worked example arithmetic Passed Independent exact recomputation of expected values, posteriors, and both break-even points
Public boundary Passed loopx check on both RFC files using an empty isolated registry
Links and diff hygiene Passed Matched equation, section and relative link anchors, git diff --check, and semantic comparison of both mirrors

Frontend and visual evidence

The change is RFC prose.

Type and area

Documentation update in the public architecture RFC. All PR commits include DCO sign-offs.

@loopx-agent loopx-agent left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Reviewer: model_agent · gpt-6.2-sol · OpenAI · runtime_reported · xhigh

动机 / Motivation

设计长程工作策略的操作者要先明确计划层面想达成什么,再判断哪项证据会改变当前选择。 同一目标下直接扩大、测试替代机制和先验证的后续机会不同。没有共同预算和选择准则时,探针的作用很难与最终成果连接。 提案给出可复核的选择规则、单 Todo 的三类验收结果,以及有预算约束的独立算例。 只验收提案、算例与单 Todo 验收设计。代码、实际调度器吞吐和代理采纳尚未由此 PR 验证。 后续试验可用事先匹配的负载条件判断所选路线是否兑现业务验收。

改动思路 / Approach

把元目标放在已有 Todo 的 Stride 计划和理由中,用可负担的探针来决定下一轮扩大、转向或保留未决。 采用既有 Todo、探针和同源回顾能够表达这项试验,无需新增调度授权或执行门。 变更限于两版架构 RFC 的提案,运行时所有者和预算准入沿用既有契约。 把探针成本与延迟计入,信息价值只通过后续选择计算一次。计划合并、冲突处理与选择价值的区分可追溯到原始文献的元目标分类。 Wilensky,§2.2

具体改动 / Changes

准确头 c59a475,分叉基线 55bae0a,当前主干整合预演 55bae0a。2 个文件,+233/-0。 核对的仓库契约 对应 plan-objective 区分元目标与具体问题和动作。single-todo-trial 给出单 Todo 条件与三类结果。same-authority 保留现有证据和授权所有者。逐段核对中英文定义、表格、决策成本、作用域、后续动作与权限。独立校对了仓库内四处链接锚点,并以准确分数复算条件概率、预算消耗、先验/后验决策和盈亏边界。先定义如何试验,把运行效果与是否采纳的结论留到通过试验之后。

关键代码讲解

对主干的风险 / Risks

未发现阻塞代码问题。算例的先验、效用和探针准确率属于设定。公开文档应避免把正净值当成实际吞吐或采纳证据。 对照复算在准确率等于 11/15 或成本等于 2 时与直接扩大持平;低信息量探索没有正的决策收益。正文涵盖证据不足及成本耗尽时由现有归属方处理。 定义规划目标、决策问题和下一动作的生命周期差别。复用已有 next_probe、previous_review、Todo 准入、结算和有权决策所有者。 双语 RFC 检查、RFC 状态索引、四处链接锚点、独立算术和本地 quick premerge 通过。 代入的信号准确率和效用是示例设定。真实峰值下的观测、采用、正确性和完整成本留给所定义的试验。

我的整体评价 / Verdict

批准当前准确头。在各用 10 个预算单位的算例中,独立用有理数复算得到直接扩大 6 分、直接转向 5 分、先探针再选 6.75 分。0.75 分是模拟净增量。中英两版的条件概率和两处盈亏边界一致。 对长程工作,把下一次选择的执行和结果接回证据,而非按探针数量计分。对于此文档,已经具备可重复的选例与验收办法。 保留当前实现边界。变更限于两版架构 RFC 的提案,运行时所有者和预算准入沿用既有契约。

English verdict: APPROVE at c59a475. The RFC defines a bounded one-Todo trial, preserves existing decision authority, and has a correctly recomputed worked example. No runtime or scheduler performance qualification is implied.

This branch has not been deployed

No deployments
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.

2 participants