Skip to content

RFC: converge mechanisms around method feedback and learning - #8749

Merged
loopx-agent merged 4 commits into
mainfrom
codex/mechanism-consolidation-rfc-20261114
Nov 14, 2026
Merged

loopx-agent merged 4 commits into
mainfrom
codex/mechanism-consolidation-rfc-20261114

Conversation

@cocolord

@cocolord cocolord commented Nov 14, 2026 •

Copy link
Copy Markdown
Member

LoopX already has method choices, predictions, execution observations and followthrough records. Mechanism convergence should connect them so evidence improves subsequent work and accepted outcomes within comparable complete budgets. This bilingual RFC defines that decision loop together with the execution ownership and structural migration it needs.

  • Center the target architecture on then-visible state and remaining budget → feasible conditional continuations and prior expectations → actual execution → feedback testing those expectations → revised continuation.
  • Audit seven responsibilities and five existing paths. The current Agent or explicitly selected Controller chooses methods; Stride, Yardstick/Explore, Decision Context and Core supply their distinct specifications, evidence and authority.
  • Define real structural convergence: reshape internal APIs when necessary, migrate callers together, and retire duplicate normal producers. Historical readers and public/durable-state compatibility retain explicit consumers and recovery.
  • Stage policy improvement through current-model feedback, forecast calibration and evidence-supported value/allocation learning. Separate forecast accuracy, adoption, causal comparisons and matched-budget outcomes; include data, inference, deployment and maintenance costs.
  • Treat Reward Memory and Skills as knowledge/procedure carriers. A decision-application receipt is an optional repair for a demonstrated caller-specific relation gap, not a prerequisite for the first useful loop.
  • Walk through a synthetic budget-sensitive choice, counterevidence changing real work, and an unchanged terminal result. Preserve prior forecasts, late/failed/unknown outcomes, original authority and once-only recovery.

中文主线已改为“状态与预算—选择与预期—执行—反馈—下一决定”,并同步迁移、回退、验收与演进顺序。本文复用现有预算规划、动作价值证据和 Controller 规范;新增设计明确标注为 AI 辅助提案,待维护者核实确认,不归为任何个人已确认观点。

Source audit/integration baseline: 8d7d234c6f55fa45fa5db0ad5a156dcaaf726d32. Open #8266, #8484 and #8516 are separately pinned in the inventory; their runtime delivery and test claims are not borrowed.

Validation on a505ddc5bae32aacc06a04fb0fec36ce39c26a92:

  • Existing docs-governance smoke, generated status-index check, staged/exact-head bilingual gates and diff whitespace check passed.
  • Strict MkDocs build passed. Both rendered pages contain the expected tables and resolve the four planning/action-value anchor links. Existing informational link diagnostics elsewhere remain outside this change.
  • Local/source links, English/Chinese heading/table and diagram-topology parity, semantic consistency and public-boundary scan checked.
  • Relative to the integration base, only the RFC pair, catalog and bilingual status indexes change. The update from the prior PR head changes three documentation files.

This is a design-only PR. No runtime behavior, defaults, training, experiments or Skill publication change; no new empirical method or learning benefit is claimed. Design acceptance, implementation and effect qualification remain separate.

Signed-off-by: lusendong.6789 <lusendong.6789@bytedance.com>

@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-6.1-sol · OpenAI · runtime_reported · reasoning_effort=max
Reviewed commit: bf96cbdd36745d5f3b9fa0bc3a33142cdd7de2bf (merge base 1cdd9bbefd1d849ac8aa733bd3d80bc13ad0b722).

[P2] 合入后两份 RFC 索引的 Accepted 计数会过期。 在主干已合入 #8752 的版本上合并当前提案。 两个 RFC 都把计数从 50 改成 51。Git 的合并树保留了两行新条目,两个计数仍是 51,索引生成器期望 52 并拒绝通过。 把分支更新至当前 main,再用生成器重建两份 STATUS 索引,重新运行索引检查和 docs-governance smoke。 索引计数 主干 #8752

动机 / Motivation

维护者要决定哪些执行路径值得合并,以及一次任务里学到的方法能否帮助下一次工作。 准入、恢复和结算分散在不同入口的规范里,经验能被召回却可能缺少后续采纳及结果证据。 例如研究任务的代理指标上升、独立验收停滞,下一回合需要校准方法,并让同类任务能检验这次经验是否有效。 如果只看回执或召回记录,后续工作可能沿用无效的方法并继续消耗试验、模型和人工时间。 新增 RFC 把路径收敛和经验复用放在同一份设计里,沿用各项已接受规范的决策权。 设计要求追踪实际采用、独立结果和完整成本,保留当下未证实的收益。 本次只新增双语规范、目录和索引,实施和实测仍由已有规范所列的工作承担。

改动思路 / Approach

让现有 Controller 规范继续管准入和结算收敛,Target/Stride/Explore 继续管选择和测量,Reward Memory 负责有范围的经验复用。直接把所有判断放进一个流程会造成重复权限,也会把方法有效、成果验收和经验贡献混为一谈。一份轻量关系说明有价值,因为维护者可以沿原始证据检查后续工作及成本,再决定是否扩展。 执行规范 经验效用规范

具体改动 / Changes

审阅依据为 scripts/generate_rfc_status_index.py,修订 e96857f62d3cd546b215ed3d0823f877d5cbb29c。双语正文、目录和两份生成索引是全部改动。

  • decision_ownership 已体现在设计中。§5 保留原准入、历史恢复和结算身份,继续由现有负责人推进收敛。 机制映射
  • experience_evidence 已体现在设计中。§5.3 和 §9–§11 要求看到后续实际工作、独立结果和成本。召回记录与成功结果相邻时,效用仍需要归因依据。 验收表
  • effective_rfc_indexes 未满足。主干和提案都新增一项 Accepted RFC,自动合并后两种语言的总数只增长一次。 生成规则

对主干的风险 / Risks

PASS · 独立 PR 头通过生成索引、双语校验与文档 smoke,98 处新增本地链接在审阅树上可以解析。主干更新到 859a23e 后重新模拟合并,双语索引仍复现同一失败。 在固定审计修订核对了 39 个被引用的源码或能力路径,核对了现行默认继承、结算边界与经验效用约束。

FAIL · 两个 RFC 都把计数从 50 改成 51。Git 的合并树保留了两行新条目,两个计数仍是 51,索引生成器期望 52 并拒绝通过。 两个 RFC 名目均在索引中;本地重新生成双语索引后检查通过。对当前主干、PR 头、未修改的模拟合并树分别运行索引检查,随后在模拟合并树中重建并复查。

UNVERIFIED · 文档与生成索引在本地验证。实际部署、跨任务应用和净收益需要后续按原规范验证。 托管 CI 未查。

我的整体评价 / Verdict

REQUEST_CHANGES。设计保留了长程反馈的归属和验证条件,从独立验收得到的负反馈仍能用于下一次合法决策。当前合并会让用户读取的目录计数和文档校验失真。查看了 Markdown RFC 和生成索引,改动未触及 App、CLI 或 Lark 界面。把分支更新至当前 main,再用生成器重建两份 STATUS 索引,重新运行索引检查和 docs-governance smoke。

English verdict: REQUEST_CHANGES at bf96cbd. Regenerate both RFC status indexes after updating against current main. The clean merge keeps both new Accepted entries but preserves a count of 51 instead of 52, so the repository's index check fails.

Comment thread docs/architecture/rfcs/STATUS.md Outdated
[中文版](STATUS.zh-CN.md) is the semantic mirror of this file.

## Accepted (50)
## Accepted (51)

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.

[P2] 合入后两份 RFC 索引的 Accepted 计数会过期。 在主干已合入 #8752 的版本上合并当前提案。 两个 RFC 都把计数从 50 改成 51。Git 的合并树保留了两行新条目,两个计数仍是 51,索引生成器期望 52 并拒绝通过。 把分支更新至当前 main,再用生成器重建两份 STATUS 索引,重新运行索引检查和 docs-governance smoke。

Signed-off-by: lusendong.6789 <lusendong.6789@bytedance.com>
…8749-architecture-20261114

Signed-off-by: lusendong.6789 <lusendong.6789@bytedance.com>
Signed-off-by: lusendong.6789 <lusendong.6789@bytedance.com>
@cocolord cocolord changed the title RFC: converge execution mechanisms and qualify experience reuse RFC: converge mechanisms around method feedback and learning Nov 14, 2026

@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-6.1-sol · OpenAI · runtime_reported · reasoning_effort=max

复审提交 a505ddc5bae32aacc06a04fb0fec36ce39c26a92,目标基线和实际合并基点 8d7d234c6f55fa45fa5db0ad5a156dcaaf726d32。原索引阻断已经修复:当前真实基线及合并树无冲突;英文和中文 Accepted 均为 52,索引生成检查、文档治理 smoke 与 RFC 配对检查通过。

动机 / Motivation

当代理指标还在上涨、权威验收没有改善时,研究或工程执行者需要判断是继续搜索、先排查指标,还是留够时间做最终确认。 旧路径各自记录计划、结果或召回;单看这些记录,维护者很难判断一次反证是否改变了下一步实际工作。新提案要求把当时的预期、执行、独立结果和下一次选择串起来。 缺少这条联系会重复投入预算,也容易把存在回执或预测更准误判成实际成果变好。 设计给出了可验收的分工和一个具体例子:剩余 90 分钟可先验证反例;如果只剩 20 分钟、确认需 25 分钟,则保留已验收产物,按原规则继续可行工作或停止。 这五个文件只交付设计及中英文索引;现有产品、运行时、权限、默认值和预算不变,不启动实验、训练或发布 Skill。 仍待现有 Target/Stride/Explore owner 在一个真实任务里读回方法采用、独立结果与后续工作;同预算收益要另由原实验 owner 做匹配验证。

改动思路 / Approach

Quota 仍决定是否准入,Turn 持有执行、恢复和一次结算;当前 Agent 或已选择 profile 的 Controller 比较方法,由原 owner 核验范围和结果。 §5 的调用顺序把目标、当时证据、预期、实际执行与下一次选择联系起来。以前的记录和选择者足以启动第一轮;只有具体调用者无法关联经验和决定时,才考虑补 application 证据。当前任务的反馈不绕长期记忆;预测校准、经验采用与策略收益分别评价。

具体改动 / Changes

规范依据:docs/architecture/rfcs/TEMPLATE.md,修订 8d7d234c6f55fa45fa5db0ad5a156dcaaf726d32。

  • design_not_delivery:中文与英文均保留 Accepted 的设计状态、Proposal 的交付状态;§12 的用户、预算、迁移方案与学习资格仍由原归属确认。
  • synchronized_index:两种语言均同步登记。独立用索引生成器核对 head 和实际合并树,保留 #8752 与本 RFC 的两条新增行。
  • §5.3–5.6给出实际做法,包括反馈未支持原预期如何缩小验证范围,以及预算不够时为何不能预报成果。C1/C4/C9 后续迁移需要真实消费者和旧路径退出证据;不把重复路径永久藏在统一门面下。

对主干的风险 / Risks

旧问题:旧版与 #8752 同时各把计数从 50 写到 51,合并树留下两行新 RFC,但生成器期望 52。 当前 head 与合并树已通过生成索引校验与 docs-governance,且 RFC 配对检查通过。旧阻断重现后核对了当前合并树;负的设计反例也得到明确处理:确认要 25 分钟但只剩 20 分钟,不能把未确认的候选计入成果;迟到结果归原实验,pending 不作零,未选方案收益未知。 §9保留旧回执语义、一次结算与独立验证,要求冻结匹配对照以后再判断收益。这是接受为可认领的设计依据的评审。实现、实际反馈采用、学习效果及生产资格未测;依据 Goal 策略未查询托管 CI。

我的整体评价 / Verdict

APPROVE。作为独立设计切片,文档能解释什么证据应改变下一步,也明确了何时可以有依据地继续;所有改动为 RFC 和索引,现有长程准入、执行与产品体验保留。仍待现有 Target/Stride/Explore owner 在一个真实任务里读回方法采用、独立结果与后续工作;同预算收益要另由原实验 owner 做匹配验证。 这是接受为可认领的设计依据的评审。实现、实际反馈采用、学习效果及生产资格未测;依据 Goal 策略未查询托管 CI。

English verdict: APPROVE at a505ddc for the design-only RFC. Both generated indexes and the bilingual gate pass against the current integration tree. Implementation, measured gain and production activation remain separate.

@loopx-agent
loopx-agent merged commit 4d23178 into main Nov 14, 2026
28 of 40 checks passed
@loopx-agent
loopx-agent deleted the codex/mechanism-consolidation-rfc-20261114 branch November 14, 2026 16:52
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