Skip to content

Shorten the Stride replan example within the existing CLI budget - #8325

Open
loopx-agent wants to merge 1 commit into
mainfrom
codex/compact-stride-authoring-example-20261108
Open

loopx-agent wants to merge 1 commit into
mainfrom
codex/compact-stride-authoring-example-20261108

Conversation

@loopx-agent

@loopx-agent loopx-agent commented Nov 7, 2026 •

Copy link
Copy Markdown
Collaborator

Change

Shorten the example that turn plan supplies when Stride requests a replan. The example still describes the execution, strategy, and evidence axes and a bounded outcome probe. Its fields, actions, current references, authoring rules, and validation retain their existing contract.

Evidence

Normalized JSON characters on the real CLI Base Head
Small task 10,553 10,553
Crowded task requiring replan 21,985 21,867
Multi-agent task 10,300 10,300
Crowded growth over small 11,432 11,314

The existing crowded-growth budget is 11,400 characters (9,300 plus 60 for each of 35 items). Base exceeded it by 32. Head has 86 characters of headroom. This isolates the inherited growth failure recorded in the progressive Goal narrative proposal. #8323

  • The full CLI output-budget smoke passed on head. All 66 focused Python tests, 42 TypeScript tests, and control-plane typechecking passed. The actual CLI on the verified wheel reproduced all three head results and the replan example.
  • The Standard canary ran 4 direct checks and 10 selected checks without a mechanical failure. The change-quality receipt and independent review are pending.
  • The same slice is tracked in C1. It saves 118 characters in that lane's matched crowded case. C1's cross-base smoke fails at the same fixture import before and after this change. #8266

The measurements cover the replan envelope. Live model usage remains open.

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

阅读拥挤 replan 计划的 Agent 需要可用的三轴示例和当前证据绑定。 用真实 File CLI 在三类计划中比较渲染与预算,并把两端实际示例送入原 validator。 小型及多 Agent 计划分别为 10,553 和 10,300 字符。拥挤计划由 21,985 降为 21,867,增长 11,432→11,314,低于原 11,400 上限。118 字符全部来自示例中的 9 处措辞。 独立 PR 只覆盖当前 Stride 示例的措辞和原 CLI 门槛。C1 中同一修改的集成资格另行评估。 验证覆盖真实源码 CLI、源码预算和隔离 validator。尚无模型采用与整轮成本的实测。

改动思路 / Approach

当前策略产生三个轴、五个 next_probe 字段和精确绑定。模板只示范写法,执行与结算仍走现有 owner。 原 authoring contract 定义三个轴、完整探针和当前约束。示例的对象、引用、verify 行为及所有校验字段逐一对照,只需调整示例表达。 相邻整理已限定为示例内容。无需改动共同的 Stride 编解码、外部效果或验证门。

具体改动 / Changes

基线为 8a4d4f9dab17b01d6fe3546936155b4d718fc20d。现有 Vision 协议定义三轴的 plan/reason、verify 的角色、完整 probe 字段和当前绑定。 基线规范

  • 计划门槛。 小型及多 Agent 计划分别为 10,553 和 10,300 字符。拥挤计划由 21,985 降为 21,867,增长 11,432→11,314,低于原 11,400 上限。118 字符全部来自示例中的 9 处措辞。
  • 示例语义。 两端用实际计划提供的示例做原 validator 对照。删除任一轴 plan/reason、五个 probe 字段或两项绑定共 13 种输入均拒绝。无效上下文引用也拒绝。关闭策略时 owner 返回空,当前有效示例通过。

关键代码讲解

拥挤计划才包含此示例。主干重复的示例说明超过既定增长预算 32 字符。候选仅缩短示例中的可复述语句,修复原增长检查。 示例生成 · 原字段回归

对主干的风险 / Risks

压缩只触及示例叶值。独立核对了保持当前方法、一次完整探针、同条件相关负例与完整可归因成本,原规则仍要求依据新增证据决定下一片。

候选完整 CLI 预算 smoke 通过。57 项 Python、53 项 TypeScript 通过。实际 CLI 缺字段回归两端各 4 项通过。 Quick premerge 的 4 项直接检查和 2 项选中检查通过。 DCO 和当前主干合并树核对通过。 未测实际 Agent 采用、后续执行或完整成本。 远端 CI 未取样。

语义与 CI 对齐

主干和候选走同一预算 oracle。原 35 项增长单位和上限逐值比对。

我的整体评价 / Verdict

同意当前提交的示例压缩。 原输出门槛、实际绑定及校验已在此提交上复验。测得的字符差适用于这三个固定夹具。C1 采用相同示例的效果按其完整公开入口单独审查。

English verdict: APPROVE. The Stride example preserves all binding, scope, and evidence requirements and restores the existing crowded-plan growth budget. Model adoption and full-turn cost remain unmeasured.

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.

1 participant