[loopx-agent]

Goal And Delivered Outcome

Complete thin heartbeat prompts with five runtime capability flags, a 320-character scope, Codex App transport and an explicit CLI route emit 4,851 normalized characters. The old 4,500-character presentation allowance incorrectly reports that supported combination as over budget. This changes only that allowance to 4,900; the emitted body and arguments remain identical.

Intended base: main.

Author Declaration

Written by: OpenAI GPT-6.1 Sol, model agent.

Implemented against docs/development/testing-and-quality.md#budget-failure-decisions at a36e01c32b0a290a96466c766629a0b01850f823:

Criterion	Disposition	Evidence
Compare the same complete input before changing a budget	implemented	The real CLI produces the same 4,851-character body with the old and new allowance; all non-budget fields are identical.
Preserve useful instructions, routes and original failures	implemented	Full scope, capability flags and settlement/recovery instructions are asserted; the historical 4,500 failure remains characterized.
Preserve hard limits and detect further growth	implemented	The 4,000-character visible-Goal limit and existing JSON/body fixture ceilings are unchanged; oversized routes still exceed the new allowance.

Scope And Continuation

Complete within the diagnostic-calibration scope. This does not reduce output, change execution authority, or establish model behavior or runtime recovery. The existing bootstrap validator also consumes this allowance: a synthetic complete 4,629-character bootstrap changes from rejection to acceptance, while the larger 13-scope case still rejects with the unchanged error. This is a presentation-policy validation change, not transport or execution authority. Merge qualification remains on hold for the separate standard-runner timeout below.

Validation

• Tested revision: 3b562b1030c1924817e135387b1efc713618a863.
• Run state: completed; merge hold retained. Input classes: synthetic, public_fixture.

Check kind	Result	Evidence / limitation
unit / real_entrypoint	passed	117 tests across prompt support, brief/notification contracts, full-scope JSON, complete capability routes and runtime-profile CLI round trips.
regression_parity	passed	Same real CLI fixture: 4,851 characters and identical body/arguments; only the allowance and its diagnostic result change.
static	passed	Diff hygiene, Python compilation, public-boundary scan and bilingual-RFC gate; no paired RFC changed.
integration	failed	Standard premerge's output-budget smoke timed out at 120 seconds; remaining selected checks passed. The same immutable base a36e01c also times out at the same 120-second outer bound with the same empty output signature. The full owning smoke separately passes in 176.660 seconds with its existing timeouts unchanged. The original timeout and separate merge hold are retained.

Full owning base/head differential: passed, exit 0. The standard 120-second premerge smoke remains a separate pre-existing merge hold; it is not relabeled as passing. No installed-runtime or model-behavior improvement is claimed.

Frontend / Visual Evidence

UI impact: none. No layout, documentation chrome or rendered prompt text changed; screenshots and viewports are N/A.

Type of Change

• Bug fix
• Documentation update
• Test update

LoopX Area

• Control plane

Technical Direction

Existing heartbeat output-budget diagnostics; no new runtime, owner, protocol or vocabulary.

Shared-authority RFC fixture impact

N/A: no authority-provider or migration claim.

Boundary Checklist

• Public diff and description contain no private state, credentials, raw traces or local operating paths.
• No maintainer-owned benchmark work is duplicated.
• The change is scoped to this reproduced diagnostic gap.
• No paired RFC changed; the local bilingual gate passed.
• UI impact is none.
• The commit has a DCO sign-off.

[loopx-agent]

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

使用完整能力声明和长 scope 的心跳调用者会遇到错误的超限诊断。

同一合法调用保留五项能力、320 字符 scope 和显式 CLI 路由，旧预算把 4,851 字符判为超限，新预算按 4,900 字符正确接纳。

已验证正文与所有非预算字段一致，诊断由超限变为合规，更长路由仍报告超限。

本 PR 不缩短提示、不证明模型行为或运行恢复，也不修改执行权限和宿主硬上限。

标准 premerge 的 120 秒 smoke 超时仍是独立合入等待项，完整 owning smoke 已通过。

改动思路 / Approach

唯一薄提示预算已有统一 owner，校准该展示策略比删除有用指令或另建路由特例更简单。

本 PR 只交付薄提示预算、对应回归测试和预算取舍说明；原自动订阅切换修复、业务消息与其他预算不在此 diff。

普通 heartbeat-prompt 先生成完整正文，再由统一预算 owner 归一化路径并计算展示诊断。另一个真实调用者是 --bootstrap：已有校验会读取同一 within_budget，因此本次校准也会接纳原来处于 4,500—4,900 区间的 bootstrap；这项间接行为已独立验证并明确披露。它没有新增权限、重试、持久状态或第二套消息消费者。保留旧上限会继续误报；删减能力、结算或恢复要求则会损害信息价值，另建路由特例也没有必要。

具体改动 / Changes

评审依据为 docs/development/testing-and-quality.md#budget-failure-decisions，固定修订 a36e01c32b0a290a96466c766629a0b01850f823。先读取该版本原有三条规则，再检查本 PR 的解释新增，避免用刚修改的说明为实现自行背书。

原规则	本 head 的处置	实际证据
1 同口径测量	implemented	同一完整 CLI 夹具为 4,851 归一化字符；原 4,500 失败保留，正文原始长度 5,143 与归一化口径分列。
2 信息价值与语义	implemented	正文及所有非预算字段一致；能力、完整路径、结算与未知效果恢复要求均保留。
3 取舍披露和复验	implemented	4,900 留 49 字符余量，增加超长路由反例；4,000 宿主硬上限和冻结资格阈值不变。

关键代码讲解

INTERFACE_BUDGET_CHARS 是已有模式预算的唯一 owner。仅 thin 从 4,500 调到 4,900，其他模式没有跟随增长。代码注释、两个既有期望值和开发说明同步更新，完整 diff 为四个文件 +64/-4，没有引入新的生产模块或公共协议。

build_interface_budget 继续使用原路径归一化和长度比较，返回真实长度与预算判断，不截断正文。生产调用来自 builder 和 bootstrap 校验。新增实际 CLI 测试覆盖五项能力、320 字符 scope、显式路由及更长路由仍超限；独立 bootstrap 对照中，11 个 scope 的 4,294 字符两边均通过，12 个的 4,629 从拒绝变为接纳，13 个仍以原错误拒绝。

对主干的风险 / Risks

最强反例是加预算掩盖了正文丢失或绕过硬边界。相同正文摘要、非预算字段比较、完整指令断言、超长路由反例及既有宿主限制测试共同覆盖它；117 项定向测试通过。直接运行完整 owning smoke 也通过，耗时 176.660 秒，原有内部探针限时未变，没有缩小负载。

标准 premerge 仍为 FAIL：唯一失败项是输出预算 smoke 在 120 秒外层限制下超时，head 记录为 122.621 秒。未修改的固定基线 a36e01c32b0a290a96466c766629a0b01850f823 运行同一命令和 120 秒限制，也在 120.100 秒超时，两边 stdout/stderr 均空。该 smoke 的脚本、负载和两个 240 秒探针未由本 PR 修改，新增两项测试也不在该 smoke 的前置调用列表中；结合完整 smoke 和变更不变量的独立通过，将此项归为 pre_existing_unrelated，但单独保留合入等待。没有用后一次成功覆盖失败，也没有提高外层限时。远端 CI 按当前策略未查询。

语义与 CI 对齐

本次复用既有展示预算词汇，明确区分 ordinary 诊断、bootstrap 校验和 4,000 字符宿主硬上限。没有授予执行额度或动作权限，未改变状态分类、范围选择及结算义务。回滚只需恢复原常量和对应测试说明；历史失败仍需保留。49 字符余量不是所有未来路由都可容纳的保证。

我的整体评价 / Verdict

在 3b562b1030c1924817e135387b1efc713618a863，本次是有用且可独立回滚的 justified_increment：long_horizon 和 user_experience 均 improved，消除了完整调用的误导性尺寸反馈，没有增加人类交互轮次或新的运行状态。一个已有常量承担策略，代码规模与故障范围相称；没有新增默认关闭能力、兼容协议或权限 owner。没有渲染布局或提示正文变化；检查的是 CLI 诊断元数据和同一 bootstrap 校验路径，因此截图不适用。 未验证展示范围为 none；模型遵从、运行部署和性能改进不在此结论中。

APPROVE 仅表示本 diff 的价值、语义和变更不变量已核验；标准 smoke 的既有超时仍是独立 merge hold。当前不据此合入或安装，也不把该校准当作业务任务完成。

English verdict: APPROVE 3b562b1; complete thin-route calibration is validated, with a separate pre-existing 120-second smoke merge hold.