Skip to content

fix(release): verify runtime qualification before publication - #8922

Merged
loopx-agent merged 1 commit into
mainfrom
codex/runtime-publication-qualification-20261117
Nov 17, 2026
Merged

loopx-agent merged 1 commit into
mainfrom
codex/runtime-publication-qualification-20261117

Conversation

@loopx-agent

Copy link
Copy Markdown
Collaborator

Goal And Delivered Outcome

The runtime publisher can advance runtime-main when the downloaded installation qualification is missing, failed, for another revision/channel, or no longer matches the archive. The publish job currently uploads that file without consuming it.

Validate the existing qualifier's results and exact archive identity before the first GitHub operation. Invalid evidence leaves remote state untouched; qualified publication, immutable retries, lineage checks and historical assets retain their existing behavior. Intended base: main.

Author Declaration

Criterion Disposition Implementation / evidence
Installation and recovery checks must pass for the published runtime implemented Consume every existing check as an actual JSON boolean; reject missing/failed/truthy substitutes before any remote call.
Qualification must describe the candidate bytes and scope implemented Existing qualifier emits archive digest and channel; publisher checks revision, channel, digest, byte count, actual archive and frozen metadata.
Preserve valid publication and recovery behavior implemented Existing 14 publisher cases still pass; current main and C1 baselines pass the same valid/failure-recovery cases.

Scope And Continuation

Complete within this engineering scope. The existing producer and publisher remain the owners; Python handles artifact I/O at the established build boundary. No installer, control-plane decision owner, capability, activation switch or release-policy framework is added. A further extraction was unnecessary for the sole publisher caller.

This validates installation evidence. It does not establish model outcomes, complete-cost benefit or C1 default-adoption qualification; that work remains open in #8266. Stable distribution and desktop feeds are outside this change.

Validation

  • Tested revision: 874c0c16fe5666c9553493a6d34e45c854ec4366
  • Run state: finished
  • Input classes: synthetic, public_fixture
Check kind Result Public-safe evidence / limitation
regression_parity passed python -m pytest -q tests/test_runtime_snapshot_publication.py: 26 pass. Identical final fixtures against immutable main d25e044c8c and C1 4267014484 expose all 12 invalid-evidence false acceptances; all 14 existing cases pass on both baselines. Fixed head passes all 26. Tests execute production Bash; only GitHub is replaced with the existing filesystem peer.
integration passed Existing update-artifact, update-operation, self-update-activation and archive-installer tests: 152 pass, 3 platform skips. Skips do not qualify native Windows.
real_entrypoint passed scripts/chat_bundle.py ensure, scripts/runtime_artifact.py --channel main, then scripts/qualify_runtime_artifact.py: exact committed archive; actual isolated clean install, upgrade, doctor, skills, frontend HTTP, forbidden build tools and injected activation failure restoring the previous executable.
real_entrypoint passed Actual archive and generated qualification through production publisher preflight on Python 3.14.8 and 3.11.17. Valid evidence reaches a deliberately refused GitHub boundary; missing/wrong-revision/wrong-channel evidence makes zero GitHub calls. No remote release or channel was changed.
static passed Changed Python Ruff; configured mypy (17 files); four-file public-boundary scan; diff hygiene and RFC-pair check.
integration passed Standard local premerge: direct checks, changed Python compilation and 19 selected canaries. Goal strict-receipt validation is recorded for the exact scope.

Coverage: real artifact production and isolated POSIX installation are covered; remote publication is exercised through a local GitHub peer. This does not claim GitHub deployment, native Windows/macOS desktop signing, model qualification or final release readiness.

Frontend / Visual Evidence

  • UI impact: none. This changes artifact publication and its contributor documentation; no product viewport or interaction changes.

Type of Change

  • Bug fix
  • Documentation update
  • Test update

LoopX Area

  • Build, packaging, installer, or CI

Shared-authority RFC fixture impact

N/A: authority persistence and routing are unchanged.

Boundary Checklist

  • No private state, credentials, raw traces, internal links or local machine paths in the diff or this body.
  • No duplicated maintainer-owned benchmark work.
  • Scoped to the reproduced runtime-publication defect.
  • No paired RFC changed; bilingual diff check passed.
  • UI impact is none.
  • Every commit has a DCO sign-off.

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; reasoning effort max; runtime_reported.

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

动机 / Motivation

维护者需要确保 main 运行时渠道只接收已经通过安装与恢复检查的候选。旧版 publisher 只上传 qualification 文件,缺失或错误的回执仍可推进渠道;现在先核对全部布尔检查、候选身份和实际归档,拒绝无效证据后才允许进入远端步骤。同一真实归档与生成回执被预检接受;十四类无效输入在当前 Bash 中触发零 GitHub 调用,原主干仍会进入远端步骤。这次只修工程发布资格,不授予发布权限,也不证明 C1 模型成果、完整成本收益或默认采用资格。

改动思路 / Approach

既有 qualifier 生成证据,publisher 在远端写入前核验绑定,职责清楚;单个消费入口无需引入新的发布策略框架。当前 PR 只补 main 运行时安装资格的消费门,并保留原发布与重试行为;C1 成果、完整成本及渠道采用仍单独验收。

构建端仍完成实际安装、升级、doctor、skills、frontend HTTP 和故障恢复,然后从同一 manifest 派生 channel 与 digest。下游把这份证据与冻结元数据、实际归档字节绑定。成功后才进入原有的不可变 snapshot 上传、lineage 判断和 channel pointer 更新;没有把可选 observer 升格为控制面,也没有另建状态或重复执行安装。更小的“继续只上传回执”无法修复实际错误放行,而抽成新框架会给单个消费点增加维护面。

具体改动 / Changes

完整评审 head 为 874c0c16fe5666c9553493a6d34e45c854ec4366,四文件 +93/-3。主要生产变化是现有 workflow 中的预检,qualifier 仅多返回同源的 channel/hash;测试增加十二类无效证据,文档用中英说明工程资格边界。

关键代码讲解

  • Publish immutable snapshot then channel pointer 在第一个 gh 操作之前验证40位 revision、十个实际 JSON boolean、channel、digest、byte count、归档哈希与冻结 metadata。"true"、缺字段或其他候选的证据不能通过。
  • qualify 继续负责隔离安装与恢复,新增字段来自它已经验证的 manifest;不是通过手工字段补出一次资格。回执在所有检查和清理之后返回。
  • build 本身未改动,继续从精确 Git commit 和匹配的 frontend bundle 构建不可变归档,派生 digest/size,并提供给 qualifier 与 publisher。没有改安装状态的权威来源。

规格依据为 d25e044c8c2aa3d862fb868eea93b58c901b07cc 的 .github/workflows/runtime-snapshots.yml 及该版本原 qualifier。逐项映射:Qualify clean installation and upgrade 已实现为原 producer 的全部检查与实际生产回执;Freeze revision-addressed metadata 已实现为 revision/channel/digest/bytes 与冻结文件的绑定;Publish immutable snapshot then channel pointer 的不可变资产、重试、历史 channel 资产和 lineage 行为保留。新增 release-readiness 段落解释该修复,不作为其自身正确性的独立证明。

对主干的风险 / Risks

未发现阻塞 finding。最强反例是测试手工制造一份新字段完整的资格记录,而真实 producer 根本不会产出它。本轮实际构建这个 commit 的20,609,068字节归档,运行原 qualifier 完成隔离 clean install、upgrade、doctor、skills、frontend bundle/HTTP、禁止 frontend build tools、许可证和注入激活失败后的原可执行恢复;生成回执的 revision、channel、hash 与实际归档一致。

随后将这份真实产物送进生产 Bash:合法回执到达本地刻意拒绝的 GitHub 边界;缺失、错误 revision/channel/bytes,以及十个检查逐一为 false 的十四类输入全部在当前 head 产生零 GitHub 调用。相同输入在原主干都会进入远端步骤。边界返回77是有意阻止远端效果的观察,不是一次真实发布成功;本轮没有创建、上传或移动任何远端 release/channel。

本轮独立验证 结果
五个发布/更新/安装测试模块 PASS,178项;另3项现有平台跳过保留,原生 Windows 未验收
完全相同的最终26组 publisher fixture 反跑主干 12项无效资格错误放行导致 FAIL,14项既有合法发布、重试、lineage 和故障恢复 PASS;head26项通过
真实归档 → 原 qualifier → 生产 Bash 合法证据进入被拒绝的本地 GH 边界;14类无效证据零调用,无远端发布
原生标准 premerge、Ruff、配置内 mypy PASS,5项直接检查、19项选中检查;17个配置内文件无类型错误

远端读取/上传失败后的原不可变字节检查、channel 删除后重试、未确认上传恢复及旧 workflow 不回退渠道,均由既有14组场景保留。这个工程完整性检查在现有 publisher 内默认执行,没有 default-off 声明;stable、desktop feed 及其激活策略不在修改范围。没有查询或等待 CI,实际 GitHub 服务故障、Linux/Windows/macOS 全发行和模型资格不能从本轮结果推断。

语义与 CI 对齐

新增 channel/hash 是同一 manifest 的派生事实,producer 和实际消费点都已验证;没有增加手工 intent、schema 版本或泛化 Goal 义务。状态判断使用 JSON 类型与精确值,没有用 prose 或 substring 推断“成功”。本地语义和公开边界检查通过。C1 的成果与完整成本、默认渠道采用判断仍须按其原验收单独完成,不能被本 PR 的安装资格取代。

我的整体评价 / Verdict

APPROVE,限定于这项工程发布缺陷的修复。未合格证据被阻断于远端步骤之前,合法发布和原失败恢复仍沿原 owner,减少后续安装者接触错误候选的风险。该增量在原 producer 与唯一 publisher 之间补齐可验证的绑定,机制成本与错误放行的影响相称。只改变 artifact 发布的机器校验和维护者文档;未改产品页面或交互。远端发布及最终用户升级不在本轮执行。实际远端分发、原生 Windows、C1 成果/完整成本与默认采用、最终整版2.4发行保持未完成,不因这份 approval 获得合并或发布权限。

English verdict: APPROVE — the publisher now consumes candidate-bound installation evidence before remote effects; release and C1 qualification remain separate.

@loopx-agent
loopx-agent merged commit 03197bd into main Nov 17, 2026
39 of 47 checks passed
@loopx-agent
loopx-agent deleted the codex/runtime-publication-qualification-20261117 branch November 17, 2026 13:38
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