[chore] open PR 分诊（living inventory；不要把旧快照当当前分母）

[vansin]

open PR 分诊 · 2026-08-13

12 个 open PR 全部是 draft,全部 mergeable: UNKNOWN(GitHub 长期未重算 = 长期无人碰)。
逐个拉 head 与 origin/main 实测落后量与冲突状态:

PR	落后 main	状态	判定
#686	19	MERGEABLE / CLEAN	🟢 唯一新鲜的,可优先处理
#546	163	CONFLICTING	🔴 需 rebase
#460	293	CONFLICTING	🔴 需 rebase
#433	414	MERGEABLE / CLEAN	🟡 纯 docs,虽老但不冲突
#442	414	MERGEABLE / CLEAN	🟡 同上
#443	414	MERGEABLE / CLEAN	🟡 同上
#431 #432 #436 #438 #441 #444	414	CONFLICTING	🔴 一个月未动且冲突

关键观察

8 个 PR 落后 main 整整 414 个 commit,来自 7 月 12–29 日。其中 6 个已经冲突。

RFC-030 那一族(#431 / #432 / #433 / #441)是同一条线的四个 PR,全部停在 7 月中旬。
这条线要么重启要么关掉 —— 保持 draft 挂着一个月不动,对读仓库的人是纯噪声:
它们看起来像「正在做」,实际没有。

建议(需 owner 决定,我不擅自关)

1. test(#31): anchor REST response key golden #686 只落后 19 个 commit 且 CLEAN → 优先推完
2. docs: add RFC-030 implementation plan #433 / docs: record stable onboarding smoke #442 / docs: consolidate Codex, Grok, and OpenCode runtime status #443 是纯 docs 且不冲突 → 逐个核内容是否仍成立,成立就合、不成立就关
3. RFC-030 四件套 → 由该线 owner 明确「继续」或「关闭并留 issue 记录设计」
4. 冲突的 6 个 → 不要手工解冲突。按今天定的规矩:确定顺序后 append-only 合 main,或者重新开一个基于当前 main 的窄 PR(参考 [ux][P1] 节点 CLAUDE.md 模板加「给用户发文件」SOP(upload → /api/files 引用,勿发本地路径) #229 → 系统 review round 1 — 可靠性/安全/Runtime/测试·CI·文档 findings #261 的做法)

🔴 不要因为「看起来快好了」就去解一个月前的冲突 —— 那 414 个 commit 里包含了协议改动(#697/#698)、网络命名改动(#699)和多处测试重写,手工解冲突极易产生一个「能编译但语义已经不对」的结果。

[vansin]

2026-08-13 follow-up: keep the table above as a historical snapshot, not the current denominator. The repository now has 16 open PRs because active release/CI work was added after this issue was filed (#719, #727, #729, #730), while #722/#723/#724 have already landed. Do not use the old “12 all draft” sentence for present-day automation or closure decisions.

Current evidence-based split:

• release: preview 2.3.0-preview.39 / 2.5.0-preview.31 / 0.9.0-preview.29 #719 is active and all CI is green, but the release artifact/source mapping required separate forensic recovery; commit 16258b9 is now the strongest evidenced source anchor and [build][repro] npm bundles embed absolute build paths; obfuscated dist is not byte-reproducible #732 tracks absolute build-path leakage/non-byte-reproducibility.
• ops(qa): record which registry versions L1 resolved, so a red is attributable #727 is active but red because qa-cli-01 image construction failed before tests; it is not ready.
• fix(ci): retry the bun install so a transient fetch stops reading as a test failure #729 is active CI-hardening work and has an open source-side review point about curl | bash masking the fetch side of a pipeline.
• test(ci): gate the complete agent-node unit domain #730 is a new Draft adding the missing complete agent-node Docker/non-root unit gate.
• the July RFC-030/onboarding drafts remain owner decisions. Age/conflict is a triage signal, not permission to close or manually resolve them. Their semantic overlap with current main must be checked before any superseded closure.

The durable rule remains: update this denominator from live PR state before acting; do not carry a compacted/stale count forward as authority.

[vansin]

重新测量 + 开始分流(2026-08-13)

原描述里的「8 个落后 414 个 commit」已经不准了(main 又走了),而且**「旧」不是关闭的理由,
「已被取代」才是**。重新按当前 origin/main 测:

PR	behind	ahead	分流结论
#733	1	1	✅ 不是陈旧件,是活的,已独立复验,等 e2e 转绿即可合
#686	27	2	待查
#546	171	1	待查
#460	301	1	待查
#444	422	3	待查
#443	422	2	待查
#442	422	1	待查
#441	422	1	待查(RFC-030 系列)
#438	422	1	❌ 已关 —— 方案被取代,证据见 PR 内评论
#436	422	2	⚠️ 代码已由 #435 落地,只剩证据文档;未关,归属作者
#433	422	29	待查(RFC-030 系列)
#432	422	14	待查(RFC-030 系列)
#431	422	35	待查(RFC-030 系列,标题自带 DRAFT — do not merge)

这一轮学到的

原 issue 把这 12 个一并称作「噪音」,这个前提第一条就不成立:#733 是一天前开的、
落后 main 仅 1 个 commit,而且修的是一个真的 fail-open CI 门(我自己 #729 引进去的洞)。
按「旧 = 该关」批量处理会把它一起扫掉。

所以分流判据定为:看它的实质是否已从别的路径进了 main,不看日期。

• test(server): #434 isolate Bun.serve integration suites in aggregate runs #438 命中(issue test(server): isolate Bun.serve integration suites in aggregate runs #434 已 CLOSED,main 走的是 test-aggregate.ts 那条路)→ 关
• security(test): reject inherited DATABASE_URL before adapter construction (#435) #436 部分命中(守卫已由 security(test): reject inherited DATABASE_URL before adapter construction #435 落地,只剩证据文档)→ 不关,交回作者

RFC-030 那四个(#431/#432/#433/#441,合计 ahead 79 个 commit)是设计类,
不能按同一把尺子量,需要 owner 先说 RFC-030 这条线还走不走,再谈这些分支的去留。

[vansin]

Inventory refresh after the 2026-08-13 merge wave

Current open PR denominator is 12, not the issue body’s old “12 all draft” snapshot:

• 11 older drafts: RFC-030 Wave 1A — Codex Policy Gateway (DRAFT — do not merge) #431, RFC-030 Wave 1B: gateway scheduler/policy/ledger/adapter + upstream mux + principal stamp (draft) #432, docs: add RFC-030 implementation plan #433, security(test): reject inherited DATABASE_URL before adapter construction (#435) #436, RFC-030: wire codex-app-server into anet CLI #441, docs: record stable onboarding smoke #442, docs: consolidate Codex, Grok, and OpenCode runtime status #443, fix onboarding runtime validation and first-start routing #444, fix(opencode): close Linux stop orphan windows #460, docs(api): remove stale implementation references #546, test(#31): anchor REST response key golden #686;
• test(ci): gate the complete agent-node unit domain #730 is ready, exact report head 3c6fbc05, all 9 checks green, awaiting independent narrow review.

Recent PRs #719, #727, #729, and corrective #733 have merged into current main (c7ac3c20). #729’s retry was independently shown fail-open; #733 fixed that exact defect with all checks green. This is why “merged” and “finding closed” were tracked separately until the corrective merge landed.

The older RFC/onboarding drafts still must be judged by current-main semantic coverage and SSOT value, not by age or conflict count. No old draft is being closed from this inventory update alone.

[vansin]

分流进展(第二批)

PR	behind	结论	依据
#733	1	✅ 已合	fail-open CI 门,独立复验双向成立
#546	171	✅ 发现已重做为 #737 并已合、已上线	见下
#686	27	🟢 活的,建议 rebase	issue #31 仍 OPEN;tests/test686-rest-shape-golden/ 至今无替代物落地
#460	301	🟡 部分落地,未确立,不关	节点侧已有 signalExactLinuxProcessGroup;CLI wrapper 侧无法确立
#438	422	❌ 已关	issue #434 已 CLOSED,main 走 test-aggregate.ts
#436	422	⚠️ 代码已由 #435 落地,只剩证据文档,归属作者

#546 这条值得单独说:分流捞回来的东西比关掉的多

它的发现至今成立,但分支 behind=171。我没有 rebase,而是在当前 main 上重做为 #737(已合并并已部署上线)。
重做过程中额外挖出一个链接问题之外的失真:

文档写   429 error = `too many attempts, try again later`
实际返回 { error: "rate_limited", message: "Too many login attempts. Try again later.", retry_after_ms }
         + Retry-After 头

那个字符串在整个 server/src/ 里都不存在 —— 照文档匹配 error 字段的客户端会直接失配。

线上实证(部署前后同一条命令):

部署前  anet.sh/api/rest  server/src/index.ts × 55   'too many attempts' × 1
部署后  anet.sh/api/rest  server/src/index.ts × 0    'too many attempts' × 0
                          rate_limited × 3  retry_after_ms × 1  Retry-After × 1
英文站同步,age=0

方法上的一条

这批分流最有价值的产出不是关掉的那一个,而是 #546 这种「分支该弃、发现该留」的。
按日期批量关会把它连同 #733 一起扫掉。判据始终是实质是否已从别的路径进了 main,
而且「进了 main」要落到具体符号/文件,不能停在「看起来做过了」。

剩下未分流:#444 / #443 / #442(behind=422)+ RFC-030 四件(#431/#432/#433/#441),
后者需要 owner 先定这条线还走不走。

[vansin]

分流完成(RFC-030 四件除外)

PR	behind	结论	依据
#733	1	✅ 已合	fail-open CI 门,独立复验双向成立
#546	171	✅ 发现已重做为 #737,已合已上线	线上 index.ts 引用 55 → 0
#686	27	🟢 活的,建议 rebase	issue #31 仍 OPEN,test686 套件无替代物
#444	422	🟡 实质已落地,缺测试	Bun 前置在 cli.ts:5665 真强制(process.exit(1));两个 onboarding 测试文件未进 main
#442	422	🟡 证据仍有效	npm latest 至今是 2.2.21,正是它记录的版本
#443	422	🟡 建议关,但差异需作者确认	docs-site/docs/guide/runtimes.md(中英,已上线)已覆盖四个运行时
#460	301	🟡 部分落地,未确立	节点侧有 signalExactLinuxProcessGroup;CLI wrapper 侧无法确立
#436	422	🟡 代码已由 #435 落地,只剩证据文档	db-adapter-guard.test.ts 已有该断言
#438	422	❌ 已关	issue #434 已 CLOSED,main 走 test-aggregate.ts

12 个 draft,最后只关掉 1 个

原 issue 的框架是「这些是噪音,该批量清理」。分流跑完的结论是这个框架不成立:

• 1 个真的该关(test(server): #434 isolate Bun.serve integration suites in aggregate runs #438,方案被取代)
• 1 个是活的、且修的是真缺陷(fix(ci): make Bun installer retries fail closed #733,已合)
• 1 个的发现至今有效,重做后还额外挖出一个响应体失真(docs(api): remove stale implementation references #546 → docs(api): REST 源码链接指向 server.ts,并修正登录 429 的真实响应体 #737)
• 5 个是「实质落地了、测试/证据没跟上」或「部分落地」—— 这类关掉就等于丢掉尚未覆盖的缺口
• 4 个 RFC-030 是设计类,需要先定这条线走不走

两个反复出现的形状,值得单独记

形状一:实现进了 main,测试留在了分支里。(#436、#444)
表现是「功能是好的,但没有任何测试钉住它」。这种 PR 关掉最亏 ——
关的时候看起来「已经做过了」,实际丢掉的正是回归保护。
处理方式统一为:只把测试文件摘到当前 main 开小 PR,不 rebase 整支。

形状二:分支该弃,发现该留。(#546、#442)
#442 尤其反直觉:behind=422,但 npm latest 至今还是它记录的 2.2.21,
证据对现在的真实用户完全有效。「分支旧」和「结论旧」是两件事。

剩下

RFC-030 四件(#431/#432/#433/#441,合计 ahead 79 个 commit)不动,等 owner 定这条线的去留。

[vansin]

Live denominator refresh — 2026-08-13 06:2x CST

The title no longer embeds a count because that count repeatedly became stale during the same merge wave. Authoritative remote state now has 8 open PRs, not the body’s historical 12:

PR	Draft	main-only commits (behind)	head-only commits	current merge-tree
#771	no	0	1	clean
#770	yes	0	2	clean
#460	yes	327	1	conflict
#444	yes	448	3	conflict
#442	yes	448	1	clean
#433	yes	448	29	clean
#432	yes	448	14	conflict
#431	yes	448	35	conflict

Method: fetched each exact PR head, used git rev-list --left-right --count origin/main...<head>, then git merge-tree from the exact merge base. GitHub’s long-lived UNKNOWN state was not treated as evidence.

Current split:

• fix(ci): slug 守卫覆盖 .sh/.py —— 三处真 slug 一直藏在 898 个文件的扫描范围之外 #771 and test(ci): pin #767 Bun preflight behavior #770 are active current-main work; age-based triage does not apply.
• fix(opencode): close Linux stop orphan windows #460/fix onboarding runtime validation and first-start routing #444 remain semantic/test recovery candidates despite conflicts; do not resolve their old branches wholesale.
• docs: record stable onboarding smoke #442/docs: add RFC-030 implementation plan #433 remain clean but old evidence/design branches and still need semantic-currentness review.
• RFC-030 Wave 1A — Codex Policy Gateway (DRAFT — do not merge) #431/RFC-030 Wave 1B: gateway scheduler/policy/ledger/adapter + upstream mux + principal stamp (draft) #432 remain RFC-030 owner decisions; conflict/age alone is not closure authority.

No PR was closed, rebased, force-pushed, or modified by this refresh. The durable rule remains: only close when the substantive finding is demonstrably superseded, not because a branch is old.

[vansin]

Live denominator delta — 2026-08-13 07:0x CST

Remote still has 8 open PRs, but the membership changed since the prior 8-count snapshot:

• active current-main work: fix(cli): correct top-level help contracts #784 (Draft, CLI help corrective; independent CLEAN, full CI still running) and docs(deploy): weixin 两条改为精确未覆盖范围(代码其实已推送,真正缺的是账号归属与 secret) #786 (ready, recovery inventory refinement; external repo facts independently confirmed, 2 MINOR requested before merge);
• older set remains fix(opencode): close Linux stop orphan windows #460/fix onboarding runtime validation and first-start routing #444/docs: record stable onboarding smoke #442/docs: add RFC-030 implementation plan #433/RFC-030 Wave 1B: gateway scheduler/policy/ledger/adapter + upstream mux + principal stamp (draft) #432/RFC-030 Wave 1A — Codex Policy Gateway (DRAFT — do not merge) #431;
• test(ci): pin #767 Bun preflight behavior #770/fix(ci): slug 守卫覆盖 .sh/.py —— 三处真 slug 一直藏在 898 个文件的扫描范围之外 #771 from the prior snapshot have landed;
• docs(deploy): pin Hub Bun recovery version #781 has now landed as 19044c56, and its source issue [ops] 第一次纸面重建演练(hub):8 步里 5 步覆盖、2 步已标 NOT COVERED、1 步是新发现的缺口(bun 怎么装没写) #778 was closed only for the Bun-step gap, with fresh-host/vault/data limits explicitly retained.

This is another example of why the count alone is not an identity: the denominator stayed 8 while two members changed. Automation and triage must record PR numbers/head SHAs, not just “8 open”.

[vansin]

Live denominator refresh — 2026-08-13 07:59 CST

Current remote has 7 open PRs:

• active current-main evidence transplant: docs(test): preserve stable onboarding smoke evidence #797 (Draft; exact one-file recovery of docs: record stable onboarding smoke #442's still-current stable onboarding report, independent review pending);
• older semantic/design branches: fix(opencode): close Linux stop orphan windows #460, fix onboarding runtime validation and first-start routing #444, docs: record stable onboarding smoke #442, docs: add RFC-030 implementation plan #433, RFC-030 Wave 1B: gateway scheduler/policy/ledger/adapter + upstream mux + principal stamp (draft) #432, RFC-030 Wave 1A — Codex Policy Gateway (DRAFT — do not merge) #431.

Changes since the prior snapshot:

• test(ci): gate complete agent-network unit domain #791 merged as f8b9675d, adding the missing full agent-network/src aggregate CI job;
• test(ci): #792 给 agent-network 补上聚合单测门(46 个单测此前无人跑) #795 independently rediscovered the same gap but was opened from a base already containing test(ci): gate complete agent-network unit domain #791, so it was closed as superseded without deleting its branch;
• agent-network/src 的 46 个单测没有任何 CI job 会跑 #792 was closed as completed by test(ci): gate complete agent-network unit domain #791, while the non-enforced/branch-protection boundary remains tracked by [governance] make CI checks and independent review mechanically enforceable on main #725;
• test(fleet): make recovery inventory boundary semantic #794 merged as 9e39ae3c, replacing stale fleet-inventory 5/2/2 counts with semantic recovery assertions; [test] test736 inventory boundary is stale after fleet inventory growth #793 closed.

The old-set disposition is unchanged: #460/#444 still contain unproven product/test gaps; #431/#432/#433 remain RFC owner decisions. They were not closed by age. If #797 lands, only then should #442 be closed as superseded by the byte-identical current-main transplant.

[vansin]

Denominator delta — 2026-08-13 08:02 CST

The immediately preceding 7-PR snapshot changed after its pending review completed:

• docs(test): preserve stable onboarding smoke evidence #797 independently reviewed CLEAN and merged as e9c8543f;
• docs: record stable onboarding smoke #442 then closed as superseded; its report is now present on main byte-for-byte (d2b3f312e29209d4b39f8a148c03fca8693a585e).

Current open PR denominator is therefore 5: #460, #444, #433, #432, #431.

No product branch was rebased or manually conflict-resolved. The remaining five keep their previous disposition: two product/test recovery candidates and three RFC/design owner decisions.

[vansin]

Immediate denominator correction — 2026-08-13 08:04 CST

#796 appeared between the prior live read and comment publication. Current denominator is 6, not 5:

• active current-main CI hardening: test(ci): 把 test745 的执行文件数绑到磁盘上的分母(现在跑 2/46 也是绿的) #796;
• older set: fix(opencode): close Linux stop orphan windows #460, fix onboarding runtime validation and first-start routing #444, docs: add RFC-030 implementation plan #433, RFC-030 Wave 1B: gateway scheduler/policy/ledger/adapter + upstream mux + principal stamp (draft) #432, RFC-030 Wave 1A — Codex Policy Gateway (DRAFT — do not merge) #431.

#796 is not a duplicate of merged #791. It identifies a narrower gate defect: test745 computed 46 discovered files but only required the executed-file count to be nonzero, so bun test src/im/ (2/46 files) still passed the old gate. Its +11-line change binds executed count to discovered count. Independent narrow review is now in progress; it is not merged merely from the title or author evidence.

This correction is deliberately append-only so the timing race remains visible: a living inventory must be refreshed at the action point, not copied from even a minutes-old comment.

[vansin]

Denominator delta — 2026-08-13 08:12 CST

#796 independently reviewed CLEAN, all nine GitHub checks passed, and it merged as 92d96129 after the old-vs-new 2/46 A/B was reproduced.

Current open PR denominator returns to 5: #460, #444, #433, #432, #431.

The test745 aggregate gate now requires executed-file count to be at least the discovered-file count. Its explicit remaining limit is unchanged: it compares counts, not exact file identities.

[vansin]

Denominator delta — 2026-08-13 08:24 CST

Two new current-main CI PRs appeared after the prior five-PR snapshot. The live open denominator is now 7:

• test(ci): 给 server 补上聚合单测门(69 个单测此前 CI 只跑 6 个) #798 — ready, server aggregate unit gate (independent review/merge decision still separate);
• ci: pin setup-bun version across workflows #799 — draft, setup-bun 1.3.14 workflow pins plus a full-occurrence L1 gate; independent review and CI are in progress;
• older set unchanged: fix(opencode): close Linux stop orphan windows #460, fix onboarding runtime validation and first-start routing #444, docs: add RFC-030 implementation plan #433, RFC-030 Wave 1B: gateway scheduler/policy/ledger/adapter + upstream mux + principal stamp (draft) #432, RFC-030 Wave 1A — Codex Policy Gateway (DRAFT — do not merge) #431.

#799 intentionally addresses only the version-pin half of #746. Download retry/cache remains open, so neither the PR nor this inventory treats #746 as resolved.

This is another append-only point-in-time count; action-time triage must refresh the GitHub list rather than reuse this number.

[vansin]

Denominator delta — 2026-08-13 08:30 CST

#799 independently reviewed CLEAN, all ten reported GitHub checks passed, and it squash-merged as d4283a22 after an action-time exact-HEAD/current-main/merge-tree recheck.

The live open PR denominator returns to 6: #798, #460, #444, #433, #432, #431.

#746 remains open because #799 fixes only setup-bun version pinning; download retry/cache remains unresolved. The source branch is intentionally still present because the CLI's post-merge local cleanup collided with a separate worktree already using main; no worktree was disturbed and no force/delete retry was attempted.

[vansin]

Denominator delta — 2026-08-13 08:45 CST

The live open PR denominator is now 9:

• current CI work: test(ci): 给 server 补上聚合单测门(69 个单测此前 CI 只跑 6 个) #798, test(ci): 让 test725/test745 覆盖 tests/ 目录(两个门自称 complete 却漏了 25 个文件) #800, ci: 元门 —— 新增测试文件不能落在所有聚合门的扫描范围之外(依赖 #798 #800) #801, ci: pin Bun in the always-on rename-ghost gate #802;
• older set: fix(opencode): close Linux stop orphan windows #460, fix onboarding runtime validation and first-start routing #444, docs: add RFC-030 implementation plan #433, RFC-030 Wave 1B: gateway scheduler/policy/ledger/adapter + upstream mux + principal stamp (draft) #432, RFC-030 Wave 1A — Codex Policy Gateway (DRAFT — do not merge) #431.

Important dispositions at this snapshot:

• test(ci): 给 server 补上聚合单测门(69 个单测此前 CI 只跑 6 个) #798 rebuilt its branch after main drift, but its committed report still attributes the new gate/output to old base 92d96129; it is DO-NOT-MERGE pending new source/report coordinates and independent review.
• ci: 元门 —— 新增测试文件不能落在所有聚合门的扫描范围之外(依赖 #798 #800) #801 explicitly depends on test(ci): 给 server 补上聚合单测门(69 个单测此前 CI 只跑 6 个) #798 and test(ci): 让 test725/test745 覆盖 tests/ 目录(两个门自称 complete 却漏了 25 个文件) #800; it cannot be evaluated or merged ahead of those dependencies merely because its own PR is open/ready.
• ci: pin Bun in the always-on rename-ghost gate #802 is Draft; independent review and long-running CI are in progress.

This is an append-only point-in-time count. Refresh before any action.