Skip to content

Commit 00856a0

Browse files
committed
Merge remote-tracking branch 'origin/dev' into codex/tmp-4676-merge
# Conflicts: # structure/transports/inventory.md
2 parents 1daf4f6 + 43cd1ad commit 00856a0

60 files changed

Lines changed: 3429 additions & 210 deletions

File tree

Some content is hidden

Large Commits have some content hidden by default. Use the searchbox below for content that may be hidden.
Lines changed: 18 additions & 31 deletions
Original file line numberDiff line numberDiff line change
@@ -1,4 +1,4 @@
1-
# 022 — 전체 스위트를 macmini-cf로 보내기
1+
# 022 — 전체 스위트를 원격 러너로 보내기
22

33
이 라운드에서 테스트가 13분씩 걸리고 멈춘 것처럼 보인 원인과, 그 해결책을
44
기록한다. 다음 세션이 같은 함정에 빠지지 않도록.
@@ -34,48 +34,35 @@
3434
양방향 검증: 경쟁 러너가 있을 때 20초간 아무 테스트도 시작 안 함,
3535
opt-out을 켜면 즉시 실행(1 pass).
3636

37-
### 2. 원격 오프로드 — macmini-cf
37+
### 2. 원격 오프로드
3838

3939
무거운 스위트는 로컬에서 돌릴 이유가 없다. 사용자가 병렬로 작업 중이면
4040
서로를 방해할 뿐이다.
4141

42-
```
43-
Host macmini-cf
44-
HostName ssh-macmini.lidgeai.com
45-
User junny
46-
ProxyCommand /opt/homebrew/bin/cloudflared access ssh --hostname %h
47-
```
48-
49-
10코어 arm64. bun은 PATH에 없고 `~/.bun/bin/bun`에 있다(1.3.9).
50-
체크아웃은 `~/ci/opencodex`에 새로 만들었다.
42+
이 문서에는 러너의 호스트명, 계정, 접근 구성, 체크아웃 경로를 기록하지
43+
않는다. 이 정보는 저장소 밖의 운영 구성에서 관리한다.
5144

5245
**브랜치를 옮기는 방법이 중요하다.** `git push`로 보내면 로컬 pre-push 훅이
53-
전체 스위트를 돌린다 — 피하려는 바로 그 상황이다. 번들을 쓴다:
54-
55-
```bash
56-
git bundle create /tmp/x.bundle origin/dev..<branch>
57-
scp /tmp/x.bundle macmini-cf:/tmp/
58-
ssh macmini-cf 'cd ~/ci/opencodex && git fetch /tmp/x.bundle <branch>:<branch> && git checkout <branch>'
46+
전체 스위트를 돌린다 — 피하려는 바로 그 상황이다. 그래서 git bundle로
47+
브랜치를 전송한 뒤 원격에서 네 가지 게이트를 실행했다. 접속 대상과 경로가
48+
포함된 전송·실행 명령은 운영 인프라를 노출하므로 여기에 기록하지 않는다.
49+
러너에서 사용하는 일반 검증 명령은 다음과 같다.
50+
51+
```sh
52+
bun run test
53+
bun run typecheck
54+
bun run lint:gui
55+
bun run privacy:scan
5956
```
6057

61-
실행:
62-
63-
```bash
64-
ssh macmini-cf 'export PATH=$HOME/.bun/bin:$PATH; cd ~/ci/opencodex && \
65-
nohup sh -c "bun run typecheck && bun run lint:gui && bun run test && bun run privacy:scan" \
66-
> /tmp/ocx-gate.log 2>&1 &'
67-
```
68-
69-
주의: macmini에 `rg`가 없다. 로그는 `grep`으로 본다.
70-
7158
## 결과
7259

73-
| | 로컬(4개 경합) | macmini-cf |
60+
| | 로컬(4개 경합) | 원격 러너 |
7461
|---|---|---|
7562
| 전체 스위트 | 13분+ (미완료) | **162초** |
7663
| load average | 10.2 | 2.4 |
7764
| 결과 | — | 6210 pass / 0 fail |
7865

79-
푸시는 `--no-verify`로 했다. 게이트를 건너뛴 게 아니라 **원격에서 이미 통과**한
80-
것을 로컬에서 다시 돌리지 않기 위해서다. typecheck / lint:gui / test /
81-
privacy:scan 네 가지 모두 macmini에서 green이었다.
66+
푸시는 `--no-verify`로 했다. 원격에서 실행한 네 가지 — typecheck /
67+
lint:gui / test / privacy:scan — 가 모두 green이었기에 같은 검사를
68+
로컬에서 다시 돌리지 않았다. 원격 실행은 이 네 가지에 한정됐다.
Lines changed: 47 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,47 @@
1+
# 2.58.0 release train
2+
3+
Open. This unit carries the 2.58.0 release from the current `dev` tip through promotion and
4+
publication, and records the evidence each step actually produced.
5+
6+
## Why this unit exists
7+
8+
The 2.58.0 line accumulated in one day: a stabilization round that closed send-budget accounting,
9+
third-party Responses compatibility, Cursor tool-marker and overflow handling, safe teardown and
10+
configuration reporting, model discovery and capacity reporting, and a dependency-audit bump, plus
11+
the native control stack landing behind default-off flags. That is more surface than a patch
12+
release, and the promotion path is the same one 2.57.0 used, so the sequence is written down before
13+
it runs rather than reconstructed afterwards.
14+
15+
## Sequence
16+
17+
| Step | Gate | Evidence to record |
18+
| --- | --- | --- |
19+
| Land the native control stack | Each layer replayed onto the current `dev`, verified byte-identical against its pre-rebase diff, flags default-off | PR numbers, exact heads, per-layer CI |
20+
| Freeze the candidate | `dev` tip with `package.json` at 2.58.0 | Candidate SHA, its push run |
21+
| Full-platform regression | `ci.yml` dispatched on the candidate with `lane=all` | All nine Windows shards individually, Linux shards, macOS legs |
22+
| Move `dev`'s version line | `dev-version-bump.yml` opens the bump to the next minor | Bump PR and merge commit |
23+
| Promote to `main` | Promotion PR from the candidate | Merge commit, `enforce-target` red by design |
24+
| Prove the release SHA | CI on the merge commit | Run id and conclusion |
25+
| Publish | `release.yml` with version, `tag=latest`, `dry-run=false`, `expected-sha` | Run id, publish line, provenance, GitHub release |
26+
| Promote to `preview` | Promotion PR, version line resolved to `main` | Merge commit, empty diff against `main` |
27+
| Registry propagation | Registry read after publish | Whether availability was confirmed or still pending |
28+
29+
## Rules this train follows
30+
31+
No local suite, typecheck, build, install, or `ocx` invocation is used for any gate. Every claim
32+
comes from hosted CI at an exact SHA. `scripts/release.ts` is not run locally for the same reason:
33+
its preflight runs the suite. The workflow it would dispatch is dispatched directly instead, with
34+
the same inputs, so the published artifact is produced by the same job.
35+
36+
A cancelled job is not a failure and not a pass: it produced no result. Re-running it is recovery.
37+
Nothing is merged or published on a red gate that describes a real defect, and no budget is widened,
38+
retry added, or platform skipped to make a gate green.
39+
40+
## Known open items carried into this release
41+
42+
The contributor readiness gate holds two fork pull requests that are otherwise verified; their
43+
authors have the evidence and the remaining step is theirs. #4800 is not merged: the lane proved the
44+
provider retry policy it edits is unreachable from the Responses passthrough lane, and #4893 records
45+
the real defect. The native stack ships with both feature flags default-off, so an installation that
46+
does not opt in sees no behaviour change from it.
47+
Lines changed: 72 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,72 @@
1+
# wp2 — the 2.58.0 release
2+
3+
Published. 2.58.0 is on `main` and `preview`, tag `v2.58.0` names the release commit, the GitHub
4+
release exists, and `npm publish` returned success with a signed provenance statement. Registry
5+
propagation is tracked at the end.
6+
7+
## Sequence, with evidence
8+
9+
| Step | What happened | Evidence |
10+
| --- | --- | --- |
11+
| Land the native control stack | Five layers replayed onto `dev` one at a time, each verified before its push | #4782 → `519db59b31`, #4858 → `3ec1af6209`, #4861 → `fa26404d7b`, #4911 (= #4864 head `77c65e1a99`) → `f671934f02`, #4912 (= #4868 head `5c79b218a6`) → `5061f2c956` |
12+
| Freeze the candidate | `dev` tip `5061f2c956`, `package.json` 2.58.0 | — |
13+
| Full-platform regression | `ci.yml` dispatched with `lane=all` | Run `35247708168`: all nine Windows shards, all four Linux shards, `gates`, `docs site build`, keyring and npm-global on all three platforms, `storage policy`, `api usage`, `docker smoke` green. macOS capacity cancellations were the only non-green entries. |
14+
| Move `dev`'s version line | `dev-version-bump.yml` run `35253998264` opened #4916, merged as `4655d32f88` | `dev` now 2.59.0, which is what `release.yml` requires before it will publish 2.58.0 |
15+
| Promote to `main` | #4915 merged as `6fe4cd0de8` | Merge commit from the frozen candidate; tree identical to `5061f2c956`. `enforce-target` red by design, as for #4829 and #4694. |
16+
| Prove the release SHA | `Cross-platform CI` push run on `main` at `6fe4cd0de8` | Green after one re-run for a cancelled macOS shard. `release.yml` refuses to publish without this, and its first dispatch (`35254221472`) failed exactly there. |
17+
| Publish | `release.yml` run `35257765967`, version 2.58.0, tag `latest`, dry-run false, `expected-sha=6fe4cd0de85d63b8cdd0c3552e5e8883c0a029ee` | `validate-dispatch` and `publish` both success. Publish step ends `+ @bitkyc08/[email protected]`; provenance in the sigstore transparency log at logIndex 2879474742. GitHub release `v2.58.0` created 17:40:35Z, tag points at `6fe4cd0de8`. |
18+
| Promote to `preview` | #4917 merged as `48e1ddba0b` | `git diff origin/main HEAD` empty; no version-line conflict this time |
19+
20+
## Three things worth recording
21+
22+
**A green pull request is not a green merge result.** #4824 and #4817 each stayed under
23+
`tests/server/server-combo-failover-e2e.test.ts`'s file-size cap on their own branch and each
24+
reported an honest green. Merged in sequence they summed to 4207 lines against a cap of 4166, so
25+
every pull request built on that `dev` failed the ratchet with the same single offender. Exact-head
26+
CI is structurally unable to see "what happens after a sibling touches the same file", and
27+
`updateBaseline` only ever lowers a cap, so no tool could absorb it. #4908 moved one test into a
28+
sibling file, byte-for-byte, the way `d3ca5522db` did for the same situation.
29+
30+
**Dispatching a stale ref runs that ref's CI definition, not the current one.** Two evidence
31+
dispatches of draft fork trees failed on Windows and looked like a systemic defect in `dev`. The
32+
branches pre-dated #4876, and `workflow_dispatch` reads the workflow file from the dispatched ref,
33+
so those runs used a `ci.yml` without `OCX_TEST_NO_QUEUE` and reproduced the batch-serialization
34+
bug #4876 had already fixed — visible in the log as `waiting for test run pid … to release the user
35+
lock` followed by eight minutes with no test result. Merging current `dev` into the evidence branch
36+
removed it and all nine Windows shards passed. #4901 records the correction; the earlier claim of a
37+
Windows test-harness class was withdrawn.
38+
39+
**Rebasing a stacked child drops whatever lived only in its merge commits.** Replaying #4864 lost
40+
seven lines its author had added inside a merge commit — the paragraph documenting that the owned
41+
connection has no absolute lifetime cap — and left a new test file referencing a field the rename
42+
had removed. Neither is a conflict git can report. Both layers after that were checked for the same
43+
shape before pushing, and #4868 carried the identical defect in `ws-steering-completion.test.ts`.
44+
45+
## Native control stack: what shipped, and what enabling it still needs
46+
47+
Both flags remain default-off, so an installation that does not opt in sees no behaviour change.
48+
Contracts confirmed in the replayed trees rather than assumed: the account and physical socket are
49+
preserved, a failed steering or injection is never reported as success, steering deadlines are
50+
finite (90s submission, successor and continuation; 30 minutes for a tool wait), sparse terminal
51+
output is preserved and an output that contradicts completed wire items fails explicitly, and the
52+
settings pin relaxes for exactly four generation parameters while model, routing, tools,
53+
instructions and multi-agent stay pinned.
54+
55+
Three conditions remain before recommending activation. #4850's caller-owned preview read fence
56+
matters because an injection create is that request shape and this stack pins the create-time
57+
preview result across the whole chain. Eligibility now widens from injection-only to
58+
injection-or-steering, so a public API route can own a steering channel, and the canonical ChatGPT
59+
wire for these frames is still unproven from any source — decide the two routes separately. The
60+
owned connection itself has no absolute lifetime cap; the per-stage deadlines are finite but 128
61+
responses times a 30-minute tool wait can be composed on one pinned credential.
62+
63+
## Registry propagation
64+
65+
`npm publish` succeeded and npm answered that the package "is being processed and may take a few
66+
minutes to become available". The workflow's own post-publish registry smoke read it repeatedly
67+
without confirming and recorded `verification=pending`, continuing to the GitHub release without
68+
republishing. Reads from this machine still returned `E404` for 2.58.0 and `latest` still pointed at
69+
2.57.0 several minutes after the publish, which is the same behaviour 2.57.0 showed. **Do not
70+
republish.** The publish is acknowledged with provenance; inspect the registry before announcing
71+
availability.
72+

0 commit comments

Comments
 (0)