Skip to content

Realign main on upstream v2.100.2 with the hub fixes - #9

Merged
dalsoop merged 709 commits into
mainfrom
hub/align-v2.100.2
Nov 11, 2026
Merged

dalsoop merged 709 commits into
mainfrom
hub/align-v2.100.2

Conversation

@dalsoop

@dalsoop dalsoop commented Nov 11, 2026 •

Copy link
Copy Markdown
Owner

What

Realign main on upstream v2.100.2 (lidge-jun/opencodex 61147ec6, the source of npm 2.100.2) for the k3s hub, so main becomes the single line for this repository:

main = upstream v2.100.2 + the six hub fixes from k3s/v2.99.0 + the personal changes that matter for the hub. Mac-only personal changes are removed with revert commits. Nothing is force-pushed; upstream history comes in through a merge commit (this PR is merged with a merge commit, not squashed).

Order of the commits on this branch (base 977bca0):

  1. 22 revert commits — personal changes that do not belong in the hub (table below).
  2. ci: allow the upstream test fixtures in secret detection — .gitleaks.toml allowlist for 30 made-up keys/canaries in 26 upstream test files (rule generic-api-key only, default rules extended, not replaced). Without it the secret check stops the merge commit and every PR that carries upstream history.
  3. Merge upstream tag v2.100.2 into main — conflicts and the leftovers of the old v2.89.0 release line resolved in favour of upstream: the merged tree is exactly the v2.100.2 tree plus the personal changes kept at that point (subscription dates, slow-step logging, secret detection). Only src/codex/auth-api/account-list.ts needed code adaptation plus two structure docs. The subscription dates were later removed again (step 8).
  4. The six hub fixes, cherry-picked with -x from k3s/v2.99.0 (all applied cleanly on 2.100.2):
  5. ci: build the hub image from main and hub version tags — the image workflow now runs on pushes to main and tags hub/* (plus manual). Same image name and tag scheme ghcr.io/dalsoop/opencodex:<package version>-k3s-<commit12>; a commit that already has its image reports the existing digest instead of building a second one.
  6. chore(release): stamp version 2.100.2 — upstream stamps the version only inside its release checkout (the v2.100.2 tag commit still reads 2.100.0). Stamped with upstream's own scripts/release-version-sources.ts sync 2.100.2 so the image reports 2.100.2 (the k3s initContainer checks it).
  7. style: keep two files under upstream's 2,000-line threshold — upstream's file-size ratchet (tests/ci-workflows/file-size-ratchet.test.ts) failed because src/server/index/serve-options.ts (keep-alive call) and scripts/test-layout/layout.json reached exactly 2,000 lines. Each is joined onto one line, like 0084624 did for the hub-accounts route. Formatting only; the ratchet baseline is untouched.
  8. Revert the subscription dates (033aeee) from main — upstream 2.100.2 requires a passive account list to make no native credential reads (test "stored Pool duplicate of native main > passive account list projects known exclusion without native reads", tests/codex-integration/codex-stored-main-duplicate.test.ts); the subscription read adds an auth.json read and failed it. Nothing reads the dates from the hub (/v1/hub-accounts does not project them; CodexBar takes them from the OpenAI dashboard), so the change is removed rather than reworked. (The comment fix just before it is superseded.)
  9. test: register the hub-fix tests in the 2.100.2 test layout — 2.100.2 resolves every test file through the explicit layout map; the cherry-picked hub-accounts/messages-keepalive tests are added to scripts/test-layout/layout.json (on existing lines) and hub-accounts.test.ts to tests/fixtures/test-layout-expected.json. Test layout only.

git diff v2.100.2 <head> therefore contains only: the slow-step logging, the hub fixes, k3s-image.yml, secret-detection.yml, .gitleaks.toml, the four version lines and the test-layout entries for the hub-fix tests (23 files).

No change here adds a path that sends Claude requests outside the anthropic pool (no combo, fallback or other-provider modelMap); the hub fixes touch only the Messages HTTP boundary (keep-alive) and a read-only account route.

Classification of the 26 personal commits on main (v2.89.0..977bca0, merges excluded)

All 26 stay reachable from archive/branch/main and archive/m5m1-personal-77796d8 (m5m1's personal build 77796d8 is unchanged).

Commit Subject Keep / remove Reason
3e2bd16 fix(catalog): require selected visibility for uncurated live models remove (revert) Catalog visibility for the Mac personal build. On the hub it flips the new-model default to off and can reduce a provider's visible models to its default on first discovery: a model-list change the canonical hub has never run and the devices' Codex model lists depend on.
e02a9e0 test(codex): isolate readiness fixtures from host runtimes remove (revert) Test isolation belonging to the catalog change.
e80b967 fix(catalog): preserve legacy visibility without a default remove (revert) Same catalog visibility change.
ac7cd68 test: isolate catalog acceptance from ambient runtimes remove (revert) Test isolation for the catalog change and the personal release fixture.
2a22e73 fix(catalog): persist preserved visibility decisions remove (revert) Same catalog visibility change.
0e5dcb2 fix(catalog): classify legacy visibility before arrivals remove (revert) Same catalog visibility change.
fd1669b test(antigravity): declare initialized visibility fixtures remove (revert) Fixture for the catalog change.
91e4a14 fix(catalog): preserve switches for missing active models remove (revert) Same catalog visibility change.
a90fe98 docs: bind publishing to our independently maintained repository remove (revert) Personal publishing notes (AGENTS.md) for the Mac build.
7b6dfa6 feat(install): manage our source releases and verify our repository in CI remove (revert) Personal source installer, personal-ci.yml, ownership notes.
1908d7d fix(install): validate installer types and keep ownership local remove (revert) Personal installer / personal CI.
6f55d6f fix(ci): require owner opt-in before deploying inherited documentation remove (revert) Personal-setup guard; the docs deployment workflow is disabled at repository level instead.
0579fdf chore(personal): identify our source version and synchronize README notices remove (revert) Personal version 2.89.1-dalsoop.1 and installer notices; main carries the upstream version.
3cba92a fix(install): launch updates with our bundled runtime remove (revert) Personal installer.
584ae7a fix(test): isolate release examples and shard our complete CI suite remove (revert) personal-ci.yml and personal release fixture; upstream CI is the reference.
0f889a2 fix(install): stop the managed proxy before replacing its service remove (revert) Personal installer.
65e4d9a fix(launchd): set KeepAlive true unconditionally to prevent stale dead proxy remove (revert) macOS launchd; the hub runs under Kubernetes.
033aeee Let account consumers read stored subscription dates without extra refreshes remove (revert, after the merge) First kept and re-applied in the merge, then removed: on 2.100.2 it breaks upstream's "passive account list makes no native reads" invariant (CI test failed), and no hub consumer reads the dates (/v1/hub-accounts allow-list has no subscription field).
5c9e5f2 fix(install): update the personal service without ocx stop remove (revert) Personal installer.
2982b13 fix(install): verify kickstarts and keep macOS rollback off service verbs remove (revert) Personal installer.
fb42598 fix(install): probe the configured host and require launchd's pid remove (revert) Personal installer.
7fcb587 fix(cli): skip Codex shim auto-restore in the startup-health probe remove (revert) Codex shim exists only on a Mac host; the hub image has none.
1b6e396 feat(startup): log client startup steps and model discovery that take over 1 s keep [slow-step] lines for model discovery and startup steps (stderr, ≥1 s only): diagnostics for hub start and model-list latency.
bf90129 fix(startup): write slow-step lines to stderr and leave src/cli/index.ts unchanged keep Follow-up of 1b6e396 (stderr, src/cli/index.ts untouched).
bdc5cd2 fix(startup): run the project config check off the main thread remove (revert) Scans Codex trusted projects on a Mac host (slow volumes); the hub has none.
977bca0 ci: add pinned gitleaks secret detection (#8) keep Repository secret detection (owner-approved 2026-11-10), not a Mac installer change.

Kept 3 / removed 23.

Repository settings changed with this PR (not in the diff)

  • Disabled workflows (gh workflow disable): Enforce PR target branch ([WRONG BRANCH]), PR Labeler, PR hygiene, Deploy Docs to GitHub Pages, Personal repository CI, Clean branches from closed pull requests (branch deletion is a separate, later step), Clean orphaned workflow runs (keeps run history). Release / Dev version bump / desktop installed-artifact gate / Windows signing test are disabled once they register on main after this merge.
  • Required status checks on main (classic branch protection, set in the web UI because the CLI token has no Administration permission): upstream ci (aggregate of test shards and gates) and gates (typecheck). Not strict (branch need not be up to date), no review requirement, force pushes and deletions not allowed.
  • Upstream release tags v2.99.0-preview.20261108, v2.99.0, v2.100.0, v2.100.1, v2.100.2 pushed to this repository: upstream's gates job (.github/scripts/release-history.test.cjs) verifies fixed legacy release sources 68eee1eb (v2.99.0-preview.20261108) and 6676629 (v2.99.0), which were not in this repository's object graph. No workflow triggers on these tags.

Verification

CI (required, on head 6e3254d): ci pass, gates pass (run 43088485892; all five test n/5 shards, storage policy, api usage, docker smoke, keyring, npm-global, structure gate, docs site build, desktop shell green). gitleaks pass. Non-required react-doctor fails: its PR scan covers the upstream v2.89→v2.100 GUI changes carried by this PR; git diff v2.100.2 <head> has no gui/ file.

Earlier CI rounds on this PR found three things, fixed above: the file-size ratchet (step 7), upstream's release-history gate needing upstream release tags in this repository (tags pushed, see settings), the passive-read invariant (step 8) and the test-layout map (step 9).

m5m2 (macOS, Bun 1.4.4, the commands upstream CI runs) — on head e10fa94 and, for comparison, on the plain v2.100.2 tag:

Command head v2.100.2
bun install --frozen-lockfile (root, gui/), cd gui && bun run build 0 0
bun x tsc --noEmit 0 0
bun x tsc --noEmit -p tests/tsconfig.doctor-service-memory-contract.json 0 0
bun x tsc --ignoreConfig --noEmit --strict … scripts/ci/docker-smoke.ts 0 0
bun run typecheck (bun check, 2,402 files, no type errors) 0 0
bun run privacy:scan 0 0
bun run structure:check 0 0
bun run test 124 (900 s budget, timeouts) 124 (900 s budget, timeouts)

bun run test does not complete on this Mac for the plain upstream tag either (macOS without GNU timeout, the PATH-Bun fixture trust rule and the 900 s suite budget; 358 distinct failing tests on v2.100.2 vs 215 on the head, the differences are timeouts that move between runs). The Linux CI suite above is the proof. Targeted runs on 6e3254d: bun test tests/test-layout.test.ts, tests/test-layout-tooling.test.ts, tests/ci-workflows/file-size-ratchet.test.ts, tests/server/hub-accounts.test.ts, tests/server/messages-keepalive.test.ts, tests/codex-integration/codex-stored-main-duplicate.test.ts all exit 0; tests/claude-integration/messages-keepalive-server.test.ts 3 pass / 1 hook timeout locally (passes in CI).

Independent review (a reviewer agent that did not author the change, read-only): APPROVE at 8cbad4b and again at 6e3254d. Checked: the six hub fixes are blob-identical to k3s/v2.99.0 except the formatting-only line join; no combo/fallback/modelMap or other Claude path outside the anthropic pool; the 23 removals are complete (removed files blob-identical to upstream); account-list.ts and the other subscription files are back to upstream exactly; k3s-image.yml, .gitleaks.toml, version stamp; the test-layout entries. Non-blocking notes: react-doctor reflects the upstream GUI diff only; workflow_dispatch of the image workflow can run from any ref (as before; k3s-infra pins by digest).

🤖 Generated with Claude Code

lidge-jun and others added 30 commits November 6, 2026 14:42
* fix(responses): preserve adapter body stall status

* docs(responses): describe adapter body stall failures

* fix(responses): preserve continuation stall status

* Preserve body stall across empty completion retry
Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
Carries #8187 by @ismell0992-afk.

Co-authored-by: ismell0992-afk <255305877+ismell0992-afk@users.noreply.github.com>
* Prioritize critical credits in quota headline

Co-authored-by: Rafael Moreira Lima <rrmlima@gmail.com>

* Fix quota headline selection for custom label collisions

Co-authored-by: Rafael Moreira Lima <rrmlima@gmail.com>

* Strengthen unlimited credits quota summary fixture

Co-authored-by: Rafael Moreira Lima <rrmlima@gmail.com>

* docs: record quota headline precedence

Co-authored-by: Rafael Moreira Lima <rrmlima@gmail.com>

* fix(gui): index quota headline windows by label

Co-authored-by: Rafael Moreira Lima <rrmlima@gmail.com>

* docs: clarify critical credit precedence in translations

Co-authored-by: Rafael Moreira Lima <rrmlima@gmail.com>

* docs: describe quota window projection

Co-authored-by: Rafael Moreira Lima <rrmlima@gmail.com>

---------

Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
Co-authored-by: Rafael Moreira Lima <rrmlima@gmail.com>
Carries #8081 by @rrmlima.

Co-authored-by: Rafael Moreira <rrmlima@gmail.com>
* fix(ci): rebalance Ubuntu tests across five shards

* test(ci): verify five-shard suite coverage
…sktop app (#8105)

Design system v2: redesigned dashboard, navigation and desktop windows, revised to the owner's Paper design (sidebar search and shortcuts, Home, Usage source dropdown with privacy-bounded client/device attribution). Classic dashboard kept at /classic/.

Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
Carries #8182 by @kerthans. Fixes #8181.

Co-authored-by: kerthans <112744990+kerthans@users.noreply.github.com>
…8240)

* fix(cli): retain Desktop selection when its executable disappears

* docs: record verified Desktop launcher fix and pending live recovery
* test(config): use supported conditional platform skips

* test(config): cover log rotation under a busy ownership lock
* test(config): retry busy first ownership registrations

* test(config): fail registration workers on barrier expiry

Exit before ownership registration when the go signal is still absent after the barrier deadline. A child-local deterministic clock regression fails before the fix and passes afterward without a 20-second delay. Focused gated/cold/racing controls: 4 pass, 20 assertions; typecheck2322 and file-size/diff checks pass. No product change.
* fix(sync): report skipped connected children as partial

* fix(sync): make update guidance installation neutral
* fix(sync): defer supervised client restart after apply

* fix(sync): clarify deferred supervisor restart guidance
* fix(link): report stale child connection readiness

* test(link): seed durable state for tunnel readiness

* fix(link): fence stale client status reads
* fix(link): preserve confirmed orphan token files

* test(skill): register disconnect capability domain

* fix(link): close unknown liveness narrowing

* fix(link): protect admission secrets in orphan recovery
…8243) (#8256)

Remove a Windows identity pin on refusal while another name still holds the token, name a leftover .identity pin in the nlink refusal, read durable Child state once per /readyz and only for the status routes that report it, type client_state_changed as a LinkReadiness reason, and show the 503 reason with recovery guidance in the dashboard.
* fix(tests): use native homes in launchd fixture

* fix(tests): isolate plist home fixtures

* fix(tests): account for shell quoting in plist assertion

* fix(tests): encode plist whitespace references

---------

Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
* fix(tests): pair Windows CLI update plan session

* fix(tests): isolate Windows plan fixture lifetime

* fix(tests): retain plan fixtures until verifier exits

* fix(tests): prove paired request reaches junction guard

* test: report Windows toolchain refusal gate

* fix(tests): contain Windows plan fixture process lifetime

* fix(tests): retain ownership of Windows plan child

* fix(tests): remove ambiguous descendant socket proof

* test: prove Windows fixture job exit with owned handle

---------

Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
…g (#8265)

Hosted Windows TEMP is an NTFS 8.3 alias (C:\Users\RUNNER~1). realpathSync keeps the short name while the sync process canonicalizes CODEX_HOME with realpathSync.native, so its catalog path hashed to a different backup and restoration stripped the catalog to [].
* test: quarantine hosted Windows CI failures for 2.98.0

* test: use Linux paths in systemd admission fixture
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
…a (#8259)

Mistral exports list only accepted reasoning_effort rungs per model, and native Chat folds caller ultra through mapReasoningEffort / ladder clamp.

Fixes #8254

Co-authored-by: Masih-0x3 (Steve)
Co-authored-by: JUN <243035832+lidge-jun@users.noreply.github.com>
…eads (#8261)

* fix(link): hedge leftover-pin refusal and count in-flight readiness reads

* fix(link): keep identity pins that are not links of the token
* test(messages): fix Windows fixture hooks and spend-owner lifetime

* test(config): retry busy ownership lock in first-registration race children

The inline child used by "processes that start together on an absent home" exited 1 whenever recordOwnedConfigPath threw CONFIG_OWNERSHIP_LOCK_UNAVAILABLE on a contended first-write lock (CI #8176). Match the sibling childScript retry: bounded 400 attempts, 5 ms backoff, only for that error code; other errors still exit immediately. Barrier, 60 s timeout, manifest assertions and the production 50 ms lock budget are unchanged.

* docs: identify the tracked inventory required by sparse test checkouts

* test: honor conditional skips in Bun ownership fixtures

* test: isolate and settle Windows provider fixtures

Await tracked cleanup, disable irrelevant native Codex synchronization in synthetic provider configs, and seed account sets through one real store mutation. Keep metrics stale rows first after sorting. Separate Kiro server setup with a bounded 10-second hook while API assertions retain the default 5-second limit.

* test: bound source oracle reads without extending deadlines

* test: separate fixture preparation and teardown deadlines

* test(messages): separate 529 fixture setup and async teardown

* test(anthropic): restore fixture environment when shutdown rejects

Attempt both server stops and restore both environment roots in finally. Preserve undrained trees and propagate failures. Controlled rejecting-stop regression fails before the change; all five file cases pass with 32 assertions on Windows.

* test(anthropic): cover fixture drain failure retention

* test(tray): wait for published PID in probe readiness
rit245 and others added 27 commits November 12, 2026 03:39
…README notices"

This reverts commit 0579fdf.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Personal version 2.89.1-dalsoop.1 and installer notices; main carries the upstream version.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…mentation"

This reverts commit 6f55d6f.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Docs deployment guard for the personal setup; the inherited docs workflow is disabled at repository level instead.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This reverts commit 1908d7d.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Personal installer typecheck in personal CI and personal maintainer roster.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…sitory in CI"

This reverts commit 7b6dfa6.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Personal source installer, personal CI and ownership notes for the Mac personal build.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ory"

This reverts commit a90fe98.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Personal publishing notes in AGENTS.md for the Mac personal build.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This reverts commit 91e4a14.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Catalog visibility change for the Mac personal build; on the hub it would change which discovered models are visible (new-model default off), which the canonical hub has never run and device model lists depend on.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This reverts commit fd1669b.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Test fixture adjustment that belongs to the catalog visibility change reverted here.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This reverts commit 0e5dcb2.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Catalog visibility change for the Mac personal build; on the hub it would change which discovered models are visible (new-model default off), which the canonical hub has never run and device model lists depend on.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This reverts commit 2a22e73.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Catalog visibility change for the Mac personal build; on the hub it would change which discovered models are visible (new-model default off), which the canonical hub has never run and device model lists depend on.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This reverts commit ac7cd68.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Test isolation for the catalog visibility change and the personal release fixture reverted here.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This reverts commit e80b967.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Catalog visibility change for the Mac personal build; on the hub it would change which discovered models are visible (new-model default off), which the canonical hub has never run and device model lists depend on.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This reverts commit e02a9e0.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Test isolation that belongs to the catalog visibility change reverted here.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…models"

This reverts commit 3e2bd16.

Removed from main when main is realigned on upstream v2.100.2 for the
k3s hub: Catalog visibility change for the Mac personal build; on the hub it would change which discovered models are visible (new-model default off), which the canonical hub has never run and device model lists depend on.

The commit stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Upstream v2.100.2 carries 30 made-up provider keys and canaries in 26 test
files (redaction and error-body tests). gitleaks' generic-api-key rule
reports them, which would stop the merge commit and every pull request that
contains upstream history. Allow that one rule in exactly those files, with
the default rules extended, not replaced.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brings main from the upstream v2.89.0 line to upstream v2.100.2
(lidge-jun/opencodex 61147ec, the source npm 2.100.2 was published from).

Conflicts and the differences a default three-way merge would carry from
the old v2.89.0 release line are resolved in favour of upstream: the
merged tree is the v2.100.2 tree plus the personal changes this branch
keeps for the hub (subscription dates in account listings, slow-step
startup logging, secret detection). Those were re-applied onto upstream:

- src/codex/auth-api/account-list.ts: keep upstream's upstreamAccountKey
  and collision-identity credits; add the subscription parameter and the
  main-row subscription read next to them.
- structure/runtime.md, structure/cli-management.md: upstream text plus
  the kept slow-step sentence and subscription section.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…pstream answers

A /v1/messages turn can wait locally before the upstream sends headers
(pre-stream retry backoff, same-target 429 wait, OAuth refresh, pacing).
Nothing reaches the client meanwhile, and a reverse proxy in front of the
hub cuts the request: a Cloudflare tunnel answers 524 after ~100 s without
response headers, although the hub would have answered.

Once a request whose body says `stream: true` has waited
OPENCODEX_MESSAGES_KEEPALIVE_COMMIT_MS (default 30000, `off` disables)
without a response, commit `200 text/event-stream` and send an Anthropic
`ping` event at once and every OPENCODEX_MESSAGES_KEEPALIVE_INTERVAL_MS
(default 15000, at least 1000). The hub's own answer is then spliced in:
a streamed success is relayed unchanged; any other answer becomes one
Anthropic `error` event (type from the body, else from the status), since
the status is already sent. A client cancel cancels the late upstream body.

Turns answered before the delay keep their real status and headers, so
fast refusals (401, 429 with Retry-After) are unchanged. Non-streamed
turns are never committed: a JSON client reads the status first, and a
late 4xx/5xx sent as 200 would be parsed as a message.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 76b652d)
Pushes ghcr.io/dalsoop/opencodex:VERSION-k3s-SHA (linux/amd64) from the
deployment branches with the upstream Dockerfile; gujoai k3s-infra pins
the image by the digest printed in the job summary.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 33c5457)
…throwing decorator

Review of the previous commit: after the commit the upstream body was
pumped eagerly into the output stream, so a slow client buffered the whole
answer in memory and the inner stream (and its lease / request-log
settlement) finished at upstream end instead of client delivery. The
committed stream now reads one upstream chunk per client pull. Pings stop
as soon as the hub's answer arrives, before any upstream byte.

A throwing finish/finishCommitted callback now rejects the boundary
promise instead of leaving the request hanging.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 64aa5bb)
The fleet shows hub accounts from a snapshot that a Mac builds with
`ocx account list`. On a k3s hub no Mac can run that, and a data key must
not read quota or account identity, so this adds GET /v1/hub-accounts with
its own key. Without the key variable the route does not exist.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 95c7e52)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 5cdcd10)
…imit

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 0084624)
The hub image used to come from k3s/<version> branches. main now carries
upstream plus the hub fixes, so build from pushes to main and from the
hub/<upstream version>-<n> tags, keeping the image name and the
<version>-k3s-<commit12> tag. A commit that already has its image (main
built before the tag on the same commit was pushed) reports the existing
digest instead of building a second, different one.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Upstream's release workflow stamps the version into its four version
sources only in the packaging checkout, so the v2.100.2 tag commit
(61147ec, the npm 2.100.2 gitHead) still reads 2.100.0. Stamp the same
value with upstream's own scripts/release-version-sources.ts so the hub
image reports 2.100.2, which the k3s initContainer checks against
OPENCODEX_VERSION.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Upstream's file-size ratchet fails a file that reaches 2,000 lines without a
recorded cap. On 2.100.2 the kept test-layout entry and the cherry-picked
keep-alive call each pushed one file to exactly 2,000. Join each onto one
line, the same way 0084624 kept the hub-accounts route on one line. No
behaviour change; the ratchet baseline is not touched.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… 2.100.2

Upstream 2.100.2 takes the main credits identity from the collision identity,
so the subscription dates no longer share a token read with credits. Say
that the dates read the stored main Codex tokens (still no refresh, no new
credential source). Comment and structure text only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This reverts the re-applied change of commit
033aeee ("Let account consumers read stored
subscription dates without extra refreshes"), which the v2.100.2 merge
carried onto upstream's account list.

Upstream 2.100.2 requires a passive account list to make no native
credential reads (tests/... "stored Pool duplicate of native main > passive
account list projects known exclusion without native reads"). The
subscription read adds one (auth.json through readCodexTokens) and fails
that test. Keeping it would mean changing when the dates are read, and
nothing reads them from the hub: GET /v1/hub-accounts projects an
allow-list without subscription, and CodexBar takes the dates from the
OpenAI dashboard. The change stays reachable from archive/branch/main and
archive/m5m1-personal-77796d8.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Upstream 2.100.2 resolves every test file through the explicit layout map
(scripts/test-layout/layout.json, mirrored by tests/fixtures/
test-layout-expected.json). The cherry-picked hub-accounts and keep-alive
tests were only in the fixture (or nowhere), so tests/test-layout.test.ts
and the membership oracle failed. Add the three files to the map and
hub-accounts.test.ts to the fixture, on existing lines so layout.json stays
under the 2,000-line threshold. Test layout only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@dalsoop
dalsoop merged commit 4ed5470 into main Nov 11, 2026
33 of 34 checks passed
@dalsoop
dalsoop deleted the hub/align-v2.100.2 branch November 12, 2026 02:47
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.