Skip to content

Release v2.117.0 - Rhythm Raccoon - #15767

Merged
atomantic merged 236 commits into
releasefrom
codex/release-agent-c1aea064
Nov 11, 2026
Merged

atomantic merged 236 commits into
releasefrom
codex/release-agent-c1aea064

Conversation

@atomantic

@atomantic atomantic commented Nov 11, 2026 •

Copy link
Copy Markdown
Owner

Release v2.117.0 - Rhythm Raccoon

Frozen candidate with new creative backends, improved full-song lyric and beat timing, reproducible local media tools, character comparison sheets, and calendar/Suno reliability fixes.

Frozen identities

  • Development branch: main; publication branch: release.
  • Release cutoff: c99a8e71e4d84f1f62b54c715c3f02a63efee285.
  • Release base: 33dc1ebf7271125bb20f9b9ec03a8dde6252d7ed.
  • Previous published tag: v2.116.0, peeled commit 33dc1ebf7271125bb20f9b9ec03a8dde6252d7ed.
  • Candidate branch: codex/release-agent-c1aea064.
  • Candidate head: a92217f7a39b496a7a0069183ea0ab9ff7be45fc.
  • Version/name: 2.117.0 - Rhythm Raccoon.
  • Full release comparison.
  • The metadata-only pin is the cutoff's direct child. No later main was merged or rebased into this candidate.
  • At least two substantive user-visible changes are included; the highlights and complete notes are in .changelog/v2.117.0.md.

Candidate amendments

The amendments below address evidenced candidate validation failures and confirmed partial reviewer leads. These exact commits are the only exceptions to release preparation; all will be forward-ported through the normal post-publication synchronization PR.

Commit Candidate-gating evidence and justification Focused validation
c02f08bb18 Omitted MiniMax steps produced invalid runner arguments. 12 runtime tests
02b04ff667 Floating calendar timestamps lost fractional seconds. 54 timezone tests
9832c9f0ae A timed line hid additional singing within the same vocal phrase. Lyric verification workflow regression
1fbc9716d8 Unrepairable texture cards offered Repair without the Xcode prerequisite. Degraded-state projection regression
9b2a774f9d Unrelated downloads were misreported as the requested Suno export. Suno export timeout and identity controls
7781ee5e56 Negative-only Google image requests bypassed the positive-prompt requirement. Google request validation
e0ddb80333 Negative-only Higgsfield image requests bypassed the positive-prompt requirement. Higgsfield request validation
af7012ede6 Cancellation during interpreter probing still launched package mutation. 15 package-install tests
81d8788f74 Empty lyric insertion and checking during pending saves broke sheet/report consistency. 5 rendered lyric-check tests
99fc9e1ae0 A different selected comparison partner blocked saved-pair regeneration. 119 CatalogIngredient tests
6ff2a43044 Higgsfield reference/audio requests silently discarded selected frame anchors. Hosted submission and route regressions
c3e396ba11 Supported Higgsfield end frames were omitted by music-video callers. 6 server files/378 tests; client plan/4 tests across both frame fixes
7d9a8ecca0 Pending cancellation uploaded private references; expired uploads still submitted a paid render. 2 deferred image-workflow regressions
521a8dfe66 Cancellation during geometry probing still published hosted video. 2 provider workflow regressions plus existing request test
e8d7ab808d Metal repair could clear a symlink target and rejected named Python executables. 16 installer contract tests
4297192619 Extreme reference aspect ratios allocated unbounded MiniMax conditioning canvases. 5 Python-backed tests with installed numpy interpreter
c055d8a771 A partial comparison publication permanently omitted character halves on retry. 5 comparison publication tests
c333bbde84 Incomplete or unrelated vulnerability scans could be reported as clean. 18 scanner workflow tests
7ef5acd60e Mood-board reimports let a later peer copy replace genuine local ownership. 25 origin/localization/image-record tests
4091acd4a6 Film-look fixture captured a stale 93-pixel initial viewport surface despite identical filter/frame data; production already navigates before capture. 10 cold isolated runs passed after initial navigation; strict pixel assertions and deadlines retained.
1ab22e4acf Hosted Grype 0.121 scan evidence wraps validity in db.status; the prior flat parser rejected every real scan. 18 scanner tests plus successful replay of all five hosted artifact pairs; invalid-DB/source controls retained.
6ea2472cc5 A fixture child exited before readiness while the test discarded startup evidence. Explicit ready/failure IPC preserves bounded diagnostics without raw stderr. Original exit cause remains unknown. Whole fixture file 3/3 and three focused runs passed; all assertions, isolation and deadlines retained. Non-reproduction, not diagnosis.
e3f6d226a0 JSON writeback tree guard exceeded its deadline while repeatedly walking whole-module tokens and irrelevant initializer ASTs. All 6 guard tests pass; all 307 preselected service outputs match exactly; source filters, exceptions and deadlines unchanged. No stable wall-time speedup claim.
a92217f7a3 The interrupted migration writer could naturally exit 13 before its parent sent SIGKILL; explicit IPC lifetime keeps the deliberate pause alive. Original harness reproduced natural exit 13; fixed harness stayed alive until SIGKILL; exact CI wrapper passed 9/9 with all deadlines and data/retry/cleanup assertions retained.

Review outcome

Configured reviewer: codex[gpt-6.1-terra]~opt~max=2~effort=high only. Ran 1 of 2 permitted rounds at 56ce1c687bac6dd40b9acd1c98e276c991a45ce8. The aggregate returned ok:false, REVIEW_CONTEXT_CAPACITY, so its status is INCONCLUSIVE, with no approval or completed-review attribution claimed. It covered 462 of 465 files in 25 calls; three indivisible large files exceeded the bridge's bound: the runtime image disposition ledger and the AuK/mflux-video lockfiles. Private raw response, exact diff and reviewed SHA are retained.

Its partial leads were investigated: 23 confirmed leads are fixed above, 7 rejected below. All confirmed partial review leads are addressed. No extra reviewer was substituted and no second call was spent repeating the same known capacity refusal. Full validation and hosted merge gates remain mandatory.

Rejected lead Evidence
Restore territory disclosure in UI and metadata (2 leads) Deliberate removal in the frozen model-terms change; PRD NR-11 retains informational license links and upstream enforcement, without a separate PortOS territory contract. Links remain available.
Torch 2.15 / torchvision 0.29.2 mismatch Official package metadata requires torch>=2.14.0; the pinned pair satisfies it. The frozen runtime guide records native imports, pip check, Install/Repair and textured export.
Empty migration-450 release record Supported release writers require both nonblank machine identifiers; readStore validates/projects records before persistence. An empty object requires unsupported manual store corruption. Applied migration remains unchanged.
Multipart false enables provider audio Preview, workers and builders normalize non-true to explicit false/off; no wrong provider result exists on the traced path.
Preserve dirty merged-checkout content Root AGENTS explicitly authorizes discarding the exact merged delivery's own checkout files after ownership checks; adding a fingerprint/receipt contradicts that contract.
Local render pool forbids hosted Higgsfield The pool selects this install versus peer placement. Existing tests explicitly preserve hosted fal/Grok backends; the same placement rule applies here.

Validation

  • Current committed-head full pregate: PASSED, all 16 selected stages on a92217f7a39b496a7a0069183ea0ab9ff7be45fc (tree 6e26f27097f80070dc6c75c766e0c8e40e9d941c), finished 2026-11-11T17:06:43.210Z. Test-memo evidence is reused where every input is unchanged.
  • npm run setup:db:test and the complete isolated portos_test suite passed. The separate post-diagnostic DB run passed 95 files / 867 tests / 1 existing skip; all focused regressions are listed in the amendment ledger.
  • Local full validation retains documented browser-fidelity exclusions and unavailable platform stages. Windows, Linux containment, native Docker/SuperCollider, macOS native supervision, and the full hosted release browser profile require their exact-head CI results.
  • Applicable PR workflows: CI (.github/workflows/ci.yml, requiring both CI Gate and Full CI Gate) and Runtime image security (.github/workflows/runtime-image-security.yml, because scanned recipes changed). Manual-dispatch-only diagnostics are excluded by their triggers. Exact-head hosted verification passed all 26 checks, including both CI gates: CI, Runtime image security.
  • Earlier hosted run 43041438224 failed Ubuntu prerequisite downloads before tests; one unchanged-head retry repeated Linux DNS and DB/browser package-download failures. The external dependency is tracked in Restore Ubuntu package access for required hosted CI #15772. No product-test failure or diagnosed-flake claim is substituted for those logs.
  • Historical collection-fixture exits remain unexplained in Diagnose intermittent collection fixture exits before SIGTERM readiness #15779. Explicit readiness/failure IPC fixes lost diagnostics; repeated focused runs and the full DB suite passed. Those passes are non-reproduction, not a root-cause claim.
  • The migration interruption race has causal proof: the original paused child naturally exited 13; explicit IPC lifetime kept it alive until SIGKILL. This matches documented IPC lifetime and unresolved top-level await behavior. The original full-run test omitted its exit code, so no retrospective code is invented.
  • New upstream scan matches are tracked in Review newly reported runtime image advisories against actual runtime paths #15768 (62 pgvector and 10 SuperCollider package/advisory pairs per architecture). PR scanning reports these for reachability triage; strict scheduled scans continue refusing unreviewed matches. No unsupported clean disposition was added.

Deferred to the next release

Unrelated issues, dependency updates, in-flight PRs (including the preloaded #15753), and development commits after the cutoff remain in development. No backlog sweep or dependency queue merge was performed. The required post-publication quality snapshot is a development update.

Decisions for operator review

  • Use a minor bump, 2.117.0, because the frozen range contains new user-visible features.

  • Name this release Rhythm Raccoon for the lyric/beat timing improvements.

  • Refuse unsupported anchored-frame/reference-audio combinations instead of silently dropping selected inputs.

  • Preserve the validated compatible Torch lock and the existing license, placement and disposable-checkout policies.

  • Treat incomplete optional review coverage honestly as inconclusive; fix confirmed partial leads and rely on required CI for admission.

  • Match the film-look browser fixture to the renderer navigation lifecycle, retaining every strict pixel assertion and timeout.

  • Forward-port candidate fixes and version metadata through the owned synchronization PR after workflow-verified publication.

  • Preserve existing test contracts while fixing repeated scanner work and child-process lifetime; keep unexplained historical fixture exits explicit in Diagnose intermittent collection fixture exits before SIGTERM readiness #15779.

Verified publication and synchronization

The repository delivery verifier independently confirms all eight delivery and quality obligations against the merged synchronization and development snapshot. This inspection does not persist synchronization receipts or claim a clean local review of unreviewed fixes. The only pending obligations are active-process settlement and service-owned checkout retirement after this run exits.

atomantic and others added 30 commits November 10, 2026 11:46
…e106fbbdb80feea24554da22b9e1024497b2a7034313de125dc4750

chore(release): synchronize v2.116.0 - Steady Otter into development
Use uv 0.13.6 for FastVideo, MLX-Gen and MiniMax H3 bootstrap tools. This includes the compromised CA removal and wheel entry-point, RECORD, ZIP and TAR validation fixes. Guard all bundled pins against falling below the patched release floor.
Update client build, icon and testing tools, align Playwright with the server, and refresh shared ws, redirect, query parsing and glob overrides. Regenerate all four npm lockfiles while preserving transitive major-version compatibility.
…updates-mwatgxoj/agent-fcc3f2c8

fix: update dependencies and patch bundled uv vulnerabilities
…mages (#15656)

A new Runtime image security workflow (weekly, on demand, and on PRs touching
the recipes; not part of CI Gate) builds or pulls the pgvector and SuperCollider
images on native amd64 and arm64 runners, generates a syft SBOM and a grype scan
including OS packages, and uploads both with provenance (image, digest, recipe
hash, commit, architecture, scanner and DB identity). Scanner errors or empty or
malformed output fail the run instead of reading as zero findings. The inline
Rapier npm pin is scanned too; the externally built eidoverse render parent and
its deno.lock are reported as unavailable with the owning project. The scheduled
dependency-updates prompt (v11) now consumes this evidence or names the gap.
…e patched uv (#15654)

Prompt Guard and jev now install a complete hash-pinned closure (new shared
lock, CPython 3.11+ because filelock's security fix needs it) instead of four
direct pins. The Reactor runtime installs its hash-pinned lock and bootstraps
uv 0.13.6 instead of 0.12.10. Generic package repair dispatches to a locked
Apple Silicon profile (mflux, or the new mflux + mlx_video + mlx_vlm lock) and
reports an unsupported profile before installing anything else, rather than
solving an unpinned graph. The legacy INSTALL_VIDEO path in the image/video
setup script follows the same rule.

PortOS-owned uv use (package installs, FLUX.3 runtime) now runs a private,
version-checked uv 0.13.6 bootstrapped under ~/.portos/tools, never the uv on
PATH.

Refs #15654
GitHub rejected the workflow file because the runner context is not
available in job-level env.
…heckpoint (#15652)

Lyric alignment no longer installs torch 2.2.2/2.8.0. Its private venv now
installs a universal, hash-pinned lock (torch 2.14.1, torchaudio 2.11.1,
numpy, every transitive package) with pip --require-hashes --no-deps
--only-binary, resolved with a seven-day cutoff and checked against OSV.

The alignment sidecar owns its CTC Viterbi and token merge instead of
torchaudio's forced_align, and loads MMS_FA weights itself: the checkpoint
bytes must match a pinned SHA-256 before torch.load(weights_only=True) sees
them. A substituted cache or download is rejected without deserializing it,
and a genuine older torch.hub copy is reused. Intel Macs and macOS before 14,
which have no patched PyTorch build, are refused before anything installs;
Apple Silicon requires an arm64 Python 3.11-3.14.
…audio after namesake removal (#15654)

Review round 1: install uv from a hash-pinned lock (python, Reactor-style
verification) in both the Node and shell bootstraps; skip package install when
cancelled during uv bootstrap; force-reinstall mlx-video-with-audio's own files
after removing the wrong mlx_video namesake; fail INSTALL_VIDEO=1 on Intel Macs.

Refs #15654
Google stores all-day events as offset-less midnight and Outlook stores
them as UTC midnight. Agenda, voice, and daily review compare those
strings with Date on the UTC server, so a non-UTC user sees the
neighboring day's all-day event and all-day touchpoints close at UTC
midnight. Anchor those instants in the user timezone and treat the
all-day end as exclusive.
…nd sanitize uv bootstrap errors (#15654)

Review round 2: pass --user to the targeted reinstall on non-venv interpreters,
stop INSTALL_VIDEO setup when either pip step fails instead of letting cleanup
mask it, and report a fixed uv bootstrap failure reason rather than pip stderr.

Refs #15654
…ivately (#15652)

Packages are only installed into a cleared venv, and a stamp of the lock's
SHA-256 decides reuse, so nothing outside the hash lock survives an upgrade.
A genuine shared torch.hub copy is copied into the private cache while it is
hashed instead of being loaded in place, and a verified file is published
with one atomic replace so concurrent first runs cannot remove each other's
cache.
Adding hours to local midnight used the midnight offset for the whole
day, so a daily-review bound on a daylight-saving transition day
included the next morning or dropped the last hour. All-day touchpoints
also kept the raw UTC midnight, which Tribe records as the previous
local day.
feat: add SBOM and vulnerability scans for pinned runtime container images
On an x64 host the installer launches the private uv through
/usr/bin/arch -arm64, so the uv-binary assertion only held on arm64.
Outlook stamps an all-day midnight with Z. Ambient and the day grid
parsed that as an absolute instant, so a July 4 all-day event ended at
UTC midnight and also occupied the previous local day.
fix: move lyric alignment to patched PyTorch with a digest-verified MMS_FA checkpoint
fix: install Python runtimes only from hash-pinned locks and a private patched uv (#15654)
…lt-app-bugs-mwautr16/agent-0575229e

fix: place all-day calendar events on the user's local day
@atomantic atomantic added the implementer:gpt-6-1-sol PR opened by the gpt-6-1-sol model label Nov 11, 2026
@atomantic
atomantic merged commit a344eb1 into release Nov 11, 2026
26 checks passed
@atomantic
atomantic deleted the codex/release-agent-c1aea064 branch November 11, 2026 17:20
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

implementer:gpt-6-1-sol PR opened by the gpt-6-1-sol model

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants