Exercise real BSON optional fields, conditional head writes, concurrent
attesters, reconcile preservation, replacement cleanup and fenced claims
against disposable MongoDB 7. Verify conditional BOS ranges and HTTP 412,
then drive the real API/coordinator from legacy partial receipts through
attestation, completed sync and cleanup proofs with synthetic objects.
Cover idempotent retries, local-content conflicts, conflicting object MD5,
changed identity and missing trust anchors. Keep credentials on stdin and
out of persisted Mongo records; conditionally create private objects and
clean only exact test-owned keys. Record cleanup and environment failures.
All Go packages, vet and the real Mongo race tests pass. No production
runtime code, deployment, business archive or published client is changed.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Integrate claude/happy-kilby-4915b3 and make its receiver gate deterministic:
fill both file slots before pausing, fail on hold timeout, accept buffered
sender progress, and reject new reservations, cancellation or retries.
Add eight fixture regressions. Reproduce the original delayed-reservation
false failure, then verify the corrected real CLI scenario. macOS runs
10 pause repetitions and the full 26-scenario matrix; offline Linux x86_64
runs 10 pause repetitions using the unchanged published 1.1.6 binaries.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Separate portable releases from the native update channel with embedded
Ed25519 trust roots, bounded verified downloads, immutable release objects
and a compare-and-swap stable pointer. Keep installation manual.
Split the Linux CLI's privileged installer from service-user IPC and state
access. Reverify signed plans as root, validate both private archives before
stopping systemd, retain baseline recovery and confirm live binary readiness.
Preserve configuration, user, state and enablement through upgrades.
Validate 317 macOS and 320 Linux workspace tests, Clippy, publisher race
tests, and 18 real systemd update/rollback/failure scenarios. No field
devices or POC API are modified by this implementation.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The native pause-resume scenario failed with both old and new builds
because the harness could not reliably see an upload in flight. Pause
itself behaved correctly.
Client progress counts bytes handed to the socket, and the daemon writes
it to SQLite once per second. Loopback buffers absorb about 1-3 MiB of a
body (sometimes over 5 MiB on Linux), so a 4 MiB PUT read as 0 or as
complete in nearly every sample. The trigger fired on the 3rd or 4th PUT
at best, later or never under load, and never on Linux. With
FILE_CONCURRENCY=2 and PUT_CONCURRENCY=1, the file queued behind the
in-flight PUT is already dispatched and finishes too, so a late pause
left nothing paused.
The fixture now holds the first direct body after 64 KiB until the
harness has paused. pause-resume uses the 12 MiB default size, so a held
body still reads as partial. The scenario asserts that the held object
arrives intact and that every received object was reserved before the
pause. The daemon already follows the documented contract: pause stops
new dispatch and lets in-flight objects finish.
pause-resume: macOS arm64 10/10, Linux arm64 10/10 (old harness 0/5 on
Linux). Full native matrix on macOS: 26/26.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The workspace mount had no readahead or parallel transfer options, so
bosfs read a file back at ~3-5 MB/s. A800 checkpoint publication hashes
each 2.7G file twice through the mount, taking ~1h per checkpoint and
tripping the 1800s per-file limit. Baidu's recommended general options
raise reads to ~370-390 MB/s (verified on 192.168.80.136); the live PV
was patched the same way.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Bind recovery to the original receipt and scope, verify the held file and MD5, persist strict restart checkpoints, and atomically refresh identity without reusing cloud authorization. Cover direct retry, replacement protection, stale journals, and transactional failures; record Linux validation limits.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Add isolated current/legacy CLI and cleanup API comparisons, legacy MD5 planning coverage, and reproducible evidence. Record stale deployment identities and the need to refresh source receipts after failed cleanup restoration.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Validate the dataset's 78-dimensional actions, export split records instead of the leakage report, and render selected training episodes with supported template fields. Add regression coverage for record round-tripping, route counts, and empty reports.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The leakage validator resolved an upstream dataset-split-manifest artifact to the
unsplit episodes file named in its sourcePath, so every row lacked a split. It now
reads the manifest itself when no SOURCE_PATH is set. ego-stage gains a media mode
that lists staged videos as a manifest, and ego_workflows.py adds the cpu_extra
workflow covering 13 more CPU operators.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Builtin operator images were rebuilt four times after 2026-06-29 but the
catalog kept the first digests: recording a release was a manual step
that nothing ran. Once the old tags were pruned from CCR, 66 builtin
operators pointed at manifests that no longer exist.
- build-poc-images-kaniko.sh now runs record_operator_releases.py after
operator builds (POC_KANIKO_RECORD_RELEASES, default true): it resolves
each pushed tag to its digest, writes release.json with the stable
:builtin tag, regenerates the catalog and repository manifests, then
runs verify_operator_image_digests.py on the built operators.
- verify_operator_image_digests.py fails when any pinned catalog digest is
missing from the registry (usable in CI / before deploys).
- sync_live_builtin_operator_images.py promotes catalog digests to a
running platform (the API keeps live digests over the catalog), only
after confirming each digest exists; dry run unless --apply.
- scripts/lib/registry_client.py parses repo:tag@sha256 references and
handles CCR's Bearer token challenge.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add "cpu_data" (Parquet summary, action/state quality, episode
boundaries, language instructions, embodiment metadata, split manifest)
and "cpu_video" (10s slicing, motion clipping) workflows that need no
GPU. The ego-stage operator gains an "episodes" mode that rolls this
dataset's per-subtask episodes.jsonl up to one row per episode with a
length, which the split manifest needs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Define "full": stage -> LeRobot overview -> Parquet schema -> metadata
consistency -> frame alignment -> SAM3.1, joined with left-camera video
QC into mask collection -> ProPainter hand removal -> upload. Workflow
inputs bind the two root nodes. Drop the ineffective runner-image
override (builtin nodes are re-resolved from the catalog by the worker)
and add pin-operator to repin builtin operators to a pushed digest.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add the ego-stage lightweight operator: it downloads presigned BOS files
into the run (or a stable /workspace/ego-staging path), collects another
node's outputs, and registers files as artifacts with ids. Operators such
as SAM3.1 and hand removal only read local paths, and their own images
emit artifacts without ids, which the current worker rejects.
Rework scripts/ego_workflows.py around what the current worker runs
reliably: short chains, one consumer per stage, validators on the
operator-runner image, plus commands to register the stage operator,
stop runs, read logs/run details and download results.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Video QC now receives its video list as an inline workflow input whose
items carry 1h BOS presigned URLs, and WiLoR gets a fresh presigned
INPUT_URL before each run, so neither depends on the worker turning
bos:// source outputs into refs. Add logs/worker/artifacts commands for
diagnosing runs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
scripts/ego_workflows.py inspects an egocentric LeRobot dataset on BOS
through the platform API and creates processing workflows for it:
integrity checks, left-camera video QC, WiLoR 3D hand pose, and
SAM3.1 + hand removal (node settings cloned from the proven egodex
pipeline). Source and upload nodes are bound to the org's storage
connection so the worker can turn bos:// outputs into object refs.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The old shipped v1.0.0 used a 300-second whole-response deadline. Keep active transfers alive, validate ranged retries, filter BOS folder markers, and preserve accurate failure/progress reporting. Verify the Linux package with interrupted 4 GiB transfers, a 307-second stream, and the real 7.47 GiB checkpoint with matching SHA256 receipts.