0.2.0-rc.2 qualification and publication

Tracking: #77. Immutable artifact source: 91dc52292c627a6febed102a8198c56f35d3afc9. Version 0.2.0-rc.2, Python 0.2.0rc2, wire protocol 1.0. This update corrects native Usagestat quota reads and qualifies the released Usagestat 2.0.0 integration. It preserves the earlier RC's provider and platform scope. All nine public release downloads were fetched anonymously and matched their expected sizes/digests. The release tag points to the exact source above. Download verification

Exact packages#

All seven SDK checks passed on prometheus-agenticdriver-01: desktop, minimum runtimes, LTS clients, current Node, account images, documentation and freshly installed candidate archives. The first attempt failed at the disk-space preflight before tests. Only the dedicated SDK runner's Go cache and private Docker build cache were reclaimed; the same source then passed. Shared services, images, credentials and other runners were preserved.

npm, Rust and Go publication passed. npm used the existing hosted publish-only OIDC exception; all builds/checks and Rust/Go publication ran on Prometheus. No npm browser approval or fallback publisher credential was used.

The public npm archive matched the reviewed bytes, and a fresh registry install runs the upgraded normal host. Seven registry signatures and two attestations passed npm audit signatures; the SLSA statement names the exact source above. The rc tag is 0.2.0-rc.2; latest remains 0.1.0 and alpha remains 0.2.0-alpha.6. Rust contents match the reviewed crate; Go's immutable tag points to the source above and its publisher verified a fresh public-proxy install. Python remains a reviewed GitHub wheel/sdist while the selected PyPI organization awaits approval.

Manifest · Archive checksums · CI/publication · Registry verification

Stable Usagestat and real usage#

The public Usagestat v2.0.0 Linux x64 daemon was checksum-verified: dc3338f29b111e629191ac1d196dc39c51f71ac016881dd3f960add59af3acb9, from source fbf84796d80fd5567516cec2e11b0784e7c187db.

• Native quota readback used the existing real daemon on port 6736: six snapshots, five explicitly selected administrative bindings, two ready and three unavailable accounts, including a failed snapshot with cached metrics. Fourteen native reads and five unbound-subject rejections passed without model calls or installing these administrative test bindings on an execution host.
• The packaged desktop's real worker read all six native snapshots, retained its usage settings over restart, and started with zero providers/connections in a fresh profile. The ordinary desktop installation was upgraded and booted with its original empty profile intact. Full native GUI interaction was not rerun.
• A separate loopback host on Prometheus ran the exact RC archive, with the existing native Codex and Claude accounts and a separately authenticated stable Usagestat service. An existing trusted OpenSSH route carried client traffic. Credentials stayed in private files; model credentials stayed on Prometheus.
• Fresh discovery reported Codex 0.157.0, signed-in Pro, nine catalog models; Claude Code 2.1.282, signed-in Pro, fourteen catalog models. Only gpt-6-luna / medium and claude-haiku-4-5-20251001 were selected for new execution. Discovery is not qualification of other reported entries.

These are native measurements, including runtime overhead. Cached tokens are a subset of input; the Claude estimate is not a subscription charge. Both runs completed while their catalogs were refreshed. The Codex answer met the requested brand format. The Claude draft preserved the historical handoff facts and pending PyPI approval, but appended an inaccurate word-count claim; it needs that edit before use. No email was sent. Prompts and run receipt

Automatic backend lookup succeeded before any explicit resubmission of the records. Identical resubmissions returned duplicate receipts. A Usagestat restart retained the records, and credentials for other application subjects could not read them. Capture remains separate from quota snapshots and daily-import totals. No parallel SDK usage backend was introduced.

Quota readback · Packaged desktop · Installed desktop

Existing regular LitAgent host#

The SDK-owned normal host at port 7433 now uses the exact registry RC.2. Its existing ingestion-only service at port 7436 was upgraded from the development alpha.3 binary to released Usagestat 2.0.0. All thirteen existing stored records read back unchanged, first from an isolated SQLite backup and then from the upgraded running service. Private configuration and credentials matched before and after; host/account/subject identities, grants, service enablement and saved application choices were preserved. Previous runtimes and a private backup were retained. The live database was not replaced.

The upgraded normal host returned both native catalogs with the existing application credential. This maintenance made no model calls and did not enable the app's saved disabled selections. The usual native readback daemon at port 6736 was unchanged. Upgrade receipt · Host readback

Consumer acceptance and scope#

All three apps pin the exact registry RC and have passing Prometheus CI. Each completed one real Codex/Luna workflow through its actual application, with automatic stable-backend capture matching the supplied identity and reported usage. Account identity remained masked and management stayed read-only.

Immutable consumer receipts · All five real records reconciled

Brandstorm retained the answer and selection, and rejected disabled execution before streaming. Its initial setup had one selection-propagation failure: the SDK emitted the expected event, but the outer form stayed empty until a read-only panel snapshot and repeat selection. This was not a verified use of the visible Refresh button. App follow-up #3 tracks reproduction and the connection-URL readiness in its selection bridge. No SDK-owned defect is established, and first-click onboarding is not certified. The test question also retained an old naming suffix; the answer marked that unsupported request unknown without another inference call.

LitAgent kept its supported finding as a pending proposal through restart; normal connection settings and disabled selections remained unchanged. This used public reading notes, not a new full-PDF retrieval evaluation. AI Workspace retained exactly one reviewed/edited draft after reload and repeat acceptance, and recovered its checkpoint from an app-backend restart without replay. No private mailbox/library content, external send or automatic application artifact acceptance was needed. Wider source and editorial limitations remain in each application's receipt.

Historical RC.1 retrieval, account-container, security and wider cancellation results remain prior evidence; this update does not claim to have rerun every earlier scenario. Other providers/models, API billing routes, macOS and Windows retain their separate qualification requirements. Temporary validation connections are distinct from ongoing application credentials. No application auth migration, fallback account/model/billing, default generation deadline or default inactivity timeout was added.