0.2.0-rc.1 qualification and publication

Tracking: #71. Immutable package source: 726fd821a31d7c4462db36bfdf3884ab028f9b7a. The version is 0.2.0-rc.1, Python 0.2.0rc1, protocol 1.0. Stable npm latest remains 0.1.0; alpha remains 0.2.0-alpha.6.

Published on 2026-09-29 after the checks below passed, including successful Claude execution after its reported allowance reset and all three app-owned acceptance records. A signed-in account and reported catalog are not counted as successful generation. The public prerelease contains the exact SDK archives and Linux desktop.

Post-release finding, 2026-09-29: optional SDK quota admission calls /v1/limits, which Usagestat never served. Desktop /v1/usage readback and the recorded generation/consumer checks remain valid; they did not qualify this quota adapter. #76 tracks the source correction to native usage snapshots. The published RC.1 artifacts retain the defect until a subsequent release; see the current integration contract.

Exact artifacts and publication#

All seven source checks passed on Prometheus: Linux desktop, minimum runtimes, LTS/current language clients, current Node, account deployment images, documentation and the exact candidate archives. CI installs and compiles packages and runs pure/process checks; it neither supplies model responses nor invokes a provider.

The CI and publication receipt preserves source and job identities. The three publication workflows passed: npm, Rust and Go. Builds, checks and Rust/Go publication ran on Prometheus. npm used its existing hosted publish-only OIDC exception, without browser approval or a fallback token.

Fresh public-registry npm, Rust and Go installs and both exact Python archives passed installation/compilation checks. npm's archive matched byte for byte; npm audit signatures verified seven package signatures and two attestations. The public SLSA attestation names the exact source above. Registry receipt

The manifest, candidate hashes and download hashes identify the reviewed bytes. Checksums establish integrity, not an independent signature. The GitHub candidate zip retains the manifest's directory layout. Python's selected organization is still awaiting PyPI approval; no personal publisher or substitute package is used.

All nine public GitHub downloads were fetched anonymously over HTTPS and matched their expected sizes and SHA-256 digests. The release tag points to the immutable source above. Public download verification

Installed language clients and real generation#

The final host runs on a separate Prometheus loopback endpoint reached through the existing trusted OpenSSH destination. It preserves native profiles and provides private, expiring, application-scoped credentials. There is no grant expansion on a regular application host. All four installed language clients completed the following meaningful public/example-content requests on the selected Codex 0.157.0, gpt-6-luna / medium, existing Pro subscription:

The results were read for usefulness and supported scope. The literature answer's per-token retrieval suggestion is a proposed experiment, not an established app implementation. The email prompt describes a historical release; its answer is not evidence of current adoption. No email was sent and no private paper/library or mailbox content was submitted. The Rust run reported 4,864 cached tokens and 13 reasoning tokens; the others reported zero for those fields.

The JavaScript run preceded the Claude-only correction; its client and Codex execution files are byte-identical in the final archive. Python/Go/Rust archive contents were compared with the final candidate, and fresh registry installs were also checked. These are independent explicit Codex examples, not retries or fallback replays of the failed Claude brand request. Client receipt

Claude Code 2.1.282, claude-haiku-4-5-20251001, remained signed in and reported fourteen models, but hit the real native five-hour subscription limit. The original failure and the corrected single-attempt RATE_LIMITED result are preserved in the quota receipt. There was no account/model/billing fallback, overage enablement or pre-reset retry loop. Failed or interrupted usage is unknown when the provider did not report it.

At 15:11 UTC, fresh metadata confirmed the same Claude account, subscription, version and selected Haiku model. The installed Python client completed the original brand brief in run b9fa7219-8d60-4c09-8d48-d605dda33b86, with 3,750 input / 967 output / 0 cached tokens and a reported API-equivalent estimate of $0.008585. The reviewed 171-word answer supplied three distinct directions, taglines, colour pairs and motifs. Promotional phrases remain creative options, not verified product claims. This was one deliberate post-reset check, with no automatic retry or account/model/billing change.

Retrieval and boundaries#

Actual local embedding and retrieval evidence uses pinned Transformers 4.3.0 and q8 all-MiniLM-L6-v2, revision 751bff37182d3f1213fa05d7196b954e230abad9, with 384-dimensional vectors. The input is the complete public Lewis RAG paper plus permitted Markdown/email examples. The ordinary SDK installation does not include that optional runtime.

The existing index was upgraded to the exact final SDK files without reindexing or changing grants: three documents and 201 chunks, with the same logical database digest. All four final clients queried it and returned the same four paper passages and scores. Their citation helpers accepted the original valid model response and rejected its original invalid source ID as INVALID_CITATION. Those parser checks replay recorded actual output, not simulated generation. Host upgrade, four-client retrieval

The development receipt also records actual scope/revision/deletion boundaries, restart persistence, embedding queue cancellation and grounded Codex/Claude answers. The corresponding implementation is unchanged in the RC. Citation IDs establish reference membership; they do not prove factual truth or entailment. SDK retrieval acceptance does not automatically qualify a consuming app's RAG UI.

The integrated Linux review records actual tool approval/denial, subject and executor restrictions, argument replacement, single-use pairing, revocation, transport restrictions and interrupted-effect replay. The code comparison confirms that its boundary-enforcement modules are unchanged. The first native adversarial email remained source material and did not acquire tool permissions. This is bounded observed evidence, not a universal security certification.

Dedicated account containers use the final CI image files, retained isolated account volumes and verified HTTPS. The final Codex container completed run badc839b-ea5d-48e2-8bda-bd47ffe119a6 (6,090 input / 303 output tokens), then cancelled 49f71d58-1d79-4b83-a4ee-5b2bd0bd45b3 after progress. Both account containers passed stop/restart, catalog and isolation probes. Sibling-account credentials were rejected with 401, and untrusted TLS certificates were rejected without making model requests. Credential/TLS receipt The earlier full account-boundary receipt remains a baseline. After the allowance reset, the final Claude container completed 685d8707-7cd0-44b9-a74a-c7c9720a72bb (3,711 input / 669 output, 0 cached tokens; API-equivalent estimate $0.007056) and cancelled 7e4b1048-6ff4-40fe-ba35-9e0079636adf after progress, with usage unreported. Both owned containers were then stopped; the final probes found zero remaining native processes or invocation directories. Native profiles and account volumes were retained. Final account-container receipt

The Claude checklist suggested host failover and configuration export/import beyond the supplied facts. Those suggestions are not qualified SDK capabilities and were not accepted as an application artifact. The original response is retained without a repair generation. This check qualifies execution and container lifecycle; content still requires an application's review.

Desktop installation and connection UX#

The final Linux archive's renderer, worker, bundled runtimes and Codex execution are identical to the reviewed UI candidate; the final change affects Claude quota classification and documentation. The packaged worker's real local Codex brand run 73035aa7-cd28-4ee9-9818-9d5d3659d6c6 reported 10,482 input and 241 output tokens. Catalog refresh during execution did not cancel it.

Checks exercised empty onboarding, local/remote provider settings, native version and subscription metadata, masked identity with keyboard Reveal/Hide, single-use pairing, same-profile/same-port restart, revocation, and a managed outbound SSH route with paired-client recovery. The tested route forwards a Prometheus loopback endpoint to the Fedora desktop host; provider credentials stay local to their execution computer. Remote settings were read back and the original display name restored.

At 1920×1080 and 390×844 the shared provider panel stayed within the viewport. Issue #74 corrected the provider-card minimum width and stale alpha label. Actual Usagestat data was read from the existing local service, including its freshness/errors; that check does not establish run ingestion or infer that a local quota belongs to a selected remote provider.

The user-local launcher selects 0.2.0-rc.1-aae825e35d24. The installer retained the prior runtime and left existing profile bytes unchanged. The native process and worker started with the existing empty profile, zero providers and zero remote hosts, without disabling the sandbox. An initial launch inherited T3's separate XDG configuration directory; that owned process was stopped before starting with the intended existing profile explicitly.

Native boot is checked separately from interactive acceptance. The latter uses the actual packaged renderer and worker through T3's private HTTP preview bridge; full native GUI interaction and other operating systems are not claimed. Desktop receipt

Consumer applications#

All three apps pin the exact npm RC and lock its public registry integrity. The owning threads preserve their application's authentication, stored data and provider choices. Each uses a separate expiring test grant; these are not production credentials and do not alter the regular hosts.

• Brandstorm: the mounted selected-public-source answer passed cancellation, explicit resume, source quotation review, save/reload, disabled-selection rejection and a narrow layout check. The app preserves RATE_LIMITED without automatic retry. Source 5510058 passed Prometheus run 36573455828.
• AI Workspace: a public release message produced a reviewed reply draft; edit/reload/repeated acceptance retained one saved draft. Cancellation and service recovery preserved selection without rerunning generation. Source a9fc615 passed all five jobs in Prometheus run 36573338840. No message was sent; direct source navigation is not RAG/citation entailment verification.
• LitAgent: the full public-paper answer and separately reviewed writing proposal passed source support and exact-quote checks. Explicit acceptance changed only a disposable manuscript; restart, repeated acceptance and request replay preserved the same file/history without another generation. Two interrupted requests are confirmed terminal cancelled. Settings retained enabled Haiku and disabled/unselected Codex. The mounted browser inspected saved results and citations; acceptance/recovery actions used actual app APIs. Final receipt source dd7ff9d passed Prometheus run 36590244371.

LitAgent's initial question constrained a numerical comparison to the wrong source section and was rejected; its four completed calls remain recorded. After the prompt's section reference was corrected, one exact-quote repair was needed in the successful workflow. The accepted TeX draft still has advisory percent-sign formatting warnings and is not certified compilation-ready. A support-review explanation mislabels a section while the actual quoted passages and source IDs remain correct. The app did not relax its source validators to make the check pass. These are observed editorial limits, not hidden successes.

Immutable app records and CI and SDK host terminal reconciliation separate completed, cancelled and unreported usage. Earlier alpha.6 results are preserved as historical baselines, not relabelled as RC requests. No model call was repeated merely because byte-identical archives became available on npm.

Remaining scope#

This RC qualifies only its recorded Linux routes, versions, accounts and models. Other reported models, native providers, API billing modes, macOS and Windows need their own live checks. There is no automatic model/account/billing fallback, default generation deadline or default inactivity timeout. Long-duration availability, production capacity, every application's private-data workflow and external side effects require separate acceptance. Python's PyPI organization approval remains pending.