Current qualification
0.2.0-rc.2 passed its scoped gates and is published. It corrects native quota
reads and qualifies the released Usagestat 2.0.0 ingestion contract. Python uses
0.2.0rc2; protocol stays 1.0. Artifact source:
91dc52292c627a6febed102a8198c56f35d3afc9.
The RC.2 receipt records all seven Prometheus source checks, exact registry/public-download verification, real Codex and Claude capture, and all three app-owned registry migrations and metered workflows. The normal LitAgent host and usage service preserve all thirteen stored records and existing credentials/grants; the local desktop preserves its original profile. Use RC adoption for installation and host/client upgrade order.
Current scoped limits include pending PyPI organization approval, the earlier provider/platform restrictions, and a single Brandstorm selection propagation observation whose reproduction is pending. SDK model selection emitted its event; no SDK-owned defect is proven. Do not claim first-click Brandstorm setup fully qualified. Wider retrieval, security and cancellation checks below remain the historical RC.1 baseline.
RC.1 accepted baseline
0.2.0-rc.1 passed its scoped release gates and was published on 2026-09-29.
Python uses 0.2.0rc1; wire protocol remains 1.0. Tracking:
#71.
The immutable artifact source is 726fd821a31d7c4462db36bfdf3884ab028f9b7a.
Use RC adoption for installation and connection setup, the
public release
for downloads, and the complete receipt
for artifact hashes, actual runs, source comparisons, CI and qualification limits.
Stable npm latest remains 0.1.0 and alpha remains 0.2.0-alpha.6.
Qualified scope
The RC covers the Linux x64 host and desktop, TypeScript/JavaScript, Python, Go and Rust clients, and the explicitly selected Codex/Claude subscription routes through local execution, trusted OpenSSH or verified HTTPS. The exercised models are Codex 0.157.0 / gpt-6-luna / medium and Claude Code 2.1.282 / claude-haiku-4-5-20251001. Other account-reported models remain visible, with per-connection permissions; discovery is not live qualification.
| Gate | Evidence |
|---|---|
| Real embeddings, scoped persistent vector retrieval, PDF/Markdown/email inputs and citation IDs | Actual model/document checks, final four-client searches |
| Dedicated Linux account containers, native execution, cancellation and recovery | Boundary review, final-image lifecycle |
| Execution, pairing, retrieval, credential and tool boundaries | Actual Linux checks, unchanged enforcement code |
| Three app-owned workflows, persistence, explicit selection and cancellation/recovery on the final RC | Immutable consumer receipts and CI, host terminal reconciliation |
| Fresh installation, preserved host/desktop state, local/remote pairing and SSH recovery | Packaged desktop and user-local installation |
| Exact-source Prometheus builds, fresh language installs, hashes and provenance | CI/publication, registry verification |
| Publication and public downloads | GitHub verification, release inventory |
At the RC.1 gate, all three applications pinned @agenticdriver/sdk@0.2.0-rc.1
with registry integrity; the current pins are RC.2.
Brandstorm reviewed a selected-source answer; AI Workspace saved and edited a
reply draft; LitAgent accepted a supported writing proposal into a disposable
manuscript. Their own authentication, data and saved choices were preserved.
No private library/mailbox data or external email send was required. SDK-owned
retrieval acceptance does not automatically establish vector search inside an app.
The first candidate exposed a real Claude allowance limit and incorrect generic
error classification. #75
corrected it to RATE_LIMITED; the same account/model succeeded after the
reported reset. The original failures and recovery
are preserved. No fallback account/model/billing, overage enablement or pre-reset
retry loop was used. Earlier candidate archives were superseded, not published
under the final artifact's identity.
Limits retained for the candidate
- PyPI organization approval remains pending; reviewed Python wheel/sdist files are published on GitHub. No personal publisher is substituted.
- API billing routes, other native providers/models, macOS and Windows need separate live qualification. Account-container checks assume trusted host and Docker administrators; they are not a universal security certification.
- Native desktop boot was checked separately from interaction with its actual packaged renderer/worker through T3's private preview bridge. Complete native GUI and other-platform interaction is not claimed.
- Model output needs review. Unsupported checklist suggestions, an initially rejected literature question, a quote repair and advisory TeX formatting issues remain visible in the receipts. Citation membership does not establish factual truth or entailment. A supported draft is not automatically publication-ready.
- Expiring validation grants are not production application credentials. Use scoped host/desktop connection setup for ongoing applications. Usage UI readback does not establish automatic Usagestat ingestion or account binding.
- There is no default generation deadline, inactivity timeout, account/model fallback, public relay or automatic desktop updater. Capacity and long-duration production reliability remain workload-specific qualification.
The older roadmap remains the historical planning baseline. Current provider and publication follow-ups stay in the organization project; completing this RC does not close unrelated provider or platform work.