跳转至

Historical version

This page records design, operation, or validation facts from an earlier release. It is not current V7.17.3 guidance. Start from the documentation hub for current instructions.

V7.13.0 validation record

This record covers compatibility support: frozen V3 remains the default, and new models are restricted to explicit evaluation. Code/artifact evidence is separate from native-budget and model qualification. Default activation is not claimed.

Boundary Evidence
V4 routing, budget, review, evaluation and observation 111 focused tests passed after the final Astra concurrency fix, using synthetic host/quality fixtures
Desktop component adapter 4 tests passed; a separate local read-only component/registration probe passed
Installer diagnostics 41 UX tests passed
Frozen protocols and evidence 129 affected checks passed after repair, including Matrix Hook, Review Controller, policies and acceptance mapping
Localization and Skill entrypoints Documentation projections, bilingual coverage and 16 modified language-specific entrypoints passed
Local full package/runtime Before the final Astra fix: package 739 passed, 1 skipped; runtime 231 passed. Later affected checks are separate, not an additive new full-suite total
Supported Python CI and reproducible artifacts Behavioral/packaging baseline a7c8f07 passed all seven CI jobs, covering Windows/Ubuntu with Python 3.11/3.13 and bilingual builds. Local standard-zlib archives match CI byte-for-byte. Final publication is verified against its own workflow and provenance
Desktop installation Candidate installation and verify passed with matching source, marketplace and cache digests; this is not native gate effectiveness
Native V4 dispatch Two calls in an existing task and one in an independent fresh task lacked reservations and creation receipts, while lifecycle starts were observed. Configuration is trusted, but the active pre/post-dispatch path remains unaccepted
Empirical qualification A 36-case/72-generation paired study is preregistered. All three unqualified native attempts retain their actual consumption; qualified samples remain zero. Interrupted or missing results are not scored as model-quality failures
Commit / Push / Public Release / Effective Behavioral/packaging baseline committed and pushed to the PR. Public publication requires separate tag, asset and provenance readback. Enforced budgets and GPT-6 production defaults are not effective

The packaging correction normalizes staged Windows launchers to CRLF; 20 bilingual regressions cover checkout newline variation and unchanged source bytes. Python 3.14 zlib-ng and standard zlib can emit different compressed bytes. Every archive member was compared equal, and two local standard-zlib builds match the complete CI ZIP. Different-compressor archive hashes alone do not prove different source content.

Desktop management queries prove configuration and registration, not active-task native acceptance. Active Desktop and management component digests matched; no diagnostic established the cause of missing pre/post-dispatch callbacks, so the cause remains unknown. Changed non-managed Hooks require normal trust review; trusted hashes and bypass flags cannot fabricate acceptance. No model production qualification, reasoning gains or actual billing is claimed. Compatibility support retains the old default; later default activation still requires these gates.

Baseline evidence: a7c8f07 CI, candidate PR.