Repository navigation
Releases: gatchimuchio/BLUE-TANUKI
Releases · gatchimuchio/BLUE-TANUKI
Release list
BLUE-TANUKI 1.0.0-rc.1
BLUE-TANUKI 1.0.0-rc.1 Release Notes
Scope
1.0.0-rc.1 is a release candidate, not GA. Public completion claims remain
blocked until owner GO and public_claim_allowed=true.
Shipped
- HDS-BRAIN authority kernel with Approval Gate, final-review boundaries,
Runtime Invariants evidence, hash-chain audit, output audit, and complete
history substrate. - First-party WebChat / Telegram channels and first-party Writing / Daily /
Developer Operator surfaces. - Control Center for runtime status, approvals, audit, history replay, settings,
connectors, evidence export, About, Backup / Restore, and Update readiness. - P10 recovery controls: recovery backup, restore, provider reset, connector
reset, and audit-preserving factory reset. - P11 manual update controls: release sidecar verification, pre-update recovery
backup, rollback plan generation, and manual-only update boundary. - P12 support boundary and known limitations documents.
- P13 owner-GO readiness boundary: release remains pre-GO and fail-closed
withoutdocs/ga-owner-decision.json. - Source release bundle with
.sha256and.manifest.jsonsidecars.
Not Shipped
- GA public claim.
- Signed native installer.
- Automatic updater or runtime app-file replacement.
- Public third-party plugin / skill marketplace.
- First-party WhatsApp support.
Validation Expectations
Before using this RC as a release candidate, run:
pnpm install --frozen-lockfile
pnpm typecheck
pnpm build
pnpm test
pnpm docs:check
pnpm validate:repo-health
pnpm validate:packaging
pnpm validate:ga
pnpm validate:product -- --phase P13
pnpm release:bundle
pnpm release:verifyWindows installed-app evidence and owner GO remain required before GA.
Support scope is fixed by docs/SUPPORT_BOUNDARY.md; remaining evidence gaps
are tracked in docs/KNOWN_LIMITATIONS.md.
Authority Boundary
Release metadata, update metadata, recovery backups, rollback plans, Control
Center state, release notes, and bundle manifests are evidence only:
used_for_authority=false. They cannot approve commands, classify risk, replace
HDS-BRAIN, infer consent, resume suspended work, or bypass final review.