GiviLoop 0.7.0 — First-use demo and shareable Double Check reports

GiviLoop 0.7.0 — try it, verify it, share the result

23 September 2026. This release connects the first-use demo to a repeatable before-commit review and a portable report. Web integrations retain their documented experimental status and provider access/terms limits.

New workflow

• givi demo --offline: an explicitly authored public example, no account/browser/network delivery, a fixed independent reproduction and a Markdown report.
• givi demo: the same public example through your configured web/local reviewer. Only bundled source is sent. The actual response and deterministic reproduction stay distinct; the known control is not a model-quality score.
• givi demo --finish: complete a paused demo after login/resume without sending another prompt. It checks example integrity and never executes a modified verifier or model-generated code. Project sources, preferences and the selected project review remain unchanged.
• givi report and MCP givi_export_report: local Markdown export with effective verdicts, stale warnings, recorded evidence, referenced source hashes and request/response fingerprints. Empty/unassessed findings and unsuccessful runs remain explicit. Raw source, requests and complete answers are omitted. Inspect evidence before sharing; nothing is posted automatically.
• A before-commit recipe and voluntary feedback fields for actual second use. Reports survive automatic run retention; regenerate them after changes because exported files are snapshots.

Validation and scope

New unit/integration cases cover offline execution, unchanged project state, tampered verifiers, provider-option errors, idempotent finish, stale/empty/failed reports, redaction, Markdown rendering, symlink protection, retention and MCP export. Real-Chrome fixtures cover live-shaped demo delivery plus login pause → resume → finish with one submission.

A real signed-in Claude demo completed with its actual response saved, the known empty-array bug independently reproduced and all four bundled candidate-fix cases passing. This does not certify other model claims or measure total token savings. The source was unchanged. Release assets include the package/CI results and a sanitized live-validation summary.

A second real Claude review examined the new implementation. An unexpected persisted verdict was independently reproduced as a misleading-report bug, then fixed by validating ledger decision states. A claim about an unhandled CLI JSON stack trace was disproved by a targeted test: the existing error boundary already catches it. The exported report preserves both decisions and explicitly notes the frozen pre-fix source and missing original CLI context.

The current Git review scope is working-tree/staged changes relative to HEAD, not an entire branch. The calling agent still verifies real-project findings under its repository permissions. The only verifier executed by the new demo is the installed, fixed public example; the report exporter runs no tests.

No new providers, automatic code fixes, model-output execution, community posts or telemetry were added. The five-user pilot kit is ready; recruitment, external trials and actual adoption measurements remain pending. The README video remains the labeled 0.5.0 recording.

Verified commit: cb5efa4. Successful CI run: 202 unit tests, 51 browser tests, zero production dependency vulnerabilities. The tarball below is the exact CI artifact. adoption-live-validation.json records the separate macOS live demo and implementation review. double-check-report.md contains the independently assessed findings.