Skip to content

2026-10-24 verification evidence

v0.1.2-preview.1 preparation

The release remains Preview, with no Stable publication. The initial Stable packaging run was cancelled before packaging or tag creation after the channel was clarified. Shared source and package versions are 0.1.2-preview.1.

Initial CI whitespace checks found trailing blank lines in eighteen historical verification pages. Only those blank lines were removed; the evidence was kept. Local Python checks passed 210 tests with one POSIX-only skip; documentation checks passed eight tests, Ruff passed, and route/runtime guards passed.

The first Preview release run 39995384625 failed creating a run in the deny case of the trust-retry fixture: 503 agent_store_unavailable, before the fixture installed its failure trigger. Twenty focused reruns (forty cases) passed. A new concurrency guard exercised eight simultaneous initializers on each of ten fresh databases and passed. These results do not establish the cause of the original storage failure.

The parallel CI run 39995381350 found a cancellation race: after tool shutdown, the worker can advance the run before Cancel commits its previously read version. Two deterministic regression cases returned 409 before the fix. Cancel now rereads the durable projection, preserves terminal results and retries version conflicts up to eight times without repeating tool shutdown. Five new cases cover progress, cancelled and completed results, retry success and exhausted contention.

After the initial cancellation fix, the full backend project passed 1,384 tests. After adding retry-bound cases, the eight focused cancellation/trust tests passed. The new fresh-schema concurrency guard passed independently. Logs are under %TEMP%: ru-v0.1.2-preview.1-release-failure.log, ru-preview-ci-failure.log, ru-preview-trust-repro.log, ru-preview-backend-after-cancel.log and ru-preview-fresh-schema-repro.log.

Final CI, package build and disposable-machine installer acceptance remain release gates. Earlier native UI and real-provider evidence is recorded in 2026-10-23; deterministic checks here do not repeat that evidence.

Tool-error recovery and duplicate streamed replies

The reported GUI/TUI duplication was reproduced in the shared Harness. A final answer after a failed receipt became audit-only and triggered recovery, but its already streamed text was never retracted. The next model step appended another answer. Three regression cases (process_exit_nonzero, executable_not_found, edit_no_match) failed before the fix; the accepted-reply control passed.

Completed streams now expose one idempotent retraction after redaction has flushed. The coordinator uses the existing per-step reset for rejected recovery, repetition, unusable and empty steps. Tests preserve accepted earlier prose and the final report, including a redactor-buffered tail. Guard decisions, recovery bounds and execution policy are unchanged.

Headless integration tests executed a real cmd.exe exit-7 command and a real missing-file read through the production Host/Broker composition with a scripted model. Two live hub subscriptions and the shared SSE delta parser observed the reset and retained only the final report. Terminal.Gui rendering and GUI reply accumulation tests covered replacement by a different model-step identity and preservation of earlier accepted text. Initial integration/render probes used the wrong safe-event field and invalid fixture step IDs; those test inputs were aligned with the production contract before passing.

scripts/test_dotnet.ps1 -Configuration Debug passed 7,764 tests across all 13 projects, including 1,388 backend, 1,647 CLI, 597 Core and 1,797 GUI tests. The documentation guard passed eight tests and git diff --check passed. Full matrix log: %TEMP%/ru-tool-recovery-streaming-dotnet.log.

These checks did not automate the user's WinUI window or call a real model provider. The fix is source-workspace work and is not part of the already published v0.1.2-preview.1 package.