Skip to content

Verification evidence - 2026-11-02

Historical snapshot only. See current verification.

New-session transition work

The user reported visible stuttering in both the sidebar and conversation/composer when creating a session. Deterministic regressions reproduced two sidebar problems: replacing an empty draft published two renders with an intermediate empty list, and creating a recent chat collapsed/replaced an unrelated workspace header.

Draft replacement now publishes one render. Existing workspace expansion is kept as the default for subsequent projections, and unchanged headers retain their container identity. Selecting a session still opens its previously collapsed group. Sidebar rows use repositioning without competing add/delete theme transitions; the existing Windows animation-preference callback reconfigures these transitions.

The known-new session path paints an empty transcript after the existing detach, draft-retention and preset-narrowing steps. It skips activation/history, replay, usage and run-adoption reads for a session whose create response already establishes an empty history. Side-panel layout is deferred before reset and committed with the new transcript. A new no-project chat does not display the outgoing workspace path. Creating a session while already on chat also avoids restarting navigation's model/catalog refresh work. Existing-history replay and cancellation ownership remain.

Evidence

  • Initial regressions: four expected failures in new-session-motion-red.log.
  • Sidebar behavior verifies one render instead of two, unchanged header identity, and explicit selection expanding an implicitly collapsed group.
  • Source ownership guards verify the create call selects the known-new path, that path has no history I/O, navigation is conditional, and reduced-motion preference changes reconfigure sidebar transitions.
  • An intermediate focused run passed 68 tests and failed one structural guard because the method-body helper cannot parse an if block. The known-empty presentation was given its own cohesive method; the guard targets that owner. Existing normal-history guards retain their original assertions.
  • Complete GUI suite: 2,046 passed, zero failures via scripts/test_dotnet.ps1 -Configuration Debug -NoRestore -ProjectFilter gui-winui3 (new-session-motion-gui-final-2.log), including the Debug GUI build.
  • Evidence logs are in the external review artifact directory alongside the earlier remediation-20261101 evidence.

These checks establish state/render counts, ownership and compilation. They do not measure native frame timing or prove the user's visible stutter is fully gone. No desktop automation, real-provider run, Release packaging, commit or push was performed.

Delivery checkpoint before TUI branch review

At the user's commit/push request, the complete current review-remediation and GUI working tree was verified again in one uninterrupted sequential .NET matrix: 9,564 passed across all 13 projects, zero failures, via scripts/test_dotnet.ps1 -Configuration Debug -NoRestore (delivery-20261102-dotnet.log). This supersedes phased verification as the delivery gate while retaining the earlier failure records.

Additional sequential checks passed: Python 268 passed, one POSIX-only case skipped on Windows; Ruff; route contract 54 HTTP, zero WebSocket; runtime boundary with zero interpreter files in the product; SDK typecheck, 63 tests and build. Logs use the delivery-20261102- prefix in the external review artifact directory. No Release package or native-frame acceptance is claimed.

The unpushed historical range's git log --check reported an extra blank line at EOF in the pre-existing 16820356 documentation commit. The current tree corrects that EOF; the historical commits are retained without rewriting them. Current/staged whitespace checks remain part of the commit gate.

TUI branch integration review

The user requested delivery of the existing work first, then a conditional review, merge and deletion of casluminous/tui-multi-session. The existing work was committed as 20745b4e7551a94e95db6abcecf8c26282821cc1, pushed to dev, and verified against the remote ref before reviewing the branch at 62e916d3b8f62a35c493ebb9fccef821b71b1b29.

Concurrent tabs already existed on dev. The useful additional work is the shared native/line/JSON conversation controller, shutdown/admission ownership guards, retention of tabs with pending navigation, inspection cancellation and slash-menu geometry regressions. The branch also intentionally removes terminal chat /skills and /skill; GUI and typed Host command support remain. Removed legacy renderers/history have been replaced by the existing native artwork and the shared controller rather than a second execution path. No schema, route, provider protocol or execution-authority change is introduced by this merge.

The sole text conflict was in CHANGELOG; both sets of entries were retained. Current dev git-integration removal, scratch provenance, restore/runtime fixes and GUI lifecycle work remain. The stale root MEMORY.md was excluded rather than restoring a second project history. Its useful dated typing investigation remains under this directory, linked from the history index and current TUI limitations. The proposed redraw/cadence optimizations in that report are still unimplemented; integration does not establish manual typing latency or fix them.

Integration failures retained

  • Initial full CLI run: 2,050 passed, one failed. The unchanged wordmark allocation guard observed 3,304 bytes instead of zero (merge-tui-cli.log). The isolated raster class then passed all 13 tests with the original assertion (merge-tui-allocation-isolated.log).
  • A second full CLI run observed 528 bytes in the same guard and then stalled. A managed stack captured TgTranscriptWriterTests.ChatFactorySelectsScreenCardsOnlyForTheTerminalGuiBridge blocked in ApplicationImpl.Dispose / MainLoopCoordinator.Stop, while ANSI input waited in WindowsVTInputHelper.TryRead. After checking its executable path and parent, only this task's testhost was stopped. The aborted run's 1,180 passes and one failure are not a completed suite result (merge-tui-cli-final.log, merge-tui-cli-final-stack.log).
  • The bridge fixture now enters and exits a running frame before disposal, following the existing documented Windows ANSI lifecycle rule. Its screen-card assertions remain. The allocation guard retains the same 300 render calls and strict zero-byte requirement, with per-frame and GC diagnostics collected into a preallocated array. It does not increase the allocation budget, suppress the assertion or change product rendering. The allocation observations alone do not establish their cause or demonstrate a renderer regression/fix.

Final integration checks

  • Full CLI build/suite: 2,051 passed, zero failed/skipped, using scripts/test_dotnet.ps1 -Configuration Debug -NoRestore -ProjectFilter RuAgent.Cli.Tests (merge-tui-cli-diagnostic.log). The reduced count relative to pre-merge CLI includes removal of obsolete legacy-renderer/command tests and added controller, navigation, cancellation and native layout coverage.
  • Windows ConPTY: 6/6 passed (merge-tui-conpty/results.json), including control, BMP, emoji, fragmented input, navigation and typing.
  • Python: 268 passed, one POSIX-only process-group case skipped on Windows (merge-tui-python.log). Documentation/contract guards are included.
  • The prescribed CI lint command, python -m ruff check tests scripts tools, passed (merge-tui-ruff-ci.log). An additional broader ruff check . found two pre-existing style issues in unchanged docs-site/build.py (UP017/UP032), outside that CI command; its failed log is retained as merge-tui-ruff.log.

Checks ran sequentially, with native/process tests under the shared matrix lock. The merged snapshot has focused CLI/ConPTY/tooling evidence; the earlier 13-project matrix above applies to the pre-merge delivery. No Release publish, installed runtime replacement, desktop automation or real-provider acceptance was performed.