Verification evidence — 2026-10-08¶
Historical snapshot only. Counts, commands and limitations describe that snapshot and do not establish current acceptance. Use current verification.
Right side panel delivery (2026-10-08)¶
The complete sequential Debug matrix passed all 12 projects: 5,799 passed / 0 failed / 0 skipped, including 1,304 GUI tests. The backend Debug build passed with zero warnings and errors; the two existing GUI test-source warnings remain. The final Python suite passed 204 tests with one skip.
The right side surface now has two tiers on a per-session policy: the floating summary card keeps the session title, the +/− edit totals and compact background-task rows without Stop, while its chevron opens a full-height details panel with closable Changes and Tasks tabs, a + reopen menu, a shared +added −removed header, expandable edit diffs, per-row task Stop through /tasks kill, drag resize (280–560 px, persisted to side-panel.json) and a fullscreen toggle; closing the last tab or the strip's collapse returns to the card. A title-bar toggle parked immediately left of the caption buttons hides and restores the whole surface through an InputNonClientPointerSource passthrough region and the caption RightInset. The Broker remains the only execution authority; no route, preset, approval, or provider behavior changed. New policy, projection, preference-store, layout and source-contract tests cover the state machine, ordering, caps, clamps and wiring. Native desktop acceptance (real hover/press on the title-bar button, drag resize, fullscreen rendering) was not run; the contract tests pin markup and wiring only.
Submission simplification and coding benchmark (2026-10-08)¶
The complete sequential Debug matrix passed 5,766 tests across all 12
projects, with zero failures and skips. Backend Debug compilation passed with
zero warnings/errors; the existing two GUI test-source warnings remain.
Focused submission-owner and codec tests passed 7 and 6 cases respectively.
The final Python suite passed 204 tests with one skip. Ruff over
tests scripts tools, dependency checking and the 42-route/zero-WebSocket
guard passed. The 12 embedded-help tests passed after documentation updates.
ChatSubmissionState consolidates pending attachments, queue admission,
held retry identity and accepted transfer cleanup. Cancellation during a trust
gate now releases its accepted temporary images. Tool-argument serialization
and interpretation share one codec while audience, exact identity/argument
checks, bounded approval binding and existing failure codes remain in their
owners. AgentCommands.Chat.cs decreased from 3,207 to 3,103 lines and still
contains substantial orchestration; this does not claim the remaining large
dispatcher has been fully decomposed.
The custom coding benchmark includes six Terminal-style tasks and three SWE-style issue repositories in C#, Python and TypeScript. All nine public starter smokes passed. Each task was graded from a fresh starter and reference three times: 27/27 validated pairs, with starter failures and reference passes. One full reference sweep reports 60 groups and 354 checks; SWE starters failed every regression group and passed every preservation group. Current fixture fingerprints match the final report. These are fixture checks, not a Ru solve rate. Verification covers fresh baseline SHA generation, source integrity, patch application against a fresh base, JSONL export, complete grader evidence and fake-process RPC trials. Windows verifier/RPC children enter a kill-on-close Job Object before they resume; real-process tests cover descendants after leader exit and assignment failure before child code. This bounds process lifetime, not filesystem or network access.
Live Ru/provider coding quality, token/cost acceptance, Docker/Harbor execution and official benchmark datasets were not run. Native fixture/oracle results are separate from a model score. Release publishing and native UI acceptance were not rerun.
Review remediation (2026-10-08)¶
The complete sequential Debug matrix passed all 12 projects: 5,753 passed / 0 failed / 0 skipped, including 1,271 GUI tests. The backend Debug build passed with zero warnings and errors. The fixes cover handle-bound file mutation and conditional patch rollback, incremental Office extraction budgets and nesting/alignment, native handle-list lifetime, spool generation accounting, physical Host root claims and thread-owned mutexes, starter readiness and bounded create acknowledgement reconciliation. The benchmark's 44 generated outputs are removed from Git tracking and ignored; benchmark source remains tracked.
The first matrix passed eleven projects but failed one GUI allocation regression because its process-wide allocation counter included parallel tests. A bounded counter check confirmed that unrelated thread allocations affect that metric; the synchronous extraction tests now measure the calling thread with the same 32,000,000-byte threshold. The full GUI project and then the complete matrix passed against that correction. The two existing GUI test-source warnings remain; no contract or authority assertion was weakened.
Existing-file writes are serialized and flushed with a 32 MiB rollback preimage cap; they do not claim power-loss crash atomicity. Session removal invalidates spool generations immediately, while an already-open background spill can await writer completion, bounded by the task's two-hour maximum. Release publishing, native desktop acceptance and real-provider quality were not rerun for this remediation.