Skip to content

Verification evidence — 2026-10-13

Historical snapshot only. Counts, commands and limitations describe that snapshot and do not establish current acceptance. Use current verification.

Composer background correction (2026-10-13)

Removed the remaining input, placeholder, chevron and interior fill so the composer background matches the terminal across its whole frame. Two rendered regressions first failed on the gray fill and now check every cell while empty and during focused multiline editing. Focused composer/layout/input suite: 31 passed / 0 failed (artifacts/tui-audit/uniform-background-green.log). Documentation guards: 15 passed. The earlier full CLI result below is the previous snapshot; this local palette correction did not rerun the full matrix. The full 12-project matrix and the publish/launcher smokes were later rerun for the final state: 6,213 / 0 / 0 (artifacts/tui-audit/final-matrix.log), build.bat 11/11 (uniform-release-build.log), smokes (final-publish-smoke.log, final-launcher-smoke.log).

Ru TUI visual refinement (2026-10-13)

Mint picker accents, a session/chat divider, a full-width active-session square composer with metadata on its edges, and a contextual footer with a short, thick five-cell busy wave (▄▄▄▄▄). The operator rejected the earlier rounded frame and clarified thickness rather than increased wave length. The footer preserves queue/refusal status and does not repeat the interrupt action. Multiline editing and modal/approval behavior remain covered. Frame, input and placeholder cells match the terminal background throughout. Slash suggestions span the composer width; other modal surfaces centre on the terminal. The session heading has no product-name prefix.

  • Full CLI Debug suite: 1,418 passed / 0 failed / 0 skipped; artifacts/tui-audit/width-modal-cli-final.log. New rendered checks verify full-width slash edges, terminal-colored frame cells and whole-screen modal centring.
  • Headless ANSI render probes cover 80×24, 120×30, 180×30 and 80×14; actual cell text/attributes are in artifacts/tui-audit/width-after/. PNG previews reconstruct those cells; they are not native terminal captures.
  • The first shared-package rebuild was blocked by locked publish DLLs (visual-release-build.log). A standalone CLI publish succeeded (visual-isolated-build.log); the shared build was retried after the operator authorized closing Ru (visual-release-build-final.log).
  • The retry passed 11/11 Release stages. bin/ru.cmd reaches the expected parser failure for a missing subcommand; its CLI assembly hash matches the Release publish (visual-launcher-smoke.log). Documentation guards: 15 passed.
  • The operator-selected square frame and thicker short wave were rebuilt in 11/11 Release stages (square-release-build.log), with the running Ru Host closed under the operator's prior authorization. The current launcher parser smoke and publish hash comparison pass (square-launcher-smoke.log).
  • The full-width/background/modal-placement correction passed 11/11 Release stages (width-release-build.log). The published launcher parser smoke and CLI publish hash comparison pass (width-launcher-smoke.log).
  • Physical terminal, IME and clipboard acceptance remain separate manual checks.

TUI input, popups and retained details (2026-10-13)

Scoped follow-up to d09a4c66: multiline prompt editing, cursor-aware code paste, source cursor restoration, adaptive picker/approval cards, independent retained tool/reasoning blocks, explicit copy and actual run/queue status.

  • dotnet test backend-dotnet/RuAgent.Cli.Tests/RuAgent.Cli.Tests.csproj -c Debug: 1,416 passed / 0 failed / 0 skipped. Log: artifacts/tui-audit/cli-verified.log. Initial regression probes failed for modified-Enter submission, multiline navigation/display and block retention; follow-up tests also reproduced loss of new details at both retention caps. Earlier integration failures were retained in the same artifact directory; footer allocation, attachment visibility and obsolete fixed-row assertions were corrected before the final full CLI pass.
  • .venv/Scripts/python.exe -m pytest tests/test_current_documentation.py tests/test_guidance_route_counts.py -q: 15 passed.
  • build.bat: 11/11 Release stages passed, including the shared flat publish package used by bin/ru.cmd. Log: artifacts/tui-audit/release-build.log. The launcher reached its parser and returned the expected missing_subcommand for agent without a subcommand; the flat CLI assembly hash matches its Release publish. This is a parser smoke, not an interactive terminal or Host/provider acceptance run.
  • The headless render probe completed all five scenarios: 80×24 and 120×30 chat, 80×14 picker, Shift+Enter and earlier tool details. Output is under artifacts/tui-audit/after/; comparison.html shows before/after cell text.
  • This is affected-project verification. The full 12-project result in the table above belongs to the earlier cards/session-picker baseline and was not rerun for this local presentation change. No execution-authority, provider, public contract or durable-context policy changed.
  • Headless ANSI tests exercise cells, input injection and resizing. Copy tests use an injected writer and do not touch the system clipboard. Command Prompt appearance, physical resize, Vietnamese IME and real clipboard acceptance remain manual; no live provider or native user interface was driven.

TUI cards and the session picker (2026-10-13)

The busy hint row's wave became one five-cell horizontal line (━━━━━) instead of three dots: the same aurora → brand → brand2 ramp travels along it with a tighter phase spacing.

Every Terminal.Gui popup now shares one flat card: the title with its dismiss hint, a dim header rule, the filled selection bar, footer keys and a scroll position (3/12) when a list is clipped. Approval, workspace trust, the secret key field, the document window and the slash suggestions render through that frame; the bordered floating-panel renderer, the framed secret composition and the floating-panel placement helpers were removed with their callers. The secret field keeps the tail of an overflowing mask so its caret stays visible.

/sessions (and bare /resume) opens the session card in the Terminal.Gui engine: title, message count and workspace per row, the current session marked, typing filters, Enter switches. Delete or Ctrl+D asks for confirmation (default: keep) and removes the session; the card reloads, deleting the open session moves to the newest remaining one or creates a fresh one (desktop-client parity), a partial cleanup reads as deleted with the host's retry hint, and switching or deleting is refused while a run is live. The sequential engine keeps the plain /sessions list and /resume <id>, and /delete <id> keeps its explicit-id flow.

Full 12-project matrix 6,178 / 0 / 0 (artifacts/tui-cards-sessions-matrix.log), build.bat 11/11 Release stages (artifacts/tui-cards-sessions-build.log) and the flat-package smoke PASS (artifacts/tui-cards-sessions-smoke.log). Live appearance and the interactive session flow on a real terminal remain manual acceptance.

TUI busy-hint colour wave (2026-10-13)

The busy hint row's braille spinner is replaced by a colour wave: three fixed dots whose colours travel the aurora → brand → brand2 ramp. The phase is a pure function of elapsed seconds in TuiSpinner (WaveRgbAt), repainted on a 90 ms bucket so the 18 ms ticker repaints about eleven times a second, and the sine envelope fades the trailing cells so the row reads as one drifting glow instead of three blinking dots. The interrupt copy, queue hints, scrolled-back prefix and idle workspace label are unchanged; the classic reference composition (ComposerBox.ComposeHints) tracks the same visible text, and the live activity row in the transcript keeps the braille spinner.

Tests pin the wave clock (bucket advance, negative clamp), determinism, distinct cells, seamless cycle wrap and the left-to-right travel, plus the Terminal.Gui row's visible text and its three distinct per-cell foregrounds. Full 12-project matrix 6,165 / 0 / 0 (artifacts/tui-color-wave-matrix.log), build.bat 11/11 Release stages (artifacts/tui-color-wave-build.log) and the flat-package smoke PASS (artifacts/tui-color-wave-smoke.log). Live appearance on a real terminal remains manual acceptance.

TUI wrap and live context meter (2026-10-13)

Three defects reported from a live Terminal.Gui session: prose wrapped at a half-width measure on wide terminals, the context meter only moved once a reply finished, and the reply's content repeated after it finished.

Transcript prose now wraps across the full viewport width like the classic renderer (the 100-cell cap is gone; aligned and gutter rows keep their exact columns). The Host publishes the measured request size for a step about to run as a typed agent.usage_estimate transient frame on the activity hub channel (a delta burst cannot evict it), and the CLI folds it into the context reading as a max against reported usage — the desktop meter's policy — so the meter moves while the model works and is never empty for a whole first reply. The no-argument turn stop is now a true no-op after a completion: _sawCompletion no longer survives the first stop, so a second stop cannot reprint the footer with a zeroed clock.

The repetition had a host-side origin: when a tool call failed and the model's final reply was still unverified, the host suppressed that reply (audit-only, "the user never sees the suppressed prose") and answered the run with a recovery note so the model could try again — up to four times. The deltas had already streamed, though, so the transcript showed the same answer two to four times, and the retry path had the same shape. The Terminal.Gui transcript now marks where each model step's prose begins and drops exactly those rows when the step is superseded (an *_advised completion or a retry reset); the held-back tail cannot leak into the answer that replaces it; the suppressed payload reference is never recorded and the turn-end gap check stands down for a turn that carried a suppressed reply, so the recovery can neither report truncation nor reprint suppressed prose. The durable session record for the reported run confirms the mechanism: the screenshot's "Hello" turn committed one clean reply in 5.3 s, and the following tool-using turn committed unverified_after_tool_failure_advised twice with a run.note_added each time before the user cancelled it.

Focused suites: Core 489, Client 235, CLI 1,369, Backend 1,087 — all 0 failed. The full 12-project matrix then passed 6,162 / 0 / 0 (artifacts/tui-suppressed-fix-matrix.log), build.bat completed all 11 Release stages (artifacts/tui-suppressed-fix-build.log) and the flat-package smoke passed (artifacts/tui-suppressed-fix-smoke.log). The estimate frame was verified at the coordinator boundary (published before the first delta, carrying the measured size and prompt budget) and at the wire mapping (activity channel, agent.usage_estimate); the client parser, tracker, step rewind, reset rewind and recovery stand-down have unit coverage. Live provider behavior on a real terminal remains manual acceptance.

Review remediation Release rebuild (2026-10-13)

The user requested a rebuild, commit and push after the review fixes and resource contracts were verified. build.bat completed all 11 Release stages and rebuilt the canonical flat package at gui-winui3/bin/Release/net10.0-windows10.0.26100.0/win-x64/publish. The three apphosts, compiled XAML, GUI assets, .NET runtime and Windows App Runtime are present. The old package was copied to artifacts/review-release-previous-20261013-172803 before the build cleaned output.

A flat-package smoke derived from scripts/publish_smoke.ps1 skipped its publish step and used the freshly built flat RuBackend.exe. It ran with isolated temporary data and a free loopback port, verified the descriptor and authenticated status, then shut down the process it started. The flat CLI parser also returned its expected missing_subcommand without starting a session. No desktop UI was opened and no live provider request was sent. SDK typecheck, all 32 tests and the ESM/declaration build passed; Ruff and pip check also passed.

The first post-build Debug matrix stopped at WorkspaceIdentityResolverTests.Moved_directory_requires_trust_again and Moved_identity_produces_different_lease_binding: Directory.Move reported a Windows sharing violation before either identity assertion. The resolver and these tests were unchanged; inspection confirmed its native handles share delete access and are disposed before the move. The owning process could not be identified. The entire resolver test class then passed in isolation (8/8), without changing code or weakening assertions. Both results are retained in artifacts/review-release-dotnet.log and artifacts/review-release-workspace-identity-isolated.log; isolation does not establish the source of the temporary lock.

Resource contracts after review (2026-10-13)

The four previously deferred items now have explicit operating contracts. Model request deadlines stop at response headers after 120 seconds and map to retryable provider_timeout; caller cancellation remains cancellation. Active streams retain an infinite whole-response HTTP timeout with the existing SSE idle watchdog. The model HTTP handler refuses redirects and adapters preserve the non-retryable redirect_rejected failure without another endpoint request.

Terminal capacity defaults to eight running processes per session and can be configured with RU_MAX_BACKGROUND_TERMINAL_TASKS from 1–64. Configuration is validated and retained when the authority is constructed, while runtime/tool composition remains lazy. Every launch reserves capacity before dispatch; foreground promotion keeps its slot. A full session refuses a new launch and does not stop existing processes to free a slot. Lifetime and spool bounds remain.

Export uses the shared 128 MiB serialized-envelope bound, independently of the general 8 MiB API bound. The Host measures the real JSON with its response serializer before sending success; CLI and GUI check declared and streamed bytes and parse directly from the response stream. This removes the eager HTTP buffer and intermediate body string, while format files and parsed JSON remain in memory. Both formats have round-trip regressions through SQLite, the real Host routes and the shared client, including large content with nested JSON escaping.

Behavioral RED logs are artifacts/limits-terminal-red.log (foreground admitted at full capacity), limits-provider-red.log (three unnamed redirect failures), limits-export-client-red.log (8 MiB rejection and eager buffering), and limits-export-gui-red.log / limits-export-gui-refusal-red.log (missing size refusal). The new Host measurement API initially did not compile before its implementation (limits-export-host-api-red.log). The first full matrix stopped at Request_deadline_does_not_cancel_a_stream_after_response_headers: the new test incorrectly expected no deltas for [DONE], although the existing adapter emits a completed stop marker. The assertion now checks that terminal marker; no production change was needed (limits-final-dotnet.log).

Final .\scripts\test_dotnet.ps1 -Configuration Debug passed all 6,134 tests in 12 projects (artifacts/limits-final-dotnet-rerun.log). The full Python suite passed 216 tests with one POSIX-only test skipped on Windows (artifacts/limits-final-python.log). Ruff, route-contract, Python-runtime-boundary and current-documentation guards also passed. The prior six-fix SDK evidence remains unchanged; no SDK code changed in this step. At this implementation checkpoint, no Release build, publication, native UI interaction or live provider request was performed. The later rebuild is recorded above; earlier live evidence retains its original scope.

Six confirmed review fixes (2026-10-13)

The review was reassessed before implementation: only six demonstrated defects were changed. Grep now checks the opened handle's final path against the authorized canonical path; Responses response.incomplete preserves usage and maps token exhaustion to existing recovery while content filtering remains a named failure without executing partial tool calls; SDK RPC requests are correlated before writing and obey deadlines under stdin backpressure; provider sign-out derives the exact credential account without a catalog lookup; failed or cancelled session import compensates only an empty created session, with an atomic SQLite check/delete and caller cancellation preserved in both clients; Ruff excludes deliberately defective benchmark candidate trees while retaining checks on graders and reference solutions.

Regression tests were run against the old implementation before fixes. The adapter run failed both incomplete-response cases, the backend run failed two catalog-independent sign-out cases and three import cleanup cases, and CLI/GUI each failed their caller-cancellation case. The grep handle-swap case failed on the old code. A second protected-path fixture from that exploratory run was removed because its process-wide environment mutation raced other tests; the final regression uses a deterministic junction swap and an ordinary-search control. Logs are retained as artifacts/review-six-*-red.log. Ruff initially reported 23 errors (artifacts/review-six-ruff-red.log). SDK regressions also failed before the lifecycle correction; a follow-up review caught the completed response's stale write timer, which is now covered for both asynchronous and synchronous responses arriving before drain.

Final verification ran one process/loopback matrix at a time:

  • scripts/test_dotnet.ps1 -Configuration Debug: 6,096 passed in 12 projects.
  • .venv\Scripts\python.exe -m pytest tests -q -rs: 216 passed, one Windows skip.
  • npm run typecheck --prefix sdk/ru-agent and npm test --prefix sdk/ru-agent: typecheck passed and 32 tests passed.
  • .venv\Scripts\python.exe -m ruff check tests scripts tools: passed.
  • check_aspnetcore_route_matrix.py and check_python_runtime_boundary.py: passed; 43 HTTP routes, no WebSocket and no interpreter files in the product.
  • benchmark.py validate --suite everyday: passed. self-test --suite everyday for py-config-isolation and py-stable-dedupe: each defective seed failed and each reference solution passed all six groups. Reports: artifacts/review-six-config-selftest.json and artifacts/review-six-dedupe-selftest.json.
  • git diff --check: passed.

The Python benchmark fixture fingerprints changed with grader formatting; old live benchmark results retain their original fingerprints and were not rerun. At this six-fix snapshot, provider response-header deadlines, redirect policy, terminal auto-background policy and session export size limits were deferred; the following resource-contract change resolves them. No Release package, native UI interaction or live-provider acceptance was run for these fixes; Release/smoke/live rows above describe earlier evidence.

TUI chat-flow redesign (2026-10-13)

The live activity moved from the header into the transcript: one pinned live row (braille spinner or rotating tool glyph, action, target, elapsed) that settles in place into a ✓/✗ row with its duration; the header is two rows now (title/usage + workspace). Every animation phase is a pure function of elapsed seconds (TuiSpinner), which fixes the TG spinner that previously advanced at the 18 ms meteor cadence; the hints-row spinner and the Grok-style user block (❯ chevron on the composer's input-surface fill) follow the approved design.

Focused CLI evidence first: 1,355 / 0 with 7 new tests (spinner phases, live row tone and settle, user-block fill/chevron, hints spinner, fence language label). Then the full 12-project Debug matrix recorded 6,079 passed / 0 failed / 0 skipped (artifacts/tui-chat-flow-matrix-rerun.log). The first attempt failed once at AgentReconcileEndpointTests.Applied_resolution_returns_the_run_to_running; the test passed in isolation and the whole project reran green (1,069 / 0), so the rerun is the recorded result. The Release package was re-published and passed the flat-package smoke (see the table); native terminal appearance remains manual acceptance.

Remembered approval patterns and durable folder grants (2026-10-13)

ADR 0018 landed: allow_always now remembers a bounded pattern as a durable project-scope allow rule — a terminal command prefix (resolved executable plus at most two leading non-flag argv tokens with the exact launch directory), or an approved external folder as a <folder>\** path pattern the evaluator honors for external targets (external_directory_rule_allow) — while flag-first invocations and shell-mode scripts stay exact and rule_asks / terminal_protected_path_requires_approval are refused server-side (degrading to allow_once with rule_saved=false). Deny/ask precedence, the plan read-only gate, structured protected paths, and single-use tokens are unchanged. Both interactive surfaces offer the session folder and the project remember; the CLI's plain fallback mirrors the offered choices.

Focused evidence first: Broker policy tests (a project allow grants the matched external path, another path still asks, folder patterns cover other structured tools), Host tests for the pattern derivations, and an end-to-end run proving allow_always on an external ask saves one project rule and a later session reads a different file in the same folder without a new approval. Then the full 12-project Debug matrix passed 6,046 / 0 failed / 0 skipped (artifacts/approval-patterns-matrix.log) and the Python suite passed 216, 1 skipped (artifacts/approval-patterns-python.log). These are deterministic tests, not native UI or live-provider acceptance; the Release package from the previous run is unchanged and was not re-published.