Skip to content

Verification evidence — 2026-10-12

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

Selective CLI park-reason integration (2026-10-12)

Ported the useful park-reason behavior from fix/cli-run-park-reason onto current dev, preserving existing scrolling, billing classification, diagnostics and UI. The original branch's two defects are corrected: the state-first live park now recovers the exact next committed reason through bounded finite replay, and reason deduplication resets across turns/runs. Status names are not invented causes. The three-argument SSE client method remains available for binary compatibility; the new overload requests live=0 without changing default subscriptions.

Four new renderer cases initially failed on the old code (artifacts/cli-park-reason-red.log). After correction, CLI 1,313 and Client 215 passed (artifacts/cli-park-reason-tests.log); additional recovery guards brought the final full matrix to 6,024 / 0 failed / 0 skipped (artifacts/cli-park-reason-matrix.log). New ChatAsync tests use the actual SSE client/parser and run-follow code against fake HTTP, covering state-first order, missing replay and no server cancellation. Additional tests cover exact cursor/run binding, timeout and propagation of parent cancellation. These are deterministic tests, not native terminal or live-provider acceptance. No new Release package was published for this integration; the previous release remains in the table above. Python passed 214, with 1 POSIX test skipped on Windows (artifacts/cli-park-reason-python.log).

Balanced default Dark brightness (2026-10-12)

After feedback that the brighter palette was excessive, default Dark uses sidebar #141414, canvas #1A1A1A, surface/hover #242424, selection #2E2E2E and line/selected-hover #3C3C3C. Foregrounds, accent and status colors retain their values. Light, Meteor and Starry Night palette definitions are unchanged from the preceding release. Startup XAML and runtime application match.

Scoped Debug GUI build/tests passed 1,393 / 0 failed / 0 skipped, including palette parity and existing AA contrast checks (artifacts/gui-dark-balanced-tests.log). Publication follow-up: the full 12-project matrix passed 6,011 with no failures or skips (artifacts/gui-dark-balanced-release-matrix.log); Python passed 214, with 1 POSIX test skipped (artifacts/gui-dark-balanced-release-python.log). The isolated 11-stage Release pipeline passed, then the complete canonical package was replaced with the previous package retained and all 749 file hashes verified (artifacts/gui-dark-balanced-release-build.log, artifacts/gui-dark-balanced-release-deploy.log). Actual launcher startup/listing, published CLI RPC creation/source labeling and authenticated Host status/shutdown passed (artifacts/gui-dark-balanced-release-smoke.log). Native appearance remains manual acceptance.

CLI Session sidebar section (2026-10-12)

Sidebar order is Workspaces, Recents, CLI Session. Subsequent sections have 36 px top separation (previously 28 px); the first heading also gains a small spacing adjustment. All headings retain non-hovering outer spacing. The CLI section folds, searches, pages and reorders sessions using the existing sidebar mechanisms. Source grouping does not change execution cwd, permissions, or the source when a CLI session is resumed in the GUI.

GUI and CLI creation/import requests send X-Ru-Client-Source. The Host validates gui|cli, stores it in existing SQLite session document/summary JSON, and returns client_source in session listings. Legacy/null-source sessions keep existing workspace/Recents grouping; their creating client cannot be inferred from cwd.

The final full Debug matrix passed 6,011 / 0 failed / 0 skipped, including persistence reopen, optional/invalid source HTTP behavior, import labeling, source-based projection, polling fingerprints and existing sidebar/theme checks (artifacts/gui-cli-section-matrix-final.log). The first full matrix also passed before import-source support was added (artifacts/gui-cli-section-matrix.log). Python passed 214, with 1 POSIX process-group test skipped on Windows (artifacts/gui-cli-section-python.log). The isolated 11-stage Release pipeline completed (artifacts/gui-cli-section-release.log), then the complete flat package replaced the canonical publish directory with the old package retained in ignored artifacts. All 749 files matched hashes and CLI/backend file inventory was complete (artifacts/gui-cli-section-deploy.log). Actual launcher startup/session listing, authenticated Host status/shutdown and published CLI RPC creation/listed source passed (artifacts/gui-cli-section-release-smoke.log). Native desktop interaction and visual acceptance remain manual.

GUI default theme brightness follow-up (2026-10-12)

Default Dark now uses sidebar #181818, canvas #1E1E1E, surface/hover #282828, selection #323232 and line/selected-hover #404040. Default Light uses sidebar #E2E2E2, canvas #ECECEC, surface #F5F5F5, selection #DEDEDE, line #CCCCCC and muted text #586579. Accent and semantic status colors are unchanged. Meteor's 23 palette values still match the original slate palette and Starry Night remains unchanged. Startup/runtime parity now covers both default packs.

Scoped Debug GUI build/tests passed 1,391 / 0 failed / 0 skipped, including existing AA contrast checks (artifacts/gui-theme-softened-tests-final.log). The initial test build failed with CS0136 after extending the parity test; renaming the inner locals corrected it (artifacts/gui-theme-softened-tests.log). The later CLI Session release above includes this palette refinement. Native WinUI visual acceptance remains manual; the section below records the earlier palette publication.

GUI neutral charcoal dark theme (2026-10-12)

Studio Dark uses sidebar #121212, canvas #181818, surface/hover #222222, selection #2C2C2C, line/selected-hover #383838, ink #EEEEEE and muted #AAAAAA. Accent and semantic status colors retain their existing values. Only default Dark uses this base. Dark Meteor Shower retains its original slate palette; Light and Starry Night packs retain theirs. Startup resources and runtime palette application are aligned, with added parity/neutral-hierarchy and theme-isolation guards. Meteor has an independent palette: all 23 colors match the previous default Dark palette; Light and Starry Night palette literals are unchanged (artifacts/gui-dark-default-isolation.log). The full Debug matrix passed 6,006 tests / 0 failed / 0 skipped, including 1,391 GUI tests (artifacts/gui-dark-default-matrix.log). Python passed 214 tests, with 1 POSIX test skipped on Windows (artifacts/gui-dark-default-python.log).

The isolated 11-stage Release pipeline completed (artifacts/gui-dark-default-release.log). Twenty GUI-owned artifacts were backed up, deployed to the canonical flat publish directory and hash-verified; backend/CLI files were retained (artifacts/gui-dark-default-deploy.log). Actual launcher startup/session listing and isolated authenticated published-host status/shutdown passed (artifacts/gui-dark-default-smoke.log). Native GUI appearance and interaction remain manual acceptance; no user desktop was driven.

Readable TUI feed and web activity (2026-10-12)

The screenshot-driven follow-up adds readable web queries/fetched hosts to redacted activity detail, aligns verified Markdown table blocks from raw rows, uses labeled values on very narrow frames, caps prose at 100 cells and retains reasoning behind Ctrl+O without repeating Thought summaries in the visible feed. Web queries no longer use the unredacted symbol field; oversize queries are omitted rather than split before redaction. Fetch labels contain only the host.

Full matrix: 6,004 passed / 0 failed / 0 skipped, including CLI 1,307, Core 487 and Backend 1,056 (artifacts/tui-readable-feed-matrix-final.log). Python: 214 passed, 1 skipped (artifacts/tui-readable-feed-python-final.log). The isolated 11-stage Release pipeline derived from build.bat completed (artifacts/tui-readable-feed-release.log), then the verified flat package replaced the canonical GUI/backend/CLI publish directory with the previous package retained in ignored artifacts. Complete dependency inventory and actual ru startup/session-list plus authenticated published-host status/shutdown passed (artifacts/tui-readable-feed-inventory.log, artifacts/tui-readable-feed-smoke.log). Headless control previews use sample data (artifacts/tui-readable-feed-wide.png, artifacts/tui-readable-feed-narrow.png); physical terminal acceptance remains manual.

Earlier failures were retained: a missing required inline-render parameter (artifacts/tui-readable-feed-cli.log), a narrow-table value split corrected by stacked rendering (artifacts/tui-readable-feed-cli2.log), and the deliberate fetch-host metadata change requiring a revised privacy assertion (artifacts/tui-readable-feed-matrix.log). The existing table test file was accidentally overwritten during test creation, restored unchanged from HEAD, and new tests moved to a separate layout-test file before final verification. The first Python run also caught a stale README badge/table total after the previous release; README and this record now agree on 6,004 (artifacts/tui-readable-feed-python.log). No assertions about URL credentials, query parameters, redaction or retained values were removed.

TUI Agent Activity presentation (2026-10-12)

Activity occupies the third header row beneath the workspace, retaining the three-row height. Semantic glyph/action, target and elapsed/tool count have separate tones; response deltas distinguish writing from thinking, and parked states have no spinner or running clock. Settled rows emphasize the operation, mute the target and reserve outcome/metrics space, including explicit failure.

CLI suite: 1,302 passed / 0 failed / 0 skipped (artifacts/tui-activity-tests-final.log). The initial run had 1,289 passed / 3 failed because it pinned raw web_fetch/todo labels and single-space tool formatting; updated assertions pin Fetch/Plan and the new grouping while retaining feed ordering, target and expansion checks (artifacts/tui-activity-tests.log). Headless Terminal.Gui buffer previews at requested 100×32 and 50×26 cover active and approval states, tool outcomes and narrow wrapping (artifacts/tui-activity-wide.png, artifacts/tui-activity-narrow.png). These are synthetic content rendered by actual controls, not native terminal acceptance. After the last color/label refinement, 54 focused tests passed (artifacts/tui-activity-polish-verified.log) and 15 documentation guards passed (artifacts/tui-activity-docs.log). The previous full Release matrix remains the release evidence below.

GUI/TUI review (2026-10-12)

Release follow-up: the full 12-project Debug matrix passed 5,987/5,987 (artifacts/gui-tui-release-matrix.log); Python passed 214, with 1 POSIX test skipped on Windows (artifacts/gui-tui-release-python.log). build.bat completed 11/11 stages (artifacts/gui-tui-release-build.log). The refreshed self-contained flat package passed complete CLI/backend file inventory and published-host authenticated status/shutdown plus actual ru launcher session listing and redirected startup/exit (artifacts/gui-tui-release-inventory.log, artifacts/gui-tui-release-smoke.log). Smoke data and bearer were isolated and synthetic. Native WinUI interaction and physical-terminal rendering remain manual.

A current-checkout review recorded five P2 findings. Sequential Debug checks passed 1,284 CLI tests and 1,385 GUI tests, with no failures or skips. A direct TuiTranscript probe reproduced chunk-dependent Markdown rendering and irreversible display clipping of long fenced code. The GUI findings are source-confirmed asynchronous ownership races; native WinUI interaction and physical-terminal acceptance were not run. Logs: artifacts/gui-tui-review-cli.log, artifacts/gui-tui-review-gui.log; probe: artifacts/gui-tui-review-probe/results.json. These results are scoped, not a new full matrix. The initial findings are followed by their fixes in the linked report.

The follow-up suites passed 1,292/1,292 CLI and 1,389/1,389 GUI with no failures or skips, sequentially (artifacts/gui-tui-fix-cli-verified2.log, artifacts/gui-tui-fix-gui-verified.log). The renderer probe now produces identical final rows across chunk partitions and retains the long code tail (artifacts/gui-tui-review-probe/fixed-results.json). The complete renderer/bridge CPU probe, including final reflow of 4,000 fragments with 1,000 history rows, measured 7.8 ms / 1,427,032 allocated bytes (artifacts/gui-tui-stream-perf/result.txt); this is synthetic CPU evidence, not native terminal FPS. Test adaptations and earlier failures are recorded in the report.

Compact TUI chrome and block cursor (2026-10-12)

Latest refinement: the five-row composer now has its real default placeholder (Ask anything…), metadata on row index 3 and bottom padding. The empty-chat footer shows only ctrl+p commands; active-session scroll/expansion hints remain. All 48 focused composer/chrome/scroll tests passed (artifacts/tui-composer-clean-tests.log). Release CLI publish/deployment checked every publish file's presence; the actual ru launcher started and exited zero with redirected /exit (artifacts/tui-composer-clean-startup.log). The layout capture is artifacts/tui-composer-clean-home.png; earlier offsets below are history of the previous iterations.

Metadata alignment follow-up: preset/model/provider moved down exactly one terminal row (row index 4) within the five-row composer. All 16 focused layout, composer and copy checks passed (artifacts/tui-meta-offset-tests.log). Release CLI was rebuilt and deployed to the default desktop publish directory; all five CLI file hashes matched the new publish output (artifacts/tui-latest-cli-build.log). The running backend was retained. Subsequent launcher testing found that the destination package was missing CLI dependencies, including Client and SessionInterchange. Copying only the five CLI files had not established package completeness. The full CLI publish was checked against the destination and 156 missing dependency/resource files were restored; no publish file remained missing. ru agent session list --json then exited zero from C:\Users\USER (artifacts/tui-launcher-session-list.log). An initial ru --help probe returned the parser's expected unknown-command/usage response, not an assembly-load exception; the parser does not support that flag. The actual no-argument ru launcher also started and exited zero with redirected /exit input (artifacts/tui-launcher-chat-startup.log), without creating a model run. This verifies chat startup, not native terminal painting.

Follow-up: the four-row composer felt too tight, so it now uses five rows, with one blank row between input and model metadata. All 16 focused composer, layout and copy checks passed (artifacts/tui-five-focused.log). The CLI is published separately without cleaning the running backend's package (artifacts/tui-five-publish.log); the buffer preview is artifacts/tui-five-session.png. The four-row evidence below records the preceding iteration.

The composer shrinks from six to four rows; metadata moves up one row and the transcript/panels reserve the corresponding smaller footer. Header title/usage and workspace/activity occupy two rows above a dim separator. Idle activity uses the secondary Ru colour. The composer input explicitly uses SteadyBlock; the focused/edited TextField retains that style in the headless driver.

The first focused run failed nine old row/colour assertions; a subsequent compile check found that TextField's documented UpdateCursor is not public. The test now exercises the public insertion-point property after real focus. All 41 targeted chrome/layout/copy/cursor checks passed (artifacts/tui-compact-focused-final.log). The full matrix was not rerun for this visual refinement. Hardware cursor appearance and native IME remain terminal-dependent acceptance, separate from the buffer captures.

Documentation guards passed 19 tests (artifacts/tui-compact-docs.log). The default build.bat stopped during clean because the running Release backend held publish files; some unheld files had already been removed (artifacts/tui-compact-release.log). Isolated self-contained CLI, backend and GUI publishes then succeeded (artifacts/tui-compact-publish-{cli,backend,gui}.log). The combined package is under artifacts/tui-compact-publish/flat; 618 missing files were restored to the default publish directory without replacing existing files or stopping the backend. The default CLI DLL hash matches the new package. Headless session/home captures are artifacts/tui-compact-{session,narrow,home}.*.

TUI header, streaming and activity repair (2026-10-12)

The uncommitted context sidebar was removed in favour of a compact header: title/usage, workspace and live activity above a full-width conversation. Frames below 20 rows retain the single heading row. The screen timer reads existing renderer status and the renderer uses the transcript's actual viewport width. Terminal.Gui now enables its tool/diff card path; sequential output does not.

Two failing regressions reproduced completion metadata being attached to the last started tool instead of its activity ID, and starting compaction being labelled completed (artifacts/tui-feed-red.log). A further regression reproduced transient failure overriding durable success (artifacts/tui-outcome-red.log). Tool metadata now updates by activity ID, with bounded settled-ID retention for late/duplicate frames; web/todo tool activities participate in proposal order. The durable result owns outcome. An SSE durable-sequence anchor is not used as a tool identity. Three scripted snapshot failures used the same activity ID for different tools; the fixture now assigns distinct IDs and retains its ordering, detail and one-row-per-result assertions (artifacts/tui-feed-snapshot.log).

The streaming probe appends 4,000 four-character fragments after 1,000 history rows into the actual headless transcript. Before repair this took 10,931.9 ms and allocated 2,960,974,208 bytes (artifacts/tui-perf-before.txt). Dirty-tail reflow now occurs at draw/scroll boundaries, and long words split in one pass. The repaired probe, including deferred reflow, took 14.5 ms and allocated 65,501,504 bytes (artifacts/tui-perf-after-final.txt). This is a synthetic CPU workload, not a native terminal frame-rate measurement. The regression includes deferred reflow and a 96 MiB allocation budget. Cell-buffer tests check that scrolling from long user rows to short prose clears old text/background, and shortening activity labels leaves no old header text.

The first broad focused run hung and its own testhost was stopped. A detailed run with a 30-second per-test hang guard passed all 285 cases without producing a hang sequence (artifacts/tui-header-diagnosis.log); the initial hang was not reproduced. Native Windows Terminal rendering/IME acceptance remains separate from these headless checks.

Final scoped CLI verification passed 1,282 tests / 0 failed / 0 skipped (artifacts/tui-responsive-cli-final.log). build.bat passed 11/11 stages (artifacts/tui-responsive-release.log). Documentation guards passed 19 tests (artifacts/tui-responsive-docs.log). Session captures at 80×24, 120×30 and 160×40 use synthetic data in artifacts/tui-header-{narrow,session,wide}.*. The preceding 5,962-test full matrix below belongs to the earlier layout snapshot; it was not rerun for these CLI-only changes.

Earlier OpenCode-inspired TUI layout (2026-10-12)

Ru's palette and landing animation are retained. The empty screen centres the wordmark/prompt; active sessions use conversation gutters, a filled prompt and a read-only context sidebar at 120×26 or larger. Below that size the session stays in one column. This earlier sidebar design was subsequently replaced by the header described above. The captured figures below belong to that snapshot.

The initial layout tests reproduced the old box/full-width arrangement. The first CLI run passed 1,254 tests and failed 13: twelve old visual assertions and one RPC backpressure fixture race. The fixture now waits for the writer to consume the response before filling its bounded queue; its assertions remain. Subsequent runs found an obsolete box assertion, a Terminal.Gui callback signature mismatch during compilation, and a long placeholder overlapping metadata at 20×12. These were corrected rather than dropped from coverage. Ten focused layout/wrapping checks then passed (artifacts/tui-opencode-edge.log).

The final Debug matrix passed 5,962 tests across 12 projects, including 1,271 CLI tests, with no failures or skips (artifacts/tui-opencode-matrix-final.log). The first Python run passed 213, skipped one and failed the README badge guard: the badge's URL-encoded count had not been updated alongside the visible count. The badge was corrected; the final Python run passed 214 and skipped the existing POSIX-only test on Windows (artifacts/tui-opencode-python-final.log).

Headless ANSI-driver captures use synthetic session data: 80×24, 120×30 and 160×40 sessions plus a 120×30 landing screen. The PNGs are rendered from the actual cell/colour buffers, with corresponding TXT/JSON files under artifacts/tui-opencode-{narrow,session,wide,home}.*. They are layout evidence, not a real-provider run or native Windows Terminal/IME acceptance.

Selective meteor branch integration (2026-10-12)

Selected changes from feat/cli-meteor-braille-landing at 5b738997 were adapted onto dev a7504f00, rather than resolving by replacing current TUI files. Landing/header/resize, payment-error mapping and narrower diagnostic classification are included. Dev's single Terminal.Gui engine, keyboard owner, bounded transcript, immutable queue submissions and profile-v2 policy remain. Scroll/wheel navigation extends that owner; follow-tail notifications keep hints consistent during busy rendering and clear. Workspace-busy hints clear only after Host acceptance or a successful session change, never on ordinary typing.

The isolated feature review reproduced three new hint/banner regressions; its three duration-locale failures were inherited from the old base and already fixed on dev. The port initially failed compilation before the busy-banner interface existed (artifacts/meteor-port-red.log). All six integration/banner regressions then passed (artifacts/meteor-port-regressions.log); 35 focused TUI cases passed (artifacts/meteor-port-focused.log). Resize tests now assert that the test driver actually reaches the requested dimensions; all 24 chrome cases passed (artifacts/meteor-port-resize.log). Provider tests additionally cover Responses streamed payment errors as non-retryable.

The full Debug matrix passed 5,952 tests across 12 projects, with no failures or skips (artifacts/meteor-port-dotnet.log). The Python suite passed 214 tests, 1 skipped (artifacts/meteor-port-python.log, POSIX-only skip on Windows). build.bat passed 11/11 Release stages (artifacts/meteor-port-release.log). Documentation guards passed 19 tests and git diff --check passed. No real-provider coding benchmark or native terminal resize/IME acceptance was performed. Headless Terminal.Gui drawing is separate evidence from a real Windows Terminal session.

TUI engine simplification (2026-10-12)

The classic ANSI composer/key loop and chat resize reconstruction are removed. Terminal.Gui owns interactive input and screen rendering; the text fallback owns only line reads, bounded history and append-only output. Redirected stdin also selects that fallback. Shared command completion no longer depends on an interactive line reader. Credential input retains no-echo and bounded paste, using the existing clipboard reader rather than a second native implementation. Plain input-loop exit/EOF/cancellation, quote preservation, prompt echo and redirected-input selection have regression coverage. Existing tests for removed classic rendering helpers were retired; Terminal.Gui and run/approval/input/ queue/session/JSON coverage remain.

The new contracts initially failed compilation before the replacement input existed (artifacts/tui-simplify-red.log). Intermediate compile runs identified remaining references to the retired resize callback. The first complete CLI run passed 1,228 tests and Client passed 214 (artifacts/tui-simplify-cli.log). After adding further regressions, a string assertion used culture comparison for an ESC character and falsely failed (1,231 passed, 1 failed; artifacts/tui-simplify-cli-final.log); exact ordinal comparison corrected it. The final complete CLI suite passed 1,232 tests / 0 failed / 0 skipped (artifacts/tui-simplify-cli-green.log). Documentation guards passed 19 tests; git diff --check passed. build.bat completed all 11/11 Release stages (artifacts/tui-simplify-release.log) and refreshed the flattened distributable. No application was launched. The full 12-project matrix and Python suite were not rerun at that CLI-only step; later full runs are recorded in the table above. This is deterministic coverage; native terminal resize/IME and a live-provider coding task are unmeasured.

Workspace terminal redesign (2026-10-12)

Profile v2 makes the canonical launch directory the Standard terminal gate, independent of executable, language, script form or network metadata. Literal argv stays literal; user-installed executable PATH locations resolve without an install-location allowlist. Protected/external cwd, explicit rules, Plan, long-call approval and process bounds retain their documented behavior. The unused PowerShell AST approval classifier and SDK are removed. Saved v1 session selections upgrade only for new runs; historical runs stay pinned. See ADR 0017.

Intermediate runs exposed outdated policy/fixture assumptions and an execution receipt bug for approved external cwd; these were corrected. The full Debug runner passed all 11 backend projects, then failed one GUI assertion of the old Standard subtitle (1,384 GUI tests passed). Updating that assertion to the new contract and rerunning the complete GUI project passed all 1,385 tests. Together, current coverage is 5,970 tests across 12 projects, with no remaining failures or skips. Logs: artifacts/terminal-redesign-dotnet.log, artifacts/terminal-redesign-gui-final.log. The final focused HTTP/conformance run passed 41 tests (artifacts/terminal-redesign-http-focused.log).

The full Python tooling suite passed 214 tests, 1 skipped (artifacts/terminal-redesign-python.log); the skip is POSIX-only on Windows.

build.bat passed 11/11 Release stages (artifacts/terminal-redesign-release.log). The flattened distributable changed from 1,087 files / 448,634,294 bytes to 749 files / 353,677,576 bytes (about 428 to 337 MiB). No PowerShell SDK or System.Management.Automation assembly remains; publish dependency manifests also contain neither package. Footprint records: artifacts/terminal-redesign-release-before.json and artifacts/terminal-redesign-release-after.json. No user application was launched. Documentation guards passed 19 tests and git diff --check passed.

These are deterministic policy, persistence, execution and presentation checks. No native GUI acceptance or live-model benchmark was run; approval reduction does not establish improved model latency or task completion.

Standard workspace Python files (2026-10-12)

Historical scoped policy, superseded by ADR 0017 above.

The user authorized a modest Standard relaxation. Existing workspace .py entries (optional -u/-B) are candidates for the appended WorkspaceScript class; Python analyzer version 2 marks them. Evaluation and execution check the canonical entry location; explicitly named interpreters must also live in the workspace. Inline code, arbitrary modules, package installation, arbitrary shell forms, protected/external targets and explicit ask/deny rules keep their existing gates. Scripts have current-user capabilities, not OS confinement. See ADR 0016.

Two regression cases failed before implementation (945 passed, 2 failed; artifacts/standard-python-red.log). The focused Broker run then passed 947 tests (artifacts/standard-python-broker.log). With added preset/execution-leg coverage, the final full Debug matrix passed 6,009 tests across 12 projects, with no failures or skips (artifacts/standard-python-dotnet.log). An initial documentation guard rejected the ADR's missing front matter; the metadata was corrected and all 19 guards passed. No native/provider PPTX acceptance was run.

The full Python suite passed 214 tests, 1 skipped (artifacts/standard-python-full-python.log); the skip is POSIX-only on Windows. build.bat passed 11/11 Release stages (artifacts/standard-python-release.log), including a fresh flattened distributable with the GUI fixes. No application was automatically opened and no live credentials or user session data changed.

Startup previously assigned an empty row collection and only rendered sidebar sections when a session poll changed its fingerprint. An empty initial result therefore left Workspaces, Recents and persisted workspace pins invisible until a local action changed sessions. The startup regression guard failed before the fix (artifacts/sidebar-startup-red.log: 1 failed, 1,384 passed). Startup now renders after resolving the app root and before starting the Host; unchanged polls also repair a missing row collection while retaining existing rows.

The final scoped Debug GUI run passed 1,385 tests / 0 failed / 0 skipped (artifacts/sidebar-startup-gui.log). Documentation guards passed 19 tests; git diff --check passed. An initial command attempted an unsupported -TestFilter runner argument and did not execute tests; the preserved red and green runs above used the supported GUI project filter.

Native restart acceptance and the user's deleted PPTX run were not reproduced. The user confirmed the run used standard; no execution preset or approval policy was changed. Its model latency and shell failures remain unmeasured.

Live transcript activity boundaries (2026-10-12)

Three regression inputs reproduced the live-only orphaned sentence suffixes (artifacts/live-activity-boundary-red.log: 3 failed). The old splitter moved the boundary backward to a word or line edge, while text rendering could still have preceding deltas queued. Live insertion now renders pending preceding text, freezes its full event-ordered extent and settles that slice; subsequent text keeps a new live tail. Earlier queued activities are drained before accepting later deltas, preserving ordering across model steps. The obsolete boundary heuristic and its tests are replaced by behavioral live-segment coverage.

The scoped scripts/test_dotnet.ps1 -Configuration Debug -ProjectFilter gui-winui3 run passed 1,384 tests / 0 failed / 0 skipped (artifacts/live-activity-gui-final.log); documentation guards passed 19 tests. This is deterministic render-model and source-wiring evidence; native GUI acceptance, real-provider streaming, Release publish and the full backend/Python matrices were not rerun for this local presentation fix. The full matrix was not rerun at that step; later full runs are recorded in the table above and their dated sections.

Architecture simplification (2026-10-12)

GUI session identity, open-request ownership, run monitoring and turn completion now belong to ChatSessionPresentation. Tests cover overlapping switches, outgoing-session bootstrap, stale monitor callbacks and completion admission. Same-run accumulator rebinding preserves event deduplication; two regression tests failed before the repair (artifacts/session-presentation-dedup-red.log).

Tool-batch classification, pause and cleanup share one implementation across initial execution and retained approval/input paths. Proposal persistence still precedes each evaluation, and approval refresh does not execute tools. Three regression cases verify that a later evaluation failure revokes fresh approvals, discards prepared executions and executes nothing.

CLI turn cancellation, active/parked/followable run identities and ordered unique assistant payload references belong to ChatRunSession. Behavioral tests cover failed Continue, parking, interruption, overlapping turn admission and reset. The existing command/rendering file remains large; this change isolates run lifetime rather than completing a terminal presentation redesign.

The first GUI implementation run passed 1,377 tests and failed three source guards that required the previous inline ownership (artifacts/session-presentation-gui.log). Those guards now check the deliberate ownership seam, with behavior covered by the new presentation tests. Final full Debug verification passed 5,997 tests across 12 projects, with no failures or skips (artifacts/architecture-simplification-dotnet.log). Documentation guards passed 19 tests. Native UI acceptance, live providers, coding benchmarks, Release publish and the full Python suite were not rerun in this change.