Verification evidence — undated¶
Historical snapshot only. Counts, commands and limitations describe that snapshot and do not establish current acceptance. Use current verification.
Verification¶
The table below records the latest full-matrix and release-pipeline evidence; entries under it are history or scoped checks. Figures are maintained by hand. Run the scripts for authoritative current figures --
scripts/test_dotnet.ps1,check_aspnetcore_route_matrix.py,pytest tests,build.bat-- rather than quoting this table as a live measurement. Historical per-snapshot counts are preserved verbatim in their sections and must not be read as current totals. Sections below that still describe voice/STT, VTube Studio, screen observation, device worker v1, or the retired 13→4 tool matrix are history, not the current contract: the current route contract has 46 routes, 18 tools, and an in-process model runner. Đợt 2026-09-13 (media teardown): repo chỉ còn coding core —/vts/*+/observation/*(11 + 8 route) gỡ cùng adapter VTS/observation, contractVtsContracts/ObservationContractsvà nhánh v1vts.*;/voice/*+/voice/stream+/voice/transcribe,src/ru_runtime,OmniVoice, adapterAudio/*/Speech/*và device worker v1 gỡ nốt (managed Worker chỉ còn/worker/v2;GET /statusphục vụ trực tiếp từ host settings); GUI bỏ voice/STT/subtitle (VoiceView, subtitle overlay, nút mic); settings retire nhómvoice_*/auto_*/subtitle_*+gemini_vision_api_key; promptruntime_contextcònmood/hardware/references; projectRuAgent.Runtime.Corevà extra[voice]/mypy gỡ khỏi build/CI. XemCHANGELOG.mdmục Removed. | Surface | Current status | | --- | --- | | Public route | 46 HTTP + 0 WebSocket perbackend-dotnet/public-route-contract.json— 25 additive/agent/v1/*plus 21 non-Agent. Every entry isaspnet-direct; the media routes are no longer mapped. Guard:scripts/check_aspnetcore_route_matrix.py, whichci.ymlruns on every push. | | .NET tests | 13-project Debug coverage: 6,884 passed across final project runs, 0 failed / 0 skipped (2026-10-19, working tree including the subagent Git preparation fix and pending directory-handle/watcher fixes). Broker 1,064; CLI 1,470; Client 241; Core 518; Persistence 269; Hosting 41; Protocol 152; Adapters 228; Model 91; SessionInterchange 43; Updates 25; Backend 1,228; GUI 1,514. Evidence:artifacts/subagent-git-failure-2026-10-19/verification.json. The initial CLI run hung in Terminal.Gui input shutdown and was aborted; unchanged source/assertions passed on a PTY rerun. Stack and both outcomes are retained. This is combined project evidence, not an uninterrupted green runner. | | Python suite | 216 passed, 1 skipped (2026-10-19), isolated staging snapshot. POSIX process-group case skipped on Windows. Log:artifacts/subagent-commit-2026-10-19/python.log; initial metadata failure retained separately. | | Subagent inbox recovery fix | 2,171 passed across four affected Debug assemblies (2026-10-19): Core 518, Persistence 273, Backend 1,228, Protocol 152. Nine inbox cases include suppressed audit replies, no queued messages, pending native calls and missing payloads. A private snapshot of the reported failure reproducedpayload_model_forbiddenat inbox delivery; the same replay passed after the query fix. Evidence:artifacts/subagent-drive-failure-2026-10-19/. These are scoped checks after the full-matrix baseline above; no real-provider task outcome is claimed. | | TypeScript SDK | 61 passed / 0 failed, typecheck and ESM/DTS build passed (2026-10-19), isolated staging snapshot. Logs:artifacts/subagent-commit-2026-10-19/sdk-{test,typecheck,build}.log. | | Release build | Local 11/11 stages passed (2026-10-20 local time), self-contained win-x64 GUI, Host and CLI in the flat runnable layout, including inbox recovery, repeated-wait and partial-web-evidence fixes and pending working-tree fixes. Temporary layout diagnostics were removed. Log:artifacts/subagent-live-gui-2026-10-19/release-build-delivery.log. No installer or remote release published. | | Managed runtime smoke | PASS (2026-10-20 local time) against that exact rebuilt flat Release tree: authenticated/statusand clean shutdown using an isolated data root. Log:artifacts/subagent-live-gui-2026-10-19/release-smoke-delivery.log. Native GUI and real-provider evidence is recorded separately below. | | Live coding benchmark | Everyday baseline 16/18 unattended passes across six C#/Python/TypeScript tasks; matched three-task follow-up 7/9 → 7/9, mean duration 251.1 → 172.2 s. Kilostepfun/step-5-preview-free, saved high reasoning, Standard, 300 s deadline; both matched phases had zero approvals. Two follow-up timeouts remain. Earlier mixed hard diagnostic was 1/10; these are custom corpus results, not official dataset scores. | | Plugins | No media plugin remains in the repo: voice/STT, VTube Studio, screen observation and the native overlay are removed, andRU_MEDIA_SURFACESis no longer read. Seedocs/architecture.md, "Coding core and plugins". | |git diff --check| passed |
Deterministic release matrix¶
# Run from %RU_ROOT%.
dotnet restore gui-winui3.Harness\RuApp.Harness.csproj --nologo
.venv\Scripts\python.exe -m ruff check tests scripts tools
.venv\Scripts\python.exe -m pytest -q
.venv\Scripts\python.exe scripts\check_aspnetcore_route_matrix.py
.venv\Scripts\python.exe -m pip check
.\scripts\test_dotnet.ps1 -Configuration Release
dotnet build gui-winui3.Harness\RuApp.Harness.csproj -c Release --no-restore --nologo
cmd /c build.bat
CI and local verification share scripts/test_dotnet.ps1, which recursively
discovers *Tests.csproj under backend-dotnet and gui-winui3.Tests, excluding
generated bin, obj, and out paths. It runs them sequentially and stops on
missing roots, discovery errors, an unavailable SDK, or the first test failure.
It restores by default; use -NoRestore only after restoring the projects.
Protocol and all three Runtime test projects are included.
The test cleanup removed the unused Python VTS/overlay package and its tests, including the test that could open a real VTube Studio connection. Managed C# VTS tests and opt-in native acceptance remain separate. Build and launcher unit tests replace only their external process boundary and never open RuApp.
Ru Agent tool hardening evidence (2026-09-05)¶
- The canonical inventory contains exactly 13 model-visible definitions. A
table-driven acceptance matrix
verifies every available definition has a schema-identical binding and that
ModelEnabledmatches canonical exposure. - DuckDuckGo HTML result redirects are decoded to their final targets and then
revalidated by
WebPolicy. A live read-only smoke returned 10 bounded citations for each oflatest dotnet release official docsandVietnam weather today; the first targets were Microsoft Learn and EaseWeather respectively. The adapter has no fallback search provider. - Sequential Debug .NET matrix:
2,548 passedacross all 13 test projects; Python suite:417 passed; route contract:64 HTTP + 1 WebSocket; documentation checks:4 passed; releasebuild.bat: 13/13 plus 2/2 native overlay tests. - Live volume mutation was intentionally excluded. MCP and connector libraries remain unadvertised and are not counted as shipped Agent tools.
Multimodal image attachment evidence (2026-09-03)¶
Historical evidence below describes the then-current Worker path and request cap. The product now uses in-process model generation and a 26 MiB JSON run body cap. The current decoded image limits remain four images, 4 MiB each and 20 MiB total; picker and clipboard intake share normalization. Current wiring is documented in architecture.
- End-to-end multimodal image attachment pipeline implemented across WinUI 3, ASP.NET Core Public Host, Worker v2, and LLM providers (OpenAI Responses, Anthropic Messages, and OpenAI-Compatible).
HarnessPayloadKind.UserImage = 11dual audience (AuthenticatedClient | ModelContext) persisted atomically in SQLiteagent_payloadstable.POST /agent/v1/runsbody cap bumped to 1 MB for runs withimage_attachments. Strict validation enforces max 4 images, max 4 MB decoded per image, max 20 MB total decoded, allowed MIME types (image/png,image/jpeg,image/webp), strict base64 encoding, and model vision capability. Non-vision models fail closed withvision_unsupported.- Worker v2 wire protocol supports
"images"in start payload;LoopbackModelStepRunner.LoadRunImagesAsyncfeeds decrypted payload data URLs intoModelRequest.Images. - Client-side
ImageAttachmentResizerdownscales images exceeding 2048px on max dimension and compresses them to JPEG (0.85 quality) or PNG before sending. Chat bubbles render thumbnails up to 240px and provide click-to-zoom in aContentDialog. - All 13 test projects passed (2,438 tests, 0 failed, 0 skipped). Both Debug and Release builds succeeded with 0 warnings and 0 errors. The detailed record is retained in git history.
Cross-system hardening evidence (2026-09-02)¶
- Workspace external-change detection resets per-workspace baselines between runs, resets per-path observation state when a new baseline is registered, skips reparse-point traversal, hashes through a bounded streaming pass, and keeps watcher debounce state isolated per registration. Debounce callbacks re-read the latest baseline before emitting; watcher callbacks also ignore paths escaping through reparse boundaries without leaking exceptions, and watcher buffer errors trigger a bounded baseline rescan. The stale-dedup, lifetime, and race regressions and eight monitor cases pass.
- Activity logs now mark event IDs only after a successful append, preserve duplicate state when clear fails, and read/trim/migrate JSONL through bounded streaming paths. Malformed migration rows are skipped independently. The ActivityLogStore and migration slices pass, including fault-injected retries.
- Activity preview sanitization now bounds the raw text entering regex/privacy processing and clamps caller-provided preview lengths (including tiny limits), preventing oversized malformed activity payloads from causing avoidable allocations.
- Legacy broker session descriptors and harness/runtime descriptors are rejected before parsing when they exceed their bounded input caps; bounded streaming also closes the metadata-check race. Oversized fixtures pass without allocating the full payload.
- Session files and settings use bounded streaming reads; oversized or concurrently growing local state falls back to the existing safe-empty path.
- SQLite workspace roots, policy documents, model-context payloads, and journal envelopes now validate BLOB lengths before allocation/decryption. Workspace restore hashes the current file in a streaming pass instead of loading it wholly into memory.
- Local session/settings JSON and the VTube Studio token file use bounded streaming reads before parsing; malformed or oversized local state falls back to the existing safe-empty behavior. Session messages are capped at 1 MiB and persisted titles are bounded to the UI's 50-character contract before they can be split or displayed; session writes preflight aggregate content size before serializing a document that cannot fit the 16 MiB file cap.
- The STT HTTP endpoint now bounds request-body buffering (including chunked requests) before forwarding audio to the worker; oversized requests are rejected without a worker call. The private speech runtime handler repeats the same 8 MiB cap before JSON parsing so direct worker invocation cannot bypass that boundary, and the Google recognizer rejects provider transcripts over 64 KiB before whitespace normalization/splitting.
- Runtime and VTS relays now apply the configured request-body cap themselves, including GET routes with unexpected bodies; oversized relay requests stop before the worker is dispatched. The direct VTS and observation-control handlers also cap their JSON bodies before parsing, so tests/compositions that invoke a handler without the router keep the same boundary.
- Policy and workspace-root cryptographic paths now treat missing or malformed key material as explicit fail-closed errors and never dereference/clear a null key buffer.
- Windows Credential Manager reads now validate blob size, pointer, and UTF-16 alignment before allocation; writes reject unreasonably large secrets before entering the native API.
- Declarative skill package loading now reads manifest/instructions through a bounded streaming path, so a file that grows after the initial metadata check cannot bypass the package size limits.
- GUI file attachments now re-check and stream image/text payloads through bounded readers; clipboard image preview and upload use independent streams, and a file that grows beyond the text cap is rejected instead of being added with empty content. Unsupported binary files are rejected instead of being queued with no transmission path. Because Agent v1 still ignores its legacy image field, image sends now fail explicitly in the GUI rather than silently disappearing; Agent v1 rejects the legacy field for external callers too, and the local UI-theme preference reader is capped as well.
- Watcher debounce timers remove themselves after firing, preventing a long- lived registration from retaining one timer per changed path indefinitely.
readwithop=readnow performs offset/window arithmetic inInt64, avoiding a 2-GiB boundary overflow that could incorrectly clear thetruncatedflag; it also rechecks the open handle's final length so concurrent appends are reported as truncated instead of being mistaken for a complete read.- Workspace-change restore checks the target and every existing parent for reparse points both before hashing and immediately before replacement, so a junction/symlink cannot silently redirect the restore path outside the trusted workspace boundary.
- Durable run Git state now caps attributed path count, path length, and JSON column size on both write and rehydration; malformed persisted path sets fail closed instead of being deserialized without bounds.
- Agent payload reads now reject oversized client-visible payloads explicitly; they no longer silently truncate JSON or base64 content into a corrupt value.
- OpenAI-compatible streaming responses now use the bounded SSE reader, so a
provider cannot force an unbounded
ReadLineAsyncallocation with one giant line; line, event, and total response caps are enforced before JSON parsing. - OmniVoice process state now has a dedicated synchronization boundary around
handle replacement/disposal;
IsRunningcannot observe a disposed handle, andStartAsyncrechecks disposal after acquiring its lifecycle gate. - OmniVoice's private process-log drain now consumes bounded char chunks rather than allocating an entire stdout/stderr line before discarding it.
- Agent REST, session and SSE client adapters now share an 8 MiB streaming response cap before JSON/error parsing, including chunked responses, so malformed/hostile local HTTP peers cannot force an unbounded allocation.
- The Agent SSE client now reads lines from bounded byte chunks instead of
ReadLineAsync, preserving strict UTF-8 and CR/LF handling while rejecting oversized lines before allocation; the first-line UTF-8 BOM behavior remains compatible with the previousStreamReaderimplementation. - Session ID validation now enforces the documented ASCII
a-zA-Z0-9_-alphabet instead of accepting arbitrary Unicode letters. - Provider catalog refreshes now use a per-provider-scope single-flight gate; a slow or unavailable endpoint cannot hold the global cache lock and block unrelated provider scopes. Canceled refreshes restore the prior attempt timestamp before propagating cancellation, so the next request is not falsely throttled.
- Diagnostic and activity JSONL readers now consume fixed-size text chunks and discard overlong records before materializing strings, while continuing to later valid records in the same file.
- Gemini Vision now caps screenshot request bytes and response bodies before decoding/parsing, including chunked responses without a trustworthy length.
- Activity outbox ingestion also skips malformed scalar rows and continues with later events, instead of leaving a processing file stuck for every retry.
- CLI/client JSON response parsing disposes
JsonDocumentinstances promptly; the full .NET matrix remains clean after the ownership changes. - The diagnostic endpoint fixture writes to a prior-day file so its seed data cannot race the live writer's current-day append handle on Windows.
- Sequential
scripts/test_dotnet.ps1 -NoRestore:2,392 passedacross all 13 test projects; Python suite:413 passed; route contract:62 HTTP + 1 WebSocket(historical snapshot — current contract is 64 HTTP + 1 WebSocket perbackend-dotnet/public-route-contract.json); documentation/route checks:8 passed;git diff --checkclean.
Test-suite cleanup evidence (2026-08-31)¶
- Python:
402 passed(audit baseline:510); unused VTS/overlay tests, synthetic file-tool simulations, and redundant guards were removed or consolidated, with new build/launcher/runner behavior checks added. - .NET Release:
2,009 passedacross all 13 projects, including backend633and GUI252. The GUI reduction from339removes cosmetic snapshots, retaining three focused resource/accessibility contracts and behavior tests. - Ruff, Mypy (
39source files),pip check, and the56 HTTP + 1 WebSocketpublic route contract all passed (historical snapshot — current contract is 64 HTTP + 1 WebSocket perbackend-dotnet/public-route-contract.json). - The first .NET run encountered a business-fixture JSONL cleanup lock. Only
those business tests now inject a discarding diagnostic emitter; the
89focused business/diagnostics tests and subsequent full matrix passed. The real writer's bounded shutdown wait still needs investigation. This cleanup does not claim that lifecycle bug fixed. - Cleanup verification did not publish a fresh GUI, automate the interactive GUI, perform real voice/VTS/cloud acceptance, or uninstall packages. Commit/push is a separately authorized Git handoff, not runtime acceptance.
Codebase optimization evidence (2026-08-31)¶
- Phase 1 (L1 Diagnostics & N1 Native Checks):
StructuredDiagnosticLogWriterlifecycle:IDiagnosticLogSink,StopAsync(Drained/Cancelled/Incomplete)andCompletionensure quiescence before gate cleanup. Tested withDiagnosticWriterShutdownTests(7 cases covering normal, honor, ignore, flush, multi-stop, noop, failure bounds).- Native overlay Release tests: replaced assert with
CHECKintest_check.h, registered--self-test-failurewithWILL_FAILin CMake, Spout probe made opt-in (RU_BUILD_NATIVE_PROBES=OFF). Pure tests pass (2/2) in Release. - Phase 2 (N2 Pixel Transform, P1 GUI Refresh, C1 Render Coalescing, O1 Outbox, F1 File Search):
- Native fused pixel path:
PrepareRgbaForOverlayfuses green matte, RGBA->BGRA, and premultiplication into a single loop. - GUI single-flight coordinator:
RefreshController<T>prevents overlapping network fetches, cancels stale requests on deactivation, and drops old-generation results. 8 tests passing inRefreshControllerTests. - ApiService cancellation: added
CancellationTokentoGetSessionsAsync,GetDiagnosticStatusAsync,GetDiagnosticLogsAsync, tested inApiRefreshCancellationTests. - ChatView render coalescing:
AgentReplyAccumulatorowns unicode deduplication and string accumulation, tested with 10k synthetic deltas inAgentReplyAccumulatorTests(5 tests passing) and scheduling inAgentReplyRenderSchedulingTests. - Outbox optimization:
HarnessOutboxStore.ReadNextWorkAsyncwithOutboxWorkunion (Idle/Transition/Delivery) avoidsBEGIN IMMEDIATEon empty queues and prioritizes transitions before deliveries. Tested inHarnessOutboxTests(79 tests passing). - Bounded file search:
ReadSearchTextprobes binary prefix (8000 bytes) and reads bounded UTF-8 stream in a single open forFileSearchAdapterandHarnessToolGateway. Tested inFileDiscoveryAdapterTestsandHarnessFileSearchTests(6 tests passing). - Phase 3 (B1/W1 Architecture Candidates):
- Preserved rejection of Broker and Worker extractions due to distinct framing, concurrency, and security invariants; all Broker and Worker tests pass without degradation.
- Phase 4 (D1 Dependencies & D2 Documentation):
- Removed unused
ddgsfrompyproject.toml. - Aligned the
voiceextra with the evidence-backed isolated provider line (torch>=2.8,<2.9,torchaudio>=2.8,<2.9,omnivoice>=0.2.1,flask>=3.1,<4; accepts the installed2.8.0+cu128CUDA build including local tags) and recorded consumer inventory, version authority, and verification policy indocs/operations/python-dependency-ownership.md. - Setup script preflight verified via
setup_omnivoice_env.ps1 -WhatIf;.venvpip checkpasses but is not fresh-install proof. Three metadata cases (tests/test_python_metadata.py) plus two documentation cases formed the five-case focused run; metadata checks do not import Torch. - Review remediation wave (six findings, 2026-08-31):
- Logger/refresh lifetimes:
DiagnosticWriterShutdownTests10 cases +DiagnosticWriterHostShutdownTests1 +RefreshControllerTests15 all green; shutdown/disposal quiescence verified by coordinator race harness (7 cases) and independent review. - Sidebar health single-flight via the production
RefreshControllerseam +HealthStatusService; productionAgentReplyRenderSchedulerused by ChatView and tested directly (AgentReplyRenderSchedulingTests). - Native unchanged-status suppression via
status_presentation.hwith failure-safe commit; pure tests pass and a forced-decision mutation fails atoverlay_options_tests.cpp:70, proving the tests bind to production. - Historical aggregate gates, 2026-08-31 remediation checkpoint (sanitized shell: no
ZCODE_DESKTOP_CONTEXT_PROMPT_ENABLED, Windows-firstPATH; see the SDD ledger note about host-env redaction andtimeoutresolution): - Python:
405 passed;ruff check .clean;mypy srcclean;pip checkclean. - .NET Release across all 13 test projects:
2,080 passed, 0 failed(Broker 488, RuBackend 650, Harness.Core 210, GUI 303, Cli 119, Adapters 84, Persistence 79, Worker 49, Client 42, Hosting 36, Protocol 11, Core 6, Connectors 3). build.batfull release pipeline13/13including native overlay tests2/2.
These deterministic checks do not establish the optimization plan's missing
before/after work counters, allocation measurements or median/p95 benchmarks.
Performance acceptance remains open in the optimization plan's SDD ledger (scripts/measure_resources.ps1), independently of bug-fix tests.
Provider/lifecycle review remediation (2026-09-01)¶
- Eight protected-bootstrap cases prove canonical credential isolation for OpenCode/Kilo/Custom and no cloud-key lookup for local providers. The missing OpenCode/Kilo cases failed before removal of both key fallbacks, then passed.
- The client process regression failed when Dispose killed its synthetic Host, then passed after client-owned job/kill behavior was removed. The helper is an isolated hidden test process, not RuBackend or any running user process.
- Startup cancellation checks before/after awaits suppress continuation even when a delayed operation ignores cancellation. Voice-status HTTP requests receive the startup token. No interactive WinUI acceptance was performed.
- Fault-injected policy/preset HTTP requests formerly exposed a synthetic secret/path; they now return safe bounded messages. Request cancellation is preserved. Focused backend checks passed 19 cases; startup/API checks 9.
- The initial repair did not publish/launch the GUI, call live providers, mutate credentials, install packages or claim performance/live acceptance.
- Final sequential gates on this working tree: Python
405 passed; .NET Release2,100 passedacross 13 projects (Broker 488, CLI 119, Client 45, Connectors 3, Harness.Core 210, Persistence 79, Hosting 36, Protocol 11, Runtime.Adapters 84, Runtime.Core 6, Worker 49, RuBackend 661, GUI 309).ruff check src tests scripts tools,mypy src(39 files),pip check,62 HTTP + 1 WebSocketroute validation andgit diff --checkpassed. Actual WinUI Release/x64 compile passed with zero warnings/errors usingdotnet build gui-winui3/RuApp.csproj -c Release -p:Platform=x64 -p:RuntimeIdentifier=win-x64 --no-restore. That compile did not update the shortcut's published GUI artifact; see the separately authorized release follow-up below.
Provider catalog refresh evidence (2026-09-01)¶
The provider catalog operating contract is current; completed design and
implementation details remain in Git history.
Host discovery (ProviderCatalogService with bounded 15 s timeout, 4 MiB cap, 5000-item cap),
capability parsing (ProviderCatalogParser for Kilo opencode.variants allowlist vs bare-ID unknown),
per-model reasoning persistence (ProviderReasoningPreferences), typed additive GET /models snapshot,
GUI polling/throttle/single-flight (ProviderCatalogController) and protected per-run Worker handoff
(AgentProviderSnapshotFactory + ProviderRequestOptionsValidator) are covered by controlled fake-handler/clock/credential
tests only — no live provider credentials or generation and no published GUI launch.
Focused and sequential gates on this uncommitted working tree after native OpenAI/Anthropic provider implementation (commit/publish remain separately authorized):
- Backend
RuBackend.Tests683 passed(including provider catalog/cache/parser, OpenAI/Anthropic capability registry2026-09-01, catalog request factory withAuthorization: Bearer/x-api-key+anthropic-version: 2023-06-01,store:false/max_tokens8192, endpoint, malformed preference, neutral startup and run-pinned snapshot cases, plus fail-closed persisted-provider/API alias handling without OpenCode fallback). - GUI
gui-winui3.Tests330 passed(including provider-catalog cases covering contract, no fabricated defaults, reasoning options, stale/empty vs error, single-flight, throttle, polling, disposal, partial saves, andOfficialProviderTestsfor OpenAI/Anthropic selector rows, DTO wire names, and model-save routing). - Worker
RuAgent.Runtime.Worker.Tests55 passed(including Worker v2 reasoning-option stripping, trusted-root bootstrap requirement, protocol dispatch viaIModelProviderResolver, cancellation cleanup, and internalper-stepbootstrap startup without provider fallback). - Adapters
RuAgent.Runtime.Adapters.Tests169 passed(including protected-bootstrap lifecycle with protocol/output-cap,BoundedSseReaderTests16,OpenAiResponsesModelProviderTests24,AnthropicMessagesModelProviderTests31). Regression coverage proves raw-byte line rejection before whole-stream consumption, CR-only framing, strict native roles before HTTP, duplicate OpenAI terminal rejection, matching SSE event/type pairs, known-event/ping mismatch rejection, and safe malformed scalar/container mapping. - Full
scripts/test_dotnet.ps1Release2,234 passedacross all 13 projects (Broker 488, CLI 119, Client 45, Connectors 3, Harness.Core 210, Persistence 79, Hosting 36, Protocol 11, Adapters 169, Core 6, Worker 55, RuBackend 683, GUI 330). A local managed-host smoke test also kept Backend, Worker, and Broker alive and returned authenticated HTTP 200 from/settings. - Python
409 passed(ruff check src tests scripts toolsclean,62 HTTP + 1 WebSocketroute contract clean,git diff --checkclean,test_current_documentation.py2/2, doc links valid). dotnet build gui-winui3/RuApp.csproj -c Release -p:Platform=x64 -p:RuntimeIdentifier=win-x64 --no-restoreclean..gitignorefixed:models/→/models/soRuAgent.Runtime.Adapters/Models(BoundedSseReader, OpenAiResponses, AnthropicMessages, resolvers) is no longer ignored. Anthropic: matchingevent: pingframes are ignored, mismatches involving a known event fail closed, and unknown future event pairs remain forward-compatible.
Boundaries: mypy src passed for 39 source files; native/microphone/speaker/VTube Studio/observation live acceptance remain pending.
Provider catalog coverage proves bounded HTTP behavior, capability-registry intersection, native reasoning serialization and protocol dispatch, not liveness of OpenAI/Anthropic/Custom/Ollama/LM Studio endpoints or billing.
No live OpenAI/Anthropic key, catalog refresh, generation, publish, or push was performed here; those are separate user-authorized acceptance steps.
Visible GUI startup/catalog smoke on 2026-09-01 launched the current Release/x64
RuApp.exe without automating controls. RuApp, Public Host, Worker and Broker
all remained responsive; authenticated /status and /settings returned 200.
The selected Kilo catalog refresh returned status: ok, one model and one
matching metadata entry, stale: false, with no error. The window remains open
for user inspection. No prompt, generation, paid request, voice/device action,
or reasoning-setting mutation was performed, so those live paths remain pending.
Grouped provider model picker evidence (2026-09-01)¶
The grouped picker operating contract is current; completed design and implementation details remain in Git history.
Sequential deterministic gates on this working tree:
- Backend
RuBackend.TestsRelease706 passed, including the strict atomic-selection expansion table (seven aliases), fail-closed rejection of unknown aliases, blank/oversized models, extra properties, mixed provider-specific model fields, and byte-for-byte unchanged settings after every rejection with no runtime load, plus per-provider catalog endpoint coverage: a configured but inactive provider serves its own catalog, a provider without a credential fails ascatalog_unauthorized, and an unknown alias fails asprovider_not_active. - A follow-up endpoint change (
GET /modelsfor any configured provider, active or not) re-ran the full backend suite at706 passedand was verified against the locally running managed Host with non-secret state only: configured providers returnedstatus: okwith usable catalogs (32 and 363 usable rows) and unconfigured providers failed closed withcatalog_unauthorizedorcatalog_unavailable. - GUI
gui-winui3.TestsRelease354 passed, including grouped projection ordering/visibility/stale/unavailable/search cases, grouped controller composition (independent per-provider publishing, coalesced refreshes, invalidation with late-result rejection, due-only polling, disposal without post-dispose publication), exact two-field body serialization for Kilo/OpenAI/Anthropic, andSettingsSavedraised only on a success envelope. - Python
409 passed;62 HTTP + 1 WebSocketroute contract unchanged;git diff --checkclean; the documentation/route-matrix pairing passed4/4. dotnet build gui-winui3/RuApp.csproj -c Release -p:Platform=x64 --no-restorepassed with zero warnings and zero errors.- Fallback audit: every
fallback/provider-retry hit ingui-winui3andbackend-dotnet/RuBackendis unrelated to model selection (playback-device, subtitle-geometry, mutex, port-config, encoding, and lease-registry naming). The grouped picker contains no provider/model/key/protocol fallback.
Boundaries: no GUI launch, no live provider catalog/generation, no credential or settings mutation, no publish, and no push. Interactive grouped-picker acceptance (visible order, searches, keyboard navigation, themes, cross-provider save with a simulated failure) remains a user-run manual gate.
Authorized release follow-up (2026-09-01)¶
- The user requested commit/push and a refreshed GUI publish. Fresh sequential
gates passed: Python
409(including four new build-script regressions), .NET2,100across all 13 projects, and native CTest2/2. - Artifact inspection caught an existing build-script defect: XAML copy used
obj/Releasewhile x64 publish emittedobj/x64/Release, and copy failures were ignored. Four behavioral regressions failed before the fix, then passed; all 12 build/launcher script cases passed. Clean and copy now use x64, and missing/copy-failed XAML fails the release before the native build. - The corrected
build.batcompleted13/13, publishing GUI, CLI, managed Worker/Core, Broker and backend. The GUI is at thestart_dev.batRelease path shown below. Its self-contained runtime config, PRI and all 11 loose compiled XAML files were verified; GUI/Client/CLI/Worker/Broker/backend DLLs and PRI matched their build outputs by SHA-256. - Ruff, mypy (39 files), dependency and
62 HTTP + 1 WebSocketroute checks passed. No GUI was launched, no user Ru process was stopped, and no live provider/audio/VTS acceptance was claimed. The old rebuildable GUI publish output was replaced; user settings, credentials and session data were untouched.
The route checker must remain at 62 HTTP routes and 1 WebSocket
route (/voice/stream; 41 non-Agent plus 20 additive /agent/v1/* plus
POST /host/shutdown). The
deterministic Harness is contract evidence. A successful build proves only compilation,
publish output, and the incremental native overlay build/test invocation; it
does not prove that a GUI, microphone, speaker, VTube Studio model, or Windows
audio endpoint worked.
The managed runtime topology is deterministic: one C# Worker serves the Agent
loopback endpoint (/worker/v2) and the separate device endpoint
(/worker/v1). Worker tokens are distinct; the startup handoff is
credential-free and provider credentials are carried only by per-step
single-use DPAPI handoffs, never by the device graph,
legacy Python configuration, descriptors, logs, or public responses. The
production Python generation/worker modules and compatibility launch branch
are retired; the test-only fixture under scripts/harness is not a production
fallback. Retirement-only selectors and probes have been deleted.
The Phase 1 measurement command times deterministic prompt/tool operations in
small batches to amortize sub-millisecond scheduler noise, then measures their
allocation peak in a separate tracemalloc pass. It compares p50, p95, and
peak bytes with the checked-in baseline using the ten-percent gate plus a
bounded 0.05 ms timing-jitter floor; it never loads a model or contacts a
device.
The Worker v2 model inventory is deterministic and should stay aligned with
AgentRuntimeCoordinatorFactory.RuntimeCapabilities: read, write, edit,
bash, web_search, and web_fetch. The single
Agent profile receives the advertised list; policy fallback behavior is
verified separately. files.apply_patch remains unadvertised. Time, machine
status, app discovery, and optional volume operations use terminal tools;
notes use write with op=write.
Typed commands (/compact, /help, /skills, /skill, /tasks) use the shared
Protocol parser and are covered by CLI/Host tests. Web search/fetch are
enabled standalone managed capabilities backed by Google HTML primary with
fixed free fallbacks (DuckDuckGo HTML, Bing, Mojeek); each call remains bounded, redirect-safe, policy-checked, and marked
untrusted. Workspace SKILL.md bodies are untrusted prompt guidance only
(skill tool + /skills + /skill <name>); they never execute tools or
grant scopes, paths, or approvals. Connectors and app.open remain unadvertised.
Session deletion regression evidence (2026-08-31)¶
The GUI client regression test uses an isolated loopback HTTP listener that
returns HTTP 200 with status: "error". ApiService.DeleteSessionAsync must
reject that response, and the WinUI delete handler must not refresh or create a
replacement session after application-level failure:
# Run from %RU_ROOT%.
dotnet test gui-winui3.Tests\gui-winui3.Tests.csproj --no-restore --filter "FullyQualifiedName~SessionDeletionApiTests"
dotnet test gui-winui3.Tests\gui-winui3.Tests.csproj --no-restore
dotnet test backend-dotnet\RuBackend.Tests\RuBackend.Tests.csproj --no-restore --filter "FullyQualifiedName~Session"
The pre-cleanup verification passed the focused regression (1/1), the then
full GUI suite (339/339), and the backend session slice (39/39). These
behavior tests remain; the cleanup only removed cosmetic GUI checks. No
production session or user data is used by the regression listener.
The desktop launcher and release build share one GUI artifact, and the release
build publishes the managed Worker before the backend:
gui-winui3\bin\Release\net10.0-windows10.0.26100.0\win-x64\publish\RuApp.exe.
backend-dotnet\RuAgent.Runtime.Worker\bin\Release\net10.0\win-x64\publish\RuAgent.Runtime.Worker.exe.
backend-dotnet\RuAgent.Runtime.Worker\bin\Release\net10.0\win-x64\publish\RuAgent.Runtime.Core.dll.
build.bat fails if any of these managed artifacts is absent and verifies that
the GUI runtimeconfig contains includedFrameworks, proving the publish is
self-contained. start_dev.bat refuses to run when the Worker Release artifact
is missing instead of silently selecting a stale Debug executable; it passes the
workspace directory to a short startup probe so an immediate RuApp.exe
failure remains visible. A clean exit code from a second launch is accepted when
the existing single-instance GUI remains alive. Normal Windows shortcuts target
the published RuApp.exe directly to avoid opening a console window;
start_dev.bat remains the manual diagnostic launcher for visible artifact and
early-startup checks.
Phase 2 parity was deterministic and status-only. The shared fixture
tests/fixtures/python-csharp-phase2-device-baseline.json was exercised by
RuAgent.Runtime.Worker.Tests/Phase2ManagedDeviceParityTests.cs; those tests
covered v1 types/sequences, settings shape, voice control ordering, stable
errors, STT normalization, VTS code sets, observation bounds, cancellation,
and forbidden-value redaction without starting a provider or device. The
fixture and its tests were removed with the media/voice/VTS subsystems on
2026-09-13; this passage is kept as record only.
Ruff and pip check are required dependency/tooling gates; do not hide
findings with permanent ignores or continue-on-error. (Mypy was retired
from CI together with the Python runtime; test_ci_contract.py asserts it
stays out of the workflow.) Live microphone, speaker, volume, and VTube
Studio acceptance were the explicit user-run checks of the removed media
surfaces.
Phase 2 entry evidence before Python retirement (2026-08-30)¶
This is the deterministic evidence retained from the checkout before the Phase 3 retirement changes:
| Gate | Result |
|---|---|
| Python reference/tooling suite | 501 passed |
| C# Release matrix | 2,085 passed across 13 test assemblies (14 restored projects including the Harness build) |
| Route contract | 56 HTTP + 1 WebSocket |
| Runtime-boundary scan | 865 files; schema 1 reported reachable_count=2 (a startup-path label, not graph analysis) |
| Historical backend deterministic Harness | contract / passed; owned cleanup green |
| Historical Agent GUI Harness | 4/4 supported Agent scenarios passed |
| Release build | Local 11/11 stages passed (2026-10-19, Subagents GUI redesign), self-contained win-x64 GUI, Host and CLI in the flat runnable layout. Log: artifacts/subagent-readonly-2026-10-19/release-build.log. No installer or remote release published. |
Managed startup is now unconditional. The legacy runtime setting, parser, and special retirement error were removed in the first legacy cleanup wave.
The logging benchmark remains an optional latency probe. Its percentile tail is scheduler-sensitive on this Windows host; a noisy run is recorded as an environmental measurement failure and does not change the deterministic runtime gate. The per-writer enqueue tests remain the blocking logging guard.
Phase 2 manual acceptance record (2026-08-30)¶
The deterministic release gates are automated and do not stand in for native device behavior. The following checks are intentionally pending a user-owned Windows run. Automated verification may launch isolated RuApp, Worker, and Broker processes, but it does not touch the physical microphone, speaker, volume endpoint, VTube Studio, or provider process:
| Check | Status | Recorded evidence |
|---|---|---|
| Microphone capture and no-speech handling | pending user-run | not collected |
| OmniVoice stream, cancellation, playback, and subtitle order | pending user-run | not collected |
| VTube Studio catalog, action, and reconnect | pending user-run | not collected |
| Opt-in Gemini observation | pending user-run | not collected |
The authorized 2026-08-30 run passed all four deterministic Agent GUI scenarios and the backend contract Harness. Its historical retirement probe was deleted in the first legacy cleanup wave. No native microphone, audio, VTube Studio, or Gemini quality claim is made here.
When these checks are run, record only pass/fail, duration, and bounded reason codes. Never copy prompts, transcripts, audio, credentials, paths, provider responses, or device values into this document or its artifacts.
Historical Phase 3 retirement evidence (2026-08-30)¶
The retirement changes were verified sequentially before the test-suite cleanup:
| Gate | Result |
|---|---|
| Python suite | 216 passed, 1 skipped (2026-10-19), isolated staging snapshot. POSIX process-group case skipped on Windows. Log: artifacts/subagent-commit-2026-10-19/python.log; initial metadata failure retained separately. |
| Ruff | passed |
| Mypy | passed (50 source files) |
| Dependency check | pip check passed |
| Python compile check | compileall passed |
| C# Release matrix | 2,086 passed across 13 test assemblies |
| Route contract | 56 HTTP + 1 WebSocket |
| Runtime-boundary scan | 864 files; schema 1 reported reachable_count=2 (a startup-path label, not graph analysis) |
| Historical backend deterministic Harness | 48/56 HTTP passed, 8 external skipped, 1/1 WebSocket passed; cleanup green |
| Historical Agent GUI Harness | 4/4 supported Agent scenarios passed |
| Managed startup smoke | /status 200, wrong bearer 401; Worker and Broker ports listened and shut down cleanly |
| Release build | Local 11/11 stages passed (2026-10-19, Subagents GUI redesign), self-contained win-x64 GUI, Host and CLI in the flat runnable layout. Log: artifacts/subagent-readonly-2026-10-19/release-build.log. No installer or remote release published. |
git diff --check |
passed |
The retired selector and no-op live diagnostic were subsequently hard-deleted. The GUI voice runner and native microphone, speaker, OmniVoice, VTube Studio, and Gemini checks remain separate acceptance boundaries; the Agent GUI scenarios above exercise the deterministic Agent path only.
Prompt composer contract checks¶
Run these focused checks when changing prompt context, composition, Worker v2, or receipt persistence. Run them sequentially. They cover the canonical seven layers, strict context/receipt validation, root-only instruction loading and final-path escape handling, the C# wire codec, internal persistence, and public SSE cursor suppression.
# Run from %RU_ROOT%.
.venv\Scripts\python.exe -m pytest -q tests/test_python_runtime_boundary.py
dotnet test backend-dotnet\RuAgent.Harness.Core.Tests\RuAgent.Harness.Core.Tests.csproj -c Debug --nologo --filter "FullyQualifiedName~AgentPromptReceipt|FullyQualifiedName~SystemPromptContract"
dotnet test backend-dotnet\RuBackend.Tests\RuBackend.Tests.csproj -c Debug --nologo --filter "FullyQualifiedName~RuInstructionLoader|FullyQualifiedName~RuPromptContextProvider|FullyQualifiedName~WorkerV2Prompt|FullyQualifiedName~WorkerV2Contract|FullyQualifiedName~WorkerV2Client|FullyQualifiedName~AgentSseReconnect"
dotnet test backend-dotnet\RuAgent.Harness.Persistence.Tests\RuAgent.Harness.Persistence.Tests.csproj -c Debug --nologo --filter "FullyQualifiedName~HarnessPayloadStore|FullyQualifiedName~HarnessProjectionTransaction"
The retired deterministic Harness artifact and fixture Worker are no longer
generated. Contract evidence now comes from the managed .NET suites and keeps
prompt text, receipts, raw instruction contents, tool arguments, filesystem
paths, and provider secrets out of public events. model.prompt_composed remains
an internal event and must stay absent from CLI/WinUI rendering.
This fixture evidence does not prove raw Worker v1 route parity in the default managed mode. The managed composition now binds every allowlisted private route to an explicit C# handler. Status/models expose only the active typed settings (provider credentials remain host-owned), runtime-scoped VTS state/config and canonical token persistence, and catalog/action/connect use the managed adapter, voice/STT use typed providers, and observation control is explicitly disabled until its managed capture and reaction loop exists. Overlay enable is fail-closed while its process owner is missing. The generic private-route fallback is no longer reachable from the managed composition. Public provider-backed model catalogs are now Host-owned (see the 2026-09-01 catalog evidence); the private Worker model status is not their discovery source. VTS overlay process/semantic actions and observation capture/reaction still require a dedicated managed-runtime integration gate before claiming full production parity.
On 2026-08-30, a bounded smoke of the freshly published managed runtime used
isolated temporary roots and loopback ports: the authenticated Public Host
/settings and Broker /health checks returned 200, while the Agent
/worker/v2 and Device /worker/v1 listeners returned the expected 400 for
non-WebSocket probes. The owned process tree was stopped and the probe roots
were removed. This confirms managed startup/listener wiring only; it does not
replace native microphone, OmniVoice, VTube Studio, or Gemini acceptance.
Test ownership and documentation integrity¶
Python tests cover documentation and route contracts, reference fixtures and repository maintenance tooling. Managed C# suites cover the Public Host, Broker, in-process model-step service, settings migration, session deletion, privacy and cancellation. Do not keep a copied implementation in a test as evidence about production.
Documentation checks validate local links and ADR metadata, not exact wording, ADR counts, colors or historical deletion lists. GUI structural checks are not visual acceptance; keep real responsive-policy and interaction tests.
# Run from %RU_ROOT%.
.venv\Scripts\python.exe -m pytest -q tests/test_current_documentation.py
The retired Python deterministic backend/GUI Harness, its fixture Worker, and Phase 1 Python prompt-reference benchmark were removed. Current verification uses the managed .NET suites, the filesystem-only Python boundary scanner, and the release build/native overlay checks above. Live provider, audio, microphone, subtitle, and VTube Studio acceptance remain explicit user-run gates.