Verification evidence — 2026-10-04¶
Historical snapshot only. Counts, commands and limitations describe that snapshot and do not establish current acceptance. Use current verification.
GUI chat create-response compatibility (2026-10-04)¶
The actual shared-client flow against an isolated HTTP Host reproduced
unknown_field: Unexpected field provider_backend after run creation. The Host
uses the run-detail projection for create responses; the create parser had not
accepted its two nullable provider fields. A regression now exercises fresh
session creation, preset selection without a snapshot token, AgentApiClient
run creation and run detail. Create/detail share type and length validation for
provider metadata, and unknown fields remain rejected.
GUI tests also reproduced acceptance being dropped when a fast stream completed before queued UI dispatch. Run ownership now survives normal stream completion; teardown invalidates stale callbacks. Composer errors remain visible without a run, and first-turn acceptance uses the newly created session ID.
After compiling the changed projects, the final sequential Debug matrix with
-NoBuild -NoRestore passed 5,326 tests in all 12 projects, including
1,175 GUI tests. Native GUI Enter handling and real-provider quality were
not driven; these checks cover the actual HTTP client and controller lifecycle.
build.bat then passed all 12/12 stages and refreshed the self-contained
flat Release distribution, including the corrected shared client and GUI.
Agent hardening and CLI workflows (2026-10-04)¶
The final sequential Debug matrix passed all 12 projects: 5,319 passed / 0 failed / 0 skipped. The original failing gates and later integration failures were retained and diagnosed before this complete green run. Workspace Trust migration tests now reconstruct the older schema without leaving later provider/table/column artifacts; their trust assertions remain strict. GUI tests still emit the existing CS8625 and xUnit2031 warnings in test sources.
build.bat exited 0 at stage 12/12. All four executables are present in
gui-winui3/bin/Release/net10.0-windows10.0.26100.0/win-x64/publish/,
with the CLI also published under
backend-dotnet/RuAgent.Cli/bin/Release/net10.0-windows/win-x64/publish/.
The SDK typecheck and its 20 tests passed.
After retiring the completed plans and updating current docs, the Python guards passed 171 tests with one skip. The route checker confirmed 41 HTTP + 0 WebSocket, with every mapped contract route declared. All 12 HelpDocsTests passed against the updated documentation, including inventory and excerpt caps.
The capability and acceptance map records the delivered Harness/TUI behavior, operating caps and remaining projects. Automated coverage includes actual Host/Harness/Broker coding-loop scenarios with a scripted provider; it does not measure real-model quality or native terminal key/resize/clipboard behavior. No native terminal/desktop UI or real provider was driven during this delivery.
Model working environment baseline (2026-10-04)¶
scripts/evaluate_agent_workflows.ps1 -ValidateOnly validated all twelve scenario
oracles against workspace SHA-256
d5de862ab22f1e8f9b5960f40a8a1e8d143f4de34139c4392dde33f10a92560e.
-RunScripted -ScenarioId recover_edit_no_match built the CLI and Public Host,
started an isolated loopback Host/Broker with a temporary data root and workspace,
and invoked RuAgent.Cli agent run --preset standard --workspace <temporary-workspace>
--json --trust-workspace <fixture-input>. The Broker returned edit_no_match for
the first edit; the scripted corrective edit returned ok, changed only seed.txt
to fixed, and the run completed. The run used 3 predetermined scripted model
steps, 2 edit calls, 1 failed tool call and 0 approval requests. This does not show
that a model read or followed the recovery note. Tokens, cost, provider/model
quality, and live-provider results are unknown/not run. The other catalog entries
retain their explicit not_run or unmet statuses; this is wiring evidence, not
a task-quality score.
The focused Python runner/fixture tests and documentation/route guards passed 26 tests. Baseline headless Harness + model conversation tests passed 21 tests. The route contract was not changed. After Tasks B–F add supported workflows, extend the runner's explicit scenario dispatch and add a concrete offline oracle for each newly supported case; do not infer passes from unrelated tests.