Skip to content

Verification evidence — 2026-10-19

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

Subagent live GUI regression checks (2026-10-19, scoped)

Native WinUI Automation used the signed-in Kilo provider and nvidia/nemotron-3.5-super-vl:free with high reasoning. The original gold-price prompt completed in root run_7a6bd61bc4ca217253490fec43cec33d; its first child attempt completed and published a readable report. The parent's follow-up attempt was cancelled on parent completion, as designed. Both attempts retained the entire parent provider/model/protocol/options selection. The parent successfully repeated wait timeouts without exhausting repetition-repair notes.

Earlier real runs exposed repeated-wait blocking and final-answer suppression when one fetched page was empty despite other readable pages. Scoped guard fixes retain exact receipt bindings, fail-closed denial/unknown-result handling and existing family budgets. Core 535, Backend 1,229 and GUI 1,514 tests passed (3,278 total). The partial-web exception has nine behavioral cases; it does not establish that generated financial figures are accurate. A prior live run also continued after audit-only recovery without the reported inbox runtime crash.

Native checks covered session switching/reopening, completed result version selection, read-only details and reader resizing. Numeric layout diagnostics and an unchanged-source comparison confirmed that Markdown already reflows; speculative layout changes and temporary diagnostic hooks were removed. Retained failure logs, run metadata, pinned-selection comparisons and screenshots are under artifacts/subagent-live-gui-2026-10-19/. These are scoped evidence, not a claim that every provider or every GUI action has been tested.

At the end of this check (2026-10-20 local time), the main Stop button cancelled run_6a9eea932a768af05cba5949bbf10c60 while its first provider step was pending. The run became terminal Cancelled, and the composer returned to Send. No child had been created after more than two minutes, so this is evidence of cancellation during provider wait, not a live Stop-of-active-child case. Child cancellation on parent completion was observed separately in the successful gold-price run above; family Stop coverage also remains in the deterministic Host suite. Provider latency and all provider/model combinations were not eliminated by these checks.

Git delivery validation (2026-10-19)

The commit snapshot includes subagents, the read-only clients/GUI and the primary session-deletion outcome handling required by the child cleanup retry tests. Independent directory-handle hardening and watcher callback repairs remain outside this snapshot, untouched in the user's working tree. That scope explains the Protocol/Backend count difference from earlier working-tree evidence.

A detached checkout received the staged patch for verification. Final project runs cover all 13 .NET projects (6,874 passed); Python is 216 passed / 1 skipped, and SDK is 61 passed plus typecheck/build. The initial CLI run had three existing TgCopyParityTests return empty buffers; isolated cases passed with both pipes and PTY, and the entire CLI rerun passed unchanged. Timing/order/global terminal state remains a suspect rather than a proven cause. Logs are retained. Isolated Host compilation also exposed that cleanup tests require the full primary-deletion outcome dependency; it was included without weakening cleanup assertions. The initial Python failure identified a malformed plan status marker; metadata was corrected before retiring completed plans into Git history.

Evidence: artifacts/subagent-commit-2026-10-19/ (manifest, retained failures, per-project logs, staged-tree identity and SDK/Python results). Prior Release and native UI acceptance above belongs to the earlier complete working tree; Release and live-provider generation were not repeated in the isolated commit checkout. Completed design/implementation records are retained in the feature commit's history; current operation and authority are documented in the ADR and operations.

Subagents GUI redesign (2026-10-19, scoped)

The session card now opens a dedicated Subagents panel with stable branch glyphs, grouped states and a version-bound Markdown reader. Narrow windows use an overlay. Users retain observation/export only; no model/provider/authority contract changed.

  • GUI project suite: 1,514 passed, 0 failed / 0 skipped via scripts/test_dotnet.ps1 -Configuration Debug -ProjectFilter gui-winui3.
  • Release: 11/11 stages passed, exact flat-package smoke PASS.
  • Route/guidance guards: 22 passed. Diff whitespace guard passed.
  • Native Release inspection: card/list/detail, wide/narrow layouts, Back/Escape, historical version switching, Stop/root retention, Markdown export and read-only controls passed using existing isolated fixtures. No new provider generation. Eight GUI preference files restored and owned processes closed.
  • Luna review's unknown/waiting-state finding was fixed with coverage. Initial compile/menu/spacing corrections and scoped limitations are recorded in the design/delivery record.

Evidence: artifacts/subagent-gui-redesign-2026-10-19/ (gui-final.log, release-build.log, release-smoke.log, docs-guards.log, native JSON/tree/PNG files). This does not replace the preceding full-matrix count or claim live-provider acceptance.

Detail refinement after user feedback: fresh GUI 1,514 passed, Release 11/11 and exact-package smoke PASS in artifacts/subagent-detail-polish-2026-10-19/. Native inspection verifies the compact header, task details/export disclosure, conditional version picker and expanded/narrow reader. An expanded-mode composer layer defect was found and fixed; final screenshot/pixel checks confirm it. UIA FromPoint was inconclusive because it returned the native input sink; the unsupported probe is retained and is not counted as passing evidence. Six saved interaction checks and the final read-only probe passed. Preferences restored, owned GUI/Host closed; no new provider generation. See that folder's native notes.

Read-only subagent revision (2026-10-19)

The current user contract removes role-model settings and public child mutations. New children inherit exact parent pins unless the user explicitly requests another model in the main prompt. Only signed-in provider catalogs can supply that model; absence/ambiguity/catalog failure produces a named refusal before child allocation. The parent interprets explicit user intent; a prompt mention alone is not consent.

Deterministic RED/GREEN covers version components (4.5 cannot become 45), untrusted provider IDs, signed-in catalog availability, original prompt evidence, legacy spawn hash compatibility, generic child run write rejection, retired user queues and preserved uncertain execution. Public route count is 45, with only two subagent GET endpoints. SDK/RPC/CLI and the native tree expose viewing only.

Native UI Automation inspected the freshly built Release app with isolated data: role settings absent, child mutation controls absent, result ValuePattern read-only, and completed history still viewable. Screenshots and assertions are in artifacts/subagent-readonly-2026-10-19/. Eight GUI preference files were restored; owned GUI/Host processes closed. No new provider generation was attempted. The Settings view reported a provider connection failure, so deterministic alternate model selection is not presented as live provider acceptance. The attempted Luna review never started (pending_init) and was cancelled; no independent review outcome is claimed.

Historical pre-read-only subagent delivery (2026-10-19)

The five tools, five root-owned HTTP controls, RPC/SDK, CLI tree and WinUI tree/role settings are implemented; usage and result/application status remain separate. See operations and the implementation plan.

  • The full 13-project matrix passed 6,861 tests. A later description-only clarification of agent_wait passed Protocol 151, Core 518 and 48 focused Host tests, including real native adapter fixtures, prompt/digest projection, both-role controls and four parked parents. Schemas, versions and execution policy did not change.
  • Fresh Luna review found a Git metadata output-cap mismatch and a WinUI disposal ordering defect. Both were reproduced, fixed and rechecked. Ordinary Git retains 256 KiB; internal snapshot listings use a bounded 8 MiB. Earlier failures remain in the delivery evidence directory.
  • The final live fixture pair used models.dev:kilo, nvidia/nemotron-3.5-super-vl:free, chat_completions: single 23.84 s; delegated 55.80 s. Both passed their oracle unattended with unchanged checks/HEAD. Both children completed, exactly one coding file was applied, and all persisted provider/model/protocol pins matched. Log: delivery-live-guidance.log under artifacts/subagent-implementation-2026-10-17/.
  • An earlier live root finished before review completed, correctly cancelling it. The tool guidance now explains that a wait timeout is not completion. QA also corrected an alias/backend comparison and now requires completed roles, not merely spawned ones. Earlier setup failures (private-data ACL on D:) and QA reports are retained; they are not counted as provider failures or strict passes.
  • One small fixture pair is functional evidence only. Delegation was slower here; reported usage is incomplete, so neither cost savings nor quality/speed gains are established. Native interaction was initially deferred; the later authorized WinUI acceptance below records the direct GUI checks and corrections.

Historical pre-read-only native subagent GUI acceptance (2026-10-19)

The user explicitly authorized GUI control. Windows UI Automation drove the actual Release app with isolated QA business data and a real configured provider. Evidence: artifacts/subagent-gui-qa-2026-10-19/ (screenshots, UI trees, native assertions and logs).

  • Reopened completed conversations, switched sessions and inspected both role rows, actual models, immutable result versions and applied/unapplied state. View and native FileSavePicker export matched the exact immutable result text (525 characters).
  • Saved a Coding-only role override through Settings; Review inheritance and the parent provider/model stayed unchanged. Reopened Settings and restored inheritance.
  • Sent an actual run through the chat composer, waited for coding/review completion, sent a review follow-up and stopped only that child. The parent and coding sibling remained intact. Queued Apply while the parent awaited input, verified no parent write, then submitted its answer: the action applied exactly one file and completed. The external oracle passed; the check-file hash and parent HEAD were unchanged.
  • Direct testing exposed a missing tree on reopened history and after Stop released the controller. Read views now reuse and retain the existing session-owned root selector; session reset clears it. Follow-ups also retained accepted text: they now clear only after success, reset on child/session changes, and disable on ended roots. Both defects have before/after native evidence. Root Stop now keeps its view/export tree; switching back restores the correct two-child applied tree.
  • Final GUI project: 1,504 passed / 0 failed (gui-stop-retention-tests.log). Release rebuilt 11/11 and smoke passed again. Wide and 980x700 native windows were inspected; this is targeted subagent acceptance, not a complete theme/DPI or screen-reader certification.
  • The owned GUI/backend were closed. All eight backed-up GUI preference files were restored byte-for-byte or removed if originally absent. Real business settings, history and credentials were not modified. Nothing was committed or pushed.