Verification evidence — 2026-10-09¶
Historical snapshot only. Counts, commands and limitations describe that snapshot and do not establish current acceptance. Use current verification.
SQLite business-persistence consolidation (2026-10-09)¶
Schema v18 consolidates settings, conversation documents/source ledgers and
Activity Timeline in the existing agent_harness.db. Legacy imports and markers
commit together; original files stay untouched. Session listing uses indexed
summaries rather than filesystem caches/budgets. Activity deduplication survives
restart without a read warmup. The unused file-outbox poller is removed, and
legacy activity backups are excluded from diagnostic retention. API keys remain
in Credential Manager; encrypted Harness payloads and the conversation outbox
retain their existing boundaries. See ADR 0013.
The first three new tests failed against the old file stores. Eight migration regressions now cover database-only writes, restart, import markers, clearing and deletion, settings credential stripping, concurrent appends, rollback/retry after an unreadable legacy file, source-ledger retention through pruning/rewind, and existing Harness deletion tombstones. The final focused Business run passed 167 tests, including strengthened checks of actual SQLite rows after rejected settings updates and globally ordered listing with more than 2,000 legacy files.
Sequential Debug coverage:
| Project | Passed |
|---|---|
| Broker | 898 |
| CLI | 1,278 |
| Client | 206 |
| Harness Core | 468 |
| Harness Persistence | 176 |
| Hosting | 58 |
| Protocol | 81 |
| Runtime Adapters | 195 |
| Runtime Worker | 69 |
| Session Interchange | 43 |
| Backend | 1,037 |
| GUI | 1,297 |
| Total | 5,806 |
The initial Persistence run failed 22/176 tests because historical-version fixtures retained tables introduced by the latest schema. Version-18 creation now handles existing tables/indexes, and the validator still requires their columns and indexes; the Persistence rerun and later full invocation passed. The first full invocation then stopped at Backend with 52 cleanup failures (985 passed): four provider fixtures did not release their newly used SQLite pool before deleting the temporary root on Windows. Scoped per-database cleanup fixed this; the targeted Backend rerun passed all 1,037, and the remaining GUI project passed 1,297. Provider assertions were preserved. GUI compilation still reports existing CS8625 and xUnit2031 test warnings.
Commands and retained local evidence:
.\scripts\test_dotnet.ps1 -Configuration Debug
.\scripts\test_dotnet.ps1 -Configuration Debug -ProjectFilter RuBackend.Tests
.\scripts\test_dotnet.ps1 -Configuration Debug -ProjectFilter gui-winui3
dotnet test backend-dotnet/RuBackend.Tests/RuBackend.Tests.csproj -c Debug --filter FullyQualifiedName~RuBackend.Tests.Business
.venv\Scripts\python.exe -m pytest tests -q
artifacts/sqlite-full-dotnet.log preserves the stopped initial invocation;
artifacts/sqlite-backend-rerun.log, artifacts/sqlite-gui-tests.log and
artifacts/sqlite-pytest.log preserve the successful completion evidence.
Python passed 211, with 1 skipped. Release publish, native UI interaction,
real-provider behavior and migration of the user's live data were not run.
Migration is exercised with isolated temporary roots; each live store imports
its old files on first access by the new build.
Office attachment retirement (2026-10-09)¶
Built-in DOCX/XLSX/PPTX extraction, its Open XML dependency, extraction-only fixtures/tests and the OOXML benchmark source are removed. GUI picker/paste/drop and CLI attachment/project-file intake now accept images and text/code only; unsupported formats are refused before text decoding. Workspace documents can be processed through installed terminal tools or optional skills; no Office converter or authoring suite is bundled. Earlier Office verification entries below describe the retired implementation.
The new CLI rejection checks first failed on the old implementation (seven
failures). Sequential focused Debug runs via scripts/test_dotnet.ps1 passed:
CLI 1,278, GUI 1,297, client 206, with zero failures/skips. The GUI
test build emitted warnings in the unchanged CodeHighlighterTests.cs (CS8625)
and AgentActivityUiSourceContractTests.cs (xUnit2031). Restored client, CLI and
GUI dependency graphs contain neither DocumentFormat.OpenXml nor its framework
package. Documentation/count/link guards passed 19 tests; git diff --check
passed for the affected files. The full 12-project matrix, publishing and native
picker/clipboard interaction were not run for this scoped removal.
Step 5 Preview Free coding pilot (2026-10-09)¶
The live pilot uses models.dev:kilo, stepfun/step-5-preview-free and
chat_completions, with default reasoning, a 1,000,000-token catalog context
and Ru's effective 8,192-token output cap per model step. Each custom task gets
one fresh Git workspace/session, the standard preset, explicit Workspace
Trust, a 1,800-second agent deadline and a 120-second verifier deadline. The
runner never approves tool requests or continues parked runs. This is native
Windows execution of the custom public corpus, not an official benchmark score
or an OS sandbox.
CLI and Host Debug builds passed with zero warnings/errors. Live
attempts used the fresh Debug CLI, whose automatic Host discovery selected the
existing published Release Host/Broker. The manifest records those actual
runtime binaries separately from the Debug compilation check; checkout HEAD
must not be interpreted as their build revision. All trials keep this runtime
selection. The Host reports 1.0.0+eb3a00fd5843c28bb66374337b8fe8fe46864f4e;
the backend commit between that revision and checkout 2e0ca685 only retires
Worker class/namespace names.
The first nine
attempts were invalid: the adapter cancelled waiting_history_user before
provider execution. A minimal live probe progressed automatically through that
state and completed with the requested model. The adapter now polls automatic
history persistence within the original deadline; a regression failed before
the fix, and the focused benchmark/RPC/process suite passed 30 tests after
it. Tests retain cancellation for approval, user-input, reconciliation,
unknown-state and timeout outcomes, including a stalled history write.
The full Python suite passed 211 tests with one skip; the .NET matrix and
release distribution were not rerun for this Python-only adapter correction.
Ignored local artifacts are under
artifacts/ru-step5-preview-free-20261009-run1 (excluded pilot and diagnostic)
and artifacts/ru-step5-preview-free-20261009-run2 (corrected runner). The latter
manifest records checkout/binary/runner fingerprints and run conditions;
per-task reports retain safe trajectories, fixture fingerprints, patches and
external grader results. Kilo's public gateway catalog advertised zero input
and output prices for this exact free model at preflight. Actual billing and
unreported token usage remain unknown.
All nine corrected trials finished in 60.99 minutes wall time (60.82
minutes summed agent time). Strict reward and unattended completion were both
0/9: Terminal-style 0/6, SWE-style 0/3. Four runs failed with
empty_model_reply, three stopped for tool approval and two parked after
provider_rate_limited. Every trajectory caught up to its authoritative event
cursor, but none reported token totals. The empty-reply cause was not isolated.
| Task | Agent minutes | Stop | Candidate grading |
|---|---|---|---|
cs-atomic-ledger |
4.85 | Empty model reply | 0/7 groups; no source change |
cs-single-flight-cache |
4.96 | Empty model reply | 5/8 groups; no source change |
py-dependency-scheduler |
6.10 | Empty model reply | 3/6 groups; no source change |
py-stream-parser |
8.19 | Terminal external-mutation approval | 6/6 groups, 152 checks passed |
ts-generation-pager |
5.04 | Empty model reply | 2/6 groups; no source change |
ts-transaction-store |
3.36 | Unknown executable approval | 4/7 groups; no source change |
cs-lease-renewal |
20.50 | Provider rate limit | Invalid: changed immutable project file |
py-migration-rollback |
6.71 | Terminal external-mutation approval | Invalid: extra root verification script |
ts-session-replay |
1.10 | Provider rate limit | 2/7 groups; no source change |
Only 1/9 admitted candidates passed the complete verifier; its unfinished
run still receives reward zero. Partial passing groups on unchanged starters
are not agent improvements. Two separate diagnostic grade-patch runs applied
only exported source changes to fresh SWE baselines, excluding forbidden files:
Python migration passed all four regression and two preservation groups; C#
lease passed four of five regression and both preservation groups, but still
failed expiry/takeover and rejection of old mutations. These diagnostics do not
repair the original submissions or change their rewards. The local summary.md
and analysis.json retain this distinction alongside the original report.json.
This pilot exposes provider/harness reliability, approval and instruction-scope failures under these exact conditions. It does not establish a pure algorithmic solve rate or a ranking against other coding agents.
Small C# task and explicit free-model switch (2026-10-09)¶
A separate diagnostic reduces the task to one shipping-threshold bug in a
fresh, tiny .NET 10 repository. Its existing console verifier checks below,
exactly at and above the threshold, an empty order and a negative subtotal.
Only src/Shipping.cs may change; the baseline fails the exact-threshold check.
This is a smoke task, not another scored Terminal/SWE corpus attempt.
The initial models.dev:kilo / stepfun/step-5-preview-free attempt parked
with provider_rate_limited after 56.09 seconds, before any tool calls.
At the owner's request, the atomic settings picker then explicitly selected
models.dev:openrouter / nvidia/nemotron-3.5-lightning:free and read back the
persisted selection. Both Ru's catalog and OpenRouter's public model API
listed the selected model; the latter advertised tool support and zero input
and output prices. This is an explicit configuration change, not a runtime
fallback. Default reasoning and the standard preset remain in use.
The first fresh Nemotron attempt ended with Ru status completed after
71.45 seconds, but failed task acceptance: it made three successful list
calls, never read source content, edited no file and ran no test command.
Client-visible final text contained malformed tool markup. The independent
post-run verifier still failed the exact-threshold check; immutable files and
Git status were unchanged. A completed run is not evidence of a solved task.
A second fresh attempt with the same model and prompt stopped after
160.84 seconds at waiting_approval /
terminal_external_mutation_requires_approval. It corrected an empty-path
list call after malformed_call, successfully read both source and tests,
then proposed the exact required dotnet run --project
tests/Shipping.Tests.csproj command with no network intent. The Broker required
approval before dispatch; the runner cancelled the parked run without granting
permission. No source changed, and the independent verifier still failed.
Neither Nemotron attempt encountered a provider rate limit, but unattended
task acceptance is 0/2. These two runs do not isolate the cause of the first
attempt's malformed final text or establish a model-quality comparison.
Unlike the earlier nine-task pilot, this diagnostic explicitly pins the Debug
Host and Broker through RU_AGENT_HOST_EXE and RU_AGENT_BROKER_EXE, with the
Debug CLI. The Broker Debug build passed with zero warnings/errors. Ignored
artifacts under artifacts/ru-small-csharp-20261009 retain the failing baseline,
selection evidence, binary hashes, per-attempt metadata and verifier results.
Native UI behavior, actual billing and larger-task quality remain unmeasured.
The focused documentation/count guards passed 13 tests. No product source
changed for this selection and diagnostic; the full .NET matrix was not rerun.
Welcome trim, light meteor retirement and terminal mark (2026-10-09)¶
The complete sequential Debug matrix passed all 12 projects: 5,821 passed / 0 failed / 0 skipped, including 1,326 GUI tests. The Python suite passed 204 tests with one skip.
The empty-state hero now keeps only the workspace readout and the greeting: the subtitle and the three starter prompts (with their click handler and button style) were removed by owner direction. The Light Meteor Shower theme was removed from the theme enum, the preference store, the title-bar toggle cycle, the backdrop gate and the Settings dropdown; a stored light-meteor preference folds to Light instead of resurrecting the theme. The terminal tab's mark is the >_ prompt set in the app's mono font, replacing the boxed CommandPrompt glyph that read as a badge at chip size. Contract tests pin all three decisions; no route, preset, approval, provider or backend behavior changed.
Side panel revisions and interactive terminal (2026-10-09)¶
The complete sequential Debug matrix passed all 12 projects: 5,817 passed / 0 failed / 0 skipped, including 1,322 GUI tests. The Python suite passed 204 tests with one skip.
The right side surface gained its final shape: only the Files tab opens by default; the summary card's Changes and Tasks rows (whole-row click targets, Codex-style) open those tabs; the add control opens a real interactive terminal built from ConPTY plus vendored xterm.js (MIT) in WebView2, running the user's PowerShell in the session workspace — a user-only surface the Agent can neither see nor call, killed with the app or a workspace change, with the WebView2 runtime that ships with Windows 11. Tabs drag-reorder, the strip hovers as one block and scrolls without a visible bar, tab switches and file-tree rows animate, and the collapse control draws the same panel glyph as the title-bar toggle. A Windows quirk was found and fixed on the way: a parent with redirected standard handles hands those stale values to its ConPTY child and breaks the console attach; declaring empty standard handles in the child's startup info restores it (verified in both redirected and no-console hosts). The Broker remains the only execution authority; no route, preset, approval, or provider behavior changed. Native terminal rendering, resize and Unicode acceptance remain manual.