Skip to content
AI usage dashboard

Native interactions, incremental archives and scale acceptance

2026-10-04, version 0.2.1; cwd D:/workspace/projs/github/owent/llm-usage. Windows 11 Pro x64 10.0.26300; Ryzen 9 9950X3D, 16 cores/32 threads, about 125 GiB RAM. Node 24.21.0, Rust 1.98, Tauri CLI 2.12.0, WebView2 154.0.4258.53; lockfiles identify dependencies. This preserves the earlier batch that day. Later query acceleration/resource/parallel-collection results are in current acceptance and remaining execution results; Plan.md owns active status. Temporary logs/executables/synthetic DBs stay under root build/query-next/, build/plan-finalization/ or build/plan-completion/; private source bodies are not copied.

Implemented behavior

  • Retention cleanup/clearing share background maintenance with precommit cancellation. Atomic state decides whether cancellation or commit wins; SQLite progress callbacks interrupt expensive SQL. Full rollback preserves events/revisions/retained positions/summaries without rescanning. Cancellation is rejected after commit begins; accepted cancellation disables the button until completion. Incomplete backups delete only owned files. Exhausted filename attempts fail before modifying existing snapshots.
  • System-task failure tests inject only at OS calls. Actual AppState/SQLite checks verify intent persisted before calls, queryable errors, no collection from residual triggers after disabling, and retry recovery. Manual-only scanning defaults off; enabling restricts both GUI/background default-source and quota discovery.
  • Charts use arrows/Home/End, Shift continuous selection and Enter queries with existing mouse ranges/model/Agent/cost interactions. Ten languages save through actual controls; HTML lang follows them, settings controls have accessible names and forms wrap by container width. CSS scaling does not verify OS DPI or assistive technology.
  • Daily summaries/Agent distributions/day charts merge rows, retaining result groups/active days only. Sessions/durations group by exact local-day boundaries in SQLite, then merge complete source/session identities across days without adding daily DISTINCT counts. Hour/DST/unknown-field/retained-archive coverage rules remain unchanged.
  • Schema remains 11. Opening old DBs adds attribution-exclusion/source-coverage/detail-coverage indexes without clearing data/rebuilding summaries/changing revisions. Covering indexes add disk/write costs. Final query/startup tests use indexed databases; startup results do not measure first index installation on old databases.
  • Each connection caches at most four summaries with 2 MiB retained string/vector capacity; this is not process memory. Keys include complete filters/selection/timezone/dates/grouping/ today/retention limits. The same read transaction checks revision/data_version/total_changes; same-connection writes/other-connection commits invalidate, failed/oversized queries are not cached. Two read-only connections are reused under locks, with short-lived fallback when busy; completed queries leave no read transactions.
  • When only today’s changes advance global revision, stable history is no longer materialized repeatedly. The transaction stores date/policy/completed-partition revision/row counts; completed weekly/monthly/yearly partitions select the latest of their day boundaries. Late data/deletion/missing archives/date-policy changes/invalid markers/future revisions trigger complete rebuilds. Clearing removes markers; cancellation rolls markers/statistics back. Highest source versions/unchanged row counts do not establish compatibility.
  • Collection polls state every 500 ms, returning to ten-second checks when idle. Accepted, pending tasks retain completion waiting; short collections completed between checks still refresh the UI, avoiding delayed notices after returning to idle polling.

Cache references: SQLite data_version, rusqlite 0.40.2 total_changes. A small cross-connection WAL check confirmed data_version remains with the held read snapshot; after that transaction ends, the next read sees external commits. Rust regressions separately cover same/external-connection writes without revision changes.

Engineering and native checks

Command/method Exit/results Scope
npm run verify 0; Rust 852/frontend 20/scripts three; five explicit/environment tests ignored by default 177 Markdown files, assets/types zero errors/warnings, fmt/Clippy -D warnings, requirement/integration tests and frontend build
npm run build:desktop 0; NSIS 3,810,444 bytes, 3.63 MiB Final release/embedded frontend; not installed
npm run test:headless 0; 11 checks, final three events/75 tokens Actual release/SQLite/isolated nonempty synthetic sources/persisted rules
npm run test:desktop – –runs 20 0; 13 checks, 20 small-data startups, P95 724.3 ms Actual WebView2/Tauri IPC, mouse/keyboard/offline/restart/exchange/ten languages/system minute task
npm run test:browser 0 Edge simulated IPC, five pages/selections/interactions/ten languages/themes/filters/pagination/short-collection handoff/idle polling
Explicit Windows tests 0; one COM task/one HKCU test-key round trip Test-namespace registration/query/drift repair/repeated deletion; actual permission denial unverified
ai-maintenance quick_validate.py 0; Skill is valid Static format only; description unchanged, model triggering/quality evaluation unverified

Logs in build/query-next/: verify-final.log, desktop-build-final.log, browser.log, headless-final.log, native-final.log. Headless/native results respectively: build/plan-completion/headless/1791056946337/ and native/1791056947059/result.json. Frontend JS 1,065.74 kB/gzip 357.03 kB; CSS 51.73/9.50 kB. Current-version explicit Windows native_task_roundtrip/registry_roundtrip used cargo test –offline –ignored, logged as windows-task-final.log/windows-registry-final.log. Python -X utf8 avoided system-default misreading of Chinese during Skill validation.

The 13 native regressions save every language and use CSS scales 1/1.25/1.5/2, checking no horizontal overflow, visible Save, accessible visible settings forms and unchanged statistics across languages. A current-user minute task ran without a GUI: after closing the UI, a third synthetic event was added. OS execution time/Ready/exit 0 and read-only SQLite before reopening confirmed three events/75 tokens. Restart retained results; cleanup deleted only the test task. Actual ordinary-user execution and injected failures are recorded independently.

Complete queries

Existing bench_v20 synthetic DBs were populated through actual EventPipeline/transactions/ daily-hourly aggregation in batches of 50,000. Coverage: 366 days/50 models/20 Agents/20 source instances. Sessions repeat across days; 20% of each source’s events reuse two large session keys, while different sources stay independent. Multi-Agent synthetic instances test grouping load only. Final executable reads existing DBs through actual query_summary, including metrics/sessions/durations/coverage/model-Agent distributions, asserting total calls/group counts rather than timing a standalone SUM/COUNT.

Each group clears connection caches before 20 complete queries, followed by 20 hits. Details: 100 queries of 200 rows. Complete-query P95 is sorted item 19; details P95 item 95. OS file caches were not cleared; misses refer to application summary caches. Repository-root commands, all exit 0:

Terminal window
cargo build --offline --manifest-path desktop/src-tauri/Cargo.toml -p llm-usage-core --release --example bench_v20
desktop/src-tauri/target/release/examples/bench_v20.exe build/plan-finalization/million-1 1000000 --query-only
desktop/src-tauri/target/release/examples/bench_v20.exe build/plan-finalization/ten-million 10000000 --query-only
Events Main DB/WAL MiB Miss P50/P95 ms Hit P95 ms Details P95 ms
1,000,000 906.3/0.3 926.5/964.6 0.033 1.29
10,000,000 8,034.1/0.0 4,093.8/4,177.5 0.049 1.23

Logs: build/query-next/million-final.log/ten-million-final.log. First index installation on owned synthetic DBs made Storage::open take 904.4/15,245.0 ms, including opening checks; this is not native startup. Complete empty-DB import throughput/memory was not remeasured with current indexes, so earlier-index import figures are not reused. Misses exceed the 200 ms target; cache hits do not satisfy it.

Native startup and resources

node desktop/tests/native-scale.mjs --runs 20 --idle-seconds 600 --ui-cancel exited 0 with the final release. Result: build/plan-finalization/native-scale/1791056317247/result.json; log: build/query-next/native-scale-final.log. Before startup, read-only checks resolve the actual path beneath root build/ and require one million events/20 synthetic-bench sources; maintenance IPC cannot target real Agent DBs. –help launches no GUI. Help/invalid-directory rejection/one valid native-cancellation run passed in native-scale-guard.log; these do not replace 20 independent startups/the complete ten-minute resource measurement.

Item Measurement
Million-event startup 20 runs, P95 1,223.1 ms; process launch until overview values/history chart render, including CDP waits
Cancellation One actual IPC cleanup/clear each plus settings cancellation button; events/revisions unchanged, no rescan
Idle 601.4 continuous seconds/118 samples; root process alive, static page, automatic collection paused
Process scope Root/all descendant WebViews, at most seven processes, about five-second samples; includes last visible cumulative CPU for exited processes
Private bytes Mean 333.6 MiB/peak 483.9 MiB; 180 MiB target unmet
Working set Peak 524.4 MiB, independently reported from private bytes
Idle CPU 0.309% of one logical core, below this machine’s 1% target; not divided by 32 threads

Startup uses the indexed million-event DB with mixed first/later OS caches. Idle follows two IPC and one UI cancellation on the same overview/CDP-enabled application; it does not represent fresh startup/tray/headless states. This development machine is not the proposed four-core/ 16 GiB reference hardware. Five-second samples do not measure instantaneous allocator peaks. Observed connected-WebView requests had no external HTTP/script exceptions; this is not a whole-process network audit.

Incremental refresh on the million-event DB

node desktop/tests/native-incremental.mjs --runs 20 --enforce-budget exited 0. Result: build/query-next/native-incremental/1791056283753/result.json; log incremental-final.log. Node SQLite online backup, minimum Node 22.16, copies the read-only million-event DB to a fresh isolated directory, first checking all sources are synthetic/bench. Baseline DB/native Agent files stay unchanged; actual Node version was 24.21.0.

Discover/consume one synthetic Codex record, then append 1,000 records in each of 20 rounds. Timing starts at actual collection-button click and ends with today’s visible updated card, including source reads/parsing/commit/daily aggregates/tiered retention/polling/native IPC/UI updates. Each round also verifies exactly +1,000 calls/+15,000 tokens, advancing revision and DB event counts; final repeat scans add no usage. Synthetic versions do not verify native new client versions.

  • P95 1,386.7 ms, maximum 1,406.3 ms; passed the two-second limit, final 1,020,001 events.
  • Visible cards used actual IPC/query results, without injected replacements. Observed requests had no external HTTP/script exceptions.
  • Stable-archive skips separately regress late data/deletion/missing archives/date-policy/ abnormal revisions. An isolated complete materialization took 1,770.7 ms; two checks with stable history/advancing global revision took 147.5/147.7 ms and materialized zero rows. These do not replace complete incremental-refresh timing.
  • ingest_runs start/end use the task’s common time; their difference cannot measure separate parsing/aggregation durations.

Final documentation and remaining conditions

Final checks exited 0: 177 Markdown files/zero issues; 95 local references/anchors in 12 changed documents; Skill static format/git diff –check. Read-only cleanup found zero test tasks/application processes/test startup entries; temporary files stay outside Git status. Outputs in build/query-next/: docs-final.log, links-final.json, skill-final.log, diff-final.log, residue-final.json. Engineering tests do not replace reference checks; simulated IPC does not verify native behavior.

Miss queries/whole-process memory still miss targets. Proposed hardware, full-import peaks with current indexes, wake counts, OS DPI/assistive technology, macOS/Linux native GUI and ongoing CI remained unverified at this stage. Installation was skipped under earlier instructions; upgrade/uninstall/logout/rollback unverified. Default Gemini/Qwen local paths were absent; Agents were not launched to manufacture samples. Other source-version/telemetry coverage, tray/power/file watching/application time limits/retries remain in Plan.md. Passing this stage does not complete every planned item.