Source-rule upgrades and SDK file integration

2026-10-06, version 0.2.1; uncommitted worktree based on c2c8f6f1c44f3dc9e398842d9dca2c93da76680c. Cwd: D:/workspace/projs/github/owent/llm-usage; no commit/push/release. Requirements: data rules, telemetry configuration and authentication. Native environments/official download hashes/independent statistics are in container sources; installation baseline is in lifecycle acceptance.

This preserves tested artifacts/results for Qwen/OpenCode/credentials/Orca. Subsequent native samples, rule corrections and final Windows/Linux rereads for eight M8 clients are in M8 container samples.

OpenCode per-record versions and old processing positions

Previously, the highest database session.version selected the whole file’s parsing basis, including effects from empty sessions. A 50,000-row limit with a 60-second overlap window could skip unprocessed old rows. Now only sessions owning step-finish usage affect file status; each event uses its owning session’s version. Only native-verified 1.18.34 is known_version; others remain latest_fallback. Cursors retain updated time, stable row ID and incomplete-window start. Only a complete valid scan finishes the rule-update marker. Unknown-version/invalid-row states persist across incremental windows; SQL type errors do not block other valid records.

Old summaries’ parse_basis was omitted from metadata normalization, so parser/basis upgrades still conflicted. Compare complete canonical/legacy summaries and allow only parsing-reference changes. Token/model/quality/ownership/revision changes retain ordinary conflict handling; diagnostic history remains and summaries/checkpoints commit transactionally.

Eleven OpenCode/ten metadata-upgrade tests and shared MiMo reading regressions passed, covering real registry discovery, unchanged old cursors/complete summaries, mixed versions, empty sessions, same-millisecond pagination, repeat reads, invalid-row recovery and checkpoint rollback. Native 1.18.34 main-loop data still lacks default title requests; differences do not create events.

Qwen 0.25.0 SDK file

Verified fixed official source 6788c035698a0ada471c958d1e789e96c6cddd9b against actual files. FileExporter writes consecutive pretty JSON objects. resource._rawAttributes is a key/value array; LLM span identity is in _spanContext, kind=0. Read only this version’s per-call qwen-code.llm_request, without adding api_response logs, HTTP spans or metrics. Object/byte/ count/time limits apply; incomplete/oversized objects retain their starting position. Invalid objects do not hide subsequent valid calls.

Native main-loop input/output: 10,226/2; automatic background memory: 5,639/556; cache read: three. API response, LLM spans and CLI independently totaled two calls/16,423 tokens. SDK supplies no cache-write value or provider reference establishing whether reasoning is included in output. Leave cache write unknown; derive total only with known reasoning=0, and keep total unknown for nonzero reasoning. A shared session does not identify parent/child calls.

Choose SDK or native data by host/user/session/local day; choose trace/span copies within the same host/user. Retain native/copy records with exclusion reasons. Other SDK versions and unknown ownership/sessions remain isolated. Retained native source/day partitions block overlapping SDK records even after detail deletion. Scope settings/exclusions/events/checkpoints/ summaries commit together; clearing removes scope identities. Partial coverage adds historical coverage notices without degrading valid-source health.

Eight requirement tests cover both import orders, repeat scans/copies, different hosts/local days, retained partitions, unknown ownership/versions, nonzero reasoning, incomplete/oversized objects, invalid objects/tokens and rollback. Read-only exporter checks and directed discovery of verified application-managed directories also passed; other supplemental output stays isolated.

An added new-directory-copy regression first exited 101: retained SDK statistics of two calls/ 16,423 became four/32,846. After detail deletion, per-call trace/span identities could no longer select copies. Now complete retained SDK source partitions are checked by host/user/local day; new output whose nonoverlap cannot be established receives qwen_sdk_partition_sealed exclusion/ coverage notices without changing retained contributions. Another test executed actual enforce_retention, reopened the DB and scanned late native/SDK copies: two calls/16,423 remain, zero details, with the persisted cleanup cutoff enforced. A test’s manual native file failed discovery layout; correcting it to actual .qwen/tmp/project/chats was a test fix. All eight tests finally exited 0.

A new deb read saved native/SDK files in offline rootless Podman: two calls, input 15,865/ output 558/cache read three/total 16,423. One native record and two SDK copies from another directory were excluded; repeat scans added no usage, Qwen/OTel health=ok. Other adapters’ format diagnostics on the initial manual-copy directory are separate from SDK parsing failures.

Source routing defects found by the actual package

The actual new package let Claude’s shared type detector claim a manual Qwen .qwen root: 12 sources/zero events/eight diagnostics. QWEN_RUNTIME_DIR instead produced one event/10,228. Allowlisted inspection of the first native user record/fixed ChatRecord source confirmed Qwen message.parts/usageMetadata/complete identity. This explicit layout/shape now directs only to Qwen; a directory name/shared type alone does not verify a format. Incorrect ownership is restored only when the entire old source has no events/daily/period/native summaries, releasing that file’s incorrect checkpoint while retaining diagnostics/enabled state/source bytes. Sources with historical contributions do not migrate. The repaired package’s manual root found one Qwen source/event and no diagnostics. Full-registry consumed-cursor/invalid-row/ retention-boundary regressions passed.

Goose first opened direct manual SDK JSON as SQLite, then failed its source with NotADatabase. It now checks the 16-byte SQLite magic first. Non-SQLite still receives unknown_format, allowing OTel to claim an explicit SDK format without duplicate adapter registration on repeat scans. Existing M8 SQLite requirement tests and targeted regressions passed.

Native Windows credential failures and correction

Reproduced credential_store_unavailable while creating credentials. Diagnostics contain only operation/numeric error code/missing-different-failed states, without credentials/targets/ authentication headers. CredWriteW succeeded, immediately followed by missing CredReadW; after 10 ms the complete content matched, still matching at 50/100 ms. Only mutation flows wait up to 50 ms on missing reads. Different content/read errors reject immediately; authentication requests still read once. Failed-creation cleanup/revocation compares complete owned content and confirms deletion. If identical content remains, Windows waits up to 50 ms; external replacement/errors/exceeded limits fail without another deletion.

After the write correction, two 100-round parallel groups passed. Earlier, a revoked credential remained across processes; another occurrence after adding deletion confirmation reproduced at round 60 and still authenticated after 10/50/100 ms. Child exit success did not establish input processing. Adding –nocapture and explicit handling acknowledgments yielded two 100-round/two-thread parallel groups passing with complete binding processing confirmed. The earlier revocation failure’s root cause remains unknown; later passes do not close it. Initial/final results remain separately under build/plan-final-push/.

Four exact test directories identified by failure timestamps contained two complete owned bindings, precisely reclaimed. Another 40 parallel create/revoke operations passed with zero owned residue; other sources’ credentials were left alone. Final checks of five exact directories, including new failures, found no pending bindings; another 40 round trips passed, zero owned residue. Official CredWriteW and CredReadW specifications do not explain this timing. The observation does not establish system-cache or thread-safety behavior.

Actual Linux screen reader

Added distribution packages Orca 48.1-1+deb13u2, Speech Dispatcher 0.12.0-5/espeak-ng, pyatspi 2.46.1-1 and xdotool to the fixed Debian GUI image; image ID 4575f4ba7437ce291ce8e1633292f232ad5899d1551c06ed20aced61f3fc212d. Ordinary UID 1000, independent D-Bus/X11, default seccomp, no network/host mounts, ALSA null audio. Debug options were checked against official Orca instructions and installed commands.

Initial actual focus events had unnamed navigation buttons; Orca spoke only button. A temporary DOM probe adding explicit names read all five items. The product therefore adds localized aria-labels to navigation/brand links, retaining names when narrow sidebars hide text. Language preparation uses/saves the actual settings form, without simulated backend IPC frontend updates. Tests with unknown language values/overbroad selectors including settings subnavigation remain recorded as test defects.

Reusable test:install:linux adds –screen-reader as an independent round after reinstalling the current package. Native Tab/Enter, actual AT-SPI button focus, five page titles and actual speech output are compared without injected names. Orca 48.1 debug.py buffers output; formal checks once failed because speech existed before the file flushed. Installed /usr/bin/orca opens this debug file with open(…, ‘w’). The test launcher changes exactly that line to buffering=1; other event/speech code stays unchanged, and changed upstream text causes rejection. Wait for actual speech and recheck AT-SPI focus. Confirm PIDs by parent/UID: Orca changes its process name, so a launch command alone cannot identify ownership. Logs retain normal-stop timeouts; final owned reader/speech cleanup uses bounded TERM/KILL/wait without claiming Orca shutdown reliability. Product assistive-technology defaults remain unchanged.

The formal current-package check exited 0: AT-SPI names and speech for Overview, Trends, Sources, Details, Settings matched actual page titles after native Tab/Enter. Brand link spoke LLM Usage. Narrow-sidebar screenshots show icons only while explicit names remain complete. Results/logs/screenshots: build/plan-final-push/; full container lifecycle: WSL build/install-lifecycle/linux/1791264897354/. This does not verify physical audibility, all controls, ten-language screen readers or Windows assistive technology.

Stage checks and remaining coverage

Final npm run verify exited 0: 927 Rust tests, core 826/app 101, eight ignored by default; 21 frontend/four script tests; Markdown 191 files, assets, Svelte zero errors/warnings, fmt/Clippy -D warnings/frontend build passed. Debian workspace: 924 tests, core 826/app 98, nine ignored by default; six explicit system-credential tests passed, keyring reclaimed. Current Windows release headless 11/receiver eight/NSIS lifecycle 12 checks passed, zero owned credential residue; NSIS output: build/install-lifecycle/windows/1791264668246/. An install prerequisite check correctly blocked a concurrent desktop test before running any checks; sequential retry after desktop exit passed, without claiming an installation defect. Final native release desktop: 17 checks/20 startups, P95 747.1 ms, exit 0; build/plan-completion/native/1791264943840/. Edge simulated-IPC language/theme/selection/filter/ pagination/narrow-layout regressions exited 0. Current Linux deb/FUSE/GTK/Orca lifecycle: nine groups/47 checks passed. Earlier scaling-factor-two/40-check results remain separate; this stage did not repeat that scale.

Final deb reread of Qwen native/SDK/direct manual copies yielded two calls/16,423, excluding one native/two SDK copies; target-source health normal, retaining one initial unrelated format-detection diagnostic. Two original OpenCode DB rereads each yielded one call/299; cache read three/zero, health=ok, parser=opencode-step-finish-parts-2, known_version; repeat scans added no usage.

Actual artifact at this stage	Bytes	SHA-256
Windows exe	9,989,120	5b3e16b2d9be5d4f25bd4886ea521a109a7d5f06ee6ddd06f2c22b23fb1e3073
Windows NSIS	3,941,612	e34aa3eebade224e865dac185a157081634a7035e2244971fda2bae2d3060076
Debian deb	5,466,678	69abe4b2463459d2969bf582d874d1e82796b14e77e5823f4ea5f38dc1adbd46
Linux AppImage	111,143,416	76b33bc5500672ba3dd6f0d115bd874fc453eac69b3283a7552da76b25137457

Both Windows/Linux builds exited 0. Old 0.2.0 package/image hashes remain in installation history. Disposable databases/scripts/raw verification artifacts/logs stay under root build/; only affected requirements are synchronized to rules/Skill.

Still pending: nonempty native samples for other products/versions, OpenCode title per-call records, Qwen cloud/main-loop cache hits, other exporter authentication/retries/sampling/ parent-child spans, new Copilot CLI/JetBrains native exports and complete outbound audits. macOS desktop/specific-hardware requirements were removed. Native macOS credentials, actual assistive technology, host login/logout and remote CI/release retain separate environment/ authorization dependencies; local builds/container exits do not establish those results.