Implementation and revised scope, 2026-10-07
Version 0.2.1; repository-root cwd, PowerShell 7/Node 24.21.0/Rust 1.98, WSL Debian/ rootless Podman 5.4.2. No commit/push/publication; historical CI does not verify this tree. User authorized specified-provider native requests, Zed setup, detail Merge, usage/cost alerts and F1 investigation, and requested removal of checks unavailable in this environment. Removed work is not marked passed.
Implementation and native sources
New detail exchange and usage/cost alerts include transactions, native IPC and ten languages. Details preserve complete normalized events, quality, source revisions/conflicts, cumulative records and archives. Imported sources disabled by default. Retention minimums, complete source-days, retained partitions and cross-host conflict rules remain enforced. Enabled cost estimates recalculate affected days in the same transaction. Alerts use the local user, independent timezone, day/month periods, tokens or occurrence-time estimates in one currency; off by default, exact micro-units, unknown-coverage notices and persistent deduplication. They do not represent bills, combine currencies or stop agents.
yq extracted specified-provider credentials inside the target process, without printing or saving them in normal configuration/logs. open.bigmodel.cn /api/coding/paas/v4 is a Coding Plan endpoint, insufficient to verify actual pay-as-you-go charges. Direct Windows probes of two models returned HTTP 200, each input 22/output 3/total 25 tokens. Isolated Podman probe of the same endpoint returned HTTP 200, input 16/output 3/total 19. These verify connectivity/usage responses, without independently verifying agent integration.
Zed 1.22.0, revision 76659a55a8c10ed355a070f8764a0b1733e3c115. Added llm-usage-zhipu provider without a key to existing settings.json, preserving other JSONC settings. Tests used isolated user-data/project, constant OK prompt and disabled tools.
| Native thread | Turns | Uncached input | Cache read | Output |
|---|---|---|---|---|
| glm-5.3-flash | 2 | 10,874 | 10,816 | 58 |
| glm-5.3 | 1 | 10,822 | Unreported | 3 |
Actual threads.db uses zstd/DbThread 0.3.0. The fixed ChatCompletionUsage mapper subtracts cache read/write from prompt_tokens. Thread cumulative records add positive request deltas; request fields retain the last request of each user turn. Cumulative values do not become individual events, and the last thread model does not receive all historical usage. Fixed source: usage mapper, API key, Windows credentials. Combined uncached input 21,696/cache read 10,816/output 61; cache write/full total unknown. Adapter coverage expands only to this verified provider/format; ACP/other providers/hosted interfaces remain separate.
Stored the key in Windows credentials using fixed-source target name, Bearer username, UTF-8 blob and persistence type; complete readback matches. A separate owned Zed process without runtime key environment variables successfully requested glm-5.3-flash: native cumulative uncached input 10,819/output 48. This thread is independent of the three turns above and excluded from their expected readback. User provider configuration/system credential retained; all owned test processes closed. Existing JSONC keys unchanged; public changes and normal Zed configuration contain no actual key.
First packaged-app full-registry readback found two empty MiMo/DSH instances, though Zed alone owned the physical file. Fix routes manual files only on complete threads.db table/ column checks, without hiding malformed files by suffix. Old-instance recovery uses each adapter’s real discover path, including MiMo parent-directory promotion. Hide only incorrect registrations lacking history; preserve other errors/diagnostics. Ten source-routing regressions pass.
First Zed capabilities lacked required fields and were rejected; completed fields then passed. Existing shared process did not inherit new key; used an owned isolated process. GUI does not support CLI –new. Chinese input method affected automated command input; after correction, obtained these samples. Failed attempts are not source acceptance.
Removed scope and remaining conditions
| Scope | Checked results and action |
|---|---|
| Command Code | Official 1.74.3 container first mounted an npm package missing dependencies; switched to full install. Official providers.json environment references, probe/glm models and –local-only retries still exit 3 requiring cmd login. No product account; remove nonempty-sample work |
| Claude Code | Original official install/request attempt timed out within limits and is retained. Later user-authorized npmmirror download matched official integrity. Debian/Podman 2.1.197 completed native Zhipu two-model main loop and saved-data reads; work restored/completed, record. Other versions/protocol limits retained |
| Gemini/Antigravity CLI | Only OpenAI/Anthropic-compatible provider credentials supplied/verified, without Gemini protocol credentials or product login. Independent SDK does not verify IDE protobuf; remove current product-interface samples |
| Amp/Qoder/Kiro | Official routes need account-managed Model Routing, Qoder login/PAT and Kiro authentication respectively; credentials absent. BYOK does not remove product-login requirements; remove nonempty samples |
| Droid/Grok Build | No usable Droid Airgap distribution. Grok official entry still unreadable this round; no substitution with third-party namesakes. Remove samples without claiming product unsupported |
| F1 IDEs | Current-user/machine installation registry and JetBrains/TRAE/Cursor/Windsurf candidate paths contain no testable installation. Other F1 local token files unverified; remote account reports still excluded. Remove from this round |
| Other existing-source versions/protocols, archives, migration, auxiliary/failure/cancellation/more loops | Corresponding nonempty records/distributions/reliable ownership absent. Existing samples verify only tested scope; remove expansion from this round, retain matrix limits |
| New exporters/CLI/JetBrains OTel/trace SQLite | Account/install/individual usage records absent. Local VS Code candidate session-store.db is not treated as a trace database. Remove expansion; synthetic receiver checks cannot verify actual exporters |
| Cause of earlier Windows revoke anomaly | First failure lacks data to reconstruct old timing. Fixes/200 passing parallel rounds retained without inferring old cause. Remove unreproducible investigation; cause still unknown |
| Whole-process outbound capture | Nonadministrator, without complete process-event capture facilities. WebView requests/TCP polling cannot verify short connections, DNS or all children. Remove complete capture, retain observable scope |
| More DPI/full screen-reader/host login-logout/OS wake/minimal-WebView diagnostics | User cancelled first four; minimal page lacks independent facilities, resource limits already measured. Remove active acceptance work |
| Main branch/signing-notarization/Release | User reported complete; removed from outstanding work. This round does not republish/reverify external releases; historical results retain their dates |
Official references checked for this round: Command BYOK, Zhipu Claude, Amp routing, Qoder authentication, Kiro authentication, Antigravity, Droid Airgap. Existing container sources/limits are in historical M8.
Validation record
Temporary scripts/logs/permitted-field extraction/check databases stay in root build/plan-20261007/ or native scripts’ owned build/plan-completion/, excluded from version control. New tests first had source-key, known/unknown expectations and string syntax errors; after fixes, detail five/alert five/Zed one tests pass. First full check failed export-error string type inference; removing unnecessary conversion fixed it. Packaged-app readback initially assumed a new database already had app_settings; inserted complete minimal settings before continuing, then found/fixed empty source instances above. Later passes do not replace first failures.
| Command/check | Result/log |
|---|---|
| npm run verify | Exit 0; Rust 1,018 passed, 8 native-platform tests ignored; frontend 22/scripts 4, clean types; verify-complete.log |
| cargo test -p llm-usage-core –test budgets | Final five pass including rollback preserving alert state and daily/monthly identity; budgets-final.log; repeated runs not added as new tests |
| cargo test -p llm-usage-core –test source_routing | Ten pass: full registry/Zed old-instance recovery and malformed manual DB diagnostics; zed-routing.log |
| npm run test:browser | Edge simulated IPC passes, ten languages, alert enable/threshold/coverage and deduplication after disable; browser.log |
| npm run test:desktop – –runs 3 | Final executable’s actual WebView2/IPC: 19 pass; three first-screen samples P95 850.42 ms; native-complete.log, build/plan-completion/native/1791346995836. Earlier two rounds 1,103.38/812.27 ms retained separately |
| npm run test:headless | New DB/actual executable/isolated synthetic sources: 11 pass, three events/75 tokens; headless-final.log |
| npm –prefix desktop run tauri – build –no-bundle | Exit 0, Windows release executable; build-complete.log, no new package/publication |
| node build/plan-20261007/zed-readback.mjs | Final executable new/old DB each seven pass. Two thread aggregates match table; zero individual events, physical file assigned only to Zed, source bytes unchanged, repeat revisions stable; two old empty sources recovered. zed-readback-complete.log/zed-readback-recovery.log |
| npm run lint:md / git diff –check | Exit 0; 206 Markdown files, no whitespace errors; lint-complete.log |
| Configuration/key/temp-file inspection | Existing Zed JSONC keys unchanged; actual key absent from normal settings/versioned changes. final-audit.json; git status has no task temp files; owned test processes exited |
Unified verification includes actual business tests/builds; document checks recorded separately. No new remote CI, cross-platform package installation or signing results this round; historical results retain their tested code scope.