- Upstream: https://github.com/coulsontl/ai-toolbox
- Initial upstream baseline commit:
5e18c4a80130e6704fe38c9a4f134f6cc039c1c5 - Linear baseline commit:
f5a739568cee811bed11ab745bb5a4d16917e9e2 upstream-mainpreserves the original upstream commit graph and authorship.mainkeeps a functional commit sequence without merge commits. The initial linear reconstruction retains original author/committer identities, dates and messages with anOriginal-Committrailer. Later upstream synchronization uses cherry-picks: source authors and author dates are retained with a(cherry picked from commit ...)reference, while committer metadata records the new integration. Rewritten commit signatures are not represented as original signatures; inspect the upstream objects for original signatures.- Commit mapping records source commits, rewritten commits and integration records. Redundant merges map to the existing linear tip. Mainline integration commits retain the original upstream integration tree. Internal merge integrations apply a parent-relative difference conservatively: the unique original parent already ancestral to the most recent mainline boundary is preferred, otherwise the first parent is used. They are not claimed to be pure conflict-resolution changes.
- The linear baseline tree is identical to the recorded upstream tree. Every original first-parent boundary was checked for tree equality.
- Existing licenses, copyright notices and third-party notices remain authoritative. No upstream author is replaced or claimed as the independent maintainer.
- Historical contribution forks and experimental branches are not automatically incorporated.
Original workflow files are retained as documentation under docs/upstream-workflows/ with .txt suffixes; GitHub does not load them as workflows. Repository Actions use reviewed validation workflows with read-only repository permissions. Structural history and JSON checks run on main changes. Full frontend and Rust validation can be dispatched manually on main; its initial-validation push trigger is restricted to validation/cloud-checks. The two heavy jobs run sequentially and do not publish or deploy.
No upstream tags, releases, secrets, signing keys, R2 storage configuration or publishing credentials are copied. The initial setup did not publish independent binaries. Product identifiers, updater endpoints and release provenance require separate review before building or distributing independent packages; the 2026-10-26 preparation is recorded below. Existing upstream development and test instructions still apply.
The initial full cloud run on f660c307122e86c769c5d024d4c57802f3e4b2f4 passed: validation run. It used Node.js 22.23.3, pnpm 9.15.0 and Rust 1.99.1 on Ubuntu 22.04.
- TypeScript check passed.
- Frontend: 966 tests passed, zero failed/skipped; 307 separate browser-fixture checks and the structural layout guards passed.
- Rust: 3,118 tests passed, zero failed; 13 upstream tests remain ignored because they require real credentials, local databases/session files, installed CLIs or WSL environments. This count includes 11 passing doctests.
- These results do not verify native desktop GUI behavior, platform-specific packaging/signing, live credentials or independent update channels. See the validation backlog.
- Latest synchronized source:
019372a8e64215fb7089a8ed28206da9f934c43c. - Previous mirrored source:
74d0ee923de0cc15b5be534e7f044b762de2a646. - Candidate starts from existing maintenance
mainat4cf5ceda436b3c86ac506a71dd83cff307268a37; existing history is preserved and only single-parent commits are appended. - The 28 source commits comprise 23 separately replayed functional/documentation commits, two patch-equivalent WebSocket commits already present in this repository, and three merges. The merges are audited rather than inserted into development history. Source authors and source commit references are retained.
- PR #428's test/fix patches are already represented by
3a753736/153fc0e7; stable patch IDs match officialbff3a99b/3efe8494respectively. - Merge
4adf486badds no independent integration change. Merge4dd72321combines service-tier and pricing/collector documentation; its exact resolved module-documentation blob is used during140a6601replay. Merge46a3f990reconciles heatmap/range naming, bilingual pricing text and formatter tests; its exact four resolved blobs are used during98567a97replay. The intervening source commits do not change those four paths. - At the
46a3f990integration boundary and final source tip, all upstream application and test files match the official tree exactly. Remaining differences are only the existing Next README prefix, maintenance/history records, validation workflows, and the inert copy of the original release workflow. - Synchronization evidence records every source commit and its candidate mapping. The initial baseline evidence above remains historical evidence, not a claim that this new candidate passed those old test runs.
- This record describes a local review candidate; it does not assert that remote
mainorupstream-mainhas moved. Publishing requires a fresh remote-tip check and review/validation of the final candidate. No binary release or deployment is part of this synchronization.
The later synchronized source is f3eae1e3fdbe7b2d1f53c74c8113c20df794e009, one ordinary upstream commit after 019372a8. It is replayed as 28e1a61777576f3d50b0261b863ff8cf97ac74ce on maintenance base fcbab9f1f8b270890610c8c6f080d2edb1d13a08. All nine DSH files match the source commit exactly. The independently integrated image-history race fix, Next attribution, license/NOTICE files and release-workflow isolation are retained.
Targeted Rust validation ran on exact candidate 799d39f1665253914e16f3a40be239ad95a399e0, Ubuntu 22.04 with rustc/cargo 1.99.1. All six added session regressions were verified in the test listing before execution. Session parsing: 21 passed, 0 failed, 2 ignored; DSH configuration/catalog: 36 passed, 0 failed, 0 ignored. real_artifact_smoke and real_sessions_root_smoke remain ignored. The existing full-frontend and full-Rust jobs were deliberately skipped; this is not full-suite validation. Local focused Node checks passed 26/26 and TypeScript passed. Browser/native checks, real DSH artifacts, and a direct new input/compat form-save roundtrip remain unverified.
The dedicated CI job is manual and restricted to maintenance/sync-dsh-021-20261018; it shares the existing serial concurrency group, uses read-only repository permissions, and neither uploads artifacts nor publishes/deploys. The evidence record is dsh-sync-20261018.json.
Open boundary for follow-up: the session-list parser detects turn/start only within its first 40 lines. A first turn later than that could be mistaken for a blank session. This is a static boundary concern, not a reproduced regression in actual DSH 0.2.1 sessions. Verify a valid delayed-first-turn fixture and intended list semantics before a separate behavioral fix; the ignored smoke test repeats the same bounded check and cannot establish this edge case.
Source fd007e1a7fc54e8442d8ef9dd086d06c8c7aa0be adds two ordinary commits after f3eae1e3: OpenCode 2.0.32 compatibility and moving the tray usage line below Open Main Window. They are replayed separately as 289e17bd and 85bb777b, preserving upstream authors and author dates on maintenance base ab8d0065. The source version fields advance to 1.2.0; this is a source synchronization, not a binary release. Existing Next attribution, AGPL/NOTICE, image/MCP/Skills corrections, and release-workflow isolation are retained.
Independent review prompted an actual synthetic SQLite/file regression: malformed legacy auth.json prevented valid V2 database credentials from being read. RED run on 3780bfb9 compiled successfully and passed 29 migration tests; free_models passed nine and failed the new malformed-JSON test. Later modules did not run after that failure. The independent fix, tested as 4a5c23a7 and consolidated with its same-feature regressions as d309e32d, uses nonempty database entries when legacy JSON cannot be read, retains the legacy error without database entries, and preserves the documented database-priority union for valid JSON. It does not infer CLI generation from a migration marker or change that priority contract. Tests use temporary synthetic files and SQLite, not user credentials.
GREEN run on exact 4a5c23a7904b1b0c22140e2c6dfc85d250b12323 passed 193 targeted Rust tests, zero failed/ignored: migration 29 (all names first checked with --list), free_models 12, commands 12, markdown_agents 6, tray_support 4, MCP configuration 40, runtime locations 19, backup utilities 61, tray tests 10. Ubuntu 22.04 used rustc/cargo 1.99.1. The existing full-Rust, frontend and DSH jobs were deliberately skipped. The manual job is restricted to the synchronization branch, read-only, serial, time-bounded and does not publish or upload artifacts.
Local frontend checks passed 93 focused OpenCode tests, all 1,055 Node tests, TypeScript and three structural guards on the unchanged frontend tree. These results do not verify native GUI, real credentials, live config/database migration, WSL or packaging. The default config.json rename followed by migration parse failure remains an untested boundary, not a reproduced defect. Earlier backup GUI confirmation/password/cancel checks and DSH live-session boundaries remain open. Evidence and mappings distinguish exact tested commits from historical runs; integration requires a fresh remote-tip check and final independent review.
A synthetic JSONL → actual session importer → temporary SQLite regression proved that Codex token_usage_record.response_id was retained in the session row ID but dropped from its matching envelope. Two otherwise identical proxy calls therefore produced an extra session row; a different known response could also be incorrectly suppressed by the unique-usage fallback. Both authoritative-first and legacy-mirror-first paths now retain that identity. Codex parser revision 7 revisits unchanged revision-6 files and rechecks saved matches without relaxing the existing matching rules.
The original RED run failed all three targeted counting assertions. The final GREEN run, on 761643d22eef42d476b6f7533da7c68b0639d62d, passed 119 session-import tests, failed zero, and left seven existing local-data audits ignored. All six new test names were checked before execution. The same-feature functional and CI commits have exactly the tested tree; this documentation is an additional evidence-only change. An intermediate YAML dispatch rejection and incorrect test ledger-selection assertions are retained in the full evidence, not counted as successful validation.
The tests cover synthesized revision-6 persisted state, both migration repair directions, repeat scans, pruned/retired non-resurrection, and the no-response-ID fallback. They do not establish the trigger in upstream issue #432, actual archived rollup-cost repair, cross-provider response-ID collision handling, GUI/native behavior, live provider/WSL operation, or packaging. No release or deployment is performed. The earlier default-config rename boundary has separately been validated within the limited preflight scope recorded in opencode-rename-20261019.json; later I/O failures and concurrent external edits remain outside that guarantee.
Source synchronization targets official a6c569e127ceb57cb807b58f50492713684d71b0 with three independent commits. Official Homebrew/Scoop distribution manifests are retained only in upstream history, not replayed as Next releases. Existing Next credential rescue, migration preflight, Codex response identity/revision-7, image/MCP/Skills and restore orchestration enhancements remain intact.
Serial validation on 788face11b471ff8e9c99f4b4f89bb85c1410193 passed 1,071 Node tests, TypeScript, three structural guards and the frontend build, followed by 263 targeted Rust tests (zero failures, seven existing local-data tests ignored). Five marker tests and retained PR8/9/10 regressions were explicitly checked in --list. This is not a full Rust or GUI suite.
Filtering remains an open boundary: config and V2 marker can be excluded independently during restore, and an explicitly excluded marker is not silently restored or packaged. The new characterization tests record this behavior without declaring partial-pair restoration safe. The tested filesystem/ZIP/migration round trip covers the normal paired helper path, not full native restore, live credentials, WSL or arbitrary filtering combinations. See exact mappings, logs and limits.
This limited port groups upstream ed57bc95, only the OpenCode page hunk of 1b1cd65c, and bb161f22 as e2623e3d, retaining the upstream author and the final source commit's author date. It is not a complete synchronization through bb161f22: the unrelated, unwired observations module and the Codex identity change already covered by Next PR #10 are not imported. AGPL, NOTICE, existing Next corrections and release isolation remain unchanged.
The tests-first branch test/opencode-auth-write-guard retains 944493419edbed849d9406c665729eb075c43bfc. Its Node execution recorded one pass and seven failures after extracting the page's actual mode state into a production-consumed controller. These are regression assertions, not seven independently observed user incidents. The command's existing disk-write body was extracted into a path-parameterized helper called by both production and synthetic tempdir tests; the added path lookup is an extra fallible command precondition. The focused Rust RED run compiled and ran all six listed tests: two V1 controls passed and four guard assertions failed. The actual shared writer accepted V2 and failed/invalid mode probes, modifying synthetic auth files or creating their directory. No user files or credentials were used.
The same-feature tests and fix are consolidated as 92c3e888221bb3c011441d83c4cfec18063cdb00, tree-identical to reviewed draft a4438d58. The serial GREEN run on exact 5161c9f3120b81672df407e3bf853db02642c1c0 passed TypeScript, all 1,081 Node tests (zero failures/skips), 18 browser-fixture checks across light/dark themes, three structural guards and the frontend production build. The subsequent targeted Rust job passed 68 tests with zero failures/ignored: commands 19, migration 37, credential reads 12. All seven auth guards, three mode-lock tests and five retained source-mode tests were checked in the test list before execution. The full Rust suite and unrelated browser fixtures were not run. Subsequent evidence edits change documentation only; exact sources, checks and log hashes identify the tested tree.
The browser fixture mounts the real shared AuthConfigModal and Monaco with the production page controller; backend answers and the final write are synthetic. It does not execute the full OpenCodePage, native IPC, live credentials, real CLI, WSL, packaging or the stale onOk callback directly. Screenshots were generated within the fixture but were not retained for visual review. Lock tests use channel/time-bounded scheduling evidence, not a formal proof; the process lock coordinates the two migration entry points and auth save, not external CLI or backup/restore writers. No release or deployment was performed.
Selectively ports the gateway foundation from upstream 1b1cd65c and request-detail integration from e5d415ba, retaining source authorship and AGPL/NOTICE. This does not claim complete upstream synchronization; upstream-main remains the previously verified official-history mirror. Independent fixes replace arbitrary lossy diagnostic values with finite safe summaries (including old-detail display/export) and observe structured, bounded upstream SSE before protocol conversion. Tests establish synthetic privacy and SSE defects, not actual user credential disclosure.
Full acceptance on a89dc80d: Linux default Cargo targets passed 3,299 tests, with 13 existing environment-dependent tests ignored; frontend passed 1,081 Node tests, TypeScript, three structural guards, production build, and 12 localized real request-detail component scenarios with 92 assertions. The component uses real API wrappers with synthetic read-only IPC data. This is not the entire pnpm test browser collection, native desktop/live provider acceptance or a performance benchmark. No release or deployment was performed.
The reorganized functional tip 9a2b7c3a has exactly the tested tree. Detailed evidence retains both RED runs, the initial HTTP fixture correction, focused GREEN counts, ignored cases and exact log hashes.
Selectively ports upstream e5b51864 with original authorship, Next auth scripts/translations and AGPL/NOTICE preserved. Claude Desktop, Gemini CLI and Grok adopt shared cards, and OmO gains the compact view. A separately verified Gemini fix preserves concurrent refresh/save pending indicators; existing single-pending consumers remain compatible. The upstream mirror is unchanged and this does not claim every official commit is absorbed.
GREEN validation passed 1,081 Node tests, TypeScript, three structural guards, production build and 213 checks across selected real-component browser fixtures. The tested tree is identical to consolidated a0cace62 before evidence-only documentation. Detailed evidence separates the actual Gemini RED from a fixture dropdown timeout and records exact scope. Full-page drag persistence, real account operations, native UI and Rust were not tested in this frontend change. No release or deployment occurred.
f9c67622 adds Codex additional credits independently of subscription windows and reset cards. It retains zero and negative balances, distinguishes missing/unknown/unlimited availability, and invalidates stale balances on failed refreshes or account changes. Credits remain native units, without a dollar conversion. The upstream issue and independent upstream PR exclude Next-only fixture maintenance.
The selective test-only port of upstream 24e9093f3668f4957e1fa0829c5921e0cb590deb is 2217f6d59d86f3abe8e6bf686d9e000853fd3b1c, preserving coulsontl's authorship and the source reference. It fixes Qoder fixtures aging out of the default retention window. The separate fa29f27f253b98861a7c48c16da5580507bc2ee5 aligns each scenario's timestamps and settling clock, including its Claude comparison, instead of leaving a future fixed-clock cutoff. Existing archive regressions and production retention logic are unchanged; the new boundary table checks clock arithmetic, not a future system-clock retention run. 2c9196ad separately stabilizes synthetic SSH clicks without retrying clicks or weakening request-race assertions. This is not a complete upstream synchronization.
Final Next validation checks code/test candidate 2c9196ad5d922f705be9d552afe2aba9a52ca6ee (tree 1d46e3d290bd465ac7394a85512981aff8aa42c7) through CI-only child 622f1db15ada53d046efb58f721c55422ebf7ade. TypeScript and all 1,124 Node tests passed; the complete SSH browser suite passed twice, 55 checks each. Rust passed 3,371 tests: 3,083 library, 277 integration and 11 doctests, with zero failures and 13 existing environment-dependent tests ignored. All seven new credits tests and ten Qoder tests were explicitly selected; the seven credits tests also passed separately. The earlier same-runner comparison reproduced the same two Qoder failures on both unchanged baseline and feature candidate before the test repair.
Frontend acceptance is segmented, not one all-green pnpm test run. The earlier frontend run passed the unchanged pre-SSH suites and 95 real CodexPage credit checks before failing in SSH fixture interactions. The comparison run passed the remaining three suites and production build while its SSH comparisons still failed. The final run above closes that SSH gap. Blob comparison establishes that only the scoped Qoder/SSH tests changed after those earlier feature checks; these evidence-only documentation additions do not rerun or alter the tested code. Tests use synthetic IPC, temporary files and in-memory databases, not real credentials or sessions. Native desktop behavior, live providers, platform packaging and independent update channels remain outside this validation. No release or deployment was performed.
Baseline-only validation compiled the real Rust library and verified seven selected synthetic entry-point cases on source tree 60733be57ff9c2f128cddffb3790901f2769a7e9 (validation commit bde2ff1c197686485b8337deb08f7b6791fc3c2e). Two expected failures reproduced rejection of clearing or changing an ordinary Codex runtime root after that root and its parent disappeared, despite no required old-runtime effects in those fixtures. Five refusal controls passed. This is a regression in an unintegrated owner-admission candidate, not a demonstrated defect of current main or upstream, and the successful classification job does not mean the candidate is ready to integrate.
Recovery remains blocked on writer coordination: checking absent gateway manifests/backups, even repeatedly or alongside a final database eligibility check, does not exclude ordinary restore, backup or arbitrary-path export writers creating them before the root change is saved. A repair must retain existing cached/in-flight leases and coordinate relevant writers without silently narrowing supported export/restore destinations. No recovery escape implementation has been accepted.
Required follow-up includes refusal tests that actually reach the applied-provider and history eligibility branches, distinct cached/in-flight ownership controls, and concurrent writer/root-save tests. The present applied/history controls can stop at the earlier missing-root error; cached and in-flight cases share the invalid-cached-root check. Process-exit lease release is not a substitute for same-process last-connection lifetime tests. Full frontend/Rust acceptance for this candidate remains incomplete; native GUI, Windows and real profiles were not validated. No product integration, release or deployment is represented by this record.
A separate synthetic coverage candidate based on 0c4f9b1af37ed6aa1aea81dbb263149c3ddee2da targets six real GrokPage/GrokModelFormModal cases: editor readback, cancel, save, repeated submission while pending, save-error retry, and remount. This is additional coverage work, not a demonstrated product defect or functional fix. The first attempt stopped during dependency preparation before building or testing. A subsequent full-application import-graph build received SIGKILL before browser startup; its cause remains unconfirmed and is not recorded as proven out-of-memory failure.
A separately reviewed component/mock candidate retains the real page, modal and child components, uses synthetic IPC responses, and replaces only the application routing module with a fail-closed boundary. Its bundle completed, but Chromium aborted because Unix socket creation returned EPERM. No page loaded and zero UI cases executed, so neither the six case assertions nor their zero-routing-call assertions have passed. Observed child processes were cleaned up, and final source, original dependency and private-copy dependency seals remained unchanged; those checks do not establish UI correctness.
All six cases remain pending. Real router initialization, navigation, sharing/deep links, native application behavior and disk persistence are outside the scoped candidate. Remount would exercise only the synthetic in-page store, not a native restart. Continuous-gate wiring remains unapplied until the six new cases and four existing Grok cases are validated together. No test candidate has been integrated, and no release or deployment was performed. Existing coverage records and other pending work remain applicable.
The synchronization targets official d49d29bc9ba53991917fced26ba8fbb03c8522ad (tree 647238a38a45e9b3ca14eae473543c42d6fe23ac) from Next a18b38030c8f811481ed63d050d4c601fc29e044. The source decision ledger classifies all 66 commits since the last full upstream baseline. Functional ports remain linear and retain source authorship. Already-equivalent #442 pricing pagination, #440 recovery, additional credits and existing Next privacy/identity/concurrency fixes are not replaced or duplicated.
New runtime ports include Windows tray refresh/exit behavior, model-aware Anthropic thinking, rebuild progress/panic recovery, shared provider cards and reorder semantics, MCP JSON export, Pi MCP compatibility, npm policy recovery, and SSH/WebSocket reporting boundaries. Next-specific OpenClaw key-order handling, account identity fences, structured tool observations, Codex lineage/revision-7 reconciliation, and reviewed browser cleanup remain authoritative. Machine-local settings are no longer tracked; project instructions/skills are retained without reading local settings. Official rolling-beta publication, signing/updater infrastructure and live CDP debugging tools are not imported as package prerequisites.
The independent Windows candidate uses the Next application/data/protocol identity and a manual Next release channel, not the official application’s updater key or protocol. The manual build workflow only uploads unsigned NSIS artifacts and checksums; it never publishes. Release acceptance requires the same final product SHA to pass the separately sealed Linux suite and actual Windows build. Existing CI records above do not validate this new tree. Grok combined resource-bound coverage and native/live-profile limitations remain open; this synchronization does not claim those gaps were fixed.
Every push to this independent repository's main triggers the Next distribution workflow; a maintainer does not need to click a manual build button. Linux desktop and standalone-crate tests, the frontend suite/build, and Windows regression/package gates run serially on the exact pushed source. Heavy runs share the maintenance lane with a bounded queue. A failed gate leaves prior public previews unchanged. The workflow does not publish business applications or deploy production services.
After validation, a separate job uses only GitHub's short-lived job token with contents: write to create an immutable v<embedded-version>-build.<source-sha-prefix> preview. Tests/build jobs remain read-only. It checks the repository, full source SHA/tree, run/attempt, version, installer filename, size and SHA-256 before uploading, then downloads and rechecks the draft assets before making the release public. Existing tags/releases are never overwritten. A partial upload remains a draft for inspection; it does not replace the previous working preview.
Distribution remains unsigned and updates inside the application remain manual. Passing tests and creating a Windows installer are not proof of interactive installation or native GUI acceptance. No signing keys, persistent publication tokens, upstream updater service, app-store publication or real account data are used. The prior manual-only artifact workflow policy is superseded solely for this explicitly authorized Next Windows preview distribution.