GD4AI/obsidian-llm-wiki

The capability this issue asks for: pnpm typecheck reports zero errors under TypeScript 6.0.3, and Gate 1 passes with typescript@6 installed. Until that holds, the bump cannot land. How the tracking surface disappeared. #770 (dependabot, 5.9.3 to 6.0.3) was closed on 2026-10-07 after its Gate 1 failed, and dependabot now ignores 6.x. That was the right call for the PR — it cannot pass — but it left the work itself with no issue, so this one records it. The failure is real, not a flake. Run 35511539309 on 05825f49 fails at the Five-Gate step with 31 errors across 14 files. Four error codes: Code Count Meaning TS2345 25 argued type not assignable to the parameter TS2564 4 property has no initializer and is not definitely assigned TS2430 1 interface incorrectly extends another TS2322 1 type not assignable to target type Files, with the first error of each: src/main.ts(70,72,73,74) TS2564 settings, wikiEngine, schemaManager, autoMaintainManag

The problem The READMEs tell users what to type into Obsidian's command palette, and in ten of the eleven locales they tell them to type English. Every locale has a translated command name. src/texts/ja.ts registers cmdViewHistory: '取り込み履歴を表示', de.ts has 'Ingestionsverlauf anzeigen', ru.ts has 'Просмотр истории операций'. But docs/README_JA.md, README_DE.md and the rest quote View Ingestion History — a string that does not appear anywhere in that language's UI. So the instruction is unusable in exactly the languages that needed translating. This became visible when #780 corrected the English string (View Ingestion History → View operation history, matching src/texts/en.ts:537). That fix was right and is now merged. The other ten files still quote English, so for them it changed one wrong string into another wrong string. Scale English command names per translated README, counted by matching "Cmd+P/Ctrl+P" → "Some English Words": README

What ai 6 → 7 with the provider packages moved in step. Filed because Dependabot's PR cannot express this: #728 bumps @ai-sdk/openai-compatible alone, and that cannot compile. The measurement Rebuilt on current main with pnpm add @ai-sdk/[email protected] and nothing else touched: src/llm-sdk/openai-compat-sdk-client.ts(298,5): error TS2322: Type 'LanguageModelV4' is not assignable to type 'LanguageModel'. Type '"v4"' is not assignable to type '"v2"'. @ai-sdk/[email protected] depends on @ai-sdk/[email protected] → specification v4 [email protected] declares type LanguageModel = GlobalProviderModelId | LanguageModelV3 | LanguageModelV2 → takes v3 at the newest ai@latest is 7.0.107 The repo currently pins @ai-sdk/[email protected], @ai-sdk/[email protected], [email protected]. All three move together or none does. Why it is not a routine bump Three things in the 3.0.0 line land on code we wrote on purpose: 74d520f: feat: migrate providers to support new top-level 'reasoning' parame

What writeFileWithIntent returns void, so a caller cannot tell an update from a skip. Once LINT_WRITE_INTENT became create: false (#750, commit e6b80a96), a lint write can decline — and both fixers report it as done. The scenario is the one #750 is about. A cancelled ingest deletes a summary page during a lint run's LLM batch; the write is now correctly skipped, because create: false keeps the completion marker from coming back; and the lint log keeps a permanent record of a fix applied to a page that no longer exists. Two sites, both reporting a fix that did not happen: Alias completion (fix-runners.ts:147) returns success: true, so filled++ runs, "added N aliases (total M)" goes into results, controller.ts:302 sees filled > 0 and calls generateIndexFromEngine() plus logLintFix('Complete Aliases', …), and the completion Notice reports N/M. Retag (fix-runners.ts:624) returns kind: 'fixed' with an [old] → [new] message, where kind: 'missing' is the accurate answer — that is wha

The measurement that started this @DocTpoint rebuilt a 413-note vault unattended over three days on v1.27.1: 413 source pages and 2,135 entity/concept pages. Three numbers from that rebuild describe the shape of the graph this plugin produces: 95 % of edges connect pages born from the same note. Plain co-occurrence reproduces 94.5 % of the Related edges. Nine of ten pages have no incoming link from a page with another source. The Related cap holds with zero violations and strays sit at 1 % — the shaping is working exactly as written. Then a null model over five multi-note questions — same character budget, same model, thinking off, truth drawn from the notes and rated blind: Arm Retrieval over Result A raw notes, embedding index baseline B wiki pages, same index slightly ahead of A C the plugin's query path lost 3 of 5 questions B ≥ A is the first evidence this project has that the compiler is not a detour — the wiki stores the notes at least as well as the raw