refactor: central config, shared utils, and verified bug fixes

.vscodeignore

@@ -15,7 +15,6 @@ node_modules/**
 !node_modules/@silvia-odwyer/photon-node/*.js
 !node_modules/@silvia-odwyer/photon-node/*.wasm
 out/**/*.js.map
-out/autocomplete/**
 out/test/**
 scripts/**
 src/**

CHANGELOG.md

@@ -6,14 +6,49 @@ All notable changes to the **OpenCode Go BYOK Provider** extension are documente
 
 ### Added
 
+- **`[Autocomplete]` Known limitation (follow-up planned):** inline-completion requests are not yet wired into the Go usage tracker's cost accounting (`tracker.record()` only runs in the chat provider path). The panel does track **Suggested / Approved** counts per day (see the usage dashboard entry), but the USD cost of completions is not attributed until a follow-up ships the transport summary from the completion engine. This is tracked as a documented TODO on #136/#138 rather than an oversight.
+
+- **`[Autocomplete]` Inline code suggestions (experimental, #49).** Ghost-text completions while typing, powered by the OpenCode gateway with thinking forced off. Opt-in via `opencodego.inlineSuggestions` (default `false`); model via `opencodego.inlineSuggestionsModel` (default `qwen3.5-plus`, whose `enable_thinking=false` mode is a genuine no-reasoning path — measured ~1.5s time-to-first-token with zero hidden reasoning). Requests are tiny (10 lines before the cursor + a short suffix), debounced 300ms, time out at 3s and abort on the next keystroke. The gateway exposes no FIM endpoint, so completions emulate fill-in-the-middle with FIM tokens over `/chat/completions`. New `src/autocomplete/` module (context, prompt, throttle, engine, provider, registration) with unit tests; `scripts/probe-completion-latency.ts` measures engine latency live. All timing/size knobs are user-tunable: `inlineSuggestionsDebounceMs`, `inlineSuggestionsTimeoutMs`, `inlineSuggestionsMaxTokens`, `inlineSuggestionsPrefixLines`, `inlineSuggestionsSuffixChars`.
 - **`[Usage]` Server-accurate Go meters via the official `/zen/go/v1/usage` endpoint (#130).** The status bar, tooltip, quick-pick and usage webview previously showed locally estimated Session/Weekly/Monthly percentages that drifted from opencode.ai (issue #23) because they missed CLI, cross-device and pre-install usage. The tracker now pulls the official endpoint (upstream anomalyco/opencode#16513, verified live) with the existing Go key on startup and after each request (60s TTL cache): rolling/weekly/monthly percent + reset times are server-computed and account-wide, `spent` is derived from the authoritative percent, and Today/Yesterday + per-session spend stay device-local. Failures (401/403/404/network) fall back to the existing SQLite → tracked estimates. The key is only ever sent as the Authorization header and never logged or persisted. New pure module `src/goUsageSync.ts` with unit tests. Documented in `docs/issues/62-20260812-pr132-go-usage-server-sync.md`. PR [#132](https://github.com/ltmoerdani/opencode-copilot-chat/pull/132) by [@Fahad090NP](https://github.com/Fahad090NP).
 
 ### Changed
 
+- **`[Internal]` Central configuration + shared utilities (no behavior change).** Every tunable value in the codebase — URLs, timeouts, limits, storage keys, setting keys and defaults — now lives in one dependency-free module `src/config.ts`, so future tuning happens in a single file. Existing modules import from it and re-export the names their callers rely on. A new pure `src/utils.ts` replaces the near-identical helpers that were copy-pasted across modules: `isRecord` (was defined 6×), `firstString`, `compactErrorCode`, `positiveNumber`, `getErrorMessage`, `formatUsd`, `formatTokenCount`, `formatRelativeTime`, `escapeHtml` and the two cancellable delay variants. Dead code removed: `getOrCreateProfile`, `formatGoUsageTooltip`, `formatGoUsageLanguageStatusDetail`, `createUsageDataPart`. Both stream extractors (`OpenAiResponseExtractor`, `AnthropicResponseExtractor`) now share a `BaseResponseExtractor` instead of duplicating reasoning accounting and think-tag filtering. The verification scripts (`verify-estimate-token-count`, `validate-models`) import the real production logic instead of drifted copies. Sanity tests added for the new modules (`utils.test.ts`, `config.test.ts`, 235 unit tests total).
 - **`[Tooling]` Strict-but-sane lint stack + intelligent pre-commit gate (PR #129).** ESLint keeps the type-aware rules that catch real bugs (`strict` + `strictTypeChecked`: `no-unsafe-*`, `no-unnecessary-condition`, `no-floating-promises`) and drops the pure-`stylistic` layer that fought prettier. `npm run lint` / `bun lint` now ends with a Tests step (compile + unit tests), making it the single "is everything green" command. A new intelligent pre-commit gate (`scripts/staged-lint.ts`, `npm run lint:staged`) lints staged files **plus their direct import dependents** (resolved from the real import graph) instead of either the whole tree (slow) or only the staged files (blind) — so changing a module can never leave type-aware errors in its consumers. Measured: config-only commit ≈1s, `src/` commit ≈10s; full-tree lint stays in CI. Branch also carries the unified `scripts/lint.ts`/`format.ts` runners, `editorconfig-checker` + `shellcheck` + `tsconfig.check.json` type-check (covers `scripts/`, which the build tsconfig never did), script renames (`*.mjs`/`*.mts` → `*.ts`), eslint 10.8.1 + `@types/node` 26.2 (supersedes dependabot #91), and husky PATH fixes. Post-review refinements (5 commits): all 217 `void describe/it/test` prefixes dropped from test files (`22e04b7`), `@ts-expect-error` allowed for proposed-API workarounds while `@ts-ignore` stays banned (`5246434`), repo moved to TypeScript-first config (`eslint.config.ts`, typed scripts via `tsx`, `76570cc`) and standard extensions (`514a63f`), markdownlint config renamed to `.json` (`c817871`). Documented in `docs/issues/61-20260812-pr129-strict-lint-stack-precommit-gate.md`. PR [#129](https://github.com/ltmoerdani/opencode-copilot-chat/pull/129) by [@Fahad090NP](https://github.com/Fahad090NP).
 
 ### Fixed
 
+- **`[Usage]` SQLite reads no longer depend on the `sqlite3` binary.** The zero-usage mystery was the Android SDK's `sqlite3` (`~/Android/Sdk/platform-tools/sqlite3`) being on the PATH only when VS Code launches from a shell that exports it — desktop-launched windows silently lost all CLI history (Today/Yesterday/Codebase = 0 while the fetched quota kept working). The CLI history is now read through Node's built-in `node:sqlite` first (zero external dependencies, retried twice on busy WAL states), falling back to the `sqlite3` binary resolved from PATH **plus** known locations (system, Homebrew, Android SDK). Failures are logged with the exact error to the "OpenCode Go Usage" output channel.
+
+- **`[Usage]` Panel polish + chat-completion charts.** Hovering anywhere on a chart (not just on points) highlights the nearest day with a guide line + dots and a cursor-follow tooltip; on the Models tab the tooltip lists every model's spend for that day. Two new tabs — **Suggested** and **Approved** — chart inline chat completions with whole-number axes and honest tooltips (the day's suggestion/approval counts, not chat token totals); both series share the exact same day buckets, so hovers never resolve to undefined. Acceptances are detected with a bounded heuristic: VS Code's stable API exposes no inline-completion acceptance event, so committing a ghost text is recognized by the document insert starting exactly at the suggested position with a matching multi-character text (30s window, cleared on first match — see `matchesAcceptance`). The default chart window is **Lifetime**, switchable live via the **Window** button (Week → 14 days → Month → Lifetime). The panel brand shows only the profile name, the legend swatches are square and text-aligned, and axes keep round equal tick steps.
+
+- **`[Usage]` Full-page usage panel.** The usage webview is now a complete dashboard (inspired by the reference page): subscription meters as animated rings, stat chips for Today / Yesterday / Codebase, an interactive chart area with Spend / Requests / Tokens tabs (all line charts) and a By-model tab (overlapping colored line charts per model by spend, with a legend and hover tooltips for the model · day · spend/tokens/requests breakdown). Data follows the cursor with a tooltip on every chart element. Chart window: `opencodego.usageChartDays` (default 14 days). The all-time workspace row is labeled **Codebase** everywhere (panel, quick-pick, status-bar tooltip); the old "Top model" / "Models used" chips are gone. Styled action buttons sit in the panel's top-right corner: **Set targets**, **Rename** (when multiple profiles exist) and **Refresh** — wired through `vscode.postMessage`, and a refresh keeps the active tab (data is pushed in place, the page is never reloaded; the background loop still runs at `usageRefreshIntervalSeconds`). Chart axes use round tick steps (1/2/2.5/5 × 10ⁿ) so gaps are always equal and clean (no more 1.25/2.50/3.75 divisions).
+
+- **`[Usage]` Realtime updates.** The status bar, tooltip and usage panel now refresh on a configurable background cadence (`opencodego.usageRefreshIntervalSeconds`, default 60s, min 5s), so terminal-side OpenCode CLI usage, server meters and midnight day rollovers appear without waiting for the next chat request. Changing any usage-view setting repaints the UI immediately, and the refresh interval itself applies live on the next tick.
+
+- **`[Usage]` Instant startup + no more slow usage reads.** The last successful server-usage snapshot is now persisted (`opencodego.serverUsage.v1`) and restored on window start, so the status bar / tooltip render real meters immediately instead of 0s until the network refetch lands (the TTL-guarded refresh still runs in the background). The CLI database read — a blocking `sqlite3` call over a multi-GB file — is memoized for 3 seconds, so a burst of UI refreshes (status bar + tooltip + panel) pays the query cost once.
+
+- **`[Usage]` Compact number formatting everywhere.** Token AND request counts render as `1.2T` / `1.2B` / `1.2M` / `12k` (with correct unit escalation at rounding boundaries — no more `1000.0M`), and dollars as `$1.23M` / `$1.50K` / `$12.30` with sub-cent precision (`$0.0004`) so tiny spend never collapses to `$0.00`. Applied to the status-bar tooltip, tooltip card, quick-pick and usage panel.
+
+- **`[Usage]` Today/Yesterday now merge the OpenCode CLI history with extension-tracked usage.** The CLI database (`~/.local/share/opencode/opencode.db`) stores per-message `cost` + `tokens.{input,output,reasoning,cache}` + `path.cwd` + timestamps, so users who also run OpenCode in the terminal get their real combined daily totals (terminal + VS Code) instead of VS Code-only estimates. Configurable via `opencodego.usageTodayYesterdaySource` (`auto` / `cli` / `extension`).
+- **`[Usage]` New "Codebase" row replaces the old "Session (est)" estimate.** Shows all-time usage for the CURRENT workspace (matched per-directory from the CLI history `path.cwd`), forever by default (`opencodego.usageCodebaseWindowDays`, 0 = all history). Toggle with `opencodego.usageCodebaseRow`.
+- **`[Usage]` Local tracked data is permanent.** The "Reset tracked usage data" action is removed and the 31-day entry cutoff is gone (only the hard 2000-entry cap applies) — the data today/yesterday/codebase rely on can no longer be deleted or silently expire.
+- **`[Usage]` New display options:** `opencodego.usageRollingSessionMeter` (hide the server 5-hour rolling meter in detailed views) and `opencodego.usageDayBoundary` (`utc` / `local` midnight for the Today/Yesterday rows).
+
+- **`[Internal]` `sanitizeToolSchema` emitted `type: "object"` via a ternary whose branches were identical** — collapsed to the constant.
+- **`[Internal]` Anthropic stream extractor wrote usage fields onto itself** instead of the request summary — the real summary was already updated by the SSE data callback, so the duplicate (and wrong-object) write is gone.
+- **`[Internal]` `formatModalityBadges` duplicated its Audio branch** under two conditions that covered all cases — collapsed.
+- **`[Internal]` `formatRelativeTime` rendered 26h as `26h`** instead of `1d 2h` — the days branch now takes precedence.
+- **`[Internal]` Debouncer ignored debounce-window config changes until the provider was recreated** — a pending run is now rescheduled when `delayMs` changes, so `opencodego.inlineSuggestionsDebounceMs` applies on the very next keystroke.
+- **`[Internal]` Numeric settings sanitized** — a misconfigured string `temperature`/`maxTokens`/timeout value can no longer reach the request body (was a potential HTTP 400 upstream).
+- **`[Internal]` `resolveModelMetadata` ignored the bundled `temperature: false` fallback** — kimi-k2.7-code sent an unsupported `temperature` on cold start / failed models.dev fetch and depended on the runtime 400-retry to save every request. The fallback chain now includes the bundled value.
+- **`[Internal]` Modality-based vision detection treated any non-text modality as vision** — an audio/pdf/video-only model advertised `imageInput: true`, so VS Code forwarded image attachments the model cannot process. Vision now requires an actual image modality.
+- **`[Internal]` The Zen `freeOnly` model filter was only applied on the live fetch path** — when the model-list fetch failed (or was cancelled), the bundled/cached fallback list could show paid models to free-only users. The filter now applies on every path.
+- **`[Internal]` The Go-gateway "thinking off" workaround misfired for Kimi K2.7 / MiniMax M3** — those route through chat-completions with an Anthropic-style `thinking` block, which the detector didn't recognize, so genuine chain-of-thought was emitted as visible text. Detection now covers every reasoning channel (`bodyRequestsThinking()`).
+- **`[Internal]` `setManualSpentTargets` over-counted re-edited monthly targets on the SQLite path** — the raw-SQLite monthly sum never contains the baseline, so subtracting it (meant for the summary fallback) inflated the manual target (target $60 with $12 raw spend produced a $72 baseline).
+- **`[Internal]` `buildSummaryFromRows` reported `sqliteAvailable: false`** for SQLite-backed summaries (only masked by the enriched builder overriding it).
+- **`[Internal]` Error classification hardened against hosts without a global `DOMException`.**
+
 - **`[Provider]` Duplicate OpenCode Go/Zen models after setting a per-model option (e.g. `reasoningEffort`) in the model picker (#131).** Setting a per-model option makes VS Code write a settings-only group (no `apiKey`) into `chatLanguageModels.json`; VS Code then resolves that group with `configuration: {}`, and the SecretStorage fallback ran on both the groupless call and the group call, serving every model twice (7 → 14 in the report). A group call whose `configuration` is present but carries no API key is now treated as a per-model configuration group and returns `[]`, leaving the groupless call as the single source. Per-model settings still apply at request time via `modelConfiguration`. Documented in `docs/issues/64-20260813-issue131-permodel-config-duplicate-models.md`. PR [#135](https://github.com/ltmoerdani/opencode-copilot-chat/pull/135) by [@Fahad090NP](https://github.com/Fahad090NP).
 
 ## [0.5.2] — 2026-08-11

README.md

@@ -348,6 +348,16 @@ Your API key and BYOK group settings are kept, so re-enabling (or the
 `Re-add to Language Models` action) restores everything. A window reload is
 required after toggling.
 
+### ✍️ Inline Code Suggestions (Experimental)
+
+Ghost-text completions while typing, powered by the OpenCode gateway with **thinking forced off**:
+
+- Opt-in: `"opencodego.inlineSuggestions": true` (requires a window reload).
+- Model: `opencodego.inlineSuggestionsModel` — defaults to `qwen3.5-plus`, whose `enable_thinking=false` mode is a genuine no-reasoning path (measured ~1.5s TTFB, zero hidden reasoning). Reasoning models (e.g. `deepseek-v4-flash`) burn 100+ reasoning tokens even with thinking off and are not recommended.
+- Tuning knobs (all optional): `inlineSuggestionsDebounceMs` (300), `inlineSuggestionsTimeoutMs` (3000), `inlineSuggestionsMaxTokens` (128), `inlineSuggestionsPrefixLines` (10), `inlineSuggestionsSuffixChars` (300).
+- Requests are tiny (10 lines before the cursor + a short suffix by default), debounced 300ms, time out at 3s, and are aborted on the next keystroke.
+- The gateway exposes no FIM endpoint, so completions use fill-in-the-middle emulation with FIM tokens over `/chat/completions`.
+
 ### 🛠️ Smart Routing & Reliability
 
 - **Native endpoint routing** per family (see [Models](#-models) table)

eslint.config.ts

@@ -17,20 +17,30 @@
 // via `--max-warnings 0`. The only rules disabled here are ones whose noise
 // outweighs their value (see the scoped overrides below).
 
-import { readFileSync } from "node:fs";
 import { defineConfig } from "eslint/config";
 import tseslint from "typescript-eslint";
 import yml from "eslint-plugin-yml";
 import jsonc from "eslint-plugin-jsonc";
+import { gitignorePatterns } from "./scripts/gitignore";
 
-const gitignore = readFileSync(new URL(".gitignore", import.meta.url), "utf8")
-  .split(/\r?\n/)
-  .map((line) => line.trim())
-  .filter((line) => line && !line.startsWith("#") && !line.startsWith("!"));
+// eslint.config.ts is loaded as ESM by ESLint (via jiti) but lives outside
+// both tsconfig projects, so the editor's inferred CommonJS project does not
+// see @types/node's `import.meta.dirname` (declared only for node16/nodenext
+// modules). It exists at runtime; declare it explicitly to silence the
+// editor diagnostic. Matches @types/node's own declaration, so type-checked
+// consumers merge cleanly.
+declare global {
+  interface ImportMeta {
+    dirname: string;
+  }
+}
 
+const gitignore = gitignorePatterns();
 // Files not covered by tsconfig (which only includes src/), type-checked via
 // the default project so strictTypeChecked rules still apply to them.
-const nonProjectFiles = ["eslint.config.ts", "scripts/*.ts"];
+// eslint.config.ts is ESM-only (loaded via jiti) and lives outside both
+// tsconfig projects; scripts/*.ts are covered by scripts/tsconfig.json.
+const nonProjectFiles = ["eslint.config.ts"];
 
 // The typescript-eslint `config()` helper is deprecated; ESLint core now
 // provides `defineConfig()`. We replicate the helper's `extends` expansion
@@ -51,6 +61,9 @@ export default defineConfig([
       parserOptions: {
         projectService: {
           allowDefaultProject: nonProjectFiles,
+          // scripts/ (and eslint.config.ts) are type-checked via the default
+          // project; keep the cap above the script count as it grows.
+          maximumDefaultProjectFileMatchCount_THIS_WILL_SLOW_DOWN_LINTING: 64,
         },
         tsconfigRootDir: import.meta.dirname,
       },