LingHub AI Router — VS Code Extension

Cost-optimized multi-LLM orchestration for developers. Route requests intelligently between OpenAI, Anthropic Claude, Google Gemini, AWS Bedrock, and local models (Ollama / vLLM) — all from within VS Code.

Links: GitHub · Issues · Platform docs

Included

• Sidebar Webview UI
• Fast / Cheap / Smart / Max modes
• Non-stream and stream chat
• Cost and context display
• Multi-file agent prompt flow
• AWS Spot / vLLM OpenAI-compatible provider with routing and cloud fallback
• Diff / Apply
• Unified diff patch preview / apply
• AutoFix loop
• Git branch / commit / push / rollback
• GitHub pull request creation
• Context selection engine

Setup

1. Open this folder in VS Code.
2. Run: npm install npm run compile
3. Press F5 to start the extension development host.

Tests (extension logic)

Pure TypeScript helpers (patch parsing, glob, PR policy) are covered by Vitest:

pnpm test               # once
pnpm run test:watch
pnpm run test:coverage  # Vitest + @vitest/coverage-v8（参照された src のみ集計）

VS Code 結合テスト（Extension Host、@vscode/test-cli + @vscode/test-electron）:

pnpm run test:vscode    # compile 後に stable VS Code を起動（初回はダウンロードあり）
# プローブ系をスキップ（CI で AWS/Ollama を有効にしたワークスペース向け）
# LINGHUB_VSCODE_TEST_SKIP_NETWORK=1 pnpm run test:vscode

From the monorepo root (linghub-platform/):

pnpm run test:linghub-vscode              # Vitest only (from monorepo root)
pnpm run test:coverage:vscode-extension
pnpm run test:vscode-extension         # Extension Host（GUI / xvfb が必要な場合あり）

テストファイルは src/test/suite/*.mocha.ts（*.test.ts は Vitest 専用）。結合テストのコマンド選別・カバレッジの詳細は docs/operations-guide.md §11。

Package extension (VSIX for internal install)

npm run vsix

Produces linghub-platform-vscode-<version>.vsix in the repo root (server/, web/, etc. are excluded via .vscodeignore). Install in VS Code: Extensions: Install from VSIX…. See docs/operations-guide.md §2.

For downloadable assets via GitHub Releases, push a tag like vscode-extension-v0.9.2 (must match package.json version). The workflow .github/workflows/vscode-extension-release.yml builds and uploads the VSIX automatically.

Optional Marketplace publish automation is available in .github/workflows/vscode-extension-marketplace.yml (requires repository secret VSCE_PAT).

Required settings (cloud router)

Set these for the primary LingHub Router (OpenAI-compatible) endpoint:

• linghub.apiBase — e.g. https://your-router/v1/chat/completions
• linghub.apiKey

Review gate (apply / AutoFix review) always uses this cloud router only — never AWS vLLM.

Optional: AWS vLLM (OpenAI-compatible)

When enabled, eligible requests can go to your AWS-hosted vLLM first, with automatic fallback to the cloud router.

Routing flags

• linghub.routing.preferAwsForCheap — fast/cheap modes may use AWS
• linghub.routing.preferAwsForSmart — smart mode may use AWS (max never uses AWS)
• linghub.routing.allowAwsForAutofix — single-file AutoFix/repair generations only (default off)
• linghub.routing.longPromptCharThreshold — larger payloads force cloud
• linghub.routing.cloudTimeoutMs / cloudTimeoutAutofixMs — cloud router timeouts

Fallback behavior

• If AWS fails (timeout, HTTP error, invalid JSON), the extension posts a visible status and retries on the cloud router (if configured).
• Audit events record llmBackend (cloud-router | aws-openai-compatible | local-llm), llmLatencyMs, and optional llmFallbackFrom / llmFallbackReason (no prompts or secrets).

Local LLM setup (OS-specific)

• Command palette: LingHub: Open Local LLM Setup Guide (OS を選択) — QuickPick for macOS / Windows / WSL / Linux, or follow linghub.localLLM.setupTargetOs (auto uses runtime detection including WSL env vars). Shows a Markdown preview (Ollama install is not automated).

Known limitations

• Local Ollama is used for non-stream routes only; streaming stays AWS → cloud (SSE wire format).
• Streaming fallback is supported; very old vLLM builds may stream differently — if stream fails, use non-stream mode or disable AWS for stream.
• AutoFix with review gate on still requires cloud credentials even if generation used AWS.
• Cost figures for AWS are labeled estimates (AWS_VLLM_ESTIMATE bucket), not invoice-accurate GPU billing.

Optional

• linghub.githubToken
• linghub.defaultBaseBranch

Operations & production hardening

See docs/operations-guide.md (Japanese) for:

• Configuration validation, JSON export for tickets, and AWS health probes
• Settings for patch limits, protected paths / globs, PR title/body policy, test command override & timeouts
• Security audit Output channel (linghub.security.auditChannelEnabled) and audit JSONL

Customer-facing security narrative: docs/security-policy.md.

Important

This remains an MVP-oriented repo; tune workspace settings and tenant policy for your environment. The operations guide lists the knobs that matter for production.