feat: support custom OpenAI-compatible API endpoints (#2135)

* feat: support custom OpenAI-compatible API endpoints (Ollama, LM Studio, etc.)

Add configurable base URL and custom model name to allow using any
OpenAI-compatible API provider. Models are now fetched dynamically
from the configured API endpoint instead of a hardcoded list.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: add error handling for unreachable API providers and model fetch failures

Show user-friendly error messages when the API endpoint is unreachable,
API key is invalid, or model is not found. Add retry button for failed
model loading. Includes translations for all 11 languages.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat: add Test API connection button, changelog and documentation

- Add sendTo handler 'testApiConnection' in main.ts for server-side
  API connectivity testing
- Add test button in admin jsonConfig using sendTo
- Add WORK IN PROGRESS changelog entry in README.md
- Add documentation for custom API support in docs/en and docs/de

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* refactor: remove Custom model name field, models are fetched dynamically

The custom model name config field is unnecessary since available models
are now fetched dynamically from the API endpoint and shown in the
dropdown. Removed from config, i18n, dialog logic and documentation.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: remove invalid icon from test API connection button

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* build: recompile backend to include testApiConnection handler

The sendTo handler was only in src/main.ts but the compiled build/main.js
was missing it, causing the test button to spin indefinitely.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* refactor: move all API calls server-side via sendTo to avoid CORS issues

- Add 'chatCompletion' sendTo handler for proxying chat requests
- Model loading already uses 'testApiConnection' sendTo handler
- Remove OpenAI SDK usage from browser (no more dangerouslyAllowBrowser)
- All API communication now goes through the adapter backend,
  which avoids CORS issues with local providers like Ollama

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: build frontend, fix TypeScript error type and disabled prop

- Fix error state type from boolean to string | false
- Fix disabled prop to use !!error for boolean coercion
- Rebuild frontend assets with server-side API changes

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: restore admin/tab.html and admin/custom/ removed by build

The frontend build process incorrectly deleted tab.html and the custom
components directory, breaking the script editor tab. Restore original
build artifacts - source changes remain in src-editor for future builds.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: rebuild frontend correctly with manual copy and tab.html generation

The automated build task failed on Node 25 due to deprecated rmdirSync.
Built editor with vite directly and manually copied assets + generated
tab.html with socket.io and monaco script injection.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: add Content-Length header and increase timeout to 600s for chat requests

Local LLMs (Ollama) need Content-Length to process the request body
correctly and more time to generate responses with large prompts.
Also explicitly set stream: false for compatibility.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: strip LLM thinking artifacts from code generator responses

Local models (Ollama) often output <think> blocks, <|endoftext|> tokens,
and reasoning commentary alongside code. The response parser now:
- Strips <think>...</think> blocks
- Strips special tokens (<|endoftext|>, <|im_start|>, <|im_end|>)
- Intelligently extracts code from unstructured responses
- Removes trailing LLM commentary after code

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs: add recommended providers (Gemini free, DeepSeek) and note about ChatGPT

- Add Google Gemini as recommended free provider with setup instructions
- Add DeepSeek as affordable alternative
- Add guidance for local models (14B+ recommended)
- Note that free OpenAI/ChatGPT API no longer works for code generation

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: strip 'models/' prefix from Google Gemini model IDs

Google's API returns model IDs like 'models/gemini-2.0-flash' but the
chat completions endpoint expects just 'gemini-2.0-flash'.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* chore: cleanup orphaned build assets and update changelog

- Remove old OpenAiDialog/ScriptEditor assets no longer referenced
- Update WORK IN PROGRESS changelog (remove 'custom model name',
  add server-side proxy and artifact stripping)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix: allow empty API key for local providers like Ollama

Local providers (Ollama, LM Studio) don't require authentication.
The API key check now only rejects requests when no key AND no custom
base URL is set. The Authorization header is omitted entirely when
no key is configured.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* style: fix prettier formatting for sendTo calls

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* Removed useless models

* Removed useless models

* Fixed some minor possible errors

* Fix linter

---------

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: GermanBluefox <dogafox@gmail.com>

Author: 3 people

.gitignore

@@ -17,3 +17,4 @@ iob_npm.dode
 
 #ignore .commitinfo created by ioBroker release script
 .commitinfo
+/.claude/settings.local.json

CLAUDE.md

@@ -0,0 +1,93 @@
+# CLAUDE.md
+
+This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.
+
+## Project Overview
+
+ioBroker.javascript is a JavaScript/TypeScript/Blockly script engine adapter for the ioBroker home automation platform. It runs as a daemon adapter that executes user-written scripts in a sandboxed environment, with a web-based editor UI (Monaco + Blockly) and an admin configuration panel.
+
+## Build Commands
+
+```bash
+# Full build (backend TypeScript + both frontends)
+npm run build
+
+# Backend only (TypeScript → build/)
+npm run build-backend
+
+# Editor UI only (src-editor → admin/)
+npm run build-editor
+
+# Admin config UI only (src-admin → admin/custom/)
+npm run admin-build
+
+# Install all dependencies (root + src-editor + src-admin)
+npm run npm
+```
+
+## Testing
+
+```bash
+# All tests (declaration checks + integration)
+npm test
+
+# Individual test suites
+npm run test:declarations    # TypeScript declaration compilation checks
+npm run test:integration     # Mocha integration tests (test/testFunctions.js etc.)
+npm run test:scheduler       # Scheduler/cron tests only
+npm run test:package         # Package structure validation
+
+# Run a single test file directly
+npx mocha test/testScheduler.js --exit
+```
+
+Test framework is Mocha, configured via `.mocharc.json`. Tests are in `test/`.
+
+## Linting
+
+```bash
+npm run lint          # Lint backend (root) only
+npm run lint-all      # Lint root + src-editor + src-admin
+```
+
+ESLint uses `@iobroker/eslint-config` (flat config in `eslint.config.mjs`). Prettier config extends `@iobroker/eslint-config/prettier.config.mjs`. The root ESLint config only covers `src/` — frontend directories have their own configs.
+
+## Architecture
+
+### Three codebases in one repo
+
+1. **Backend adapter** (`src/` → compiled to `build/`): Node.js TypeScript, compiled to CommonJS. Entry point is `src/main.ts`. Runs as an ioBroker adapter daemon.
+
+2. **Script editor UI** (`src-editor/`): React + TypeScript app built with Vite. Contains the Monaco code editor, Blockly visual editor, script management sidebar, log viewer, and OpenAI integration. Built output is copied to `admin/` by `tasks.js`.
+
+3. **Admin config UI** (`src-admin/`): React + TypeScript app built with Vite. Provides adapter configuration via `@iobroker/json-config`. Built output goes to `admin/custom/`.
+
+Each frontend has its own `package.json`, `node_modules`, and ESLint config. The build orchestration in `tasks.js` handles cleanup, npm installs, compilation, and copying built assets to `admin/`.
+
+### Key backend modules (`src/lib/`)
+
+- **`sandbox.ts`** (~244KB) — The script execution sandbox. Provides all global functions available to user scripts (`getState`, `setState`, `on`, `schedule`, `createState`, `httpGet`, etc.). This is the core of the adapter.
+- **`scheduler.ts`** — Cron and astro-based scheduling engine (sunrise/sunset events via suncalc2).
+- **`protectFs.ts`** — Filesystem access control for sandboxed scripts.
+- **`debugger.ts`** — Script debugging support.
+- **`mirror.ts`** — Mirrors scripts to/from the filesystem.
+- **`javascript.d.ts`** (~140KB) — Complete TypeScript type definitions for the script API, used for editor autocompletion.
+- **`typescriptTools.ts`** — TypeScript compilation for user scripts (via `virtual-tsc`).
+
+### Build orchestration (`tasks.js`)
+
+Custom Node.js build script using `@iobroker/build-tools`. Supports granular steps via CLI flags (`--0-clean`, `--1-npm`, `--2-build`, `--3-copy`, `--4-patch`). The `npm run build` command runs backend compilation first, then `node tasks` for the full frontend build pipeline.
+
+### Adapter configuration
+
+`io-package.json` defines adapter metadata, native configuration defaults, instance objects (astro variables, debug states), and ioBroker dependency requirements (js-controller >=5.0.19, admin >=7.6.1).
+
+## TypeScript Configuration
+
+- Backend targets ES2022, compiled to CommonJS (`tsconfig.build.json`)
+- `@types/node` major version should match the lowest supported Node.js version
+- Frontend apps target Chrome 89+ (Vite build config)
+
+## CI/CD
+
+GitHub Actions (`.github/workflows/test-and-release.yml`) runs lint, build, and adapter tests across Node 18/20/22/24 on Ubuntu, Windows, and macOS. Releases are automated via `@alcalzone/release-script` with NPM publish and GitHub release creation.

README.md

@@ -25,6 +25,14 @@ Executes Javascript, Typescript Scripts.
 ## Changelog
 <!--
 	### **WORK IN PROGRESS**
+* Added support for custom OpenAI-compatible API endpoints (e.g. Ollama, LM Studio, Google Gemini, DeepSeek, OpenRouter)
+* Added configurable base URL in adapter settings
+* Models are now fetched dynamically from the configured API endpoint
+* Added "Test API connection" button in adapter settings
+* Added error handling with user-friendly messages for unreachable providers
+* Added retry functionality for failed model loading
+* All API calls (models + chat) are proxied server-side to avoid CORS issues with local providers
+* Strip LLM thinking artifacts from responses (for local models like Ollama)
 -->
 ### 9.0.18 (2026-01-11)
 * (@GermanBluefox) Corrected an error message with `lastSync`
@@ -40,7 +48,7 @@ Executes Javascript, Typescript Scripts.
 ### 9.0.10 (2025-07-27)
 * (@klein0r) Added Blockly block to format a numeric value
 * (@GermanBluefox) Fixing some blocks in blockly: cron, time
-* (@GermanBluefox) Added new block: "unconditional return"
+* (@GermanBluefox) Added a new block: "unconditional return"
 * (@GermanBluefox) Type definitions for TypeScript were updated
 * (@GermanBluefox) Corrected bug with deleting of sub-folders