integration issue

[Bug]: doctor --fix nano-banana migration creates invalid models.providers.google (missing baseUrl and models)

`openclaw doctor --fix` crashes with `Config validation failed: models.providers.google.baseUrl: Invalid input: expected string, received undefined` when migrating legacy `skills.entries.nano-banana-pro` config to `models.providers.google`.

App Router _rsc navigations intermittently fail with "The router state header was sent but could not be parsed" when Proxy matcher is broad on Vercel

[Bug]: Qwen 3.5 stops working after upgrade to v0.18.0

[Bug]: Gateway becomes unresponsive ~60s after startup on 2026.3.31

After upgrading from 2026.3.28 to 2026.3.31, the gateway starts successfully (binds port, loads plugins, connects WhatsApp) but becomes completely unresponsive within 30-60 seconds. The process cannot be stopped gracefully (SIGTERM times out after 30s, requires SIGKILL). Consistent across multiple restart attempts. Rolled back to 2026.3.28 — immediately stable, no issues.

[Bug]: Gateway health check fails (1006) + Ollama model timeout during onboarding on Windows

I encountered multiple issues while setting up OpenClaw on Windows using the QuickStart flow with Ollama. After completing onboarding and restarting the gateway service, the health check fails with a WebSocket error. Additionally, the embedded agent fails due to a timeout when using the default Ollama model.

cannot import name 'HybridCache' from 'transformers

[BUG] API Error: Stream idle timeout - even small file tasks fail on Claude Code Web

[Bug]: ERR_UNSUPPORTED_ESM_URL_SCHEME on Windows — plugin loader passes raw drive-letter paths to import()

OpenClaw version: 2026.4.5 (3e72c03) Node.js version: v22.22.2 (also reproduced on v24.14.0) OS: Windows 11 (10.0.26200) Running openclaw onboard on native Windows crashes immediately after selecting a model. The ESM loader rejects the module specifier because it receives a raw Windows path (C:\...) instead of a file:/// URL. Steps to reproduce 1. Install openclaw globally on Windows: npm install -g openclaw 2. Run openclaw onboard 3. Select Ollama as the provider 4. Select any model (e.g. ollama/gemma4:e4b) Error Error [ERR_UNSUPPORTED_ESM_URL_SCHEME]: Only URLs with a scheme in: file, data, and node are supported by the default ESM loader. On Windows, absolute paths must be valid file:// URLs. Received protocol 'c:' Root cause The plugin/extension loader in dist/loader-BkajlJCF.js resolves plugin paths to absolute Windows paths (e.g. C:\Users\...\extensions\openclaw-web-search) and passes them directly to jiti / dynamic import(). On Windows, ESM requires these to be file:///C:/... URLs. Relevant lines in dist/loader-BkajlJCF.js: - Line ~2526: mod = getJiti(safeSource)(safeSource); - Line ~2213: const runtimeModule = getJiti(runtimeModulePath)(runtimeModulePath); Also in dist/io-CS2J_l4V.js: - Line ~161: mod = getJiti(contractSource)(contractSource); Suggested fix Wrap paths with pathToFileURL() before passing to dynamic import: import { pathToFileURL } from 'node:url'; // Before mod = getJiti(safeSource)(safeSource); // After mod = getJiti(safeSource)(pathToFileURL(safeSource).href); Workaround Running openclaw under WSL2 avoids the issue since Linux paths don't have drive letters.

[Bug]: web_fetch blocked when DNS returns FakeIP/special-use address

web_fetch blocks when a public domain resolves to a special-use IP (FakeIP), returning "Blocked: resolves to private/internal/special-use IP address", even though the domain is public.

[Bug]: 【Bug修复】OpenClaw Gateway模式下newapi自定义提供商报network connection error

[Bug]: LMCache does not work with vLLM 0.17.0 (Qwen3Next)

[Bug]: Filename Becomes URL-Encoded String After Sending Files in Lark/Feishu

When sending a file (via attachment or upload) in Lark/Feishu, if the original filename contains non-ASCII characters (e.g., Chinese, Japanese, emoji, or special symbols), the file is received with a URL-encoded filename (e.g., %E6%B5%8B%E8%AF%95.txt) instead of the original human-readable name (e.g., 测试.txt).

[Bug]: Interactive browser tool fails despite healthy gateway (Docker / Hostinger image)

Interactive Browser Tool Fails Despite Healthy Gateway (Docker / Hostinger Image) Environment Image: ghcr.io/hostinger/hvps-openclaw:latest OpenClaw version: 2026.3.8 (3caab92) Deployment: Docker (Hostinger VPS) Model: openai/gpt-5.2 OS: Linux VPS Problem Interactive browser automation fails even though both the OpenClaw gateway and browser controller are running. When prompting: open nytimes.com the agent responds: “I can’t open nytimes.com in the interactive browser because the OpenClaw gateway isn’t running on this host.” It suggests running: openclaw gateway install openclaw gateway start However this is incorrect because the gateway is already running inside the container. Verification Gateway is reachable: curl http://127.0.0.1:18789 Result: 18789_http=200 Browser controller is reachable: curl http://127.0.0.1:18791 Result: 18791_http=401 Logs [gateway] listening on ws://127.0.0.1:18789 [browser/server] Browser control listening on http://127.0.0.1:18791/ [browser/service] Browser control service ready When browser tool runs: [tools] browser failed: timed out Browser Runtime Check Chromium present: /usr/bin/chromium Playwright installed: ~/.cache/ms-playwright/chromium-* Steps Already Tried container restart openclaw doctor openclaw doctor --repair full browser state reset recreated browser profiles verified gateway + browser controller ports confirmed Playwright + Chromium installed None resolved the issue. Expected Behavior The browser tool should open the page in the managed browser. Actual Behavior Agent incorrectly reports gateway not running and falls back to HTTP fetch. Notes Other OpenClaw features work normally: Telegram bot cron jobs web fetch/search skills The issue appears limited to interactive browser tool invocation.

ADAPTER command fails to load mmproj (vision projection) file - 500: unable to load model

[Bug] Gateway timeout after 120ms blocks antfarm cron creation

Antfarm workflow cron creation fails with 'gateway timeout after 120ms' error when using OpenClaw CLI fallback, preventing workflows from running autonomously.

MiniMax image-to-image generation fails with "minimax edit does not support resolution overrides"

MiniMax image-to-image generation fails with "minimax edit does not support resolution overrides"

[Bug]: openai-codex enters the generic OpenAI websocket path and falls back to HTTP

`openai-codex` appears to enter the generic OpenAI websocket transport path on current `main`, but the websocket connection target is still the generic OpenAI Responses websocket endpoint. In an isolated smoke test on merged `upstream/main`, this causes the websocket attempt to fail with HTTP 500 and fall back to HTTP transport. The chat/tool flow still succeeds after fallback, so this is not a total functional failure. The issue is that the `openai-codex` provider seems to be treated as websocket-eligible without clear evidence that the generic OpenAI websocket endpoint is the correct transport target for the Codex OAuth path.

Docs: langchain-huggingface README missing [full] extra — HuggingFaceEmbeddings silently fails without sentence-transformers>=5.2.0

The `langchain-huggingface` README only shows: ```bash pip install langchain-huggingface ``` But `sentence-transformers` is an **optional dependency** (under `[full]` extra, requiring `>=5.2.0`). Users who follow the README and try to use `HuggingFaceEmbeddings` hit an `ImportError` at runtime with no guidance on how to fix it.

WhatsApp: creds.json corruption on every reconnect + overly aggressive heartbeat timeout

Two related issues with the WhatsApp channel causing constant reconnect churn on headless/VPS deployments.

Feature request: PARAMETER think false in Modelfile for Qwen 3.5 / thinking models