#gpu-setup
Sorted by views, then solution_desc, solution, and root_cause length (desc).
2568 issues
After upgrading to OpenClaw 2026.3.13, the local loopback gateway behaves inconsistently. This is broader than the already reported status/probe `missing scope: operator.read` issue. In my case, the same gateway runtime shows: - successful local webchat/control-ui RPC for some connections - `missing scope: operator.read` for other local RPC calls - intermittent websocket handshake timeout / closed before connect - occasional failed first restart attempt before successful second start - some CLI commands failing completely with `gateway closed (1000 normal closure): no close reason` So this looks like a wider local auth/scope/session instability, not only a misleading status/probe message. Environment - OpenClaw version: 2026.3.13 - OS: Debian/Linux - Node: 22.22.1 - Gateway mode: local - Gateway bind: loopback (ws://127.0.0.1:18789) - Gateway service: systemd user service - Browser server enabled locally - Telegram channel enabled and healthy
Setting `ask: off` / `security: full` in exec-approvals.json has no effect on Windows; every exec call returns `exec denied: allowlist miss` even after gateway restart.
After upgrading from 3.24 to 3.28, the gateway failed to start up. openclaw doctor --fix says: **Invalid config: channels.feishu: invalid config: must NOT have additional properties** How can this issue be fixed?
On macOS, OpenClaw 2026.3.13 completes the browser auth step for OpenAI Codex OAuth but fails at code-to-token exchange, while 2026.2.6-3 succeeds under the same proxy setup.
After updating to **2026.4.5**, third-party memory plugins fail to initialize with: ``` memory-lancedb-pro invalid config: embedding: must have required property 'embedding' 1 plugin(s) failed to initialize (validation: memory-lancedb-pro) ``` This does not occur on 2026.4.4 with the same config.
The OpenClaw gateway HTTP server binds to 127.0.0.1 inside the Docker container and is not reachable via Docker host port forwarding. Accessing the gateway from inside the container succeeds (401 Unauthorized), but requests to the host's forwarded port fail with "Connection reset by peer".
Telegram cannot start because the plugin fails to load at startup due to missing grammY-related runtime dependencies. This is not a Telegram token / webhook / network issue. The failure happens during module loading before the Telegram runtime can start.
When routing Azure GPT-5.2 chat requests through LiteLLM, `vector_store_ids` can leak into provider payload via `extra_body`, causing Azure to reject the request with: `litellm.BadRequestError: AzureException BadRequestError - Unknown parameter: 'vector_store_ids'`
`openclaw doctor` reports `Telegram: not configured` in the channel probe section, even though the Telegram channel is fully operational.
The Telegram extension requires `grammy`, `@grammyjs/runner`, and `@grammyjs/transformer-throttler` at runtime, but none are listed in `package.json` under `dependencies`, `optionalDependencies`, or `peerDependencies`. They're only present as runtime `require()` calls in bundled dist files. This means every auto-update or `brew upgrade` wipes these packages, making the CLI completely unusable — all commands fail immediately with `PluginLoadFailureError`.