GPU compatibility

[Bug]: openai-codex OAuth runtime fails on 2026.4.9 with 403 HTML; 2026.3.28 works

On the same Ubuntu VPS and openai-codex OAuth setup, OpenClaw 2026.4.9 fails to return runtime chat replies in Control UI, WhatsApp, and WeCom, while 2026.3.28 works.

Bug in Release Notes (v2026.3.31): 'openclaw flows' is actually 'openclaw tasks'

[Bug]: embedded run failover decision because of timeout

After upgrated to 2026.4.1, the gateway display the following message after input some message in openclaw webpage: 17:48:38+08:00 [agent/embedded] embedded run failover decision: runId=fa92620f-4ad8-40d4-9855-9f82a5d42ac8 stage=assistant decision=surface_error reason=timeout provider=custom-127-0-0-1-8088/NVIDIA-Nemotron-3-Super-120B-A12B profile=- ... and then the LLM running is idle, the openclaw stopped working.

Add a bypass features to bypass the @onlyOn-based decorator

[Bug] 2026.3.13 Windows: ACP runtime unavailable after update — acpx 0.3.0 vs pinned 0.1.16, expectedVersion:any workaround not effective

After upgrading to OpenClaw 2026.3.13 on Windows, the ACP runtime backend is unavailable. This appears to be a recurrence of #43997, but with two compounding issues specific to Windows.

[Bug]: [Bug]: Requests abort after ~1 minute with Request timed out before a response was generated despite agents.defaults.timeoutSeconds set to 18 minutes

OpenClaw aborts requests after about 1 minute with Request timed out before a response was generated even though agents.defaults.timeoutSeconds is set to 18 minutes. <img width="1050" height="537" alt="Image" src="https://github.com/user-attachments/assets/963947c8-09cf-4902-88f4-76ea46eb40e4" />

[Bug]: The hit rate of prefix caching in Qwen3.5 35BA3B is very low, always less than 0.1%

[Bug] Telegram getUpdates 409 conflict when multiple bot accounts poll from same gateway

When running a single gateway with multiple Telegram bot accounts (8 in our case), the gateway produces persistent `409 Conflict` errors on `getUpdates` — even though every account has a **unique bot token**.

Webchat: Maximum call stack size exceeded when adding photo/image attachment

Enhancement: Allow vault/exec secret references in auth.profiles apiKey

Allow vault/exec secret references (e.g. OpenBao, HashiCorp Vault) to be used as the `apiKey` source in `auth.profiles`, the same way they are already supported in `tools`, `channels`, `gateway`, `skills`, and `messages`.

[Bug]: 2026.3.28 bundled feishu channelConfigs schema overrides openclaw-lark plugin schema even when bundled plugin is disabled

When upgrading from 2026.3.24 to 2026.3.28, Gateway refuses to start with: ``` channels.feishu: invalid config: must NOT have additional properties channels.feishu.accounts.jeeves: invalid config: must NOT have additional properties channels.feishu.accounts.ivy: invalid config: must NOT have additional properties channels.feishu.accounts.crichton: invalid config: must NOT have additional properties channels.feishu.accounts.marple: invalid config: must NOT have additional properties ``` The external plugin `@larksuite/openclaw-lark` (2026.3.26) registers channel id `feishu` with a full config schema that includes keys like `replyMode`, `footer`, `botName`, `blockStreaming`, `markdown`, `dedup`, `threadSession`, and `uat`. However, the bundled `@openclaw/feishu` plugin also registers channel id `feishu` with a `channelConfigs` schema (new in 2026.3.28) that does **not** include these keys and sets `additionalProperties: false`. The bundled plugin is explicitly disabled (`plugins.entries.feishu.enabled: false`), but its schema still takes priority during config validation.

[Bug]: task-manager worker falls back to bundled SKILL.md instead of workspace skill

`task-manager` workers can resolve `SKILL.md` from the bundled OpenClaw install path instead of the current workspace path, which causes startup to fail with `ENOENT`.

[Gemma 4] mm_token_type_ids required for text-only fine-tuning - should default to zeros

Bug: `openclaw gateway restart` leaves gateway down for extended periods without error

The `openclaw gateway restart` command sends SIGTERM to the gateway process but does not verify that the gateway successfully restarts. This can leave the gateway down for extended periods with no error indication or automatic recovery.

[Bug]: WebSocket Gateway Probe Timeout on Loopback

Running openclaw gateway probe returns "Connect: failed - timeout" even though the gateway is fully functional (HTTP works, WebSocket connections work). Environment: OpenClaw: 2026.3.13 OS: Linux 6.12.62+rpt-rpi-2712 (arm64) - Raspberry Pi 5 Gateway: loopback (127.0.0.1:18789) Steps to reproduce:

quwen3-next:80b failed to load 500 server error on ollama v0.17.7

[Bug]: Local model support broken by three compounding failures in model validation, provider routing, and error classification

[Bug] Auto-update subprocess crashes before restart handoff: piped stdout/stderr breaks when bootout kills parent gateway

The built-in auto-update consistently kills the gateway on macOS with no recovery, leaving the service unloaded indefinitely. The root cause is that `resolveCommandStdio` forces piped stdout/stderr for the update subprocess, creating a fatal dependency on the gateway parent process — which is then killed by `launchctl bootout` during the restart step. This is the underlying mechanism behind several related reports (#54861, #57379, #58041, #40811, #48992) that describe symptoms (service unloaded, multi-hour outages, KeepAlive not respected) without pinpointing why the restart handoff never completes.

Feature: per-model Ollama API options in config (e.g. think: false for Qwen 3.5)

Control UI on local loopback (127.0.0.1) still requires pairing in 2026.3.13, despite docs saying local connections auto-approve

Control UI on local loopback (`127.0.0.1`) is still requiring pairing in `2026.3.13`, even though the docs say local connections are auto-approved.