Material checked on 2026-11-14. Error strings are quoted verbatim so you can match them against your own log.
Split the failure by the first log line
| First key error | Where it is stuck | Check next |
|---|---|---|
post-start inbound/tun ... auto-redirect: setup nftables: flush nftables followed by operation not supported | TUN was created, but the system rejected an automatic redirection rule | Keep the full log; check kernel, nftables, auto_route, auto_redirect and the rules your system actually accepts |
permission denied, or missing permission to manage networking | Privileges or container capabilities | Check the running identity, the container's network-management capability and how it was started |
network: missing default interface with no fatal nftables error above it | Default egress or the network namespace | Check the default route and interfaces; this warning alone is not a verdict on nftables support |
Log points at route.rule_set download or parsing | Remote rule set loading | Use a remote rule set checklist instead |
| Core started, ordinary HTTPS requests fail | Connection, DNS, routing or node | Verify a real request and the egress |
What auto_redirect actually requires
The official TUN configuration documentation states that auto_redirect is Linux-only, requires auto_route to be enabled at the same time, improves TUN routing through nftables, and needs kernel support for nfqueue for pre-matching. So the option depends on what the system's rule engine can do — but the documentation does not promise that every kernel and nftables combination supports it.
One case that can be checked
In sing-box issue #4588 the reporter used OpenWrt 23.05-SNAPSHOT on Linux 5.4.213, arm64, nftables 1.0.8, sing-box 1.14.2. The configuration created a TUN inside an isolated network namespace; with auto_route and auto_redirect enabled, the log first showed the TUN coming up and then reported operation not supported while installing the nftables rules. The reporter compared the same TCP reject rule at different hooks with nft -c and raised the question of kernel support ranges.
The issue was still open at the time of checking, with no general fix published by a maintainer. It therefore cannot be used to claim that every Linux 5.4 device, every OpenWrt build or every sing-box version fails.
The same log also contains network: missing default interface. That may relate to routing inside the isolated namespace and must not be treated on its own as the root cause of the auto_redirect failure. Judge from the earliest fatal chain, your current configuration, and what your own system supports.
A recovery order that keeps an exit
- Secure a way back. Work from a local terminal or a management path that does not depend on this proxy chain. Back up the sing-box configuration, the current network settings and the firewall rules. If remote management of the device depends on the same TUN, do not repeatedly restart or flush rules on a production router.
- Record the environment. Save the output of
sing-box version,uname -randnft --version, and note whether you run on a host, in a container or on OpenWrt. Record whether bothauto_routeandauto_redirectare enabled, and whethernfnetlink_queueis loaded. A loaded module does not prove every nftables rule form is supported. - Separate static configuration from system application. Run
sing-box check -c config.jsonon a copy to validate syntax and fields. Passing the check only means the static configuration parses — it does not mean the kernel accepts the nftables rules generated at startup. Then use the startup log to find the firstFATAL, and do not count the repeated netlink errors after it as separate faults. - Compare under control. If you can preserve the original configuration and have a local way back, compare startup with
auto_redirecton and off in an isolated environment. Turning it off changes how traffic is taken over, so "the process starts" must not be read as "all device traffic is proxied correctly". Do not swap the system's nftables tables on a production device. - Act on the evidence you have. If the error genuinely lands on a rule form the system does not support, check the firmware, kernel and sing-box version compatibility information first, then report the redacted log to the project or to the device maintainer.
When to stop
Stop testing if you cannot keep an exit path, if the device is the network gateway for the whole site, or if continuing requires flushing the production firewall. Those actions can take remote management and other devices offline. The namespace reproduction in the public issue describes that environment's error chain; it is not a configuration for everyone to copy, and the reporter's default_interface: lo is part of that isolated setup.
This page did not reproduce the fault on the OpenWrt, Linux 5.4.213 and sing-box 1.14.2 environment described in #4588, and it has not measured that any particular kernel version always succeeds. If you are using Clash Verge Rev TUN or an ordinary system proxy, the configuration and takeover path differ — do not apply these Linux rule conclusions there.
Sources
Official TUN configuration field documentation; sing-box issue #4588 (OpenWrt 23.05-SNAPSHOT, Linux 5.4.213, nftables 1.0.8, sing-box 1.14.2, still open at the time of checking). Checked 2026-11-13.
Frequently asked questions
sing-box check passes, so why does it fail at runtime?
A static configuration check is not the same as the kernel accepting auto-generated nftables rules. Keep both the check output and the real startup log — they locate two different stages.
Does 'missing default interface' mean Linux 5.4 is unsupported?
No. The isolated namespace in the public issue shows that warning together with the fatal nftables error, but the default-interface question needs its own look at the routing environment. Without the explicit auto-redirect: setup nftables chain, it is not the same fault.
nfnetlink_queue is loaded but I still get 'operation not supported'. What now?
Identify the first rejected rule and your system version. The reporter in #4588 had the relevant module loaded and still failed on a particular TCP reject rule — a module being present does not guarantee every rule form is supported.
Can I just turn auto_redirect off and keep using it?
Only as a controlled comparison, and only with the original configuration kept and a way to restore networking. The official documentation notes that this option changes Linux TUN routing; after disabling it you still have to verify device traffic, DNS and egress rather than judging by the service starting.
Does this mean my provider's nodes are broken?
The error occurs while installing TUN rules on the local machine, which is too early to judge a remote node. Fix core startup and routing first, then make a real HTTPS request.
Continue reading
Hiddify WARP: netip.ParsePrefix with an empty IP · set system dns failed: exit code 1 · Full error string index