| AI-First Design |
Experimental |
Self-describing CLI-as-API with MCP transport for AI assistants. ze create agent skill installs the ze skill in Claude Code, Codex, OpenCode and Pi, and a session-start hook in Claude Code and Codex, so an agent reads the guide the installed binary serves. |
| API Commands |
Experimental |
Peer management, route updates, RIB operations, event subscription |
| Appliance Host Name |
Experimental |
On the appliance build, Ze sets the kernel host name (what uname -n, syslog and the shell prompt show) from system host at start and on every reload that changes it, so each appliance carries its own name rather than the image's build-time ze. With system host unset, the kernel keeps the image's name. The value is validated as an RFC 1123 host name before it is applied. A sethostname refusal leaves the kernel name as it was, logs one warn line naming the value, the error and the fix, and ze doctor reports doctor-system-host-not-applied; the commit is not refused. A distro build (any build without ze_appliance) never calls sethostname: the operating system owns the name there, and system host changes only Ze's own identity. The booted proof is ./le test qemu appliance-hostname. |
| Archive Pruning |
Supported |
commit-revisions (0 to 1000, default 0) caps the config archive files kept at each archive location, pruning the oldest after each archive write; 0 keeps every file, and an HTTP archive is never pruned. Only files whose whole name matches the archive block's filename format are counted or removed. . file:// |
| ASN to Regional Internet Registry |
Experimental |
ze resolve rir <asn> and show resolve rir <asn> answer which registry holds an AS number, and that registry's whois host. Both read a delegation table the binary ships as embedded data, so the answer needs no network, no daemon and no file on disk. update resolve rir refreshes that table from the five registry delegation files into the managed store under meta/rir/delegation, and the stored copy answers whenever its generation date is later than the shipped seed's, so an upgrade carrying fresher data takes over on its own. The refresh is all or nothing: a registry that does not answer, a file the parser refuses, and a run that read no ASN record each store nothing and leave the previous table answering. An AS number in no delegated range and a table that cannot be read stay two distinct answers on every path. An appliance behind a mirror names where each file is read from, per registry, under system/rir: a registry with no block is read from the file it publishes, and a mirror is HTTPS or plain HTTP from the host itself. The stored table names the URLs that run read in its Source: lines. ./le data asn-delegation update rewrites the shipped table through the same fetch. |
| Autonomous Anomaly Response |
Experimental |
Shadow-first responder (anomaly/shape) that subscribes to anomaly-detect incidents and installs a surgical per-SOURCE firewall action (rate-limit via firewall.MatchSourceAddress+Limit, drop fallback). SHADOW is the default: it logs the would-be action and installs nothing. In armed mode each anomalous source gets its own live nft rule with a mandatory timed AUTO-REVERT (withdraws after a TTL regardless of any clear event), a global BLAST-RADIUS cap (refuses to arm beyond N), a KILL-SWITCH (reverts all + forces shadow), and an ALLOWLIST (protected sources are never armed). One mutex guards the armed map; a per-timer generation guard makes a superseded timer a no-op. Status via show anomaly shape. Separate firewall owner key isolates it from ddos/local. A ze doctor check (doctor-anomaly-shape-armed-no-firewall) warns when armed without a firewall. Prometheus: ze_anomaly_shape_armed, ze_anomaly_shape_reverted_total, ze_anomaly_shape_arm_refused_total, ze_anomaly_shape_killswitch_total. |
| Backend-Aware CLI Completion |
Experimental |
CLI auto-completion filters options based on the active backend. Config editor mode (set/delete/edit/show) and operational command mode (show/clear/monitor) hide nodes annotated with ze:backend when the active backend is not in the annotation's list. Backend names are derived from the candidate tree at each config completion call, and handed to the operational command completer after each edit. Same ze:backend annotations used by commit-time validation, applied earlier at completion time. |
| Backup and restore |
Supported |
ze data backup <file> [spare <n>] copies every key of the live store to one new 0600 blob artifact under the write lock, so a commit is in the backup whole or not at all. ze data restore <file> config [name <source-name>] commits one config from a verified artifact as a new version of this device's config and leaves credentials, identity and history unchanged. ze data restore <file> full replaces the whole tree, moves the previous tree to database.replaced-<stamp>, and records a durable intent, so an interrupted restore is refused by every opener until the same command finishes it. Both verbs are offline; while a daemon owns the store, request data backup path <file> and request data restore path <file> config run over SSH, and a restored config becomes active only when the reload accepts it. ze init --from <path-or-url> [--sha256 <hex>] creates a store from a blob, with the digest checked before a key is written. ze config edit --backup <artifact> edits a config inside a backup with no daemon. A damaged store is checked and salvaged while ze runs: ze data check and ze data repair --output <path> on the host, and show data check and request data repair path <file> through the daemon, which writes the salvage on its host with its writes held for the walk; installing the salvage needs ze stopped. The store is maintained while ze runs, beside backup and restore: ze data rm, ze data write and ze data import on the host send their work to the daemon (delete data key, update data key, update data config), which removes or writes runtime state and imports a config as a new version, through the reload for its own config. Offline, ze data import of a config the store holds commits a new active version and keeps the previous one as rollback. |
| Behavioral Anomaly Detection |
Experimental |
Darktrace-style SECURITY anomaly detector (report-only), a domain separate from volumetric DDoS. Consumes the neutral trafficfeature signals and learns each entity's own pattern-of-life via a per-(entity,feature) EWMA baseline (internal/core/stats), scores self-deviation plus peer-group rarity (prefix cohorts), and correlates multiple weak feature deviations on one entity into a single incident (capped/discounted combine, not naive sum). An entity is a SOURCE (an anomalous sender), a DEST (a distributed sink or a probed host) or a PORT; a port is identified by proto and port rather than by an address, and is scored by self-deviation alone because it has no cohort. A confirm/clear state machine debounces; confirmed incidents emit on the anomaly-detect event bus and land in a bounded recent-incident ring viewed via show anomaly detect. It takes NO action (the anomaly/shape responder acts); scoring is bounded per-entity with idle eviction. A ze doctor check (doctor-anomaly-detect-no-feature-source) warns when enabled without a flow source. Prometheus: ze_anomaly_incidents_total, ze_anomaly_active, ze_anomaly_tracked_entities (labeled dimension=source/dest/port since 2026-08-18; it was a bare gauge, so an existing query stops matching and must sum over the label). The detector's ring records confirmations only, so a separate anomaly-observe plugin subscribes to the same events and keeps the incident LIFECYCLE: a bounded ring (incident-ring-size) opened on AnomalyDetected and finalized with an end time on AnomalyCleared, or by a one-second stale sweep (stale-incident-timeout) when a source goes silent and the detector evicts it without a clear. show anomaly observe returns that list newest-first with finalized incidents included, which is where a finished incident's duration is readable. The detector, the lifecycle store and the shape responder are all compile-out-able with the ze_anomaly build tag: default-on in ZE_FEATURES, dropped from ze-stripped / bare ze_core builds, with an anomaly {} block rejected as unknown. |
| CLI Commands |
Experimental (partial) |
Protocol tools, config management, schema discovery, daemon control, AS topology graph, policy dry-run testing (show policy test). Verb-first is now the only spelling: every command starts with show, clear, monitor, request, set or delete, and the words after it are the YANG path. The bare forms are gone from the command tree, not aliased, so the dispatcher answers unknown command for daemon reload, daemon status, daemon quit, bgp summary, system memory, interface rate and the rest; the verb-first spellings are request reload, show status, request halt, show bgp, show system memory, show interface rate. Three exec verbs stay bare on purpose (stop, restart, reboot): the SSH exec middleware intercepts them before the dispatcher, which registers no key for them. show interface reshaped its keywords too: name <n> detail and name <n> counters replace detail <n> and counters <n>, and brief, scan, type, errors and rate each answer a distinct question. Scope gaps: declared commands without leaves (plan/release/fixes/spec-declared-commands-without-leaves.md). |
| CLI Default Output Format |
Experimental |
Configurable default output format via environment { cli { format { default text; } } }. Supported values: text (default), table, json, yaml, ndjson. Session override via set cli format <value> in operational mode. Explicit pipe operators (| json, | table, etc.) always win over the configured default, and over the --format flag of ze cli -c. |
| CLI Session Transcript |
Experimental |
Transcript recording of ze cli and ze config edit sessions to, written by the process that runs the session's model: the daemon for a session on a stored configuration, the client for ze cli -c. Preserves command input and output for post-disconnect recovery, with credential tokens on the command line replaced by <redacted>. Enabled via environment { cli { transcript enabled } } config or ze.cli.transcript env var. Best-effort writes never block CLI operation. . $XDG_DATA_HOME/ze/transcripts/ |
| Commit-Time Backend Capability Check |
Supported |
YANG nodes that correspond to backend-specific features carry a ze:backend "<names>" annotation. On commit (daemon reload, first-apply, and ze config validate), the walker rejects the config with the YANG path and the list of supporting backends whenever the active backend does not implement the feature -- instead of letting an Apply-time "not supported" error fire inside the backend. The gate covers interface (netlink-only bridge, tunnel and veth under the vpp backend), traffic/control (the backend-leaf CALL wired into OnConfigure/OnConfigVerify; tc-only feature annotations ship with spec-fw-7-traffic-vpp), and firewall (seven ze:backend "nft" annotations on conntrack-driven matches and nft-only action/modifier leaves). |
| Config Dependency Graph |
Supported |
ze config graph visualizes config dependency relationships. |
| Config Schema Stamp |
Supported |
Config files carry a schema version stamp. When a config written by a newer release fails to load at startup, recovery keeps it in history, walks the rollback versions newest first, skips those a newer release wrote, loads the first that parses, and re-stamps it with this release. No field is pruned. |
| Configuration |
Experimental (partial) |
YANG-modeled config with prefix limits, update groups, session resilience, duplicate list-key rejection, side-effect-free in-process plugin verifiers for static/API/CLI validation, and transactional commits using active/candidate/rollback pointers. Live external plugin OnConfigVerify callbacks run only in daemon reload/commit transactions. Scope gaps: live external plugin OnConfigVerify runs only inside daemon reload and commit transactions; YANG types bits and identityref and instance-identifier and leafref (plan/pre-release/spec-config-yang-type-.md); YANG deviation and if-feature and submodules (plan/pre-release/spec-config-yang-.md). |
| Configuration Reload |
Experimental (partial) |
Live reload via SIGHUP with automatic reconciliation. Reload stages edited config as a candidate version and promotes it to active only after runtime reload succeeds. Plugin-server transactions, config-provider roots, subsystem reload, and changed external plugin replacement roll back on failure; remaining reload safety depends on component-specific journals and privileged dataplane evidence. Scope gaps: reload safety that depends on component-specific journals; privileged dataplane rollback evidence. |
| Connected Routes |
Experimental |
Redistribute directly connected interface prefixes into BGP via redistribute { import connected }. Subscribes to interface address events; emits RouteChangeBatch on address add/remove. Reference-counted: multiple addresses on the same prefix emit one announcement, withdrawn only when the last address is removed. IPv4 and IPv6. The prefix is also a path in the shared Loc-RIB, so rib { distance { connected N } } ranks it against BGP, OSPF, IS-IS and static for the same prefix, and it makes a next-hop covered only by an interface prefix resolvable. Ze still programs no kernel route for it: the protocol declares that the OS installs its entries, so a connected winner WITHDRAWS the route Ze had programmed rather than adding a second one beside the kernel's own. |
| Connection Tracking Management |
Experimental |
Declarative conntrack configuration under system { conntrack {} }. Conntrack helpers (ALGs) assigned only when listed: each module (amanda, ftp, h323, irc, netbios-ns, pptp, sane, sip, snmp, tftp) gets a ze-owned nft table with a ct helper object per kernel helper and transport, assigned by destination port in prerouting and output at priority -5, in an inet table for a helper the kernel registers for both families and an ip table for an IPv4-only one; an empty list assigns nothing. Ze's kernel builds every helper in; on a distribution kernel the module is loaded with modprobe. show system conntrack lists the assigned helpers. User-friendly config for table sizing (table-size, hash-size, expect-max), per-protocol timeouts (TCP, UDP, ICMP, ICMPv6, GRE, SCTP, DCCP), TCP behavior flags (be-liberal, loose, max-retrans, ignore-invalid-rst), and global flags (accounting, timestamp, checksum, log-invalid). All sysctl values routed through the sysctl plugin for three-layer precedence. Dual-setting prevention rejects keys in sysctl {} that conntrack manages. On gokrazy (modules built-in), module loading is skipped gracefully. CLI: show system conntrack. Telemetry: configured-max gauge alongside existing per-CPU counters. |
| Control-Plane Policing (CoPP) |
Experimental |
Rate-limit new TCP connections to the BGP listen port (TCP/179) to protect against connection-flood DDoS. Generates an nft input-hook chain via the firewall registry: established/related sessions pass at full rate, operator-supplied trusted-source prefixes bypass the limit, new connections are rate-limited. Configurable rate, burst, protected-port (default 179), trusted-source prefix list, and over-limit policy (accept/drop, default accept). The over-limit policy rides the rate-limit term as an nftables inverted limiter. The input chain keeps an accept policy under both settings, so no CoPP config locks the operator out. Doctor check verifies the CoPP input chain is active when configured. Compile-out-able with the ze_copp build tag: default-on in ZE_FEATURES, dropped from ze-stripped / bare ze_core builds, with its config block rejected as unknown. |
| Core Diagnostics |
Experimental (partial) |
Built-in diagnostic commands replacing ss, dmesg, lsof, dig, nc, traceroute, mtr, ping, tcpdump, and pprof on gokrazy appliances: show system sockets (TCP/UDP state), show system kernel-log (dmesg), show system goroutines (dump with singleflight dedup), show tcp-check (port probe), show traceroute (ICMP path trace with per-hop RTT, IPv4/IPv6), monitor traceroute (live mtr-style continuous trace with | log | resolve and | log | origin enrichment), monitor ping (continuous ICMP ping with live stats), show capture interface (AF_PACKET live capture with BPF filters, pcap or text output), show system file-descriptors (FD counts and limits), show dns lookup/cache (DNS resolution, cache listing, selective delete, flush, stats reset), show system profile (cpu/heap pprof), show system memory (/proc/self/status OS view). monitor system netlink streams kernel route/link/address change events as JSON (replaces ip monitor). A BFD raw capture ring is started and dumped with the other protocols when show capture raw names no protocol; its protocol leaf offers only l2tp and bgp, so BFD cannot be selected alone. At startup Ze warns when it runs without root or without a required capability, and keeps running. See Production Diagnostics Guide. Scope gaps: BFD raw capture cannot be selected alone. |
| Crash Capture |
Experimental |
Automatic stderr redirect captures panic stack traces from any goroutine. Forwarded to syslog (via ze.log.destination) in real time and persisted to crash files on disk. Crash reports include ring buffer context (last 64 log entries before the panic), version, build date, uptime. Crash dir autodetected (on gokrazy, fallback chain for other platforms). CLI: show crashes, show crashes latest. Env vars: ze.crash.dir, ze.crash.keep. Kernel panics are captured too: system crash-dump enabled true plus an appliance built with image.crash-dump reserves a named memory region the kernel writes its own panic message and backtrace into, which a warm reboot does not clear. Ze harvests the record on the next boot as a kernel-kind report in the same directory, under the same retention count, listed by the same show crashes. No Ze code runs at the moment of the fault. show crashes reports configured and armed separately, because a reservation is a boot argument: three ze doctor checks report the reservation, the record store and the crash directory. A full memory image is a separate opt-in and is amd64-only. Prometheus: ze_crashes_artifacts and ze_crashes_kernel_capture_armed, so a fleet can alert on capture being silently unarmed. . /perm/ze/crash/ |
| Custom Value Validators at Startup and Reload |
Experimental |
A YANG leaf can name a custom validator with ze:validate, and the named function runs on the same walk the offline ze config validate uses, so the offline check and the daemon cannot disagree about the same bytes. The set covers address families, community ranges, event and message type tokens, MAC addresses, IS-IS NET / system-id / hostname, OSPF router-id and area-id, redistribute sources, IPv4 and IPv6 addresses and prefixes, set references, port specs, and internal plugin names. A ze:validate binding with no registered function is an integrity failure reported at schema build, so a leaf cannot silently declare a validator that does not exist. |
| Deactivate / Activate |
Experimental |
Junos-style inactive: prefix on any node (leaf, container, list entry, leaf-list value); kept in file, skipped at apply. CLI: ze config deactivate/activate <file> <path>. TUI: deactivate <path> / activate <path>. Engine-level, no schema annotation required. |
| Don't Fragment probes |
Experimental |
show ping, resolve ping, show traceroute and resolve traceroute take do-not-fragment honor-cache or do-not-fragment bypass-cache. The probe socket installs IP_MTU_DISCOVER and IP_RECVERR, the Linux kernel runs RFC 1191 and RFC 8201 path-MTU discovery, and the reply that a router refused carries status: too-big, next-hop-mtu-reported and next-hop-mtu read off the socket error queue, never parsed from ICMP by ze. A zero or an IPv6 value below 1280 is reported as no usable value. bypass-cache ignores the kernel's cached estimate so the wire answers. Without CAP_NET_RAW, ping runs on the unprivileged datagram ICMP socket when the daemon's group is inside net.ipv4.ping_group_range; ze doctor reports doctor-icmp-probe when neither socket opens and doctor-icmp-probe-unprivileged when only the datagram one does (traceroute needs the raw socket). Linux only. |
| Dynamic peer groups |
Experimental |
A group whose connection/remote/ip is dynamic accepts a session from any address inside its range list and opens its own listening socket, so a configuration naming no static peer still accepts members. Each accepted connection becomes one peer that inherits the group's whole resolved settings, its families, its attach process blocks and its per-peer plugin config. Overlapping ranges resolve by longest prefix match. max-peers bounds the group (1..100000, default 1000). |
| Environment Variables |
Experimental |
Ze-native env surface: ze.user, ze.pid.file, ze.pprof, ze.bgp.openwait, ze.bgp.announce.delay, ze.cli.format; ExaBGP-compat env keys retired 2026-04 |
| ExaBGP Compatibility |
Experimental |
Automatic config migration and plugin bridge. The bridge plugin and the ze exabgp command, which carries ze exabgp migrate, are compile-out-able with the ze_exabgp build tag (default-on in ZE_FEATURES). |
| Firewall |
Experimental |
Packet filter and NAT via the nftables backend on Linux. The abstract model's match types are the Match implementations in: source/destination address, port ranges, protocol, input/output interface with wildcard prefix, connection state, marks, DSCP, ICMP type, ICMPv6 type, TCP flags, named sets, IPv4 TTL and IPv6 hop limit below a threshold, the TCP port and destination quoted inside an ICMP error, ARP operation and sender address, and ND target address, plus two predicates only the daemon builds: link-layer unicast (VRRP) and a route into an IPsec transform (the IKE tunnel-source guard). Its action and modifier types are the Action implementations: accept, drop, reject, jump/goto/return, SNAT/DNAT with address ranges, masquerade, redirect, notrack, flow offload, conntrack helper assignment (built by system conntrack module, not by a firewall { } term), set mark/connmark/DSCP/TCP-MSS, TCP-MSS clamped to the route MTU (tcp-mss-clamp-to-pmtu, forward, output and postrouting chains only), counter, log, rate limit. The lan-services topology scenario proves tcp-mss-set and masquerade on a CPE converted from VyOS, and road-warrior proves masquerade of WireGuard client traffic. NAT exclude rules emit a Return verdict to skip translation. Global-options container maps keyword toggles (all-ping, broadcast-ping, syn-cookies, source-validation, etc.) to kernel sysctls via the sysctl plugin's default layer; explicit sysctl settings always override. IRR-based prefix-list filtering: source-asn, source-as-set, destination-asn, destination-as-set leaves in from-block resolve ASN/AS-SET references to nftables interval sets via the shared IRR PrefixStore (firewall-irr plugin). Per-interface source validation: bind an AS-SET to a customer-facing interface via irr { interface <name> { source-as-set <AS-SET>; } } to drop ingress traffic with source addresses not in the IRR-resolved prefix set (BCP 38). An IRR answer carrying no prefixes never replaces cached ones: the store keeps the last known good list, show firewall irr reports the entry as stale with the age of the data, ze doctor reports it, and clear firewall irr asn|as-set is the only path that removes prefixes. A binding with no prefixes produces no table at all, so an empty answer cannot leave an interface dropping everything. DNS-sourced address groups: a domain-group holds domain-names, and source-domain-group / destination-domain-group leaves in the from-block match the addresses those names resolve to (firewall-domain plugin). Each name is re-resolved on its own answer TTL, per family, clamped by ttl-floor, and the set is reprogrammed only when the addresses actually changed. Only NXDOMAIN empties a set: a SERVFAIL, a REFUSED or an unreachable server keeps the last good addresses, so a DNS outage is never enforced as a filter. Resolved addresses persist in ZeFS, so a restart with DNS down resumes filtering from them, and every change is appended to a bounded JSON-lines log at <config-dir>/firewall-domain-group.dns.jsonl. show firewall ruleset shows the DNS name beside each address through the show-enricher registry, adding no field to the shared set model. Component reactor wires into ze's engine lifecycle: Apply on boot and reload, rollback on failure. ze_ prefix on all kernel tables. Every backend Apply is bounded by a deadline: 10s by default, clamped to 1s. 60s (ze.firewall.nft.netlink-timeout, ceiling firewall.MaxBackendDeadline). The bound is load-bearing rather than cosmetic, because ApplyAll holds one process-wide reconcile lock across Backend.Apply, so an unbounded kernel call would stall every firewall owner (copp, policy-routes, ddos-local, the firewall engine) and not just one. A wedged reconcile is counted, never hidden: ze_firewall_apply_timeout_total, and ze_firewall_apply_duration_seconds{result} with ok, timeout, error and panic. . internal/component/firewall/model.go |
| Fleet Management |
Experimental |
Centralized config distribution over TLS |
| FlowSpec-to-Firewall Bridge |
Experimental (partial) |
flowspec-firewall plugin converts BGP FlowSpec rules into nftables entries in a dedicated ze_flowspec table. Scope gaps: RFC 8955 and RFC 8956 predicates the ledger marks Partial. |
| Graceful Listener Migration |
Supported |
Hot reload of listener endpoints (web, LG, REST, gRPC, MCP, SSH). New listener starts before old one stops; in-flight connections are drained. |
| Granular Debug |
Experimental |
Verb-first debug (set/delete/show/clear, matching VyOS syslog-level config) with per-module flag/direction/scope filtering and named profiles. CLI: ze set debug module <name> (enable), ze delete debug module <name> (disable), ze set debug module <name> flag <flag>, ze set debug module <name> scope direction <dir>, ze set debug module <name> level <level>. Hierarchical prefixes work (ze set debug module bgp covers all bgp.* subsystems). Named profiles: ze set/delete debug profile name <name>, ze show debug profile [name <name>], ze set debug active name <name> (apply). Stored in debug.zefs (separate from config). Each plugin declares its valid flags via the debug YANG registry. Not auto-applied on reboot (safety). show debug (YANG-dispatched) queries live daemon state. |
| Health Registry |
Experimental (partial) |
Aggregated component health via show health and /health HTTP endpoint (503 when any component is down). Registered components: l2tp, report-bus, ipsec, pki, bgp (session-stuck/flap/EOR), fib (sync-failure/orphan/lag), firewall (stale-table/drift audit), iface (error counters), plugins (crash/disabled), vpp (API socket probe), flow-export. The l2tp check reports healthy whenever the subsystem is running and does not probe the kernel data plane. Each check is expected to return within 1 second by contract; the registry recovers panics but does not enforce a timeout. Scope gaps: l2tp health check does not probe the kernel data plane. |
| Host Inventory |
Experimental |
Structured hardware inventory for ISP fleet monitoring: CPU (vendor, topology, hybrid P/E layout, scaling driver, frequencies, throttle counts), physical NICs (driver, PCI IDs, link speed, queue counts, firmware, rings), DMI board identity, memory with ECC counters, hwmon thermal sensors + per-CPU throttle, block devices with NVMe firmware, kernel release/cmdline/microcode/arch flags. Read-only sysfs/procfs, no daemon required. Single show host cpu/nic/... command served by the daemon when reachable, falling back to the same in-process detection when no daemon is running; JSON by default for pipeline consumption. |
| Interactive Launcher |
Experimental |
Running ze with no arguments in a terminal shows a BubbleTea menu of all commands grouped by section, with type-ahead filtering, scrolling, and drill-down into YANG verb sub-commands. Non-TTY invocations fall back to static help text. |
| Interfaces |
Experimental |
The PPPoE client answers access-network LCP renegotiation on the existing session. Standalone tests cover repeated authentication, IPCP and complete result notifications. The interface consumer reapplies changed MTU, local/peer addresses and the default route; consumer tests cover reapplication and failure handling. Target-kernel proof and wholesale interoperability proof remain outstanding. The PPPoE client asks IPCP for the access network's DNS servers (RFC 1877 Primary and Secondary DNS), adopts the unicast values the access concentrator Naks back, and keeps the session up when it Rejects them. The learned servers become the session's resolver record, replaced when a renegotiation changes them and withdrawn when the session ends, and show interface pppoe-client reports them as dns-servers. The accel-ppp interop scenarios 01-pppoe-chap-ipv4 and pppoe-ipcp-dns-rejected-ze-client prove both halves. The PPPoE client's dead-peer detection and redial hold-off are configurable on each pppoe-client entry: keepalive { interval; failures } sets how often an LCP Echo-Request is sent (1 to 3600 seconds, default 10) and how many unanswered requests in a row end the session (1 to 255, default 3), and reconnect { delay-min; delay-max } bounds the doubling wait before the next discovery (1 to 3600 seconds, defaults 1 and 60). delay-max below delay-min is refused naming both leaves. A reload that changes only these leaves keeps the running session up: the new values apply from the next session and the next redial wait. The PPPoE client carries IPv6 when its ipv6 enabled is true: it negotiates IPv6CP beside IPCP (RFC 5072), each Configure-Request carrying one Interface-Identifier. A valid identifier the access concentrator Naks back is adopted for the next request, and a fresh one is chosen when the suggestion equals the one Ze last Naked (RFC 5072 Section 4.1). Once IPv6CP opens, the session device passes IPv6 and holds the link-local address formed from the negotiated identifier. An access concentrator that sends an LCP Protocol-Reject for IPv6CP, Configure-Rejects the identifier, or never answers stops IPv6CP only: IPCP and IPv4 stay up, and a rejected identifier is logged at WARN. The accept-ra, autoconf and forwarding leaves become the net.ipv6.conf.<name>.* sysctls of the session device each time a session creates it, so the client can take its default route and a SLAAC address from the access concentrator's Router Advertisements. A dhcpv6 block runs the DHCPv6 client on the session device, with its DUID-LL and IAIDs built from the source interface unless duid is set, and pd { length } asks for a delegated prefix. A LAN unit numbers from that prefix with ipv6 delegated-prefix <pppoe-client> { subnet-id; interface-identifier }: it holds the /64 at subnet-id with the remaining lifetimes of the delegation, and its Router Advertisements follow the CE rules of the IPv6 Router Advertisements feature. Config verify refuses a subnet-id the delegated length cannot hold, two units taking one id from one source, and a source that is not a pppoe-client whose dhcpv6 asks for a prefix; the editor completes the source from that same set. ze_pppoe_client_ipv6cp_opened{name} and ze_iface_delegated_prefixes_held{name} report the state. The accel-ppp scenarios pppoe-ipv6cp-ze-client and pppoe-ipv6cp-rejected-ze-client prove the negotiated link-local, the RA default route and SLAAC, and an IPv4 session that survives a rejected IPv6CP; the pppoe-dhcpv6-pd-ze-client topology scenario proves the delegation reaching a LAN host., "PPPoE client IPv6", is the operator guide. Linux interface management via netlink: ethernet, dummy, veth, bridge, loopback, VLAN (with 802.1p ingress/egress QoS maps translating PCP to internal priority and back, and named class-of-service profiles via the cos plugin with interface-level inheritance and per-unit override/opt-out; dynamic per-subscriber CoS via RADIUS Filter-Id "cos:<name>" or vendor-specific attributes (Cisco-AVPair, Juniper ERX, Nokia, Huawei) applied on session-up with mid-session CoA changes and session-down revert; MikroTik Mikrotik-Rate-Limit VSA for shaper rate extraction ), 9 tunnel kinds (GRE, GRETAP, IP6GRE, IP6GRETAP, IPIP, SIT, IP6TNL, IPIP6, VXLAN with a 24-bit VNI over UDP 4789 per RFC 7348), each carrying an outer-header TTL of 64 when the config names none, so an encapsulated packet survives a multi-hop underlay; ttl 0 still asks the kernel to inherit the inner packet's TTL, and the default reaches netlink-backed tunnels only, because the VPP tunnel calls carry no TTL field, so a commit that names a ttl on a vpp-backed tunnel is refused by path rather than accepted and dropped , and WireGuard (declarative peers with $9$-encoded keys); DHCP (config-driven, routes, DNS, NTP discovery), NTP client (clock sync, RTC, time persistence, max-step safety cap), monitoring, migration, mirroring, per-interface offload and steering (GRO, GSO, SG, TSO, LRO, hw-tc-offload via kernel ethtool ioctl; RPS, RFS via sysfs; boolean three-state: true/false/absent), per-family reverse path filtering (rpf-check strict|loose|disable in ipv4/ipv6 containers, backed by rp_filter sysctl on Linux; IPv6 enforcement requires VPP), per-interface rate tracking (1s sampler with show interface rate, monitor interface rate, 12 ze_interface_* Prometheus gauges, web rate columns). Local privileged integration covers first-apply non-adoption, reload deletion scoped to previously managed links, and rollback of created kernel links after partial apply failure; target-runner evidence is still required before production deployment claims. Every kernel route the interface layer installs (DHCP, RA, PPPoE, PPP) carries rtm_protocol 253, the ze-iface producer id, and a delete matches on that protocol, so one producer's expiry cannot remove an operator static route. A default route learned on a unit takes metric 254 by default (route-priority, range 0. 4294966271), on an interface unit and on a pppoe-client. The NTP client is compile-out-able with ze_ntp (the show system NTP block reads it through the nil-safe registry seam) and the class-of-service profiles with ze_cos; the dynamic per-subscriber CoS handler additionally requires ze_l2tp (a BNG feature: no sessions, no dynamic CoS). . docs/guide/pppoe.md |
| IPsec CLI and Diagnostics |
Experimental (partial) |
show vpn ipsec sa/status/peer, clear vpn ipsec sa all, clear vpn ipsec sa peer <name> (the bare form is refused), monitor vpn ipsec (live SA event stream). show vpn ipsec dataplane sa/policy/drift reads Linux XFRM state and compares expected Child SAs with the SAD. VPP and noop report unsupported enumeration. SA displays carry kernel byte and packet counters, with null for unavailable measurements, and show vpn ipsec sa lists a peer's leased inner address under virtual-ip. Web page at with SA table. Health reports degraded for drift or an unknown dataplane observation. Prometheus retains the belief gauges ze_ipsec_sa_count, ze_ipsec_tunnel_up, and ze_ipsec_tunnel_degraded. Kernel gauges ze_ipsec_dataplane_sa_count{if_id} and ze_ipsec_dataplane_drift{peer} remove their series when observation fails. Show commands reach the shared pipe layer; ze help command --json publishes each command's operators. Address operators require declared address fields. ze_ipsec_virtual_ip_leases{pool} counts the virtual IP addresses live IKE SAs hold in each remote-access pool, and ze doctor warns under ipsec-virtual-ip-pool when one family of a pool is at least 90% held, and under ipsec-unauthenticated-input for a peer whose handshake in progress dropped messages that failed their integrity check, and under ipsec-peer-malformed-message for a peer whose live IKE SA received authenticated messages that do not parse. Scope gaps: monitor vpn ipsec; the web page and the Prometheus gauges have no real-path test. /show/vpn/ipsec/ |
| IPsec Data Model |
Experimental (partial) |
YANG vpn { ipsec {} } config for site-to-site VPN: ESP groups (proposals, lifetime, PFS), IKE groups (proposals, dead-peer detection, key-exchange, close-action; close-action and the dead-peer-detection action are parsed but every value produces the same behavior, and key-exchange ikev1 still negotiates IKEv2), site-to-site peers (X.509 and PSK auth, VTI bind, group references). A peer's remote-address is an IPv4 literal or a DNS name, resolved again on every connect cycle; its local-address is an address literal or absent, and commit refuses anything else (a keyword such as any, or a name). With no local-address the peer sends from the vpn ipsec interface address, waiting while the interface has none and restarting when it changes, or, with no interface, from the source the routing table picks toward the peer. Algorithm enums match strongSwan naming. Cross-reference validation (group names, PKI certificates, interface binding, local-id/CN match). Operator-authored Security Policy Database entries under vpn ipsec policy, carrying the BYPASS and DISCARD dispositions of RFC 4301 Section 4.4.1 with a selector, a direction and a total order. Named virtual-IP pools under vpn ipsec remote-access pool, each a range and range6 with DNS servers; a respond peer names one with virtual-ip-pool, and commit refuses a name no pool carries and two pools whose ranges overlap (RFC 3948 Section 5.1). Config diff detection for reload. Scope gaps: close-action and dead-peer-detection action values all produce the same behavior; key-exchange ikev1 still negotiates IKEv2. |
| IRR BGP Import Filtering |
Stub-backed |
bgp-filter-irr builds per-ASN import prefix-lists from explicit or PeeringDB-discovered IRR AS-SETs, refreshes the lists and keeps them in memory only (a restart resolves them again), and rejects received prefixes outside the selected list. Includes status, prefix inspection, dry-run checks, and manual refresh commands. |
| Kernel Routes |
Experimental |
Redistribute externally-installed kernel routes into BGP via redistribute { import kernel }. Consumes parsed route events from a shared netlink route watcher. Routes Ze installs itself are filtered out (the protocols rtproto.IsZe names: GTSM 249, fib-kernel 250, static 251, policy-route 252), as are RTPROT_KERNEL (2) and RTPROT_REDIRECT (1). A route of every other protocol is emitted as a RouteChangeBatch, for example DHCP (16), PPP/manual (BOOT=3) and admin static (STATIC=4) routes. Tracks announced prefixes; withdraws all on shutdown. IPv4 and IPv6. Shares a single netlink subscription with fib-kernel (route re-assertion). . internal/core/routewatch/ |
| Kernel Tunable Management |
Experimental |
Sysctl plugin centralizes kernel parameter management with three-layer precedence (config > transient > default). Plugins declare required defaults (e.g., fib-kernel enables forwarding), users override via config or CLI. Original values restored on clean stop. Named profiles group co-dependent tunables (dsr, router, hardened, multihomed, proxy) applied per interface unit. User-defined profiles supported. CLI: show sysctl, ze sysctl list, ze sysctl describe, set sysctl, ze sysctl list-profiles, ze sysctl describe-profile. |
| Live Credential Revocation |
Experimental |
Every credential surface reads the running config's user list, so a user the config stops declaring stops authenticating at the next reload: the web password, the web session cookie, the SSH password, the SSH public key, and REST/gRPC Bearer. The revoker is identified rather than assumed. An authentication result names the backend that produced it, and only a result the local backend granted is revoked by the local list; an empty source is not local, because reading silence as local would attach the local list's revocation to a session some other backend granted. A session a RADIUS or TACACS+ backend granted is therefore not revoked by an edit to the local users. Revocation governs new authentication; an open session survives by design. |
| Local Audit Trail |
Experimental |
Structured local audit log for config commit/discard, daemon reload, and failed authentication across SSH, web, REST, gRPC, MCP, CLI, and system surfaces. show audit filters by action, actor, surface, time range, and count. Disk-backed JSON-lines storage is used when Ze starts from a config file; stdin configs use memory-only storage. |
| Looking Glass |
Experimental |
Public BGP looking glass with birdwatcher API, AS path graphs, and BMP-monitored route display. Runs on its own configurable port, separate from the web UI. tls defaults to true. The certificate leaf names a pki { certificate <name> } entry, and the listener then serves that leaf with every intermediate the store holds. An unset leaf keeps the self-signed certificate. An optional token gates every route and is compared in constant time. An unset token leaves the looking glass open, which is the default. Pages render through templ with typed view models, not html/template. |
| Management Listener Exposure Guard |
Experimental |
One boot-time, fail-closed check refuses to start a management service on a non-loopback address without authentication. It covers the insecure web listener, MCP, gNMI, and the API server (REST and gRPC). It runs once, after every address and credential is resolved and before anything binds, so a refusal leaves nothing bound and the process exits non-zero. Fail-closed on classification too: a wildcard (0.0.0.0, ::), an empty host, and any DNS name including localhost all count as non-loopback, so remote reachability cannot be smuggled past it through a name. An unauthenticated surface that declares no resolved address is refused rather than passed by iterating zero times, which is the empty-set trap that once let an insecure web server reach. The refusal names the service, the address, and the fix, and never prints a token. For MCP the guard mirrors the server's own precedence: a token written beside an explicit auth-mode none does not authenticate, because the server builds its accept-all authenticator in that case. . 0.0.0.0:3443 |
| Memory Lock |
Experimental |
The memlock plugin locks the running ze executable's file-backed pages with mlock2(2) and MLOCK_ONFAULT, so the kernel cannot evict the daemon's own text under memory pressure and fault it back from disk. Linux only. The lock is taken in the package init(), before main(), so every ze process holds it from its first instruction. The whole mapped size is charged against RLIMIT_MEMLOCK, which the generated ze.service unit lifts with LimitMEMLOCK=infinity; a limit too small for the binary makes the lock fail, and show plugin list reports the memlock row as soft-failure carrying the cause and the remedy. The failure is soft: the daemon serves every session correctly with an unlocked executable. A ze doctor check answers the question the record cannot, before ze runs: it compares RLIMIT_MEMLOCK against the size of /proc/self/exe and warns under doctor-memlock-rlimit-low when the limit cannot hold the executable, staying silent for a process that holds CAP_IPC_LOCK. |
| Modular Deployment |
Experimental (partial) |
Config-driven plugin loading: BGP, interfaces, and FIB load only when their config section is present. Add or remove subsystems at runtime via config reload (SIGHUP). Required config-root autoload failures fail closed, reload diffs restart same-name plugins when their definition changes, and changed external plugin replacements are pre-started before the old handler is removed. Scope gaps: BGP-off plugin functional coverage (plan/post-release/engineering/spec-fixit-bgp-off-plugin-functional.md); idempotent deferred reapply at startup (plan/release/fixes/spec-startup-deferred-reapply-idempotency.md). |
| MRT Dump |
Experimental |
RFC 6396 daemon-side MRT recording with three independent streams (updates, all messages, periodic TABLE_DUMP_V2 RIB snapshots). YANG config, per-peer and direction filtering, extended timestamps, add-path aware, on-demand CLI dump, async non-blocking writes, strftime file rotation. Analysis tools: show, routes, aspath, inject, replay, convert (pcap/json), statistics, filter. The reader derives each record's AS width from the record type rather than from one file-wide guess, decodes the RFC 8050 add-path TABLE_DUMP_V2 and BGP4MP subtypes, and REPORTS what it could not decode instead of printing it as fact: a record that fails to parse renders [unparseable: <tag>] <err>, and every analysis subcommand ends with warning: N malformed MRT record(s) skipped or partially decoded; results are incomplete. The daemon MRT plugin is a BGP plugin gated under ze_bgp, with no build tag of its own: every MRT stream is BGP data and its TABLE_DUMP_V2 snapshot is filled only by bgp-rib, so ze-stripped and every build without ze_bgp drop internal/plugins/mrt and its config schema, and the internal/mrt format library plus the shared internal/core/bgp/msgtype leaf drop by dead-code elimination with the engine. In a BGP build the RIB dump always finds bgp-rib: it is auto-loaded by the same bgp config root that starts BGP. Its reactor bridges reach the engine through a plugin-registry seam so the always-on hub never imports the plugin. |
| Named Service Listeners |
Experimental (partial) |
Every service that accepts inbound connections (web, ssh, mcp, looking-glass, telemetry, REST, gRPC, plugin hub) models its listen endpoints as a named YANG list. Each entry binds its own listener on the same subsystem; bind is all-or-nothing with rollback on failure. At config parse time CollectListeners refuses two configured endpoints that share an ip:port, across services, and a dual-stack :: endpoint beside an IPv4 endpoint on the same port, whose bind could never succeed (dualStackPair; a service registered with RegisterListenerV6Only binds :: IPv6-only and is exempt). It does not fill in YANG defaults, so an endpoint left at its default is compared only by ze doctor (CollectListenersWithDefaults), and ServerEndpoint.Listen joins an IPv6 host to its port without brackets. Scope gaps: parse-time conflict check over endpoints left at their YANG default. |
| Netdata-compatible OS Telemetry |
Experimental |
Prometheus metrics from /proc and /sys (CPU, memory, network, disk, IPv4/IPv6 protocols, conntrack, PSI, cpuidle, cpufreq, ZFS, btrfs, mdstat, SCTP, IPVS, wireless, etc) named and labeled after Netdata's Prometheus exporter, so dashboards built on Netdata's metric names can read them. No test compares the names or the count against a running Netdata. Per-collector enable/disable, interval override, and prefix are scoped under telemetry.prometheus.netdata so Ze-native metrics keep their ze_* names. The Prometheus HTTP service defaults to loopback and can require HTTP Basic Auth. The Prometheus HTTP exporter (the /metrics + /health listener, telemetry config extraction, basic-auth, and the Netdata OS collectors) is compile-out-able with the ze_telemetry build tag: normal and appliance builds include it, while ze-stripped and bare ze_core builds drop internal/component/telemetry/exporter and internal/component/telemetry/collector and reject the telemetry {} config block as unknown. Metric COLLECTION (the always-on internal/core/metrics registry used by ~60 packages, plus its no-op dummy) stays linked in every build, so a no-telemetry binary still records every ze_* counter; it just cannot expose them over HTTP. |
| Operational Report Bus |
Experimental (partial) |
Cross-subsystem ze show warnings and ze show errors commands with source <name> filtering: single place to surface prefix-threshold crossings, stale route data, BGP NOTIFICATIONs sent/received, unexpected session drops, session-stuck/flap/EOR-timeout, route-count-anomaly (>50% drop), FIB sync failures/orphans/programming-lag, firewall stale-table/drift, plugin crashes, interface error counters. State-based warnings + event-based error ring, login banner reads the same source. Scope gaps: functional coverage of most named sources; empty-bus shutdown observation (plan/release/validation/spec-finish-report-bus.md). |
| Outbound source-address |
Experimental |
Optional source-address local-IP binding for outbound service connections: BMP collector, RPKI/RTR cache, flow-export (UDP), IRR whois, and the managed hub TLS client; LDP binds its transport-address as the session source (RFC 5036). Unset = OS-selected source (unchanged). Shared network.RealDialer applies the bind for TCP/TLS; flow-export binds the UDP socket. See Outbound Source Address. |
| Packet Capture and Decode |
Experimental |
show capture raw action start protocol bgp records whole BGP messages in memory, and show capture raw action dump protocol bgp format pcap writes them as a pcap under LINKTYPE_RAW (101) with a synthetic IP and TCP header around each message, so Wireshark dissects the file as BGP. The 19-byte BGP header is rebuilt at the tap, sequence numbers advance per direction, checksums are computed, and IPv6 peers are written in their own family. A truncated message declares its on-wire length, so a reader shows truncation rather than a malformed packet. Reading back: ze bgp decode pcap <file> decodes every BGP message in a capture ze wrote or tcpdump wrote, reassembling each TCP direction first; ze bgp decode pcap - and ze bgp decode - take a capture and hex lines from standard input. A hole in the capture is reported and never read across, and a file with no BGP in it exits non-zero naming what it examined. The TCP ports are fabricated as 179 at both ends, because reading the real pair costs two mutexes on the session read path, so a ze-written capture is a readable reconstruction rather than a packet-level record. When no BGP capture provider can start, the start command answers with an empty started list rather than refusing. |
| Password Weakness Warning |
Supported |
Every password-set path judges the plaintext before it hashes it and warns when the password is shorter than 8 characters or matches one of eight common passwords (password, 123456, 12345678, qwerty, admin, letmein, root, changeme, compared whole and case-insensitively). The warning is advisory: the password is hashed and set, the commit succeeds, and the exit code stays 0, so a config Ze accepted before the check existed still commits. One policy serves ze config set, ze config deactivate, the CLI editor commit, the daemon load path, the REST/gRPC/gNMI commit path and ze passwd; the CLI surfaces name the leaf on stderr or in the commit status, and the daemon paths log at WARN. The reason names the rule and never the password. |
| Path MTU diagnostic for IPsec tunnels |
Supported |
show mtu measures the path MTU to every IPsec peer's installed endpoint and to a reference address with Don't Fragment ICMP probes (a reported MTU confirmed on the wire, then a candidate ladder and a bisection when ICMP is filtered), derives each tunnel's ESP ceiling from the transform it negotiated, and reports each bound xfrm interface as oversized, tight, under-utilized, ok, down or without a usable MTU, with the set interface ... mtu commands to apply and whether the access circuit or only the peer paths are clamped. A peer whose IKE SA is up is measured a second time over the SA itself, with an INFORMATIONAL request padded to the size under test: the authenticated reply confirms the ICMP figure or refutes it where ICMP errors are filtered, and the row names its prober (ike, icmp). show mtu host <address> measures one path; exhaustive bypasses the kernel's cached estimate. The command changes nothing. Linux only. |
| Per-protocol FIB withholding |
Experimental |
rib { fib-withhold [ bgp isis ]; } names the protocols whose selected routes ze does not write to the forwarding table. The list is empty by default, so every protocol is programmed. It accepts a name only when a protocol registered it, and a commit that names anything else is refused with the names ze accepts. A withheld route is not a dropped route. It stays in the Loc-RIB, and show rib and show ecmp-groups report it with the equal-cost paths that competed for the prefix, because both commands read the RIB rather than the last group written to the FIB. It still wins the prefix on administrative distance when its distance is the lower one. redistribute still moves it into another protocol, and the protocol's own event stream still carries it. Only system-rib/best-change, the stream every FIB writer reads, drops it, and the prefix leaves the next-hop table with it, so show nexthop-table stops listing it under its next-hop. A commit that changes the list sweeps the routes already programmed: ze withdraws each prefix it has programmed for a protocol the list now names, programs each prefix whose protocol the list no longer names, and rewrites the multipath group of a prefix that lost or gained a member. The sweep is what makes the change reach a converged router, which sends no further update for a prefix. Permitting a protocol programs its prefixes the way a live change does, so a next-hop no route in the Loc-RIB covers is programmed all the same: such a gateway is on-link on an OSPF or IS-IS link whose connected route no plugin inserted, and reading it as unreachable would leave the prefix out of the forwarding table for good. An SRv6 Service SID that reaches nothing is the one refusal the sweep keeps. The setting covers every FIB writer. They all read one publish stream, so it cannot write BGP into VPP and withhold it from the kernel. A withheld protocol is no member of an equal-cost group either, so traffic does not reach its gateway through a multipath route a permitted protocol won. A prefix leaving the RIB is withdrawn from the FIB only where ze has an install outstanding for it, so a route ze never programmed produces no delete and no FIB sync error. |
| Persistent storage |
Experimental |
One live database/ tree holds checksummed keys, credentials, runtime state, and per-config version pointers. Config history is content-addressed: each version's bytes are stored once under their SHA-256, and every read verifies the hash. One process owns writes; read-only inspection remains available. Explicit config files remain authoritative on every explicit-file start. Daemon commits update the file and history through a recoverable publication intent. ZeFS blobs remain explicit offline artifacts and appliance seeds. |
| Pipe Output Limiting |
Experimental |
| first N and | last N pipe operators bound output to the first or last N ROWS. They act on rows, so a command whose answer has none refuses them by name rather than truncating a document (client-side truncation otherwise). Commands that register them as pipe filters (e.g. RIB) get server-side early termination: | first N stops the iterator at N, saving both iteration and serialization cost. | last N keeps a trailing window. JSON output includes a "pipe" metadata dict recording which data-shaping modifiers were applied. |
| PKI Certificate Store |
Experimental |
YANG pki {} config for CA certificates and device certificates with private keys. Every certificate and key leaf takes a PEM document or base64 DER, PKCS8/SEC1/PKCS1 private key detection, $9$ sensitive encoding for keys, chain validation, expiry checking, atomic reload. PEM export for IPsec and TLS consumers. ServerTLSMaterial gives a TLS listener the leaf plus every stored intermediate, so the web/API HTTPS listener (environment.web.certificate), the looking glass (environment.looking-glass.certificate) and the as112/geodns DoT/DoH listeners (tls { certificate }) can serve an operator certificate with its full chain. A configured name that does not resolve is an error, and no listener falls back to the self-signed certificate. Health check (degraded at 30 days, down when expired). Report bus warnings for approaching expiry. show pki certificates, show pki certificate name <name>, and one command per export form: pem, bundle pem and fingerprint. Shared infrastructure for IPsec, TLS, and future mutual-auth features. The package has a second half config does not reach: a local certificate authority. Ze generates one root at its first start, keeps it in ZeFS, and issues its own components' certificates from it, so a peer validates an ISSUER rather than a copy of one leaf. The root lives ten years, an issued leaf 24 hours, and every certificate backdates NotBefore by five minutes for clock skew. A listener does not outlive its leaf: it answers each handshake through plugin.ServingLeaf, which reissues once two thirds of the current leaf's life is spent. Three surfaces consume the root today: the plugin transport (the engine hands it to every plugin child as ZE_PLUGIN_CA_PEM), the managed hub's listener, and ze appliance init, which issues the appliance's web leaf from an appliance root. The operator-facing web, looking-glass and DoT/DoH listeners are unchanged and still self-sign when no certificate is named. show pki local-ca pem exports the root certificate for an operator to paste into another node's pki { ca <name> { certificate } } block; the root private key has no accessor and reaches no command, no log and no config dump. File permissions are its only protection, and they belong to the blob rather than to the key: ZeFS applies no per-key mode, and the whole store is 0600 because atomicWrite creates it with os.CreateTemp. |
| Plugin Setup Results |
Supported |
Every registered plugin records, from its own init(), what its setup achieved: succeeded, soft-failure, or hard-failure, with the reason an operator acts on. show plugin list carries that record in its outcome and reason columns, in any ze process with no daemon running, and lists a plugin that recorded nothing as unknown rather than omitting it, so a feature that is absent for an environment reason can say why. The daemon reads the same registry once, at the first statement of hub.run, and refuses to start when a plugin recorded a hard failure, naming every failing plugin rather than the first. A CLI verb never reaches that gate, so the command that reports the fault keeps working on a host where the daemon will not boot. This is not show health, which runs a probe now: this replays an outcome decided once, before main(). |
| Plugin-Declared Pipe Aliases |
Experimental |
A pipe alias names an operator chain, so show bgp | summary says what show bgp | display router-id local-as uptime peers-configured peers-established says. A plugin names one for its own commands in the pipes list of its Stage 1 registration, and the engine parses the expansion once, at registration. The alias leaves the registry when the plugin stops, so a stopped plugin can start again. The pipe layer selects and re-sequences and computes nothing, so a command that offers an alias MUST emit its aggregate fields beside its detail rows. show bgp rpki does this, and show bgp rpki | summary answers the same record as show bgp rpki summary. command help "<name>" lists the aliases a command answers to, with the chain each one stands for. A declared alias resolves over ze cli -c and over ssh, and NOT inside ze cli with no command argument, which expands the chain in the client process. |
| Plugins |
Experimental (partial) |
The peer's config decides which program sees which peer: each peer attaches a program in an attach process block, receive states what that program is fed and send what it may originate toward that peer, and a peer that attaches nothing feeds nothing. A group's block reaches every member it produces, dynamic members included. show event delivery prints the edges. RIB, route server (with reactor-native RS fast path, dynamic peers, RS-client, community filtering), graceful restart, RPKI (ASPA policy enforcement), healthcheck, community filters, prefix-list, AS-path, AS-path length, community-match and reject-asn match conditions (each registered by its own package and evaluated in process inside a match instance) , external filter plugins in any language over the versioned, batched filter-update contract, which sends a filter only the fields it declared , per-address-family filters (remove an AFI/SAFI from an UPDATE on import or export -- e.g. a FlowSpec route reflector that never advertises ipv4/flow back to edge peers via export remove ipv4/flow -- or tear down the session on a received family; applies to forwarded and originated/injected routes), route attribute modifiers (set/increment/decrement for local-preference/med/aigp, community add/remove for standard/large/extended, MULTI_EXIT_DISC removal on an import chain as RFC 4271 Section 5.1.4 requires; a modifier can carry its own match container over standard, large and extended communities, and a route that meets none of the stated values passes through UNCHANGED rather than being rejected, because the chain already drops a rejected route and a conditional modifier exists so the routes it does not touch keep flowing), ingress community hygiene on the community filter (RFC 7454 Section 11 own-Global-Administrator scrub with a function keep-list, the RFC 8195 Section 3.2 relation-to-origin tag driven by the RFC 9234 role, and the RFC 7999 propagation guard), remove-private-AS policy action, AIGP (RFC 7311), BMP (RFC 7854), interface monitoring, cross-protocol redistribute (egress; routes injected by a redistribute source replay to a BGP peer that establishes after the injection, via a re-emit-on-peer-up request), FlowSpec-to-firewall bridge. BMP Loc-RIB and some redistribution claims remain explicitly scoped by readiness review. Scope gaps: BMP Loc-RIB; redistribution claims scoped by readiness review. |
| Policy Routing |
Supported |
Policy-based routing via nftables packet marking and kernel ip rules. Steers traffic to alternate routing tables or next-hops based on L3/L4 match criteria (address, port, protocol, TCP flags, set references). Actions: accept (bypass), drop, table N (fwmark + ip rule), next-hop (auto-managed table from 2000-2999), tcp-mss clamping. Interface binding by name or wildcard prefix (e.g., l2tp*). A policy that names several interfaces matches traffic on any of them. Ze installs one nftables rule per named interface. Config reload reconciles nftables tables, ip rules, and auto-managed routes. Compile-out-able with the ze_policyroute build tag: default-on in ZE_FEATURES, dropped from ze-stripped / bare ze_core builds, with its config block rejected as unknown. |
| Prefix-limit counting mode |
Experimental |
Per-family prefix { count offered | installed; }, default offered. offered counts announcement events off the wire; installed counts the set of prefixes the family currently holds, so a peer that re-announces the same prefix does not walk toward its limit. |
| Protocol Event Capture and Replay |
Experimental |
Per-peer, opt-in JSONL capture of a BGP session's inbound protocol events, and a replay harness that feeds a capture back through the same read path. Capture tees the COMPLETE wire message at both read paths (standard and coalesced) before RFC 7606 enforcement and before coalescing, so a malformed UPDATE is recorded exactly as the peer sent it and a coalesced batch is recorded as the separate messages it arrived as. Config operations applied while a capture runs are recorded with their transaction id, with secret-bearing values redacted. Off by default: the tee costs one nil comparison and no allocation per received message. Each file is capped at capture { maximum-size } megabytes exactly (1-1024, default 100); at the cap it rotates once to <file>.1 or stops, per capture { on-limit }, so a peer uses at most twice the cap on disk. Writer backpressure sheds events rather than blocking the read loop, counts them in ze_bgp_capture_dropped_events_total, and writes the gap into the stream. ze doctor reports a capture directory the daemon cannot write. Replay: le test replay <file> drives Session.ReadAndProcess over a stub connection and an injected clock and reports the FSM transitions, the prefixes announced and withdrawn, and any NOTIFICATION the session sent. |
| Resolution CLI and Pipes |
Experimental |
Offline ze resolve tool for DNS, Team Cymru ASN names, PeeringDB prefix counts, and IRR AS-SET expansion. The | resolve (reverse DNS) and | origin (ASN/network lookup via Team Cymru) pipe operators decorate the fields a command DECLARES to hold an IP address, and are refused by name over a command that declares none. ze pipe brings these pipe operators to offline commands: ze show debug profile | ze pipe match reactor, ze show debug profile | ze pipe count, ze show debug profile | ze pipe resolve; standalone input carries no declaration, so there they walk every field whose value parses as an address. |
| REST/gRPC API |
Experimental (partial) |
Programmatic API with OpenAPI 3.1 spec, config sessions. Both transports accept multiple named listen endpoints via environment.api-server.rest.server <name> / .grpc.server <name>. REST is plaintext and therefore loopback-only; expose it remotely only behind a TLS terminator. Non-loopback authenticated gRPC listeners require TLS. Bearer token auth, per-user auth, CORS support. Both transports share one engine for identical command output. SSE and gRPC streaming are wired to registered streaming commands such as monitor event, using the same authorization and accounting path as SSH monitor commands. Completion remains future work. Each transport is independently compile-out-able: ze_rest gates internal/component/api/rest and ze_grpc gates internal/component/api/grpc, so a build can ship gRPC-without-REST or vice-versa. normal and appliance builds include both; ze-stripped and bare ze_core drop both. With a transport compiled out its server code and its config container (rest{}/grpc{}) are absent, so that block is rejected as unknown; the shared api-server { token } base and the parent internal/component/api engine stay always-on (gNMI uses the parent). Scope gaps: API completion. |
| Route Installation |
Experimental |
FIB pipeline: protocol RIB best-path tracking, shared Loc-RIB selection by admin distance, recursive next-hop resolution (max depth 8, IGP metric accumulation, cascade on NH change), ECMP grouping (equal-cost paths collected into nexthop groups, max 128 members), rich route programming (route type blackhole/unreachable/prohibit, metric, VRF table ID, MPLS labels, SRv6 SID), kernel backend via netlink multipath with route type/metric/table/MPLS-lwtunnel/SRv6-seg6 support, VPP backend via GoVPP with multi-path/route-type/metric/table support, per-producer netlink protocol ownership, crash recovery via stale-mark-sweep, external change monitoring. Selection and programming are separate stages: the shared Loc-RIB picks one winner for each prefix, and the system RIB decides whether to program it. rib { fib-withhold } naming the winner's protocol is one of the reasons the write is declined, and the Per-protocol FIB withholding row states what withholding leaves intact. The forwarding action (unicast, blackhole, unreachable, prohibit) rides one route from the protocol RIB through sysrib into the FIB backend. It is deliberately NOT part of arbitration: selection stays on administrative distance, then metric, then first-seen order, so a discard route never beats a forwarding one on the strength of being a discard. It IS part of the change comparison, because a prefix that turns into a discard with every other field unchanged must still reach the kernel. A prefix and a PATH own different fields, and the split is at the type rather than at each call site: the label stack and the SRv6 SID belong to the forwarding path, so an equal-cost member promoted over an unreachable winner is programmed with its own, while the protocol, the route type, the metric and the fast-reroute alternate belong to the prefix and come from the winner whichever member takes it. Whether ze holds an install is the PRESENCE of the prefix in its installed map, never the validity of the address there, because a member named by a device alone resolves to an invalid address. The install also records the PROOF it was made under: a next-hop inside a connected covering route resolves to itself, so the address cannot say which rule programmed it, and a cascade that finds a once-proved path now unproved is the one case that withdraws. show rib reports it as route-type, omitted for an ordinary route, and a Linux discard route names no next-hop, device, multipath or encap. CLI: show nexthop-table, show ecmp-groups. Local privileged integration covers FIB restart sweep and flush-on-stop preserving static and policyroute-owned routes; target-runner evidence is still required before production deployment claims. |
| Route Policy |
Experimental |
One policy library supplies source import chains for connected, static, OSPF and IS-IS routes and a central rib import chain for every Loc-RIB source. Rejected paths remain hidden candidates, visible through show rib candidates. A policy reload re-judges stored producer values. Named static tables bypass these Loc-RIB attachments. BGP keeps peer import/export chains. Redistribution applies a destination/source import chain before the destination's IS-IS or OSPF export chain. Each attachment checks which fields its routes carry and which values its filters can set. The system RIB runs internally or externally with one daemon authority and the same forwarding worker. The external selected projection runs no policy or cross-protocol selection. Distance updates and forwarding output require daemon acknowledgment. The real-path tests run each case once with an internal and once with an external system RIB, against a real kernel in a network namespace of their own; the four rib-import-tag cases also cross the static producer's placement. A recorded green run, not the list, is what counts toward a level. |
| RPF Lookup |
Experimental |
Reverse Path Forwarding query: longest-prefix-match against Loc-RIB for any CIDR family (IPv4/IPv6 unicast/multicast). Exposes show bgp rib rpf <family> <source-addr> command returning matched prefix, next-hop, admin distance, and metric as JSON. Generic LPM on the sharded Loc-RIB (queries all shards, picks most specific). |
| Runtime Diagnostics |
Experimental |
Production debugging via CLI and MCP: show l2tp observer (per-session event ring), show l2tp cqm (per-login echo RTT/loss buckets), show l2tp echo (current echo state), show l2tp reliable (reliable transport Ns/Nr/cwnd), show traffic control (TC qdisc/class state), show metrics pool (BGP attribute pool occupancy and dedup rates), enhanced subsystem-list (real plugin state). All auto-exposed as MCP tools for AI-assisted troubleshooting. |
| Self-Documenting System |
Experimental |
Runtime introspection of plugins, env vars, RPCs, schemas, commands. Each command declares two help texts: a one-line summary every listing prints verbatim, and a long explanation only that command's own help page prints. A YANG command node declares them as ze:help and description. A plugin declares them as short-help and description in its Stage 1 registration. No surface derives one from the other. Inside ze cli, Tab reaches both halves. The completion menu lists command names, and the second message line above the prompt carries the summary of the selected command. Tab with nothing left to complete opens that command's long explanation in its own box. A command argument, an rpc or notification leaf, a config node and an enum value declare the same pair, and every surface that names the node renders both: the CLI JSON help, the site and wiki catalogs, the MCP tool schema, the web admin and config forms, the API and gRPC schemas, the published configuration reference and the completion menu. A config node several modules declare shows every declaration's summary, joined, because nothing in the schema says which module owns the node. |
| Self-Update |
Experimental |
Platform-aware update backend. Normal Linux uses Ze self-update or passive version checking with SHA-256 verified download, atomic binary replacement via rename, .prev hard-link rollback, deterministic spread scheduling (FNV-1a per device+version), maintenance windows, server-side pause, and persisted update history. Gokrazy appliances report backend: gokrazy-ab through the gokrazy backend and keep the manual update system firmware {check,download,apply,restart,rollback} command family wired; unsupported operations report that system image updates are managed by gokrazy. Minimal (no-tag) builds return an explicit self-update unavailable in minimal build response. ze update serve standalone server remains available on Ze-managed platforms to distribute artifacts. Config: system { update-check { auto-apply true; spread 1800; maintenance-window { start 02:00; end 06:00 }; restart { time 03:00 } } }. |
| Static Routes |
Experimental |
Config-driven static route plugin with named routing tables (policy-based routing), interface-only next-hops (PPPoE/GRE tunnels), mixed ECMP (gateway + interface-only in same group), next-hops through an interface Ze creates on demand (a pppoe-client), which wait as inactive members while the device is absent and join or leave the group as it comes up or goes, with no reload, ECMP (multiple active next-hops), per-next-hop weighted load balancing, BFD-tracked failover (next-hop removed from ECMP group on session DOWN, re-added on UP), blackhole/reject, IPv4/IPv6, config reload reconciliation, and redistribute integration (redistribute { import static }). Named tables resolved via routing-table registry; non-default table routes are PBR-only (not redistributed into BGP) and are programmed directly, via netlink multipath or VPP via GoVPP. A MAIN-table route is a Loc-RIB path instead: rib { distance { static N } } ranks it against BGP, OSPF and IS-IS for the same prefix, and the FIB plugin programs the winner as RTPROT_ZE, so one writer owns the entry. Such a route needs a FIB plugin and Ze loads one: static declares that it needs a data plane, each FIB plugin declares the data plane it programs, and the engine resolves the pair against the backend interface { backend } selects, so no fib { } block is required and an explicit one always wins. |
| Storage SMART Management |
Experimental |
YANG-modeled SMART disk health: auto-enable on detected ATA/NVMe devices, periodic health polling with three-tier temperature alerting (informational warning, rate-of-change warning, critical error) via report bus, scheduled self-tests (short daily, extended weekly with day-of-week constraint), in-progress detection (skips duplicate tests), live status via show storage smart (per-device health, temperature, power-on hours, error count, NVMe percent-used/available-spare, self-test schedule). Pure ioctl (no smartctl binary, gokrazy-safe). Config reload updates intervals live. ze doctor verifies SMART accessibility when enabled. First NOS with YANG-modeled SMART management. |
| Subscriber Session Model |
Experimental |
Unified subscriber session model for L2TP/PPPoE subscribers with shared session lifecycle, auth, pool, and shaper infrastructure. Show enricher registry lets plugins contribute data to show commands: in-process plugins register via show.MustRegister in init; external plugins declare enrichers at registration (Stage 1) and handle ze-plugin-callback:enrich-show callbacks at runtime with a 2s timeout. Web service-locator pages call show.Enrich explicitly. show subscriber detail and show subscriber gain CoS profile data when the cos plugin is loaded. . internal/core/show/ |
System Readiness (ze doctor) |
Experimental |
Offline pre-start checks: runtime platform detection (gokrazy, systemd, container, plain-linux, darwin) with capability probing (read-only root, /perm, systemd, gokrazy socket, reboot, persistent storage), config syntax, YANG validation, TLS certs (missing/expired/invalid), VPP socket, kernel modules, interface existence/state, SSH host key, listener conflicts, plugin binaries, storage integrity, dangling config references, disk space (<5% free), DNS resolver reachability, clock skew (>5 min vs NTP), VPP version (Linux), VPP hugepage reservation (Linux). Platform-aware checks: gokrazy /perm writability, container read-only root, missing Ze-managed clock synchronization, platform-mismatched persistence/DNS paths, and Linux machine-id presence. Config-delivery audit: doctor-config-root-unclaimed names a configured subtree that commits and reaches no plugin and no hub handler, with doctor-config-claims-unavailable when the claim set cannot be read. Other codes added since: doctor-tls-reference (a certificate name the PKI store does not hold), doctor-pki-ca-root-missing and doctor-pki-ca-root-expiry (the local certificate authority root absent, or within 90 days of expiry), doctor-bgp-capture-directory (a capture directory the daemon cannot write), doctor-iface-ra-forwarding (a unit advertising while its device has IPv6 forwarding off), and doctor-vpp-lcp-plugin (vpp.lcp on while the running VPP loads no linux_cp_plugin.so; a failed probe degrades to a warning rather than an error). Kernel capability gate: a subsystem registers the kernel feature it needs, a predicate that reads the config tree, and a native probe (netlink or procfs, never an external binary). When the configuration uses the subsystem and the kernel lacks the feature, ze doctor reports it at error severity, ze REFUSES to start with the same message and exit 1, ze config validate fails, and a reload is refused with the running configuration still serving. A capability that cannot be DETERMINED warns and ze starts, so an unreadable probe never turns a working deployment into a dead one. There is no operator override. MPLS (CONFIG_MPLS_ROUTING, from the fib kernel plugin) and IPsec (CONFIG_XFRM_USER, from the ike engine) are enrolled. An empty vpn { ipsec { } } block installs no Security Association, so it gates nothing, warns about no kernel module and binds no IKE listener: one predicate answers for all three. Ze WRITES net.mpls.platform_labels before it programs its first label, because a label space of 0 is a dead configuration ze can repair rather than report. --json output with stable diagnostic codes and ze show diagnostic <code> for remediation. Each finding may carry a fix, the command that repairs it, drawn on its own line after the message and carried as the fix key in JSON. Checks of live daemon state register in the runtime phase and run only in the daemon, through show doctor: the local ze doctor first asks the running daemon for its show doctor answer over SSH, and when no daemon answers it runs its own checks and adds a doctor-runtime-not-checked warning naming why and the command that gets the daemon's answer. The recent-events runtime check reports the report bus's recent error events and active warnings as warning rows, so a transient problem appears in show doctor without making the box read not ready. |
Tech-Support Bundle (ze request support archive) |
Experimental |
Offline archive generator with 20 pure-Go modules (no shell-outs, gokrazy-safe): version, doctor, host, platform, config (sanitized), crashes, disk, interfaces (netlink), routes (netlink), neighbors (netlink), env, sysctl, runtime, dmesg, sockets, kernel modules, conntrack, file descriptors, DNS, firewall (nftables via netlink). Module selection (--module/--exclude), time scoping (--since), privacy-by-default (--sensitive to include secrets), reason metadata (--reason), JSON manifest (--json). Archive: ze-support-<hostname>-<timestamp>.tar.gz with one JSON file per module. No NOS vendor produces structured JSON-per-module output. None of the modules collects SMART disk health. |
| Traffic Control Lifecycle |
Experimental |
The traffic/control section of the config is now programmed at boot and on SIGHUP reload. The traffic component's reactor calls the selected backend's Apply(map[string]InterfaceQoS) in OnConfigure and OnConfigApply, with sdk.Journal rollback on apply failure. Linux default backend is tc (netlink); future backends plug in via traffic.RegisterBackend. Local privileged integration covers netlink qdisc snapshot/restore after backend restart; target-runner and reactor-level boot/reload kernel-state evidence remain open. |
| Traffic Feature Signals |
Experimental |
Neutral per-source detection SIGNALS (facts, never verdicts) derived from the observation feed by a second consumer alongside the traffic monitor: fan-out (distinct destinations), out/in byte ratio (exfiltration), destination-port entropy, new-peer, rare-port/proto, and coarse beaconing (interval regularity, bounded to periods of a few seconds by the 1s sampling tick). Computed via the shared internal/core/stats primitives (rolling window, Shannon entropy, interval regularity). Viewed via show traffic feature; bounded per-source state with idle eviction. The judgment layer (the anomaly detection family) consumes these facts. |
| Traffic Monitor |
Experimental |
Lazy consumer-refcounted aggregation service (internal/component/trafficstat) that subscribes to the observation feed and maintains a time-windowed ranked usage view (per-interface rates, top-N source/dest IPs, top ports with service names, protocol mix, 60s history). The aggregator is rebuilt on the shared internal/core/stats rolling-window primitive; severity is now a display-only CLI computation from the history facts (the neutral layer holds no verdict). Runs only while consumers are attached. Consumed by show traffic stat (one-shot JSON snapshot), monitor traffic stat (full-screen alt-screen TUI via the generic MonitorProvider registry), and ddos/detect (Depth-1: pre-computed per-interface rates instead of raw counter diffing). Includes internal/core/portname hardcoded port-to-service-name table with amplification-vector overlay for 7 known reflection ports. |
| Traffic Usage |
Experimental |
eBPF TCX per-(port, protocol) and opt-in per-IP byte accounting on operator-selected interfaces, exported as Prometheus metrics and viewed via show traffic usage [name <interface>]. IPv4 only, monitoring only (never drops or modifies traffic), Linux >= 6.6 (no-op elsewhere; needs CAP_BPF + CAP_NET_ADMIN). The eBPF programs are assembled in pure Go (cilium/ebpf asm.Instructions) and loaded from memory: no C source, no committed .o, no clang/LLVM; hand-written assembly validated by BPF_PROG_TEST_RUN tests. Per-port accounting (ingress by dst_port, egress by src_port) is always on; track-ip adds per source/destination IPv4 (off by default to bound cardinality). Configurable poll interval, stale-timeout (delete unseen series to bound /metrics cardinality), and per-map LRU max-entries (top-talker eviction). Prometheus metrics: ze_traffic_usage_ingress_port_bytes_total, ze_traffic_usage_egress_port_bytes_total, ze_traffic_usage_ingress_bytes_total and ze_traffic_usage_egress_bytes_total (track-ip only), ze_traffic_usage_map_entries. A ze doctor check (doctor-traffic-usage-ebpf) warns when enabled but eBPF/TCX is unavailable. Compile-out-able with the ze_trafficusage build tag: default-on in ZE_FEATURES, dropped from ze-stripped / bare ze_core builds, with its config block rejected as unknown. |
| Volumetric DDoS Detection and Mitigation |
Experimental |
The volumetric domain, kept separate from the behavioral anomaly detector above. ddos-detect runs two-stage detection over the traffic-monitor feed: per-source packet-rate and bandwidth p99 baselines, attack characterization, and incident confidence scoring. Three responders consume its incidents. ddos-local drops box-directed attacks with an on-host nft rule. ddos-flowspec originates a FlowSpec or RTBH rule upstream for transit attacks. ddos-flowtriq reports incidents to the Flowtriq cloud API. ddos-observe holds the incident lifecycle behind show ddos status and show ddos incidents. Clearing an upstream mitigation is the hard half, because an upstream drop blinds the box's own sensors: the leak probe narrows the announced rule to probe-rate and reads the detector's attack-ongoing bandwidth to see whether the flood still arrives. An announce is also withdrawn by max-mitigation-duration (checked once a second, 0 means no cap, default 3600s, range 0..604800), by a characterized exemption, and by a victim that reclassifies from remote to local. ddos-local enforces its own max-mitigation-duration on the on-host drop rule the same way: measured from when the rule went in, checked once a second, 0 means no cap, default 3600s, range 0..86400. |
| VPP Firewall Backend |
Stub-backed |
Registered as firewall { backend vpp }. Filter chains translate ze Match/Action types to VPP ACL rules via GoVPP binapi (source/destination prefix, port range, protocol, ICMP type/code, TCP flags, permit/deny/reflect). Connection-state established,related maps to ACL_ACTION_API_PERMIT_REFLECT (VPP reflexive ACL). NAT chains configure VPP NAT44-ED: masquerade via output-interface mode, SNAT via address pool + inside interface feature, DNAT via static mappings with tagged cleanup. SetMark and Limit actions use VPP's classify pipeline: classify tables match traffic by packet header fields, SetMark sets opaque metadata via CLASSIFY_API_ACTION_SET_METADATA, Limit creates a policer bound to the classify table via PolicerClassifySetInterface. Expression types without a faithful VPP representation are rejected at commit via firewall.RegisterVerifier("vpp", Verify): interface matches, connection marks, DSCP, sets, packet modification (connmark/dscp/tcp-mss), counters, log, chain traversal. |
| VPP Traffic Control Backend |
Stub-backed |
Registered as traffic { control { backend vpp } }. Scope is per-class policing: HTB and TBF qdiscs translate each class to a VPP policer (CIR = Rate, EIR = Ceil, kbps with round-up). An unfiltered single class binds to interface egress via PolicerOutput. A class carrying a steering filter -- filter protocol <n> or filter dscp <v> -- is instead bound to the ingress policer-classify feature: the backend builds per-family (IPv4 + IPv6) classify tables whose mask/match vectors match the packet at absolute frame offsets and steers matching packets to the class policer via HitNextIndex (byte-for-byte VPP's own policer-hit-next, verified on VPP v25.10). Offsets: IPv4 protocol byte 23 / IPv6 next-header byte 20 for filter protocol; IPv4 TOS byte 15 (mask 0xFC) / IPv6 traffic-class bytes 14-15 (masks 0x0F/0xC0) for filter dscp -- filter dscp polices the DSCP-matched traffic (police-by-dscp), it is NOT a QoS remark. Multi-class HTB/TBF is accepted when EVERY class carries a steering filter: per-interface, per-family the sessions are grouped by field mask into one table per distinct mask (a session per class -> its own policer), and distinct-mask tables are chained via ClassifyAddDelTable.NextTableIndex (real VPP v25.10 validated the multi-session steering and chain fall-through). A multi-class config with any unfiltered class, the prio qdisc (a scheduler VPP does not expose), mark filters (Linux SKB fwmark has no faithful VPP equivalent), and every other qdisc type are rejected at OnConfigVerify via traffic.RegisterVerifier("vpp", Verify). Per the exact-or-reject principle, shipping silent-no-op features (classify sessions in a detached table, or N unfiltered policers stacked on the output feature arc producing min(rates) instead of per-class shaping) is banned. Apply waits up to 5s for VPP to be reachable and returns vpp not connected after 5s on timeout. On partial-apply error the backend undoes what this call programmed in VPP before returning. Reconcile-time deletions are tolerant of stale indexes (post-VPP-restart): failures log a warning and continue instead of failing the commit. A fresh backend instance scans VPP policers with the ze/ prefix, removes undesired startup orphans, and rebinds desired policers from config. Same-process reapply also replays PolicerOutput(apply=true) for existing desired policers, repairing VPP-side output unbinds without requiring a Ze restart. Protocol-filter classify tables are reclaimed by in-process reconcile (unbind + delete by tracked index); because VPP classify tables are anonymous they cannot be reclaimed by name at startup like policers, so a table left by a previous process is inert (unbound) memory reclaimed only by a VPP restart. Real-daemon traffic-control evidence covers policer apply/bind, restart preservation, startup orphan cleanup, protocol-filter and dscp-filter classify-table attachment + reconcile removal, and multi-class per-class steering (two policers bound via one table per family). . ai/rationale/exact-or-reject.md |
| Web Interface |
Experimental |
HTTPS config editor with YANG-driven UI, CLI bar, and L2TP session management (CQM graph, event timeline, disconnect). Every page is a .templ source compiled to Go (58 in the web package, 9 in the Looking Glass), so a renamed view-model field is a compile error rather than a blank panel; regenerate with ./le repo update. The browser library is htmx 4.0.0, shared by the web UI, the Looking Glass, and the chaos dashboard. The service is default-on but compile-out-able with the ze_web build tag: normal and appliance builds include it, while minimal ze_core and ze-stripped-build builds drop internal/component/web. Self-signed HTTPS certificates are generated by the shared internal/core/selfcert helper and persisted through caller-owned storage. Every response carries the security headers, and the X-Ze-Version build banner among them. environment { hide-version true; } takes that banner off the web interface and the Looking Glass together; the default sends it, and no other header changes. Ze sends no Server header at all. |
| XFRM Interfaces |
Experimental |
Route-based IPsec via XFRM interfaces (interface { xfrm <name> { if-id <N> } }). Traffic routed through the interface is encrypted; traffic arriving is decrypted. Created and deleted via netlink. |