Skip to content

0.7.1 — 2026-11-12

0.7.1 adds OpAMP remote management (Enterprise) for signed configuration overlays, runs probes in parallel with per-probe scheduling offsets and a configurable collect_timeout, makes the first PRTG/Nagios reading wait for initial data, and accepts licences signed by delegated platform keys. It also masks more credential forms in logs, publishes senhub.agent.license.invalid at 0 on validation, fixes root-owned config and licence file permissions on Linux, restarts the Windows service after update, and accepts OTLP/HTTP JSON with logged refusals. update now installs a named older version, reports a version it cannot resolve as an error, follows auto_update.include_beta for latest and says what a dry run would do; uninstall and license remove fail when they cannot ask for confirmation. The console shows the live state of a probe changed by remote management and counts a series once its export succeeded. Subcommand help flags, OCI container labels, and http(s)://host:port OTLP output endpoints are also supported.

0.7.0 was not released: 0.7.1 replaces it, with one more fix to the masking of the configuration the agent shows and sends to a remote management server.

Before you upgrade

  • config show --redact masks more values. The redacted view, and the configuration sent to a remote management server, now show *** for every parameter a probe or output declares secret, whatever its name, and for any parameter whose name contains pass, pwd, secret, token, key, session, private, cookie, auth, credential, dsn or uri. In any other value, a URL loses its user and password (https://***@host/path) and its query values, as in the logs. A script that read such a value from config show --redact gets ***; read it from the file, or use config show --resolved on the host.
  • The redacted configuration shows *** for every OTLP header. OTLP output headers are declared secret as a whole, so config show --redact and the configuration sent to a remote management server mask every header value, including non-secret ones such as X-Scope-OrgID. Header names stay visible.
  • Logs and the console mask the same values as config show --redact. The agent log writes a masked value as *** in full, where it used to keep its first and last characters, and the probe start and reload lines mask every parameter a probe declares secret. The console no longer displays a few non-secret parameters whose name contains a secret word, such as an SNMP v3 auth_protocol or a key_file path; editing a probe keeps their stored values.
  • Linux: the process probe no longer reports kernel threads. Numbered kernel threads (scsi_eh_N, ksoftirqd/N, irq/N-…, kworker…) each became a process name, so the series count grew with CPUs, disks and interrupt lines: on one measured host, more than half of 191 names. They are now left out before the roll-up and top_n, and the agent logs once how many were skipped. Set include_kernel_threads: true to report them as before.
  • Network interfaces: oper_state takes the seven RFC 2863 values in snake_case, and the administrative state has its own attribute. snmp_poll reported lowerLayerDown; it is now lower_layer_down, and not_present joins the set. The host interfaces kept only up and down, folding dormant and lowerlayerdown into down; they now report the same seven values as snmp_poll. When the operational state is not known (no sysfs, on Windows for instance), oper_state is left out instead of being taken from the interface's administrative flag, which now goes to admin_state (up or down). A filter or an alert on oper_state = lowerLayerDown, or on down for a dormant host interface, needs the new value.
  • prometheus_scrape: a target URL with a user part or a query string gets a new target value. The target tag and the log lines carried the URL as written, user part and query values included. Such a target is now shown as its scheme, host and path, the query with each value replaced by ***, and a short fingerprint in brackets that keeps two such targets apart, for example https://h:9090/federate?match[]=*** [1a2b3c4d]. Its series start under the new target value; a dashboard or an alert that filters on the old value needs the new one. A target without a user part or a query keeps its value.
  • Keep the host:port form for OTLP outputs while a rollback is possible. An OTLP output endpoint written as a URL (http://… or https://…), accepted from 0.7.1, is not understood by 0.6.5 and earlier: there, config check reports it valid but the OTLP output does not start. Switch to the URL form once going back to an earlier version is ruled out.

Features

  • tcp_dial reads the TCP table for a target on its own host. A dial against a local port is a connect-and-close the checked service logs as an error, about 1,440 lines a day at a one-minute interval. A target that resolves to a loopback or host address is now checked from the TCP table: senhub.tcp.listener.up, senhub.tcp.listener.connections and senhub.tcp.listener.peers per port, no connection opened. mode: dial keeps the previous behaviour. See tcp_dial.
  • Remote management through OpAMP (Enterprise). The agent can report its description, health and effective configuration, secrets masked, to an OpAMP server, and apply a signed overlay from it: switch local probes on or off, change their interval and the log level, and with the v2 overlay run whole probes defined by the server and pause the agent's own update check. Every overlay is verified against pinned keys, applied by a reload, confirmed when its probes start and otherwise reverted; the local files are never modified. exec probes can only be managed if agent.yaml allows them. See OpAMP remote management.
  • Probes no longer start one after the other, and each runs at its own offset. Every probe makes its first collection at once; its next collections then follow a fixed offset inside its interval, up to 30 seconds, derived from the probe's identity: it is the same after every restart, and probes of the same interval no longer fire together on every cycle. A slow or unreachable target holds only its own probe, not the ones after it. The first data of a probe can therefore arrive up to 30 seconds after the start. Each collection of a probe that can be cancelled also runs under a deadline: five minutes, or the probe's interval when that is longer, or the probe's collect_timeout when it sets one. A collection that runs past its interval is counted and never run twice, but it is not cut before its deadline. Four new self-metrics describe the scheduling: senhub.agent.collect.in_flight, senhub.agent.collect.lag (maximum start delay since the previous read by the same reader, Prometheus or OTLP), senhub.agent.collect.overruns (collections that ran past their interval) and senhub.agent.probe.first_collection.delay; the last three carry probe_name and probe_type, so two probes of the same type are two series. See Prometheus metrics reference.
  • collect_timeout per probe. A probe entry accepts collect_timeout, a duration (90s) or a number of seconds, which replaces the default deadline of one collection, shorter or longer. Without it a collection is never cut before five minutes or its interval, whichever is longer. A probe slower than its interval delivers its data at its own pace, never two collections at once, and each late collection counts in senhub.agent.collect.overruns: that is the signal to raise the interval or lower the work. 0 means unset, while the empty string is refused. The exec probe now stops its command, with its whole process group, when this deadline ends; its own timeout still applies and the shorter of the two wins. A collection cut at its deadline is logged with the deadline and the name of the setting. See Configuration.
  • PRTG and Nagios: the first reading after a start waits for the probe's first collection. In the two minutes after the agent starts, a request for a probe that has no data yet waits up to 10 seconds for its first points before answering, so a sensor no longer reads "No channel found" (PE231) once after a restart when the probe collects within that time. A reading that arrives before the agent has started a configured, enabled probe waits for it too. A probe slower than that still reads PE231 once, answered normally rather than by a closed connection, and outside the two minutes the answer is immediate as before. Probes that receive events (syslog, event log, OTLP receiver) and probes whose first collection finished with no data never wait.
  • Licences signed by a delegated platform key. Besides licences signed by the root key, the agent accepts licences signed with EdDSA by a platform key that the root key has delegated for a month and for a set of customers; the delegation travels inside the licence. A delegated licence must name a customer inside that set, and expires at the latest with its delegation, followed by the usual 7-day grace period with a warning. Existing licences are unaffected.
  • Container images carry OCI labels. The ghcr.io/senhub-io/senhub-agent and -oss images now set org.opencontainers.image.source, .revision, .version and .title, so a registry or scanner can trace an image to its release.

Fixes

  • Built with Go 1.26.11, which fixes GO-2026-6859. A race in net/http's MaxBytesReader header handling let a crafted request crash an HTTP/1 server; the agent's own HTTP endpoints (PRTG, Nagios, Prometheus, console) are served by that package. Releases up to 0.7.0-rc.1 were built with Go 1.26.10 and are exposed; upgrading is the fix.
  • Parameters written to the agent's own log mask more credential forms. When a probe or output starts or its configuration changes, the agent logs its parameters with secrets masked. Values under an ordinary name are now masked too when they carry a credential: the values of a URL query string, the fragment of a URL, the whole user part of a URL, an authorization header value whatever its scheme (all the parameters of a Digest header included), a key-value pair whose name denotes a secret, a JSON Web Token or a cloud access key id. Query values are masked whatever their name, so a logged URL keeps its parameter names but not their values.
  • senhub.agent.license.invalid reads 0 once a licence validates. An agent whose licence validated at its first start sent no series at all, so a dashboard could not tell a valid licence from an agent that sends nothing, and an alert on > 0 had no value to recover to. The gauge is now published at 0 for each reason (validation_failed, binding_mismatch, expired_no_grace) as soon as a licence validates. A host with no licence configured still sends no series.
  • Linux: commands run as root keep agent.yaml readable by the service. license activate and the other commands that rewrite a configuration file gave the rewritten file the owner of its folder, so a package-installed agent.yaml (root:senhub) under a root-owned /etc/senhub-agent became root:root, and the service could no longer read it. A rewrite by root now keeps the owner and mode of the file it replaces.
  • Windows ZIP install: update restarts the service. The new binary was installed but the running service kept the old one until a manual Restart-Service senhub-agent. When the senhub-agent service is registered and running, and the host is not MSI-managed, update now restarts it and checks that it runs the new binary, as on Linux.
  • Probes start in parallel. The agent started its probes one after the other, so a probe slow to start, for instance one that connects to its target first, delayed every probe after it, the OTLP, syslog and SNMP trap receivers included. They now start eight at a time, receivers first, and a start that takes more than 10 seconds is named in the log; each start line gives its duration (start_ms).
  • Linux: sudo senhub-agent license activate leaves a licence the service can read. Run as root, the command wrote license.jwt as 0600 root, so the agent, which runs as senhub, could not read it and kept retrying. The file now takes the owner and group of agent.yaml and its mode, up to 0640.
  • OTLP receiver: refused requests are logged and counted, and JSON is accepted. The HTTP receiver answered 400, 413, 415 and 429 without a log line, so a sender refused for its media type or size left no trace on the agent. Each refusal is now logged at Warn, at most once a minute per signal and reason, and counted in senhub_agent_otlp_receiver_refused_total. The Content-Type is parsed rather than compared, so application/x-protobuf; charset=utf-8 is accepted, and OTLP/HTTP JSON (application/json) is accepted for metrics, logs and traces. See OTLP receiver.
  • OTLP output: an endpoint written as a URL is accepted. Previously http://collector:4318 was rejected with a parse error naming https://http:...; now the scheme is stripped, steers TLS when tls.enabled is unset, and host:port keeps working unchanged.
  • Windows ZIP install: doctor, config check and config show read the service's configuration. Without --config-path they read %ProgramData%\SenHub\agent.yaml, which a ZIP install does not have, so doctor reported a failure. They now take the path from the registered service's command line when there is one, and say which file they read and why. When no configuration is found at all, doctor says where it looked instead of reporting a failure.
  • Container Apps: the jobs are followed by default when the discovery is declared in SENHUB_PROBES. The jobs probe added next to a Container Apps discovery appeared only when the discovery was declared with SENHUB_PROBE_<NAME>_* variables, not in the SENHUB_PROBES descriptor that the container guide uses. Both forms now add it, with the same identity and resource groups. A probe of type azure_container_app_jobs you declare yourself, under any name, still replaces it, and SENHUB_PROBE_JOBS_TYPE=none still turns it off.
  • -h/--help after a subcommand prints usage and exits 0. Any -h or --help before -- now shows that command's help without running it, asking for root, or taking the flag as a value; secret wire-unit -h no longer regenerates the credential drop-in.
  • A read-only configuration directory is left alone. With /etc/senhub-agent mounted read-only (a Kubernetes ConfigMap, for instance), every configuration load tried to write the default probe fragments, their marker and the sealing key there and logged seven warnings each time. The agent now writes nothing into a configuration directory it cannot write to and says so once at start; inline secrets stay in place with one hint towards ${env:}, ${file:} and ${secret:}; and the sealing key and secrets.age go to the state directory (SENHUB_STATE_DIR, STATE_DIRECTORY, or the platform default) when the configuration directory is read-only and holds no key yet. A host that has its key next to agent.yaml keeps it.
  • The open-source edition is built once, stamped with its own commit. The senhub-agent-oss binary of the release archive reported the commit of the enterprise build, and the .deb and .rpm of that edition were rebuilt separately with another stamp. Both now carry the same public commit, and the packages hold the very binary of the archive.
  • URL secrets are now masked in every agent log line. Secrets carried by URLs are now masked in every line of the agent's own log, not only in the generic probe start and configuration-change lines. Before, a probe's own lines (for example http_check's start line listing its targets, and its failure line quoting the URL) printed the user part, query values such as tokens or keys, fragments, JWTs and access keys in clear, in the console, the log files and the logs shipped over OTLP. The defect dates back to 0.6.5 at least. If such URLs were configured, treat those credentials as exposed in older log files.
  • Console: the "Series exported" card no longer names the edition on an Enterprise agent without an OTLP output. Series are counted on OTLP outputs only; an Enterprise agent that only answers PRTG or Prometheus showed "Series metering is part of the Enterprise edition". It now says that series are counted on OTLP outputs only and that none has exported in the last hour. A series counts once its export succeeded: an OTLP output whose collector does not answer no longer shows its series as exported.
  • update <version> installs an older version when asked. The command printed "Already up to date." and exited 0 without installing anything when the version named was older than the running one, which a script read as success. It now installs it, as the documentation says, and on Linux the copy the service runs follows; the periodic update check, update latest and a version constraint such as >=0.6 still never go backwards. On an MSI-managed Windows install it refuses with a message and a non-zero code, since Windows Installer does not downgrade in place. A version it cannot parse, a version the update registry does not have, or a registry it cannot reach is now an error with a non-zero code, where the command used to say "Already up to date.". The defect dates back to 0.6.5.
  • uninstall without a terminal fails instead of exiting 0. Run from a script, a remote management tool or ssh without a terminal and without --yes, the command read nothing, printed that the uninstall was cancelled and exited 0, so the caller took the agent for removed. It now prints confirmation required: rerun with --yes and exits non-zero; nothing is removed in either case. The behaviour dates back to 0.6.5. Pass --yes for an unattended removal.
  • license remove without a terminal fails instead of exiting 0. Without a terminal and without --force, the command printed Cancelled. and exited 0, so a script took the licence for removed. It now prints confirmation required: nothing was removed; rerun with --force and exits non-zero. The behaviour dates back to 0.6.5.
  • update --dry-run no longer says the update was installed. It printed "Update installed" and asked for a restart while nothing had been installed. It now prints the version it would install over the running one and that no change was made. The defect dates back to February 2025.
  • update latest follows auto_update.include_beta. Run from the command line, update latest always resolved on the stable channel, while update without argument and update --list honoured include_beta; an agent on the beta channel never received the next beta that way and read "Already up to date.". It now resolves on the channel the configuration sets. The defect dates back to July 2026.
  • Console: a probe changed by remote management no longer shows as stopped. A probe whose interval an OpAMP overlay changed appeared stopped, with the interval of the file, while it collected at the new one. The console now shows its live state and the interval it runs at, marked remote.
  • OpAMP (Enterprise): a first start no longer warns that the replay floor was lost. A newly enrolled agent that had not yet received an overlay logged this warning at every start, since its state file did not exist yet. It now logs one information line; an unreadable state file still warns.
  • update --list lists each release once, whatever the registry's order. The registry names its newest release under an alias as well (latest, latest-beta). An entry is now an alias when its name is not a version, and each version keeps the entry named after the release, so the newest beta always appears, once.
  • The configuration sent to a remote management server masks every secret. With OpAMP enabled, the effective configuration the agent sends to the server could show a password: a connection string under dsn or uri written as ${env:...} or left unsealed, a Cookie or session header, a parameter named like pass or pwd, or a password in a URL such as an http_check target http://user:password@host/. These values are now masked, as in config show --redact.