hyperhive/nix/agent-modules/otel.nix

# OpenTelemetry wiring for an agent container: the per-agent options the
# meta flake injects (host-driven from `services.hyperhive.otel.*`), the
# environment every OTLP producer inside the container reads, and the
# collector that forwards this container's journal to the hive.
#
# Two directions, one endpoint. Metrics are PUSHED by processes in here
# that speak OTLP themselves; logs are not pushed by anyone, because
# nothing in a journal has an SDK. The forwarder is what turns the second
# into the first, and it has to run inside the container — a hive-level
# reader cannot see this journal at all (the host-side per-container
# directory is an id-mapped bind mount journald never writes into).
#
# Claude Code is one producer here, not the owner. `hive-metric` — and
# any future in-container exporter — reads the same endpoint, protocol
# and resource labels, so they are container environment rather than
# claude settings. Claude's own switches (its telemetry master flag,
# which signals it emits) stay in `claude-settings.nix`, which reads the
# options declared below.
{
  lib,
  pkgs,
  config,
  ...
}:
let
  cfg = config.services.hyperhive.agent.otel;
  userName = config.services.hyperhive.agent.user.name;
  # Hive/swarm display names, read from the per-agent options meta.rs
  # renders (NOT from `environment.variables` — those carry the same
  # names at *runtime* only, so reading them here silently yielded
  # "unknown" on every agent while the process env held the right
  # answer). `null` means the hive did not name itself; "unknown" is then
  # an honest label rather than a guess.
  hiveDisplayName =
    if config.services.hyperhive.agent.hiveName == null then
      "unknown"
    else
      config.services.hyperhive.agent.hiveName;
  swarmDisplayName =
    if config.services.hyperhive.agent.swarmName == null then
      "unknown"
    else
      config.services.hyperhive.agent.swarmName;

  # Resource labels every producer in this container stamps on what it
  # emits. `service.name` names the container's role, not one binary
  # inside it: claude's samples and `hive-metric`'s both come from this
  # agent, and their metric names already tell them apart.
  resourceAttributes =
    "service.name=hyperhive-agent,agent=${userName},hive=${hiveDisplayName},swarm=${swarmDisplayName}"
    + lib.optionalString (cfg.extraResourceAttributes != "") ",${cfg.extraResourceAttributes}";

  # The variables an OTEL SDK reads on its own, in any language, in any
  # process here.
  #
  # There is no auth header among them, and no mechanism to add one. An
  # agent exports to the hive's own collector, which is the only thing
  # holding a credential for anything upstream; nothing an agent can read
  # is a secret to the swarm. An earlier revision forwarded the
  # operator's upstream token into this container and merged it into the
  # agent's own `~/.claude/settings.json` — which handed every agent the
  # hive's credential, and was removed with the direct-export path it
  # served.
  producerEnv = {
    OTEL_EXPORTER_OTLP_ENDPOINT = cfg.endpoint;
    OTEL_EXPORTER_OTLP_PROTOCOL = cfg.protocol;
    # Force CUMULATIVE temporality — Claude Code defaults to DELTA,
    # which Prometheus/Mimir-family backends (incl. grafana-lgtm)
    # silently drop without a deltatocumulative processor.
    OTEL_EXPORTER_OTLP_METRICS_TEMPORALITY_PREFERENCE = "cumulative";
    OTEL_RESOURCE_ATTRIBUTES = resourceAttributes;
  }
  // lib.optionalAttrs (cfg.metricIntervalMs != null) {
    OTEL_METRIC_EXPORT_INTERVAL = toString cfg.metricIntervalMs;
  };

  # The exporter's NAME picks the wire protocol, so `protocol` has to be
  # read here as well as handed to the SDKs above — same expression as the
  # swarm tier's upstream exporter in nix/host-modules/swarm-otel.nix.
  logExporter = if cfg.protocol == "grpc" then "otlp" else "otlphttp";
in
{
  # OTEL stats export is configured ONCE at host level via
  # `services.hyperhive.otel.*` (see nix/host-modules/otel.nix) and
  # injected into every agent's build by the meta-flake renderer
  # (`hive-c0re/src/meta.rs::otel_config`). These per-agent options are
  # the build-time implementation surface that injection writes into;
  # they are not meant to be set directly in an agent.nix. Marked
  # `internal` so the host option is the only documented operator knob.
  options.services.hyperhive.agent.otel = {
    enable = lib.mkOption {
      type = lib.types.bool;
      default = false;
      internal = true;
      description = ''
        Export this agent's telemetry — Claude Code's stats (token
        usage, cost, tool calls) and anything pushed with
        `hive-metric` — to an OTLP endpoint. Each agent exports
        directly to the hive's collector, so it keeps working even when
        hive-c0re is down. Host-driven: set
        `services.hyperhive.otel.enable` instead.
      '';
    };

    endpoint = lib.mkOption {
      type = lib.types.str;
      default = "";
      internal = true;
      description = ''
        OTLP collector endpoint, set as `OTEL_EXPORTER_OTLP_ENDPOINT`.
        Host-driven via `services.hyperhive.otel.endpoint`.
      '';
    };

    protocol = lib.mkOption {
      type = lib.types.enum [
        "http/protobuf"
        "http/json"
        "grpc"
      ];
      default = "http/protobuf";
      internal = true;
      description = ''
        OTLP wire protocol, set as `OTEL_EXPORTER_OTLP_PROTOCOL`.
        Host-driven via `services.hyperhive.otel.protocol`.
      '';
    };

    extraResourceAttributes = lib.mkOption {
      type = lib.types.str;
      default = "";
      internal = true;
      description = ''
        Extra comma-separated entries appended to
        `OTEL_RESOURCE_ATTRIBUTES` after the built-in
        `service.name` / `agent` / `hive` / `swarm` labels.
        Host-driven via `services.hyperhive.otel.extraResourceAttributes`.
      '';
    };

    metricIntervalMs = lib.mkOption {
      type = lib.types.nullOr lib.types.ints.positive;
      default = null;
      internal = true;
      description = ''
        Metric export interval in milliseconds, set as
        `OTEL_METRIC_EXPORT_INTERVAL`. Null leaves the SDK default (60s
        for Claude Code). Host-driven via
        `services.hyperhive.otel.metricIntervalMs`.
      '';
    };

    debug = lib.mkOption {
      type = lib.types.bool;
      default = false;
      internal = true;
      description = ''
        Emit OTEL SDK diagnostics to stderr (`CLAUDE_CODE_OTEL_DIAG_STDERR=1`).
        Host-driven via `services.hyperhive.otel.debug`.
      '';
    };
  };

  config = lib.mkIf cfg.enable {
    # Two assignments, because they cover disjoint sets of processes and
    # neither implies the other:
    #
    # - `systemd.globalEnvironment` is merged into the `Environment=`
    #   lines of every generated unit, so the harness, the
    #   bash/matrix/MCP daemons and everything the bash tool spawns
    #   beneath them all get it. This is the assignment that matters:
    #   those daemons are claude's *siblings*, not its children, so
    #   nothing shipped inside claude's own settings could ever reach
    #   them, and `hive-metric` invoked from a tool call exited with
    #   "OTEL_EXPORTER_OTLP_ENDPOINT not set".
    # - `environment.variables` lands in `/etc/set-environment`, sourced
    #   by `/etc/profile` — login shells, i.e. `hivectl shell` and
    #   `hivectl choom`, which systemd does not start.
    #
    # `NIX_REMOTE` in `default.nix` is set both ways for this same
    # reason. The difference is visible on any running agent:
    # `HIVE_ASSETS_DIR` is an `environment.variables` entry and is absent
    # from `systemctl show hive-bash-daemon.service -p Environment`,
    # while `NIX_REMOTE` — the same value, also set globally — is there.
    systemd.globalEnvironment = producerEnv;
    environment.variables = producerEnv;

    # The journal this container writes is reachable from inside it and
    # from nowhere else, which is why the forwarder runs here rather than
    # the hive reading the container's journal from outside: the host-side
    # per-container directory is an id-mapped bind mount and journald
    # writes nothing into it.
    assertions = [
      {
        # Sibling of the swarm collector's storage assertion in
        # nix/host-modules/swarm-otel.nix, and it exists because the
        # failure is silent at every layer: with a volatile journal the
        # receiver below finds an empty directory, reads nothing, and the
        # collector starts clean and stays healthy forever.
        assertion =
          !(lib.elem config.services.journald.storage [
            "volatile"
            "none"
          ]);
        message = ''
          services.hyperhive.agent.otel.enable is on for agent ${userName}, but
          services.journald.storage is
          "${config.services.journald.storage}" in this container.

          The log forwarder reads /var/log/journal, which journald only
          writes when it stores persistently: with "volatile" the journal
          lives in /run/log/journal and with "none" there is none at all.
          Either way the forwarder would ship nothing while looking
          healthy.
        '';
      }
    ];

    services.opentelemetry-collector = {
      enable = true;
      # Contrib, and not a preference: `journald` is a contrib receiver.
      # The upstream default build has no way to read a journal at all.
      package = pkgs.opentelemetry-collector-contrib;
      # Runs `otelcol validate` at build time. ⚠️ A parser, not a wiring
      # check — it accepts a pipeline naming a component the build lacks,
      # and the collector then dies at startup. Same caveat as both other
      # tiers; see nix/host-modules/otel.nix for the measurement.
      validateConfigFile = true;
      settings = {
        # `journalctl --follow --lines=0` (what this receiver runs under
        # the hood) ships only what's written *after* it starts — every
        # collector start has a silent gap at the front. `storage:` below
        # persists the read cursor, so a restart resumes from where it
        # left off instead of re-opening that gap; `start_at` stays `end`
        # (its default) because a cursor already covers every run after
        # the first, and `beginning` without one would re-ship the whole
        # journal on every restart. The cursor lives in the unit's own
        # `StateDirectory` (nixpkgs' module already sets one, `%S`, below
        # as `WorkingDirectory` too) — container-lifetime, not the
        # host-persisted state bind mount: it indexes /var/log/journal,
        # itself fresh on every container recreate, so the two must share
        # a lifetime or the cursor outlives the journal it points into.
        # `create_directory` matters here for a build-time-vs-runtime
        # reason, not a convenience one: `validateConfigFile` above runs
        # `otelcol validate` in the nix build sandbox, before systemd's
        # `StateDirectory=` has ever created this path. Without it the
        # validator refuses with "directory must exist" and the build fails.
        extensions.file_storage = {
          directory = "/var/lib/opentelemetry-collector";
          create_directory = true;
        };

        receivers.journald = {
          # ⚠️ STATED, and it must stay stated: the receiver's own default
          # is the RUNTIME journal (`/run/log/journal`), which in an agent
          # container is empty — journald stores persistently here, so
          # every entry is under /var/log/journal. Dropping this line
          # leaves a collector that validates, starts, reports healthy and
          # forwards nothing.
          directory = "/var/log/journal";
          storage = "file_storage";
          # No `units` allowlist, and no user-session filter like the swarm
          # tier's: a container's journal is the harness and what the
          # harness spawns, so there is no foreign traffic to filter out and
          # an allowlist would only be a list to forget to update.

          # journald's PRIORITY carries the level every line already has;
          # without this the record reaches VictoriaLogs with
          # `severity_text: Unspecified` and the store cannot tell an error
          # from a debug line. The table — and the inverted direction that
          # makes it worth a file of its own — lives in
          # ../journald-severity.nix, shared with the swarm tier's receiver.
          operators = import ../journald-severity.nix;
        };

        # Identity the hop above cannot supply. A host-side reader can say
        # which machine a line came from; only a collector running as the
        # agent can say which agent. `hive` and `swarm` are deliberately
        # NOT stamped here even though the env above carries them for
        # in-process producers: the swarm tier upserts `hive` from the
        # receiver that accepted the sample, precisely so the label comes
        # from something the sender cannot write.
        processors.resource.attributes = [
          {
            key = "service.name";
            value = "hyperhive-agent";
            action = "upsert";
          }
          {
            key = "agent";
            value = userName;
            action = "upsert";
          }
        ];

        # The hive's collector, at the same base address every in-process
        # producer already exports to. `endpoint` is a BASE the exporter
        # appends `/v1/logs` to — the path an OTLP/HTTP receiver serves.
        exporters.${logExporter} = {
          endpoint = cfg.endpoint;
        }
        // lib.optionalAttrs (cfg.protocol == "http/json") { encoding = "json"; };

        service.pipelines.logs = {
          receivers = [ "journald" ];
          processors = [ "resource" ];
          exporters = [ logExporter ];
        };

        # An extension configured but not listed here is INERT — the
        # journald receiver's `storage: file_storage` above would name a
        # component the collector never starts.
        service.extensions = [ "file_storage" ];
      };
    };
  };
}