tessaryai/tessary:compose-1.0.2

# Production stack. Three long-running services on the `tessary` network, plus the launcher:
#   - postgres  Postgres 16, EBS-backed data volume, healthcheck-gated
#   - backend   Spring Boot JVM (layered bootJar), internal :8080
#   - frontend  Caddy serving the built static + reverse-proxying /api → backend,
#               TLS on :443 when SITE_DOMAIN is set, plain HTTP on :8000 always
#
# EVERY SERVICE HERE STARTS ON A DEFAULT `up`, with ONE named exception: `redpanda`, the opt-in
# durable ingest buffer under `--profile kafka`. It is image-only, so a registry reference
# resolves it like every other service, and it is off unless an operator asks for it by name; what
# a reader sees is what boots, plus one block that says when it does not.
#
# NO TELEMETRY COLLECTOR. A Grafana Alloy sidecar used to sit here as the egress point for OTLP
# traces, Docker logs, host metrics and profiles, together with the backend's OTEL_* and
# PYROSCOPE_* keys that only ever addressed it. It was opt-in behind `profiles: ["observability"]`
# and started on no default `up`; it now lives in a separate overlay compose file that this
# edition does not ship, merged in by an file-exists probe when it is present.
#
# Nothing about a default boot changed. What changed is the reading cost: `docker compose -f
# oci://…` shows an operator a confirmation table of EVERY variable this file interpolates before
# it will run, and that block contributed fourteen GRAFANA_CLOUD_*/LANGFUSE_* rows plus two
# hostnames naming a container the file did not define. To ship OTLP somewhere from this stack,
# point OTEL_EXPORTER_OTLP_ENDPOINT at your own collector and add it as an override file.
#
# Usage:
#   # populate .env with the values your deployment needs (.env.example lists every key)
#   docker compose pull && docker compose up -d
#   # pin every image to a different published version at once: set TESSARY_VERSION in .env
#   # (each service's own *_IMAGE key still overrides individually when that's what you need)
#
# THIS FILE IS ALSO THE PUBLISHED ONE-COMMAND INSTALL. Every release
# pushes it verbatim as an OCI artifact, so someone with nothing cloned runs
#
#   docker compose -f oci://docker.io/tessaryai/tessary:compose up -d -y
#
# and gets this exact file, EXCEPT for one generated substitution: the release stamps each
# floating TESSARY_VERSION default below with the version it is publishing, so the artifact a remote
# install pulls is pinned to that release and can never name older images. `tessaryai/tessary`
# also carries the four IMAGE tags (`<service>-<version>`, `<service>-latest`); `compose` and
# `compose-<version>` are artifact tags in the same repository and shadow none of them.
#
# NO VERSION LITERAL LIVES IN THIS FILE, deliberately. The git tag `v<semver>` that
# .github/workflows/release.yml pushes is the only source of truth for a published version; a
# hand-typed fallback here would be a second one, and the copy that goes stale first. From a
# clone the unset default therefore FLOATS to `-latest`; set TESSARY_VERSION in .env to pin.
# scripts/check-version-consistency.sh holds that no such literal comes back.
#
# FOUR RULES THIS FILE NOW KEEPS, because a remote install has no repository to resolve against
# (scripts/check-compose-artifact.sh is the gate for all four):
#   1. NO HOST-PATH BIND in a default-profile service. A relative bind resolves against the file's
#      own directory, which for a remote install is Compose's cache under ~/.cache/docker-compose;
#      a ${PWD} bind writes into whatever directory the operator happened to run the command from.
#      Both leave state outside Docker. Postgres and the launcher work dir are named volumes now.
#   2. LONG-SYNTAX `ports:` and LITERAL memory limits. `docker compose publish` loads this file
#      WITHOUT interpolation (that is what preserves ${VAR} for the consumer), and both the short
#      port syntax and any ${VAR} in a memory field fail that stricter decode.
#   3. `env_file:` IS INERT FOR A REMOTE INSTALL. publish rewrites the path to a content hash and
#      ships no blob (including the blob is `--with-env`, which would bake the PUBLISHER's .env
#      into a public artifact — never pass it). A remote operator's `.env` still feeds every
#      ${VAR} below, because Compose reads .env from the working directory; only keys that are
#      passthrough-only reach nothing. docs/self-hosting/setup.mdx documents the override-file
#      escape hatch under "Operating a stack you started from a registry".
#   4. A PROFILE-GATED SERVICE MUST BE REMOTE-RESOLVABLE: image, volumes and env only, no host
#      bind. The old `observability` profile broke this (alloy bound
#      ./observability/alloy/config.alloy, which a registry reference cannot resolve) and moved to
#      the paid overlay. `redpanda` under `--profile kafka` is the one profile-gated service now,
#      and it keeps this rule: a named volume, flags, an image.

# Pin the Compose project name so the stack is identical from any directory
# (incl. git worktrees); otherwise Compose derives it from the cwd and spawns a
# duplicate set of containers/volumes per worktree.
name: tessary

services:
  postgres:
    # pgvector/pgvector:pg16 = Postgres 16 with the pgvector extension preinstalled, so
    # `CREATE EXTENSION vector` (migration 042) works. Drop-in for postgres:16-alpine — same
    # PG 16 data directory format, so existing volumes remain compatible.
    image: pgvector/pgvector:pg16
    environment:
      POSTGRES_DB: ${POSTGRES_DB:-tessary}
      # Defaulted, not `:?must be set`. Compose refuses to RENDER on an unset `:?` key, so
      # those two turned `docker compose up -d` on a fresh clone into a failure before any image was
      # pulled — the one command the quickstart is named after. This service publishes no host port,
      # so the credential is reachable only from the `tessary` network; change it if you publish the
      # port or share the host.
      POSTGRES_USER: ${POSTGRES_USER:-tessary}
      POSTGRES_PASSWORD: ${POSTGRES_PASSWORD:-tessary}
    volumes:
      # A NAMED VOLUME by default (rule 1 in the header), not the ./.local/postgres it used to be:
      # a relative bind is meaningless to a remote install, and the point of the one-command
      # install is that nothing lands on disk but what Docker manages. Compose classifies this by
      # the value's shape — a name is a volume, anything starting with / or ./ is a bind — so
      # POSTGRES_DATA_DIR still takes a host path and still means exactly what it meant: on the
      # deployed EC2 host, /data/postgres is a dedicated EBS mount that survives instance
      # replacement. `docker compose down` keeps the volume; `down -v` is what deletes the data.
      - ${POSTGRES_DATA_DIR:-tessary-postgres-data}:/var/lib/postgresql/data
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U $$POSTGRES_USER -d $$POSTGRES_DB"]
      interval: 5s
      timeout: 3s
      retries: 30
      start_period: 10s
    restart: unless-stopped
    networks:
      - tessary

  backend:
    # tessaryai/tessary is the one published repo: one TESSARY_VERSION
    # setting moves all four service tags together, and BACKEND_IMAGE (etc.) still overrides the
    # whole reference individually when that's what's needed. Unset means `-latest`, which the
    # release repoints — never a version literal (see the header). Until a human dispatches
    # .github/workflows/release.yml for the first time, no tag exists on either registry yet —
    # build locally instead (`docker compose up -d --build`) until it does.
    image: ${BACKEND_IMAGE:-tessaryai/tessary:backend-${TESSARY_VERSION:-1.0.2}}
    # A declared ceiling, so -XX:MaxRAMPercentage in backend/Dockerfile sizes the heap against a number
    # this file states rather than against whatever the host's Docker VM happens to have. Without it a
    # 2 GB dev VM and a large production host produce two very different JVMs from the same image, and
    # the ingest write buffer's byte budget (tessary.ingest.substrate.queue-max-bytes, 64 MiB) is tuned
    # against a heap size that is then only true by accident.
    #
    # A LITERAL, not the `${BACKEND_MEM_LIMIT:-2g}` it used to be (rule 2 in the header): publish's
    # non-interpolating loader decodes every memory field as a number, so an interpolation here —
    # under `mem_limit:` or under `deploy.resources.limits.memory:`, both were tried — made the
    # whole file unpublishable. Raising it is an override-file edit now, documented in
    # docs/self-hosting/setup.mdx.
    mem_limit: 2g
    expose:
      - "8080"
      # OTLP/gRPC trace receiver. Reachable on the internal `tessary` network so the Caddy
      # ingress can route :4317 to it. Inert unless tessary.ingest.otlp.enabled=true AND
      # tessary.ingest.otlp.transport is grpc|both. NOTE: publishing 4317 to the host additionally needs
      # `EXPOSE 4317` in backend/Dockerfile (a protected path — human-owned) and a
      # host `ports:` mapping wired with the ingress change.
      - "4317"
    env_file:
      - path: ./.env
        required: false
    environment:
      TESSARY_JDBC_URL: jdbc:postgresql://postgres:5432/${POSTGRES_DB:-tessary}
      # Same defaults as the postgres service above, and they have to be repeated rather than
      # inherited: these are what the BACKEND connects with, and a bare ${POSTGRES_USER} would hand
      # it an empty credential the moment the file stopped hard-failing on an unset one.
      TESSARY_DB_USERNAME: ${POSTGRES_USER:-tessary}
      TESSARY_DB_PASSWORD: ${POSTGRES_PASSWORD:-tessary}
      SPRING_PROFILES_ACTIVE: ${SPRING_PROFILES_ACTIVE:-production}
      # Liquibase phones home to config.liquibase.com at startup for its own analytics unless told
      # not to, through its own environment key. Off, because a self-hosted instance's one permitted
      # outbound destination is the opt-out heartbeat (docs/self-hosting/configuration.mdx, "What
      # leaves your network"); an egress audit found this lookup with the heartbeat
      # already disabled. Spring's `spring.liquibase.analytics-enabled` did not reach it.
      LIQUIBASE_ANALYTICS_ENABLED: "false"

      # THE TWO KEYS THIS INSTANCE SEALS EVERYTHING WITH, AND THEY SHIP INSECURE ON PURPOSE.
      #
      # Both defaults below are PUBLIC: they are in this file, in a public repository. Anyone can
      # forge a session cookie against an instance still running the first one, and decrypt every
      # provider API key an instance still running the second one has sealed. They exist so the
      # ten-minute quickstart is `docker compose up -d` and nothing else, on a localhost box holding
      # nothing real.
      #
      # PlaceholderSecretGuard is what keeps that from reaching a deployment: set SITE_DOMAIN to a
      # real hostname with either key still at its default and the backend REFUSES TO START, naming
      # the key. On localhost it warns on every boot instead. Replace both with
      # `openssl rand -base64 32` before this instance holds anything you care about.
      #
      # Each default is base64 of an exactly-32-byte ASCII string, which is not cosmetic: SessionCipher
      # and SecretBox both require a 32-byte decode, so a "REPLACE_ME"-shaped placeholder would boot
      # green and then fail at first sign-in. Distinct per key so the guard can name which one is still
      # default. `:-` treats an EMPTY value as unset, so a .env that carries these keys blank (as
      # .env.example used to ship them) also lands on the placeholder rather than on a broken boot.
      TESSARY_AUTH_COOKIE_PASSWORD: ${TESSARY_AUTH_COOKIE_PASSWORD:-Q0hBTkdFLU1FLWluc2VjdXJlLWRlZmF1bHQtY29va2k=}
      TESSARY_SECRET_KEY: ${TESSARY_SECRET_KEY:-Q0hBTkdFLU1FLWluc2VjdXJlLWRlZmF1bHQtc2VhbCE=}

      # What arms the guard above, and the ONE variable that puts this instance on a hostname:
      # PublicOriginGuard derives the sign-in landing origin and the WorkOS callback from
      # it, refuses a value that is not a bare hostname, and refuses to start when an explicitly-set
      # origin below names a different host. TLS_MODE and ACME_EMAIL are the frontend's settings,
      # repeated here so that guard can refuse an ACME boot with no account address before Caddy
      # would fail to obtain a certificate.
      SITE_DOMAIN: ${SITE_DOMAIN:-}
      TLS_MODE: ${TLS_MODE:-acme}
      ACME_EMAIL: ${ACME_EMAIL:-}

      # Where the browser lands after sign-in and sign-out. Derived from the port this file actually
      # publishes, because the code default (AuthProperties#frontendUrl) assumes :8000 while the
      # frontend service below publishes ${HTTP_PORT:-80} — so a self-hoster who did not
      # copy .env.example verbatim got a stack that booted clean, authenticated correctly, and then
      # bounced every sign-in to a port nothing was listening on, with no warning. With SITE_DOMAIN
      # set, this localhost default is replaced by https://<SITE_DOMAIN>/ at boot; an explicit
      # value on another host refuses the boot instead.
      TESSARY_AUTH_FRONTEND_URL: ${TESSARY_AUTH_FRONTEND_URL:-http://localhost:${HTTP_PORT:-80}/}
      # TESSARY_PLATFORM_BEDROCK_ENABLED lived here until it was removed — the switch it fed had zero
      # production callers (RCA/TRIAGE never resolved through it), so removing it changes no live
      # behavior. Every RCA/TRIAGE run resolves the org's own Bedrock/mantle (or other provider)
      # credential unconditionally now; there is no ambient-identity path left to opt into.
      # Where the backend reaches the agentic-analysis launcher (the sandbox-runner service on the
      # internal network). Only used when tessary.observer.analyzer-type=agentic; harmless otherwise.
      # Set explicitly (not silently blank) so an agentic analyzer can't no-op on a missing URL.
      TESSARY_OBSERVER_AGENTIC_LAUNCHER_URL: http://sandbox-runner:8080
      # A SHIPPED PLACEHOLDER, not blank — the same treatment TESSARY_AUTH_COOKIE_PASSWORD and
      # TESSARY_SECRET_KEY get above, and for the same reason: this is a symmetric secret between two
      # services in THIS compose file, so the value only has to match itself. Blank did not mean
      # "unauthenticated", it meant BROKEN: server.js rejects every request when its own key is
      # empty (`!SANDBOX_API_KEY || auth !== ...`), so an install that set nothing got a 401 on
      # every triage and RCA run with both sides agreeing the key was "". PlaceholderSecretGuard
      # covers the real exposure — an instance still on a public default must not face the network.
      TESSARY_OBSERVER_AGENTIC_LAUNCHER_API_KEY: ${TESSARY_OBSERVER_AGENTIC_LAUNCHER_API_KEY:-Q0hBTkdFLU1FLWluc2VjdXJlLWRlZmF1bHQtbGF1bmNo}
      # RCA — same launcher sidecar (and bearer key) as the observer's agentic lane. REQUIRED:
      # the sandboxed agent is RCA's only analysis path, so a blank launcher URL fails every
      # "Run RCA" press.
      TESSARY_RCA_AGENTIC_LAUNCHER_URL: http://sandbox-runner:8080
      # Same placeholder default, repeated verbatim rather than referencing the key above: compose
      # interpolates every one of these from the HOST environment, so a default written in one entry
      # is invisible to the others. All three must carry it or the pair silently disagrees.
      TESSARY_RCA_AGENTIC_LAUNCHER_API_KEY: ${TESSARY_OBSERVER_AGENTIC_LAUNCHER_API_KEY:-Q0hBTkdFLU1FLWluc2VjdXJlLWRlZmF1bHQtbGF1bmNo}
      # The evidence door, for BOTH agent lanes. Where it has to point depends on where the agent
      # runs, and that is SANDBOX_NETWORK_ISOLATION's call (sandbox-runner below):
      #
      #   isolation OFF (default) -- the sibling agent container shares this backend's network, so
      #     the internal service name resolves and needs no public hostname. This is what lets a
      #     plain `docker compose up` run agentic RCA and Triage out of the box.
      #   isolation ON, or the E2B backend -- the agent is off this network (a microVM is off the
      #     host entirely) and can only come back in through the public origin. Set SITE_DOMAIN and
      #     this picks it up; the concatenation is safe only because PublicOriginGuard refuses a
      #     SITE_DOMAIN that already carries a scheme.
      #
      # Neither lane has a fallback: their dossiers carry the detector's claim and every row behind
      # it is fetched over MCP, so an unreachable value dead-letters every triage job and fails
      # every RCA press. It is a deployment fault and it fails loudly as one — including the
      # scheme-only `https://` an isolated install with no SITE_DOMAIN used to build, which
      # PublicOriginGuard now rejects instead of handing to the agent.
      # The default is the INTERNAL service name, which is correct precisely because isolation
      # defaults off: the agent shares this network and resolves it. It stays correct on an install
      # that sets SITE_DOMAIN, too — a sibling on the service network reaches the backend directly
      # whether or not a public origin also exists, and skips a round trip through Caddy to do it.
      # The two deployments that CANNOT use it — SANDBOX_NETWORK_ISOLATION=1, and SANDBOX_BACKEND=e2b,
      # where the agent is off this host entirely — set this variable to their public origin, and
      # PublicOriginGuard rejects the value if they set it to something an outside caller can't reach.
      TESSARY_RCA_AGENTIC_MCP_BASE_URL: ${TESSARY_RCA_AGENTIC_MCP_BASE_URL:-http://backend:8080}
      TESSARY_CLASSIFIER_TRIAGE_MCP_BASE_URL: ${TESSARY_RCA_AGENTIC_MCP_BASE_URL:-http://backend:8080}
      # Encoder classification endpoint — the standalone classify-service on ECS Fargate
      # (both values delivered via SSM -> .env). Required: signal sweeps fail loudly
      # without it; there is no fallback endpoint.
      TESSARY_OBSERVER_ENCODER_URL: ${TESSARY_OBSERVER_ENCODER_URL:-}
      TESSARY_OBSERVER_ENCODER_API_KEY: ${TESSARY_OBSERVER_ENCODER_API_KEY:-}
      # The ingest spool: `memory` (default, no broker) or `kafka`, the opt-in durable buffer
      # served by the `redpanda` service below under `--profile kafka`. Nothing depends_on it on
      # purpose: with the mode left at memory the profile never starts, and with it set to kafka a
      # broker that is still coming up simply sheds (503 + Retry-After) until it answers.
      TESSARY_INGEST_SPOOL_MODE: ${TESSARY_INGEST_SPOOL_MODE:-memory}
      TESSARY_INGEST_SPOOL_KAFKA_BOOTSTRAP_SERVERS: ${TESSARY_INGEST_SPOOL_KAFKA_BOOTSTRAP_SERVERS:-redpanda:9092}
      # The sweep, and how long the instance keeps traces and detections; 0 keeps forever.
      # Interpolated here rather than left to env_file so the registry install, which carries no
      # .env, still reads the operator's .env beside the command (rule 3 in the header).
      TESSARY_RETENTION_ENABLED: ${TESSARY_RETENTION_ENABLED:-true}
      TESSARY_RETENTION_TRACE_TTL_DAYS: ${TESSARY_RETENTION_TRACE_TTL_DAYS:-90}
      TESSARY_RETENTION_DETECTION_TTL_DAYS: ${TESSARY_RETENTION_DETECTION_TTL_DAYS:-90}
      # The opt-OUT heartbeat, interpolated for that same reason, and it matters more here: left
      # to env_file alone, an operator on the registry install who sets this to false in their
      # .env gets no error and keeps sending the ping.
      TESSARY_TELEMETRY_ENABLED: ${TESSARY_TELEMETRY_ENABLED:-true}
    depends_on:
      postgres:
        condition: service_healthy
    # "Up" means serving, and the stack says so itself: every default
    # service carries a compose-side probe, one mechanism for all four, so `docker compose ps`
    # reports `healthy` and the frontend's start waits on it. The runtime image ships no curl or
    # wget, so bash's /dev/tcp speaks the one request; /actuator/health is the exact path AuthFilter
    # admits without a session, and it is 200 only once the database indicator is UP.
    # Latency this adds to the boot clock: at most one interval (5 s) after the first UP; a success
    # inside start_period counts immediately, so start_period adds none. Unhealthy arrives no sooner
    # than start_period plus 60 retries x 5 s, 315 s, and only then do dependents stop waiting.
    healthcheck:
      test: ["CMD", "bash", "-c", "exec 3<>/dev/tcp/127.0.0.1/8080 && printf 'GET /actuator/health HTTP/1.0\\r\\nHost: localhost\\r\\n\\r\\n' >&3 && grep -q '\"status\":\"UP\"' <&3"]
      interval: 5s
      timeout: 3s
      retries: 60
      start_period: 15s
    restart: unless-stopped
    networks:
      - tessary
    logging:
      driver: json-file
      options:
        max-size: "10m"
        max-file: "3"
        tag: "tessary/backend"

  frontend:
    # Same tessaryai/tessary convention as backend above — TESSARY_VERSION moves both together.
    image: ${FRONTEND_IMAGE:-tessaryai/tessary:frontend-${TESSARY_VERSION:-1.0.2}}
    # Long syntax, not "${HTTP_PORT:-80}:8000" (rule 2 in the header): publish's non-interpolating
    # loader rejects a short-form port string outright, so the file could not be published at all.
    # `published` must stay quoted — it is a string here, which is what carries the ${VAR}.
    ports:
      - mode: ingress
        target: 8000
        published: "${HTTP_PORT:-80}"
        protocol: tcp
      - mode: ingress
        target: 443
        published: "${HTTPS_PORT:-443}"
        protocol: tcp
    environment:
      BACKEND_URL: backend:8080
      # frontend/caddy/render.sh builds the Caddyfile from these at container start, so a
      # blank SITE_DOMAIN simply means no TLS site (the inert-loopback placeholder the old static
      # Caddyfile needed is gone). TLS_MODE picks how the domain is served: acme (Let's
      # Encrypt over TLS-ALPN-01, ACME_EMAIL required and never defaulted to an address that is not
      # the operator's), owncert (a certificate mounted at /certs, see docs/self-hosting/custom-domain.mdx),
      # or upstream (plain HTTP on :8000 behind the operator's own terminator, whose addresses go in
      # TRUSTED_PROXIES so the forwarded client IP and scheme are honoured from it and from nobody else).
      SITE_DOMAIN: ${SITE_DOMAIN:-}
      TLS_MODE: ${TLS_MODE:-acme}
      ACME_EMAIL: ${ACME_EMAIL:-}
      TRUSTED_PROXIES: ${TRUSTED_PROXIES:-}
      TLS_CERT_FILE: ${TLS_CERT_FILE:-/certs/tls.crt}
      TLS_KEY_FILE: ${TLS_KEY_FILE:-/certs/tls.key}
    volumes:
      # Caddy needs to persist its ACME state across restarts so we don't
      # re-request certs on every redeploy (rate-limited by Let's Encrypt).
      - caddy-data:/data
      - caddy-config:/config
    # Waits on the backend SERVING, not merely created, so a `healthy` frontend line in
    # `docker compose ps` means the whole path a browser takes is open. Caddy's own probe is its
    # plain-HTTP site answering the SPA shell; caddy:2-alpine carries busybox wget.
    # Latency added to the boot clock: at most one interval (5 s) after Caddy is up; start_period
    # is 5 s because Caddy serves within a second of starting.
    depends_on:
      backend:
        condition: service_healthy
    healthcheck:
      test: ["CMD", "wget", "-q", "-O", "/dev/null", "-T", "3", "http://127.0.0.1:8000/"]
      interval: 5s
      timeout: 3s
      retries: 12
      start_period: 5s
    restart: unless-stopped
    networks:
      - tessary
    logging:
      driver: json-file
      options:
        max-size: "10m"
        max-file: "3"
        tag: "tessary/caddy"

  # The opt-in durable ingest buffer: a single-node Redpanda, a Kafka-API broker in one binary,
  # started only by `docker compose --profile kafka up -d` together with TESSARY_INGEST_SPOOL_MODE=kafka.
  # Internal network only, no published port; its own data volume is what makes an accepted batch
  # survive a restart. Sized for the quickstart box: one core, 512 MB, overprovisioned, and the
  # startup resource checks skipped (`--check=false`) — and fsync kept, explicitly. The image's
  # shipped redpanda.yaml carries `developer_mode: true`, and it has to stay: with it off Redpanda
  # refuses to start below 1 GiB. From that mode rpk would add `--unsafe-bypass-fsync=true` to the
  # binary's flags, which would make "persisted before the 200" untrue, so the flag is passed as
  # `false` outright and rpk keeps the explicit value (rpk still prints its dev-mode data-corruption
  # WARNING on every start; that warning is about the bypass it would have added, which is what
  # the explicit flag denies). The proof is the binary's own `Starting redpanda` line, not rpk's. Write caching is off at the cluster default and again per topic (the
  # spool sets `write.caching=false` on the topics it creates, where the broker accepts it). The
  # phone-home reporter (`enable_metrics_reporter`, which posts to Redpanda's endpoint) is off at
  # boot, so the zero-egress posture holds with the profile on as well as off; `enable_usage` is
  # the local usage-stats collector, off too. `--set` is two tokens: rpk parses it that way only.
  redpanda:
    image: docker.redpanda.com/redpandadata/redpanda:v25.2.1
    profiles: ["kafka"]
    command:
      - redpanda
      - start
      - --smp=1
      - --memory=512M
      - --reserve-memory=0M
      - --overprovisioned
      - --check=false
      - --unsafe-bypass-fsync=false
      - --set
      - redpanda.write_caching_default=false
      - --set
      - redpanda.enable_metrics_reporter=false
      - --set
      - redpanda.enable_usage=false
      - --kafka-addr=internal://0.0.0.0:9092
      - --advertise-kafka-addr=internal://redpanda:9092
    volumes:
      - redpanda-data:/var/lib/redpanda/data
    healthcheck:
      test: ["CMD-SHELL", "rpk cluster health | grep -E 'Healthy:.+true' || exit 1"]
      interval: 10s
      timeout: 5s
      retries: 12
      start_period: 10s
    restart: unless-stopped
    networks:
      - tessary
    logging:
      driver: json-file
      options:
        max-size: "10m"
        max-file: "3"

  # Agentic-analysis launcher (drives sandboxed OpenCode runs for RCA/triage/analysis/synthesis/
  # codegen). Reachable only on the internal `tessary` network at http://sandbox-runner:8080 — the
  # backend's tessary.observer.agentic.launcher-url. Deliberately NOT given `env_file: .env` (which
  # the backend uses) so it receives only its own minimal secrets; the AWS_* keys are remapped
  # from the OBSERVER_LAUNCHER_-prefixed SSM vars so they never override the backend container's
  # instance-profile creds.
  # One-shot init: Docker creates a fresh named volume as root:root mode 0755, unwritable by
  # sandbox-runner's non-root `node` user (uid 1000, the default in node:24-alpine3.24; see
  # sandbox-runner/launcher/Dockerfile's final `USER node`). This runs as root and chowns it to
  # 1000:1000 so the launcher's mkdtempSync calls (server.js's LAUNCHER_WORK_DIR) succeed. Exits
  # immediately; sandbox-runner waits on it via `condition: service_completed_successfully` below.
  sandbox-runner-work-init:
    image: busybox:1.36
    command: ["sh", "-c", "mkdir -p /launcher-work && chown 1000:1000 /launcher-work"]
    volumes:
      - sandbox-work:/launcher-work
    restart: "no"

  sandbox-runner:
    depends_on:
      sandbox-runner-work-init:
        condition: service_completed_successfully
    # Starts by default. It was gated behind the `observer-launcher` profile after the
    # 2026-07-12 OOM (four warm ONNX heads, ~1.5 GB, OOM-looping the shared host); that hazard
    # is gone — encoder /classify now lives in the standalone classify-service on ECS Fargate,
    # and this launcher is lightweight orchestration only, hard-capped below at 1536m so
    # any relapse kills this container alone, never the host.
    #
    # Ungated once RCA deleted its inline-LLM fallback, so "Run RCA" now reaches
    # this launcher on EVERY click. While the profile stayed off, each click POSTed to an
    # unresolvable `sandbox-runner` host and died as `agentic RCA transport failure`. A profile
    # that must be flipped on for the product to work at all is not a safety valve, it is an
    # outage waiting for someone to notice — mem_limit is the real guard.
    # A `build:` block, matching backend/frontend above: before this, `docker compose
    # up -d --build` built backend+frontend from source but still tried to PULL sandbox-runner,
    # so a bad or missing published tag failed as `manifest unknown` instead of building locally
    # — one of five default-profile services had no source path at all. Same shape as
    # docker-compose.dev.yml's sandbox-runner service.
    # Same tessaryai/tessary convention as backend/frontend above.
    image: ${SANDBOX_RUNNER_IMAGE:-tessaryai/tessary:sandbox-runner-${TESSARY_VERSION:-1.0.2}}
    expose:
      - "8080"
    # The launcher's own GET /healthz (sandbox-runner/launcher/server.js), unauthenticated by
    # design; node:24-alpine carries busybox wget. Same 5 s interval as the others, so a
    # launcher that exits on a config it rejects shows as unhealthy, not merely restarting.
    healthcheck:
      test: ["CMD", "wget", "-q", "-O", "/dev/null", "-T", "3", "http://127.0.0.1:8080/healthz"]
      interval: 5s
      timeout: 3s
      retries: 12
      start_period: 5s
    # `USER node` (sandbox-runner/launcher/Dockerfile) is non-root but is NOT a member of whatever
    # group owns /var/run/docker.sock on the host — on a typical Linux Docker install that's the
    # `docker` group, group-writable (0660, root:docker), and its GID varies per install (read it
    # as a container sees it, per the comment below). This grants that ONE supplementary group rather than running
    # the container as root, matching the "keep the launcher non-root" mitigation described
    # below.
    #
    # The default is 0, not the common Debian/Ubuntu `docker` GID of 999, because Docker Desktop
    # mounts the socket root:root 0660 — so 0 is the answer for every Mac and Windows install, which
    # is most people running this compose file for the first time. 999 sent them all to `EACCES
    # /var/run/docker.sock` at boot and a launcher that could not spawn a single agent container.
    # On a Linux host whose socket really is root:docker, set DOCKER_SOCK_GID to that group AS A
    # CONTAINER SEES IT: `docker run --rm -v /var/run/docker.sock:/var/run/docker.sock busybox:1.36
    # stat -c %g /var/run/docker.sock`.
    group_add:
      - "${DOCKER_SOCK_GID:-0}"
    volumes:
      # The docker sandbox driver spawns fresh, hardened SIBLING containers for each
      # agentic run by talking to the Engine API over this socket — READ-WRITE. It is the only
      # socket mount left in this file: alloy's read-only, log-only one went to the paid overlay
      # with the service. This is functionally host-root-adjacent
      # access for the launcher process itself; no flag on a spawned sibling container changes
      # that fact about the process sitting behind the socket. Inherent to this architecture,
      # not something designed around — mitigated by keeping the launcher non-root (see
      # sandbox-runner/launcher/Dockerfile's `USER node`) and adding no other host mount to this
      # service beyond this and the work dir below. Only exercised when SANDBOX_BACKEND=docker
      # (the default); a self-hoster who switches to e2b/local never needs this mount at all, but
      # it is unconditional here rather than profile-gated so switching backends is one env var,
      # not a compose edit.
      - /var/run/docker.sock:/var/run/docker.sock
      # The other half of the Docker-outside-of-Docker path pair (see server.js's LAUNCHER_WORK_DIR
      # comment for the full mechanism). A NAMED VOLUME, not the ${PWD}/.data/sandbox-runner-work
      # bind it used to be: a host path is something only the operator's own filesystem can supply,
      # which a remote install has no way to name (rule 1 in the header) — and the daemon resolves a
      # volume NAME identically from inside a container and from the host, which is exactly the
      # property the bind was there to fake. The launcher mounts the same volume into each sibling
      # container by name, addressing the per-run subdirectory with the Engine API's
      # VolumeOptions.Subpath; SANDBOX_WORK_VOLUME below carries the name so it can.
      - sandbox-work:/launcher-work
    environment:
      PORT: "8080"
      # The launcher half of the pair the backend presents. Same placeholder default, same reason as
      # the backend's two entries: an empty key here is not "no auth", it is server.js rejecting
      # every request, so it must never resolve blank just because the operator set no env var.
      SANDBOX_API_KEY: ${TESSARY_OBSERVER_AGENTIC_LAUNCHER_API_KEY:-Q0hBTkdFLU1FLWluc2VjdXJlLWRlZmF1bHQtbGF1bmNo}
      # The open default — see server.js's own Env-block comment for what each value means.
      # Unset here (falls through to the code default 'docker') rather than hardcoded, so a
      # self-hoster who opts into e2b or local only has to set ONE var, not edit this file.
      SANDBOX_BACKEND: ${SANDBOX_BACKEND:-}
      # Cut agent containers off from this network. Off by default — see server.js's own Env-block
      # comment for the trade. Turning it on means the agent can only reach the backend over the
      # public origin, so an install that sets this must also set SITE_DOMAIN and point
      # TESSARY_RCA_AGENTIC_MCP_BASE_URL at it.
      SANDBOX_NETWORK_ISOLATION: ${SANDBOX_NETWORK_ISOLATION:-}
      # The tag is agent-sandbox, not agent: this Docker tag is deliberately named to
      # match the E2B template alias tessary-agent-sandbox. Pulled lazily by the launcher itself
      # over docker.sock (server.js), not by `docker compose pull` — so `up -d --build` cannot
      # build this one locally the way it can backend/frontend/sandbox-runner above. Until a human
      # dispatches release.yml for the first time no tag exists yet, and "Run RCA" 502s
      # (server.js:765-771) — set AGENT_IMAGE to something you built or already have locally if you
      # need this before the first release ships.
      AGENT_IMAGE: ${AGENT_IMAGE:-tessaryai/tessary:agent-sandbox-${TESSARY_VERSION:-1.0.2}}
      SANDBOX_DOCKER_CONCURRENCY: "1"
      # Must be the SAME volume as the `sandbox-work` mount above — see that mount's comment. The
      # docker backend fails at boot if this is unset. The top-level volume declares this literal
      # as its `name:`, so Compose does not project-prefix it and the launcher can hand the daemon
      # the name verbatim.
      SANDBOX_WORK_VOLUME: ${SANDBOX_WORK_VOLUME:-tessary-sandbox-work}
      E2B_API_KEY: ${E2B_API_KEY:-}
      # NAMESPACED, AND VERSIONED, for the same reason AGENT_IMAGE above is both. E2B scopes a
      # template name to the project that built it, so the bare `tessary-agent-sandbox` this used
      # to say resolved ONLY for a key belonging to the Tessary team — a self-hoster who set
      # SANDBOX_BACKEND=e2b with their own key got `Sandbox.create` finding nothing. The template
      # is published public; `tessary/…` is the form anyone's key can reach.
      # Unset floats to `:latest`, which release.yml's finalize repoints; the version lives in the
      # git tag, never in a literal here.
      E2B_ANALYZER_TEMPLATE: ${E2B_ANALYZER_TEMPLATE:-tessary/tessary-agent-sandbox:${TESSARY_VERSION:-1.0.2}}
      # FULL REMOVAL of every deployment-env-var credential the launcher used to read
      # here — AWS_REGION/AWS_ACCESS_KEY_ID/AWS_SECRET_ACCESS_KEY/MANTLE_REGION, the whole
      # AGENT_PROVIDER dev-override block (ANTHROPIC_API_KEY, AWS_BEARER_TOKEN_BEDROCK,
      # OPENAI_API_KEY, GEMINI_API_KEY, GLM_API_KEY/GLM_BASE_URL, GROK_API_KEY,
      # CUSTOM_OPENAI_BASE_URL/CUSTOM_OPENAI_API_KEY). The launcher now reads NO provider secret
      # from its own process env — every /rca and /triage carries the org's own decrypted
      # ProviderCredential on the request itself (see server.js's file-header doc and
      # AgenticCredentialResolver on the Java side). MANTLE_PROJECT_ID stays below: it is an
      # attribution scope, not a secret, and has no per-org column to move to.
      MANTLE_PROJECT_ID: ${TESSARY_MANTLE_PROJECT_ID:-}
    # The limit must be BELOW host RAM to mean anything — the 2026-07-12 outage was a
    # classification burst ballooning past the 4 GB host inside a 4g "limit": the kernel
    # thrashed the page cache at the EBS IOPS cap and froze every other container. 1.5g
    # OOMs this container alone; restart: unless-stopped brings it back in seconds. Unrelated to
    # the docker backend's OWN per-sibling-container memory cap (SANDBOX_DOCKER_MEMORY_MB) — this
    # limit is on the launcher process, that one is on each spawned agent container.
    mem_limit: 1536m
    restart: unless-stopped
    networks:
      - tessary
    logging:
      driver: json-file
      options:
        max-size: "10m"
        max-file: "3"
        tag: "tessary/sandbox-runner"

networks:
  tessary:
    driver: bridge

volumes:
  redpanda-data:
  caddy-data:
  caddy-config:
  # The Postgres data volume the `postgres` service falls back to when POSTGRES_DATA_DIR names no
  # host path. NO explicit `name:` — Compose project-prefixes it, which is what keeps a clean-room
  # boot (COMPOSE_PROJECT_NAME set, as the boot checks do) from sharing or destroying a
  # developer's own database. Find it with `docker volume ls | grep postgres-data`.
  tessary-postgres-data:
  # The launcher's work volume — shared between sandbox-runner and every sibling agent container it
  # spawns. Explicit `name:` is load-bearing here, not cosmetic: the launcher hands this string to
  # the Engine API as a sibling container's mount source, and a project-prefixed name would differ
  # between what Compose created and what the launcher asks for. That also means the name is NOT
  # isolated by COMPOSE_PROJECT_NAME the way every other volume here is, so a clean-room boot must
  # set SANDBOX_WORK_VOLUME to a run-scoped value or its `down -v` reaches another stack's volume.
  sandbox-work:
    name: ${SANDBOX_WORK_VOLUME:-tessary-sandbox-work}