Files
clawmates/deploy/compose
Omar SobhandClaude Opus 5 0a2bd6f868
deploy / test (push) Failing after 1m54s
deploy / build (push) Skipped
feat(decide): cm-decide — typed calibrated decisions; Jev + local NLI backends; skill triage in shadow
A third kind of decision-maker between deterministic code and a full LLM
call: Choice / Score / Noul questions answered as probability
distributions with a confidence, behind one Decider trait, with the
composition patterns (confidence gating, composite scoring, rerank) as
code. Two backends: TypeSafe's Jev over HTTP, and a DeBERTa-v3 MNLI
cross-encoder run in-process with candle (feature nli; metal/cuda).

decide-eval measures a backend on labelled cases the way judge-eval
measures the judge. eval/skill-triage.json: 20 mission tasks × 53 skills,
75 positives, hand-labelled. Measured 2026-09-21:

  lexical overlap        AUROC 0.851  [email protected] 0.47  top-k 48/75  ECE 0.095
  jev (named wording)    AUROC 0.989  [email protected] 0.84  top-k 63/75  ECE 0.064  213 ms
  jev (plain wording)    AUROC 0.970  [email protected] 0.66  top-k 52/75
  nli mnli-base          AUROC 0.790  [email protected] 0.28  top-k 38/75  ECE 0.263  1.5 s
  nli zeroshot-v2        AUROC 0.782  [email protected] 0.43  top-k 39/75  ECE 0.054  1.2 s

The vendor's calibration claim survives our data; the local cross-encoder
ranks below keyword overlap on either checkpoint or wording and is kept
as the measured negative, not shipped. A local backend would need the
logit-readout route over the fleet's 9B model — a separate spike.

Shadow: one Jev call per phase launch (spawned, 10 s cap, silent without
TYPESAFE_API_KEY) records a skill.triage event; the Skill-Use report
carries triage_p beside each skill's Trigger verdict. It selects nothing.

Co-Authored-By: Claude Opus 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01WZb5A2kfVfjpdwSochkuHz
2026-09-21 10:08:31 -05:00
..

Clawmates on Docker Compose

A single-node deployment of the whole platform — the same images the Kubernetes/Helm path uses, with the same §15 security topology. This is the air-gapped appliance route; it is equally usable as a simple self-host.

What runs

Service Role
postgres database (named volume pgdata)
broker the secret broker — credentials never leave it; reachable only over a private socket volume shared with the server
socket-proxy allow-listed Docker API so the server can spawn agent sandboxes and nothing else
server API + streaming gateway + agent runtime + scheduler
frontend the Next.js web app

Networks core, secrets_net, sandbox_net, and engine_net are all internal: true; only edge is published. Agent sandboxes run with no network at all (egress is the browser container's alone).

Quick start

cd deploy/compose
cp .env.example .env
# Edit .env: set POSTGRES_PASSWORD and CLAWMATES_BOOTSTRAP_OWNER_PASSWORD.

# Either build the images locally...
docker compose build
# ...or load them from a signed release bundle (see deploy/airgapped/install.sh).

docker compose up -d

The server self-migrates on boot and, if CLAWMATES_BOOTSTRAP_OWNER_PASSWORD is set, provisions the first Owner + workspace once (a no-op on every later boot). Then open:

Configuration

Host- and secret-specific values come from .env (overlaid onto clawmates.toml at runtime — CLAWMATES_* env always wins). The knobs:

Concern Where Notes
DB password .env POSTGRES_PASSWORD also forms the server/broker DSN
Image tag .env CLAWMATES_VERSION latest after a local build, or a release version
First owner .env CLAWMATES_BOOTSTRAP_* password set = bootstrap on first boot
LLM clawmates.toml [llm] openai_compat (default, point base_url at vLLM/Ollama/llama.cpp) or anthropic (+ ANTHROPIC_API_KEY in .env)
Auth clawmates.toml [auth] + .env AUTH_MODE local (default) or clerk (see docs/clerk.md)
Tracing .env CLAWMATES_TELEMETRY__OTLP_ENDPOINT unset = logs only, no egress

Using Anthropic instead of a local model

In clawmates.toml:

[llm]
provider = "anthropic"
model = "claude-sonnet-4-6"

and set ANTHROPIC_API_KEY in .env.

The broker master key — back it up

On first boot the broker generates its encryption key into the broker_key volume and logs a reminder. Every stored credential is unrecoverable without it. Back it up:

docker compose cp broker:/etc/clawmates-broker/broker.key ./broker.key.backup

Operations

docker compose ps                 # status
docker compose logs -f server     # follow server logs
docker compose pull && docker compose up -d   # upgrade to a new CLAWMATES_VERSION
docker compose down               # stop (keeps volumes/data)
docker compose down -v            # stop AND delete all data

Production notes

  • Put a TLS-terminating reverse proxy in front of ports 3000/8080; the SSE gateway needs response buffering off (the Helm ingress sets this; nginx: proxy_buffering off; with long read timeouts).
  • socket-proxy mounts the Docker socket read-only and exposes only the container lifecycle verbs — verified by crates/cm-sandbox/tests/socket_proxy.rs.
  • The full bring-up is exercised end to end by scripts/rehearse-install.sh (verify a signed bundle → docker loadcompose up → assert the login page serves), which also runs in the release pipeline.