Files
archy/apps/routstr/manifest.yml
ssmithxandClaude Opus 5.5 d3b1f37d97 fix(routstr): name the sold model, and mark peer models as remote
The seller advertised "Local model (llama-server)", which the buyer's AIUI
showed as if the model ran on the buyer. The seller now names the model
("Qwen2.5 1.5B Instruct"), and the buyer appends "· via mesh peer" to
every model it lists from a peer.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-25 14:21:37 +00:00

120 lines
3.6 KiB
YAML

# Development candidate; not catalog-published or lifecycle-qualified.
app:
id: routstr
name: Routstr Provider
version: 0.5.0
upstream:
kind: github
repo: Routstr/routstr-core
description: >-
Private Cashu inference payment gateway for this node's llama-server.
Accepts the node wallet's mints and pays earnings out to the node wallet's
Lightning address. Mesh serving requires the authenticated FIPS gateway and
an explicit opt-in; installation does not publish it.
container:
build:
context: /opt/archipelago/docker/routstr
dockerfile: Dockerfile
tag: localhost/archy-routstr:0.5.0-archy1
network: archy-net
generated_secrets:
- name: routstr-admin-password
kind: hex32
secret_env:
- key: ADMIN_PASSWORD
secret_file: routstr-admin-password
# llama-server's generated key: the only credential for its private API.
- key: UPSTREAM_API_KEY
secret_file: llama-server-api-key
# The node wallet owns which mints are trusted and where earnings go.
# Re-applied on every start by archy-start.py (Routstr's saved settings
# would otherwise win after the first boot).
derived_env:
- key: CASHU_MINTS
template: "{{ECASH_MINTS}}"
- key: RECEIVE_LN_ADDRESS
template: "{{ECASH_LN_ADDRESS}}"
dependencies:
- app_id: llama-server
- storage: 2Gi
resources:
memory_limit: 1Gi
cpu_limit: 2
security:
capabilities: []
readonly_root: true
no_new_privileges: true
network_policy: isolated
ports:
- host: 8200
container: 8000
protocol: tcp
bind: 127.0.0.1
auth: local
volumes:
- type: bind
source: /var/lib/archipelago/routstr
target: /data
options: [rw]
- type: tmpfs
target: /tmp
tmpfs_options: rw,noexec,nosuid,size=128m
environment:
- DATABASE_URL=sqlite+aiosqlite:////data/keys.db
- ROUTSTR_SECRET_KEY_FILE=/data/routstr_secret.key
- MODELS_PATH=/data/models.json
- AUTO_GENERATE_NSEC=false
- ENABLE_ANALYTICS_SHARING=false
- HTTP_URL=
- ONION_URL=
- RELAYS=
- PRIMARY_MINT_URL=
- UPSTREAM_BASE_URL=http://llama-server:8080/v1
- UPSTREAM_5XX_RETRY_ATTEMPTS=0
# Below the node gateway's own limit: Routstr v0.5.0 keeps an X-Cashu
# token without refund when the upstream times out or the caller hangs up,
# so the slowest answer must still end inside every outer timeout.
- UPSTREAM_READ_TIMEOUT=240
- ENABLE_PRICING_REFRESH=false
- ENABLE_MODELS_REFRESH=false
- ENABLE_MODEL_PATHS_REFRESH=false
- PROVIDERS_REFRESH_INTERVAL_SECONDS=0
# Install-default price for the llama-server alias ($1 / $2 per million
# prompt / completion tokens). Applied once; the operator's price wins.
# Name and context are re-stated every start and must match the model file
# and llama-server's --ctx-size (apps/llama-server/README.md).
- ARCHY_MODEL_ID=archy-llama
- ARCHY_MODEL_NAME=Qwen2.5 1.5B Instruct
- ARCHY_MODEL_CONTEXT=8192
- ARCHY_MODEL_MAX_COMPLETION=1024
- ARCHY_DEFAULT_PROMPT_USD_PER_TOKEN=0.000001
- ARCHY_DEFAULT_COMPLETION_USD_PER_TOKEN=0.000002
# Process health only; does not claim a model or mint is ready.
health_check:
type: tcp
endpoint: localhost:8000
interval: 30s
timeout: 5s
retries: 3
interfaces:
main:
name: Private provider API
description: Host-local API; not a browser launch surface or public mesh endpoint.
type: api
port: 8200
protocol: http
path: /v1
metadata:
category: tools
tier: optional
author: Routstr
repo: https://github.com/Routstr/routstr-core