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>
Routstr provider — development candidate
This is the local-server foundation of
docs/routstr-fips-llama-integration.md.
It is not yet lifecycle-qualified or catalog-published. The node's FIPS
gateway (/routstr/v1/models, /routstr/v1/chat/completions) serves it to mesh
peers only after the operator sets routstr_selling on; it has not yet been
exercised between two real nodes.
Do not install it over an existing Routstr wallet/database to test it.
Package boundary
- Builds Routstr Core v0.5.0 at immutable source commit
4aa58bf380dd41f8e3ab7dc768b8e1d112d92019, using its frozen uv dependency lock. - Upstream's image workflow publishes moving
latestimages; we therefore build from reviewed source instead of inventing an upstream versioned image tag. Base-image digest pinning and a production image build remain acceptance gates. - Host API binds only
127.0.0.1:8200.auth: localprevents the app gate from opening external listeners. There is deliberately no browser launch interface. - Root filesystem is read-only;
/datacontains the database, encryption key, relative.walletdirectory and logs./tmpis disposable. - A dedicated generated admin secret enters through Podman secrets. No appliance Nostr private key is copied. Admin credential UI is not wired yet.
- Upstream is the node's private
llama-server(http://llama-server:8080/v1), authenticated with llama-server's generated key (llama-server-api-key). - The node ecash wallet owns payment settings.
CASHU_MINTSrenders from{{ECASH_MINTS}}(the wallet's accepted mints; Minibits by default) andRECEIVE_LN_ADDRESSfrom{{ECASH_LN_ADDRESS}}(the wallet's Minibits address), so earnings above Routstr's 210-sat payout floor are paid to the node wallet, where the existing claim flow redeems them. Never leaveCASHU_MINTSempty: Routstr then silently trusts its built-inmint.minibits.cashdefault. - Routstr's database copy of settings wins over the environment after the first
start, so
archy-start.py(the image's PID 1) re-applies the mints and payout address through the admin API on every start. Mint changes made in Routstr's own admin UI are therefore overwritten by the node wallet's list. - The
archy-llamamodel is priced once at install defaults ($1 / $2 per million prompt / completion tokens, plus Routstr's 1% fee). Routstr refuses to serve an unpriced model; an operator's later price is never replaced. - No Nostr identity or publication address; analytics and upstream 5xx retries are disabled.
- Upstream starts exchange-rate/background tasks; disabled pricing refresh does not establish zero external traffic. Runtime egress qualification is pending.
- TCP health means the API listener is up, not that inference/payments are ready.
llama-server connection contract
The reviewed upstream GenericUpstreamProvider accepts an OpenAI-compatible base
URL and optional API key. The initial target is llama.cpp's llama-server:
- provider base URL points at the private runtime's
/v1API; - fetch
/v1/models, then explicitly import/enable the chosen model and pricing; - enforce a stable model alias and bounded context/output/concurrency;
- verify
usagetoken accounting and the actual model's chat template; - do not advertise tool/streaming support until tested with that model;
- keep the raw runtime off external listeners and give it a dedicated upstream key.
A candidate managed runtime now lives in apps/llama-server; its source build
and lifecycle still need qualification. Private network isolation, configuration
RPC/key handoff and model acquisition UI are still to be implemented. There is
intentionally no automatic model download or default cloud upstream. Merely sharing archy-net does not
isolate the runtime from other installed apps; that is a remaining security gate.
Payment modes (observed at the pinned revision)
X-Cashu: <token> pays for one request and returns change in the X-Cashu
response header (the whole token when the request is refused). Authorization: Bearer <token> instead funds a server-side sk- balance and returns no change
until POST /v1/balance/refund, which replays the same token if repeated. Buyers
must use X-Cashu; the existing Archipelago clients are fixed by the ngit
proposal "fix(routstr): pay per request with X-Cashu so change is refunded".
Upstream routstr/balance.py has persisted refund retrieval and pending outcomes;
routstr/upstream/base.py emits millisatoshi cost headers. Neither is evidence
that an ambiguous completion can safely be resubmitted. Recovery must not issue
another payment or inference request by default.
Source-only checks
From the repository root:
bash scripts/validate-app-manifest.sh apps/routstr/manifest.yml
python3 -m unittest discover -s tests/routstr-provider -p 'test_*.py' -v
These tests do not build/start containers, contact a provider, or touch wallets.
For future disposable builds, use docker/routstr as the build context. For node
packaging the manifest references /opt/archipelago/docker/routstr, following the
normal runtime payload contract. Do not stage this on a live node for acceptance
without an explicit test plan and isolated data.