Multi-node fleet architecture: sous-api + souslet over mTLS gRPC - #1

Merged

codemug merged 36 commits into

mainfrom

Sep 1, 2026

Commits on Aug 31, 2026

docs: bring in the multi-node spec and implementation plan
codemug
committed
e710fa6
feat(souslet): add gRPC wire contract and generated code

codemug
and
claude
committed
c71ce37

Commits on Sep 1, 2026

fix(souslet): pin protoc version and fix go.mod indirect dependencies

codemug
and
claude
committed
de924af
feat(mtls): self-issued CA for per-node souslet client certs
codemug
committed
0b532ff
feat(nodecatalog): in-memory per-node state, level-triggered replace
codemug
committed
9a9bb4d
feat(grpcserver): API-side Souslet service, snapshot ingestion, correlated Send
codemug
committed
a2f6e41
fix(grpcserver): stop leaking the write-loop goroutine on node disconnect
codemug
committed
39f9bbe
fix(grpcserver): unblock Send when a node disconnects while a reply is pending
codemug
committed
d31c26c
feat(grpcclient): dispatch Envelope commands to local engine/fetch

codemug
and
claude
committed
697fa40
feat(grpcclient): cache declared footprint so Snapshot reports real WeightsGib/KvGib

codemug
and
claude
committed
45fcd3a
feat(souslet): connect/reconnect loop and the souslet binary
codemug
committed
aca0336
fix(souslet): reset reconnect backoff on a genuinely healthy connection
codemug
committed
ecabbac
feat(httpapi): route deploy/undeploy/plan through grpcserver, node-scoped

codemug
and
claude
committed
1e66a1c
fix(httpapi): guard nil gsrv/nodes, restore WarnFreeGiB for node-scoped plan

codemug
and
claude
committed
ea43778
feat(sous-api): control-plane binary, persisted node CA
codemug
committed
9e34866
feat(gateway): proxy inference traffic over the souslet gRPC connection

codemug
and
claude
committed
47b4527
fix(gateway): chunk proxied request bodies, stop relaying on client disconnect

codemug
and
claude
committed
6ce610d
feat(deploy): fetch weights first when not yet cached on the target node
codemug
committed
fa75005
fix(deploy): poll FetchCommand until done, fix Connect's catalog/conns ordering

codemug
and
claude
committed
33e5494
fix(grpcclient): HandleFetch checks Status before Start, so polling observes done

codemug
and
claude
committed
b97eb27
feat(weights): recipe-card cleanup replaces the per-node larder page

codemug
and
claude
committed
a5a5711
fix(weights): restore archived-recipe rollback protection at sous-api

codemug
and
claude
committed
b2f5f69
fix(weights): browser form wiring + active-reference guard tier

codemug
and
claude
committed
f643ff5
feat(ui): per-node dashboard cards for the multi-node fleet

codemug
and
claude
committed
98f2481
fix(ui): stop fleet-card segments from claiming false readiness

codemug
and
claude
committed
cd250ca
feat(ui): drag-and-drop recipe deploy onto node capacity cards
codemug
committed
9a59429
fix(ui): withhold drag-and-drop deploy from archived recipes
codemug
committed
0039a08
chore: remove single-node Sous, larder, and old deploy.Manager; add node-cert CLI

codemug
and
claude
committed
5bce138
docs: point README at the two-image sous-api/souslet build

codemug
and
claude
committed
1f940c8
fix(mtls): give sous-api a real server certificate, and save CA state atomically

codemug
and
claude
committed
ef0f0dd
fix(grpcserver): verify node identity, survive reconnects, and stop letting uploads block commands

codemug
and
claude
committed
76edbb3
fix(souslet): keep re-reporting node state instead of snapshotting once at connect

codemug
and
claude
committed
ef030a7
fix(souslet): allocate a real host port for node-scoped deploys

codemug
and
claude
committed
76662de
fix(gateway): wire the multi-node proxy into the real server, and enforce key scope on it

codemug
and
claude
committed
ef3a075
docs: describe souslet's port range and its continuous state reporting

codemug
and
claude
committed
bf8e6b7
fix(souslet): reject a non-positive -port-low, and drop a doc comment the port fix superseded

codemug
and
claude
committed
2e27cdc