fix(client/rust): handle provider_status arm (build break); real system-roots TLS; design doc (Client.Rust-030..032)

This commit is contained in:
Joseph Doherty
2026-06-15 02:39:11 -04:00
parent 47062c1a6e
commit b57d02cc4d
7 changed files with 442 additions and 65 deletions
+80 -12
View File
@@ -162,12 +162,73 @@ impl GatewayClient {
`stream_alarms` opens with one `active_alarm` per currently-active alarm
(the ConditionRefresh snapshot), then a single `snapshot_complete`, then a
`transition` for every subsequent raise / acknowledge / clear. The feed is
served by the gateway's always-on alarm monitor — no worker session is
opened — so any number of clients may attach. Dropping the stream cancels
the gRPC call cooperatively. `acknowledge_alarm` is idempotent at the
MxAccess layer; the returned `AcknowledgeAlarmReply` carries the native
MxStatus from the worker.
`transition` for every subsequent raise / acknowledge / clear. A fourth
`provider_status` oneof case (`AlarmProviderStatus`: `mode`, `degraded`,
`reason`, `since`) is emitted once on stream open and again on every
failover/failback so late joiners learn the current alarm-provider mode.
The CLI renders all four cases in both its one-line summary and its
protobuf-JSON output (`alarm_feed_message_summary` /
`alarm_feed_message_to_json`). The feed is served by the gateway's always-on
alarm monitor — no worker session is opened — so any number of clients may
attach. Dropping the stream cancels the gRPC call cooperatively.
`acknowledge_alarm` is idempotent at the MxAccess layer; the returned
`AcknowledgeAlarmReply` carries the native MxStatus from the worker.
## Galaxy Repository
`GalaxyClient` is a session-less metadata client (requires the
`metadata:read` API-key scope). Alongside `test_connection`,
`get_last_deploy_time`, `discover_hierarchy`, and `watch_deploy_events`, it
exposes a lazy hierarchy walker built on the `BrowseChildren` RPC:
```rust
impl GalaxyClient {
pub async fn browse(&mut self, options: Option<BrowseChildrenOptions>) -> Result<Vec<LazyBrowseNode>, Error>;
pub async fn browse_children_raw(&mut self, request: BrowseChildrenRequest) -> Result<BrowseChildrenReply, Error>;
}
pub struct BrowseChildrenOptions {
pub category_ids: Vec<i32>,
pub template_chain_contains: Vec<String>,
pub tag_name_glob: Option<String>,
pub include_attributes: Option<bool>,
pub alarm_bearing_only: bool,
pub historized_only: bool,
}
impl LazyBrowseNode {
pub fn object(&self) -> &GalaxyObject;
pub fn has_children_hint(&self) -> bool;
pub async fn children(&self) -> Vec<LazyBrowseNode>;
pub async fn is_expanded(&self) -> bool;
pub async fn expand(&self) -> Result<(), Error>;
}
```
- `browse(options)` returns the root objects as `LazyBrowseNode`s. The
supplied `BrowseChildrenOptions` filter is captured and reused when any
returned node is expanded, so a single filter set scopes the entire walk.
- `BrowseChildrenOptions` mirrors the request-level filters on the wire and
combines them with **AND**: a child appears only when it satisfies every
populated criterion (`category_ids` membership, every
`template_chain_contains` substring, the `tag_name_glob`, plus the
`alarm_bearing_only` / `historized_only` flags). `include_attributes` is a
tri-state (`None` = server default). Empty/`None` fields impose no
restriction. See
[Galaxy Repository — BrowseChildren](../../docs/GalaxyRepository.md#browsechildren)
for the wire-level semantics.
- `LazyBrowseNode` is cheap to clone — clones share state through an internal
`Arc`, so expanding one clone makes the children visible to every clone.
`has_children_hint()` exposes the server's `child_has_children` hint so a UI
can draw an expand affordance without issuing an RPC. `expand()` is
idempotent: the first call issues a paged `BrowseChildren` walk (page size
500) under an async mutex held across the await, sets the `is_expanded`
flag, and caches the children; subsequent calls are no-ops and re-hit
nothing. The internal paged loop guards against a server returning a
repeated `next_page_token` by failing with `Error::InvalidArgument` rather
than looping forever.
- `browse_children_raw` issues a single `BrowseChildren` RPC and returns the
raw reply for callers that want to drive paging themselves.
## Authentication
@@ -200,13 +261,20 @@ Rust client is therefore **pin-only** — it requires either:
- `ClientOptions::with_ca_file(...)` to pin a CA (the supported path for the
gateway's self-signed certificate; export the certificate and pin it), or
- `ClientOptions::with_require_certificate_validation(true)` to verify against the
system trust roots.
operating system's trust roots. This enables the `tonic` `tls-native-roots`
feature and calls `ClientTlsConfig::with_native_roots()`, so the handshake
validates a certificate that chains to a root the host already trusts. It does
**not** accept a bare self-signed gateway certificate — that still needs
`with_ca_file`.
With TLS enabled (`with_plaintext(false)`), no pinned CA, and certificate
validation not required, `GatewayClient::connect` rejects the connection with a
clear, actionable error pointing at `with_ca_file` /
`require_certificate_validation` rather than silently accepting the certificate.
The CLI exposes `--ca-file` and `--require-certificate-validation`.
`build_tls_config` computes the trust posture with the pure `tls_trust_decision`
helper (`None` / `PinnedCa` / `SystemRoots` / `RejectNoCa`) so the posture is
unit-testable without a live handshake. With TLS enabled (`with_plaintext(false)`),
no pinned CA, and certificate validation not required (`RejectNoCa`),
`GatewayClient::connect` rejects the connection with a clear, actionable error
pointing at `with_ca_file` / `require_certificate_validation` rather than building
a config with zero trust anchors. The CLI exposes `--ca-file` and
`--require-certificate-validation`.
## Streaming