From d91e6eb0e2fd3cc863a27a49292ab00ed24045e4 Mon Sep 17 00:00:00 2001
From: damocles <damocles@hyperhive.local>
Date: Fri, 10 Jul 2026 21:17:39 +0200
Subject: [PATCH] fix(hive-metric): blocking OTLP client (no-reactor panic),
 reject negative counters, allow negative gauges

---
 Cargo.lock              |  3 ++-
 hive-metric/Cargo.toml  |  8 ++++++--
 hive-metric/src/main.rs | 31 +++++++++++++++++++++++++++----
 3 files changed, 35 insertions(+), 7 deletions(-)

diff --git a/Cargo.lock b/Cargo.lock
index c5872403..e505921b 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1591,7 +1591,6 @@ dependencies = [
  "opentelemetry",
  "opentelemetry-otlp",
  "opentelemetry_sdk",
- "tokio",
 ]
 
 [[package]]
@@ -3425,7 +3424,9 @@ checksum = "219c5811de6525e5416c7d5d53bb656d3afdbc6c5af816e0802bcfa42dbdc1c3"
 dependencies = [
  "base64",
  "bytes",
+ "futures-channel",
  "futures-core",
+ "futures-util",
  "http",
  "http-body",
  "http-body-util",
diff --git a/hive-metric/Cargo.toml b/hive-metric/Cargo.toml
index 794d135f..d71614be 100644
--- a/hive-metric/Cargo.toml
+++ b/hive-metric/Cargo.toml
@@ -13,10 +13,14 @@ clap.workspace = true
 # OTEL SDK — only used by this crate, so kept local rather than in workspace.dependencies.
 opentelemetry = "0.32"
 opentelemetry_sdk = { version = "0.32", features = ["metrics"] }
+# Blocking (not async) reqwest client on purpose: the metrics SDK drives the
+# OTLP push from a `PeriodicReader` background thread that has no Tokio runtime,
+# so the async client panics there with "no reactor running". The blocking
+# client sends on that thread directly — and for a fire-and-forget CLI that
+# records one point then `shutdown()`s, a synchronous send is exactly right.
 opentelemetry-otlp = { version = "0.32", default-features = false, features = [
   "metrics",
   "http-json",
-  "reqwest-client",
+  "reqwest-blocking-client",
   "reqwest-rustls",
 ] }
-tokio = { workspace = true, features = ["rt", "macros"] }
diff --git a/hive-metric/src/main.rs b/hive-metric/src/main.rs
index 931509a3..48e066eb 100644
--- a/hive-metric/src/main.rs
+++ b/hive-metric/src/main.rs
@@ -24,13 +24,19 @@ use opentelemetry_otlp::{Protocol, WithExportConfig};
 #[derive(Parser)]
 #[command(
     name = "hive-metric",
-    about = "Push a labeled metric to the OTEL collector (requires services.hyperhive.otel.enable = true)"
+    about = "Push a labeled metric to the OTEL collector (requires services.hyperhive.otel.enable = true)",
+    // Let a leading-minus value parse as the `value` positional rather than
+    // being rejected as an unknown flag — negative gauges are legitimate
+    // (e.g. a delta), and it lets the counter-sign check below produce a clear
+    // domain error instead of clap's "unexpected argument '-1'".
+    allow_negative_numbers = true
 )]
 struct Cli {
     /// Metric name (e.g. `tasks_completed`, `latency_ms`).
     name: String,
 
-    /// Numeric value (f64). Floats and integers both accepted.
+    /// Numeric value (f64). Floats and integers both accepted. Gauges may be
+    /// negative; counters must be >= 0 (they only increase).
     value: f64,
 
     /// Metric kind: `counter` (cumulative sum, default) or `gauge` (instantaneous value).
@@ -52,8 +58,7 @@ enum MetricKind {
     Gauge,
 }
 
-#[tokio::main]
-async fn main() -> Result<()> {
+fn main() -> Result<()> {
     let cli = Cli::parse();
 
     // Fail early with a clear message when OTEL is not configured.
@@ -66,6 +71,18 @@ async fn main() -> Result<()> {
 
     let labels = parse_labels(&cli.labels)?;
 
+    // Counters are monotonically non-negative — the OTEL SDK silently drops a
+    // negative `add`, so a typo like `hive-metric tasks_done -1` would vanish
+    // with exit 0. Reject it loudly instead.
+    if matches!(cli.metric_type, MetricKind::Counter) && cli.value < 0.0 {
+        bail!(
+            "counter '{}' value must be >= 0 (got {}); counters only increase — \
+             use `--type gauge` for a value that can go down",
+            cli.name,
+            cli.value
+        );
+    }
+
     let exporter = opentelemetry_otlp::MetricExporter::builder()
         .with_http()
         .with_protocol(Protocol::HttpJson)
@@ -73,6 +90,12 @@ async fn main() -> Result<()> {
         .context("failed to build OTLP metric exporter")?;
 
     // SDK automatically reads OTEL_RESOURCE_ATTRIBUTES for resource labels.
+    // The metrics SDK has no `with_simple_exporter` (that's a *traces*-only
+    // API) — a push exporter is always driven by a periodic reader. For this
+    // fire-and-forget CLI the timer never actually fires: we record one metric
+    // and immediately `shutdown()`, which force-flushes the pending export
+    // synchronously before returning. So there's no background-timer flush race
+    // in practice — the single point is exported exactly once on shutdown.
     let provider = opentelemetry_sdk::metrics::SdkMeterProvider::builder()
         .with_periodic_exporter(exporter)
         .build();