Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript

docs(hcc): resolve the 4 non-blocking findings from the final audit

Alfredo Lopez37856b2b73cc
diff --git a/Makefile b/Makefile
index 7334f00fc..b92b177d7 100644
--- a/Makefile
+++ b/Makefile
@@ -944,7 +944,7 @@ test-e2e-hcc-readiness-bootstrap: ## Prove HCC readiness while its initial Host
E2E_HCC_READINESS_FAULT_INJECTION=1 E2E_EXPECTED_PRE_GATE_GATE="$(E2E_EXPECTED_PRE_GATE_GATE)" MINIKUBE_PROFILE=$(E2E_KUBECONTEXT) KUBECONTEXT=$(E2E_KUBECONTEXT) bash scripts/e2e/e2e-hcc-readiness-bootstrap.sh
.PHONY: test-e2e-hcc-watch-churn-readiness
-test-e2e-hcc-watch-churn-readiness: ## Prove HCC readiness survives sustained apiserver watch churn (PR #205 GKE livelock). EXPECT_LIVELOCK=1 reproduces the bug.
+test-e2e-hcc-watch-churn-readiness: ## Prove HCC readiness CONVERGES under sustained apiserver watch churn (PR #205 GKE livelock). Positive-only; the livelock RED lives in host-context-controller/src/k8sClient.test.ts. EXPECT_LIVELOCK=1 is refused.
@echo "Running HCC watch-churn readiness gate..."
@test -n "$(E2E_EXPECTED_PRE_GATE_GATE)" || { echo "Set E2E_EXPECTED_PRE_GATE_GATE to the gate recorded by the branch-owned pre-gate sync" >&2; exit 1; }
E2E_HCC_WATCH_FAULT_INJECTION=1 EXPECT_LIVELOCK=$(EXPECT_LIVELOCK) E2E_EXPECTED_PRE_GATE_GATE="$(E2E_EXPECTED_PRE_GATE_GATE)" MINIKUBE_PROFILE=$(E2E_KUBECONTEXT) KUBECONTEXT=$(E2E_KUBECONTEXT) bash scripts/e2e/e2e-hcc-watch-churn-readiness.sh
diff --git a/host-context-controller/src/k8sClient.ts b/host-context-controller/src/k8sClient.ts
index 465f120a8..ff4874570 100644
--- a/host-context-controller/src/k8sClient.ts
+++ b/host-context-controller/src/k8sClient.ts
@@ -207,6 +207,10 @@ type HostWatchEventType = 'ADDED' | 'MODIFIED' | 'DELETED'
type InitialConvergenceLane = 'McpServer' | 'NetworkPolicy'
type NetworkPolicySafetyCertificate = {
+ // VESTIGIAL — retained for record shape only, never read in any decision.
+ // The certificate is identified SOLELY by content revision (the PR #382 fix);
+ // the generation fields survive as diagnostics and MUST NOT gate readiness
+ // (reading them would reintroduce the channel-identity livelock).
contextGeneration: number
serverGeneration: number
contextRevision: number
@@ -1656,6 +1660,12 @@ export class McpServerWatcher implements McpServerProvider {
// Kubernetes bumps metadata.generation only on a spec change (not on the
// status writes HCC itself makes), and the uid distinguishes a delete+recreate
// of the same name. Same (uid, generation) = same desired Host state.
+ // SCOPE: unlike the McpServer/Context comparators (which hash spec+labels+
+ // annotations), this is blind to Host annotations/labels — correct today
+ // because no Host annotation/label is load-bearing for a mutation (the wake
+ // annotation drives dispatch, not template content). If a Host annotation or
+ // label ever becomes mutation-relevant, extend this to hash it, or a
+ // reconnect that re-LISTs an unchanged (uid, generation) would skip the bump.
private sameHostDesiredRevision(previous: HostCRD, current: HostCRD): boolean {
return previous.uid === current.uid && previous.generation === current.generation
}
diff --git a/scripts/e2e/e2e-hcc-watch-churn-readiness.sh b/scripts/e2e/e2e-hcc-watch-churn-readiness.sh
index 578d71b18..c894191a1 100755
--- a/scripts/e2e/e2e-hcc-watch-churn-readiness.sh
+++ b/scripts/e2e/e2e-hcc-watch-churn-readiness.sh
@@ -83,8 +83,9 @@ CHURN_MIN_AGE_MS="${CHURN_MIN_AGE_MS:-3000}"
CHURN_MIN_CUTS="${CHURN_MIN_CUTS:-3}"
READINESS_BUDGET_SEC="${READINESS_BUDGET_SEC:-360}"
STABILITY_WINDOW_SEC="${STABILITY_WINDOW_SEC:-30}"
-# Longest tolerated CONTINUOUS 503 streak under churn: 5s recovery timer +
-# re-LIST + one retried cut cycle. A livelock produces an unbounded streak.
+# Longest tolerated CONTINUOUS 503 streak under churn: immediate re-LIST +
+# re-WATCH + a jittered backoff retry if that first attempt fails. A livelock
+# produces an unbounded streak.
RECOVERY_BUDGET_SEC="${RECOVERY_BUDGET_SEC:-15}"
# 1Hz observation window under churn after first Ready (fix mode): ~8 cut
# cycles at the default 7s cadence, enough for repeated 503->200 evidence.