feat(remote-signer): fix secret modification

internal/agentruntime/charts.go

@@ -5,12 +5,14 @@ package agentruntime
 // deployments. It MUST be updated as a single edit; bumping it here
 // updates every consumer in lockstep.
 //
-// Chart 0.3.2 ships remote-signer image `v0.3.0`, which emits canonical
-// Ethereum recovery-id signatures (`v=27/28`) from `/sign/.../message`,
-// `/sign/.../typed-data`, and `/sign/.../hash`. Earlier images returned
-// `v=0/1` (alloy y-parity), which was rejected by EIP-712 / ERC-3009
-// verifiers like USDC `transferWithAuthorization` and forced the buy.py
-// caller to renormalize.
+// Chart 0.3.3 ships remote-signer image `v0.4.0`, which honours
+// `SIGNER__AUTH__TOKEN` (the bearer token the controller mints into the
+// keystore Secret and injects via env). Canonical Ethereum recovery-id
+// signatures (`v=27/28`) from `/sign/.../message`, `/sign/.../typed-data`,
+// and `/sign/.../hash` have been the baseline since chart 0.3.2 / image
+// v0.3.0 — earlier images returned `v=0/1` (alloy y-parity) and forced
+// the buy.py caller to renormalize for EIP-712 / ERC-3009 verifiers like
+// USDC `transferWithAuthorization`.
 //
 // renovate: datasource=helm depName=remote-signer registryUrl=https://obolnetwork.github.io/helm-charts/
-const RemoteSignerChartVersion = "0.3.2"
+const RemoteSignerChartVersion = "0.3.3"

internal/embed/embed_crd_test.go

@@ -875,6 +875,8 @@ func TestServiceOfferControllerSecretRBAC_Scoped(t *testing.T) {
 
 	readNames := map[string]bool{}
 	deleteNames := map[string]bool{}
+	updateNames := map[string]bool{}
+	patchNames := map[string]bool{}
 	var sawCreate bool
 	for _, r := range rules {
 		rm := r.(map[string]any)
@@ -908,9 +910,25 @@ func TestServiceOfferControllerSecretRBAC_Scoped(t *testing.T) {
 			if verbs["delete"] {
 				deleteNames[n] = true
 			}
+			if verbs["update"] {
+				updateNames[n] = true
+			}
+			if verbs["patch"] {
+				patchNames[n] = true
+			}
 		}
-		if verbs["list"] || verbs["watch"] || verbs["update"] || verbs["patch"] {
-			t.Error("serviceoffer-controller scoped secrets rule must not grant list/watch/update/patch — Secrets are create-only in the reconciler and all reads are by name")
+		if verbs["list"] || verbs["watch"] {
+			t.Error("serviceoffer-controller scoped secrets rule must not grant list/watch — all reads are by name")
+		}
+		// update/patch is allowed only on remote-signer-keystore, which the
+		// reconciler updates via backfillSignerAuthToken to add the bearer
+		// token key to keystores minted before signer auth existed.
+		if verbs["update"] || verbs["patch"] {
+			for n := range names {
+				if n != "remote-signer-keystore" {
+					t.Errorf("serviceoffer-controller must not grant secrets:update/patch on %s — only remote-signer-keystore is mutated (auth-token backfill)", n)
+				}
+			}
 		}
 		if names["litellm-secrets"] && verbs["delete"] {
 			t.Error("serviceoffer-controller must not grant secrets:delete on litellm-secrets; the code only reads LITELLM_MASTER_KEY")
@@ -930,6 +948,16 @@ func TestServiceOfferControllerSecretRBAC_Scoped(t *testing.T) {
 			t.Errorf("serviceoffer-controller must grant resourceName-scoped secrets:delete on %s for agent teardown", name)
 		}
 	}
+	// backfillSignerAuthToken (agent_wallet.go) calls Update on the keystore
+	// Secret to add the signer-auth bearer token to legacy keystores. Without
+	// update + patch the Agent stays in Provisioning and every downstream
+	// ServiceOffer condition blocks on WaitingForAgent.
+	if !updateNames["remote-signer-keystore"] {
+		t.Error("serviceoffer-controller must grant resourceName-scoped secrets:update on remote-signer-keystore for backfillSignerAuthToken")
+	}
+	if !patchNames["remote-signer-keystore"] {
+		t.Error("serviceoffer-controller must grant resourceName-scoped secrets:patch on remote-signer-keystore for backfillSignerAuthToken")
+	}
 	if !sawCreate {
 		t.Error("serviceoffer-controller must retain secrets:create for minting the per-agent API token + wallet keystore in dynamic namespaces")
 	}

internal/embed/infrastructure/base/templates/x402.yaml

@@ -143,16 +143,19 @@ rules:
   # ServiceAccount + PVC are namespace-scoped child resources.
   #
   # Secrets: the reconciler touches exactly three, by name:
-  #   - litellm-secrets        get               (read LITELLM_MASTER_KEY, ns llm)
-  #   - hermes-api-server      get/delete/create (mint the agent API token)
-  #   - remote-signer-keystore get/delete/create (mint the agent wallet keystore)
-  # Secrets are create-only in the reconciler (isCreateOnlyKind): the API token
-  # is preserved, not rotated, and the keystore is minted once and never
-  # reshaped, so no `update`/`patch` verb is needed. `litellm-secrets` is
-  # get-only; delete is confined to the two per-agent Secret names that teardown
-  # owns. `create` is split into its own rule below: resourceNames cannot scope
-  # create, and agent namespaces are minted dynamically, so create stays
-  # namespace-wide. That is an integrity surface only.
+  #   - litellm-secrets        get                      (read LITELLM_MASTER_KEY, ns llm)
+  #   - hermes-api-server      get/delete/create        (mint the agent API token)
+  #   - remote-signer-keystore get/update/patch/delete/create (mint + backfill auth token)
+  # `hermes-api-server` is create-only: the token is preserved, not rotated.
+  # `remote-signer-keystore` needs `update`/`patch` for a one-shot backfill
+  # path (backfillSignerAuthToken in agent_wallet.go) that adds the
+  # signer-auth bearer-token key to keystores minted before signer auth
+  # existed. Idempotent: presence of the key (even empty) is a no-op, so
+  # operator-rotated tokens are never clobbered. `litellm-secrets` is
+  # get-only; delete is confined to the two per-agent Secret names that
+  # teardown owns. `create` is split into its own rule below: resourceNames
+  # cannot scope create, and agent namespaces are minted dynamically, so
+  # create stays namespace-wide. That is an integrity surface only.
   - apiGroups: [""]
     resources: ["serviceaccounts"]
     verbs: ["create"]
@@ -173,8 +176,12 @@ rules:
     verbs: ["get"]
   - apiGroups: [""]
     resources: ["secrets"]
-    resourceNames: ["hermes-api-server", "remote-signer-keystore"]
+    resourceNames: ["hermes-api-server"]
     verbs: ["get", "delete"]
+  - apiGroups: [""]
+    resources: ["secrets"]
+    resourceNames: ["remote-signer-keystore"]
+    verbs: ["get", "update", "patch", "delete"]
   - apiGroups: [""]
     resources: ["secrets"]
     verbs: ["create"]

internal/serviceoffercontroller/agent_wallet.go

@@ -15,11 +15,12 @@ import (
 // Constants for the per-agent remote-signer side-stack. Image pinned by
 // digest is desirable but the chart still publishes by tag — keeping the
 // version synced with agentruntime.RemoteSignerChartVersion's notes
-// (chart 0.3.2 → image v0.3.0, the canonical recovery-id behaviour).
+// (chart 0.3.3 → image v0.4.0, the first image that honours
+// SIGNER__AUTH__TOKEN).
 const (
 	remoteSignerName  = "remote-signer"
 	remoteSignerPort  = 9000
-	remoteSignerImage = "ghcr.io/obolnetwork/remote-signer:v0.3.0"
+	remoteSignerImage = "ghcr.io/obolnetwork/remote-signer:v0.4.0"
 	// Image hard-codes /data/keystores as the default and reads its
 	// config under the SIGNER__... env namespace; values picked to match
 	// the master agent's working config in hermes-obol-agent.
@@ -34,8 +35,8 @@ const (
 	// Bearer token for the signer's REST API. Injected into the signer
 	// as SIGNER__AUTH__TOKEN and into Hermes as REMOTE_SIGNER_TOKEN —
 	// defense-in-depth on top of the agent-isolation NetworkPolicy.
-	// Signer images < v0.4.0 ignore the env, so injection is a safe
-	// no-op until the image pin advances.
+	// Honoured by signer image v0.4.0+ (the version chart 0.3.3 ships);
+	// older images silently ignore the env.
 	remoteSignerAuthTokenKey = "authToken"
 )