agent-relay-server 0.33.1 → 0.34.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-relay-server",
-  "version": "0.33.1",
+  "version": "0.34.1",
   "description": "Lightweight HTTP message relay for inter-agent communication across machines",
   "module": "src/index.ts",
   "type": "module",

package/runner/src/adapter.ts

@@ -174,7 +174,23 @@ export function profileAllowsRelayFeature(config: RunnerSpawnConfig, feature: ke
 
 export const RELAY_CONTEXT = `[agent-relay] You are connected to Agent Relay, a real-time message bus between agents and users. When you receive a relay message: read it, do what it asks, and reply through the relay when a text response is needed. Use agent-relay /react <messageId> <emoji> for lightweight acknowledgement or approval. If Relay MCP tools are available, prefer relay_reply, relay_get_message, relay_get_thread, relay_send_message, relay_upload_artifact, relay_attach_artifact, relay_agent_status, relay_find_agents, relay_spawn_agent, and relay_shutdown_agent. You never need to know or pass your own agent id — relay fills it from your token; use relay_whoami only if you need to reason about yourself. relay_spawn_targets / relay_spawn_agent / relay_shutdown_agent only appear if your profile grants spawning (a live-children quota); when present, call relay_spawn_targets FIRST for the live host/provider/model matrix + your quota, then stand up long-living child agents and shut down your own — find them later with relay_find_agents spawnedBy:me. CLI fallback: agent-relay /reply <messageId> --stdin < response.md; if a delivered message says it was truncated, fetch the full body with: agent-relay get-message <messageId>. For command details, run: agent-relay /guide`;
 
-export const PROVIDER_MESSAGE_BODY_PREVIEW_CHARS = 4000;
+// #306 — deliver the FULL message body by default. Only a pathological body beyond this
+// high cap truncates (with a get-message hint) so it can't nuke an agent's context; the 99%
+// case (incl. multi-thousand-char handoffs/reports) arrives whole in one shot, no extra
+// round-trip. The mirror flow (user typing directly) already injects full text — this removes
+// the asymmetry where agent↔agent + attachment-bearing messages were second-class.
+export const DEFAULT_PROVIDER_MESSAGE_BODY_MAX_CHARS = 24_000;
+
+// Resolve the delivered-body cap. Deployment-dependent (the right ceiling tracks the host's
+// context budget), so it's overridable via env — set on the orchestrator/host that spawns runners.
+export function providerMessageBodyMaxChars(): number {
+  const raw = process.env.AGENT_RELAY_MESSAGE_BODY_MAX_CHARS;
+  if (raw !== undefined) {
+    const parsed = Number.parseInt(raw.trim(), 10);
+    if (Number.isFinite(parsed) && parsed > 0) return parsed;
+  }
+  return DEFAULT_PROVIDER_MESSAGE_BODY_MAX_CHARS;
+}
 
 function attachmentRefs(message: Message): Record<string, unknown>[] {
   const payloadRefs = message.payload?.attachments;
@@ -280,15 +296,16 @@ export function providerAttachmentText(message: Message): string | undefined {
 
 export function providerMessageText(messages: Message[]): string {
   const replyable = latestReplyableMessage(messages);
+  const maxChars = providerMessageBodyMaxChars();
   const sections = messages
     .map((message) => {
       const subject = message.subject ? `Subject: ${message.subject}\n` : "";
       const isMemoryContext = isMemoryInjection(message);
       const canReferenceMessage = isPersistedRelayMessage(message) && !isMemoryContext && !isReactionNotification(message);
-      const shouldPreview = canReferenceMessage && message.body.length > PROVIDER_MESSAGE_BODY_PREVIEW_CHARS;
+      const shouldPreview = canReferenceMessage && message.body.length > maxChars;
       const preview = shouldPreview
         ? {
-          body: message.body.slice(0, PROVIDER_MESSAGE_BODY_PREVIEW_CHARS),
+          body: message.body.slice(0, maxChars),
           truncated: true,
         }
         : {
@@ -297,7 +314,7 @@ export function providerMessageText(messages: Message[]): string {
         };
       const truncationGuidance = preview.truncated
         ? [
-          `[truncated: showing first ${PROVIDER_MESSAGE_BODY_PREVIEW_CHARS} of ${message.body.length} chars]`,
+          `[truncated: showing first ${maxChars} of ${message.body.length} chars]`,
           `Read full: agent-relay get-message ${message.id}`,
           `Body only: agent-relay get-message ${message.id} --body`,
           `Long reply: agent-relay /reply ${message.id} --stdin < response.md`,

package/src/mcp.ts

@@ -820,10 +820,10 @@ async function relaySpawnAgent(auth: McpAuthContext, args: Record<string, unknow
     }
   }
 
-  assertComponentResourceAllowed(auth, {
-    scope: "agent:write",
-    resource: { orchestratorId: orchestrator.id, cwd: resolvedCwd, policyName, spawnRequestId },
-  });
+  // #323 — gate child spawn only on `orchestrators` (the parent's legit bound), NOT its self-scoping
+  // spawnRequestIds/cwdPrefixes/policies: those describe the child, not parent-owned resources, so
+  // gating on them makes maxSpawnedAgents unreachable for every component token (cwd checked above).
+  assertComponentResourceAllowed(auth, { scope: "agent:write", resource: { orchestratorId: orchestrator.id } });
 
   // Child runner token: a normal long-living agent that is NOT itself spawn-capable
   // (canSpawn:false → no grandchildren), stamped with authoritative lineage so it registers