fix: align agent-builder with longer pack timeouts

agent_builder/config.py

@@ -9,6 +9,7 @@ from dataclasses import dataclass
 class Settings:
     sandbox_url: str
     sandbox_timeout_s: int
+    deploy_wait_timeout_s: int
     litellm_url: str
     litellm_key: str
     litellm_model: str
@@ -24,7 +25,8 @@ def load_settings() -> Settings:
         sandbox_url=os.environ.get(
             "SANDBOX_URL", "http://sandbox.sandbox.svc.cluster.local:8000"
         ),
-        sandbox_timeout_s=int(os.environ.get("SANDBOX_TIMEOUT_S", "240")),
+        sandbox_timeout_s=int(os.environ.get("SANDBOX_TIMEOUT_S", "1200")),
+        deploy_wait_timeout_s=int(os.environ.get("DEPLOY_WAIT_TIMEOUT_S", "900")),
         litellm_url=os.environ.get(
             "A2A_LITELLM_URL", "http://litellm.llm.svc.cluster.local:4000"
         ),

agent_builder/tools.py

@@ -38,7 +38,7 @@ class ToolContext:
 async def _wait_for_live_card(
     url: str | None,
     expected_version: str | None,
-    timeout_s: float = 180.0,
+    timeout_s: float = 900.0,
 ) -> tuple[bool, dict[str, Any]]:
     """Poll ``{url}/.well-known/agent-card`` until ``version`` matches
     ``expected_version`` or ``timeout_s`` elapses. Returns
@@ -435,7 +435,11 @@ def build_tools(ctx: ToolContext) -> list[Any]:
         # so the LLM can verify the input_schema actually matches the
         # code it wrote — catches the schema/code skew that causes
         # downstream 400s on call_agent.
-        live, live_card = await _wait_for_live_card(url_, version_)
+        live, live_card = await _wait_for_live_card(
+            url_,
+            version_,
+            timeout_s=settings.deploy_wait_timeout_s,
+        )
         out: dict[str, Any] = {
             "ok": live,
             "name": name_,