feat: let builder use caller or platform llm

README.md

@@ -47,7 +47,7 @@ Three platform-managed bits land on the inner skill via `RunContext`:
 | Field | Source | Required? |
 |---|---|---|
 | `ctx.workspace.bucket` | grant minted by the orchestrator | yes |
-| `ctx.llm` | platform-scoped LiteLLM grant (Card declares `llm_provisioning=platform`) | yes |
+| `ctx.llm` | caller-selected LLM creds when available, otherwise a platform-scoped LiteLLM grant (Card declares `llm_provisioning=platform_or_caller_provided`) | yes |
 | `ctx.cp_jwt` | caller's CP JWT (Card declares `wants_cp_jwt=True`) | yes — used by `cp_deploy_tarball` |
 
 The user opts into all three when they pick `agent-builder` from the
@@ -55,8 +55,9 @@ marketplace — the dashboard already surfaces these on the agent card.
 
 ## Pricing
 
-`$0.10 / call` during the current pricing model. The deployed
+`$0.10 / call` during the current pricing model. The builder uses your
-agent is yours forever; subsequent invocations of your agent run on
+selected LLM creds when present and falls back to a scoped platform grant.
+The deployed agent is yours forever; subsequent invocations of your agent run on
 their own pricing (whatever you set when generating it).
 
 ## How to call it

agent.py

@@ -10,8 +10,8 @@ This agent only works when invoked through the platform orchestrator,
 because it needs three things forwarded from the caller:
 
   1. ``ctx.workspace.bucket`` — the user's MinIO bucket
-  2. ``ctx.llm`` — a platform-scoped LiteLLM grant token for the builder's
+  2. ``ctx.llm`` — caller-selected LLM credentials when available, otherwise
-     own DeepAgents loop
+     a platform-scoped LiteLLM grant token for the builder's own DeepAgents loop
   3. ``ctx.cp_jwt`` — the user's CP JWT, so cp_deploy_tarball can
      POST to /v1/agents/from-tarball as the user.
 """
@@ -53,7 +53,7 @@ class AgentBuilder(A2AAgent[BuilderConfig, NoAuth]):
     config_model = BuilderConfig
     auth_model = NoAuth
     tools_used = ("deepagents", "langgraph", "a2a-pack", "litellm", "microsandbox")
-    llm_provisioning = LLMProvisioning.PLATFORM
+    llm_provisioning = LLMProvisioning.PLATFORM_OR_CALLER_PROVIDED
     wants_cp_jwt = True
     workspace_access = WorkspaceAccess.dynamic(
         allowed_modes=(WorkspaceMode.READ_ONLY, WorkspaceMode.READ_WRITE_OVERLAY),
@@ -62,8 +62,9 @@ class AgentBuilder(A2AAgent[BuilderConfig, NoAuth]):
         price_per_call_usd=0.10,
         caller_pays_llm=False,
         notes=(
-            "Uses a platform-scoped LiteLLM grant for the builder's own "
+            "Uses caller-selected LLM credentials when available; otherwise "
-            "reasoning loop; deployed agent is yours forever."
+            "uses a platform-scoped LiteLLM grant for the builder's own "
+            "reasoning loop. Deployed agents remain yours."
         ),
     )
 

agent_builder/builder.py

@@ -28,8 +28,9 @@ class BuilderContext:
     workspace: Any | None = None
     grant_token: str | None = None
     # LLM creds forwarded by the orchestrator according to this agent's Card.
-    # The builder itself uses platform grants; generated agents usually use
+    # The builder accepts caller-selected creds and falls back to platform
-    # caller-provided creds. Falls back to settings for local development.
+    # grants; generated agents usually use caller-provided creds. Falls back
+    # to settings for local development.
     llm_base_url: str | None = None
     llm_api_key: str | None = None
     llm_model: str | None = None
@@ -57,6 +58,10 @@ Default generated/user agents to ``LLMProvisioning.CALLER_PROVIDED`` and
 usage. ``LLMProvisioning.PLATFORM`` is allowed only through the A2A LiteLLM
 grant path: the generated agent must still read ``ctx.llm`` and must never read
 ``A2A_LITELLM_KEY`` or any provider key directly.
+``LLMProvisioning.PLATFORM_OR_CALLER_PROVIDED`` is reserved for trusted
+platform/meta agents that should prefer caller-selected credentials and fall
+back to a scoped platform grant; do not use it for ordinary generated agents
+unless the user explicitly asks for that mixed billing mode.
 
 You have packaged DeepAgents skills loaded from ``/.agent-builder/skills/``.
 Use ``deepagent-agent-design`` before designing generated agent internals,

agent_builder/skills/a2apack-agent-authoring/SKILL.md

@@ -24,8 +24,10 @@ Use `A2AAgent` class attributes for runtime and marketplace behavior:
 For generated/user-owned agents that call an LLM, default to
 `LLMProvisioning.CALLER_PROVIDED` with `Pricing(..., caller_pays_llm=True, ...)`.
 Use `LLMProvisioning.PLATFORM` only when the user explicitly wants
-platform-paid LLM usage. In either mode, read `ctx.llm`; never read
+platform-paid LLM usage. Reserve `LLMProvisioning.PLATFORM_OR_CALLER_PROVIDED`
-`A2A_LITELLM_KEY`, provider keys, or platform secrets directly.
+for trusted platform/meta agents that must work with either the caller's
+selected creds or a platform fallback grant. In every mode, read `ctx.llm`;
+never read `A2A_LITELLM_KEY`, provider keys, or platform secrets directly.
 
 Keep each `@skill` method `async`, put `RunContext[...]` immediately after
 `self`, and annotate every public argument. Do not use `*args` or `**kwargs`;

agent_builder/skills/agent-quality-review/SKILL.md

@@ -31,7 +31,8 @@ Check these items:
 - The public schema is small and stable. No unbounded command strings, hidden
   prompt fragments, or provider credentials as public inputs.
 - The implementation reads `ctx.llm` for both `CALLER_PROVIDED` and
-  `PLATFORM` LLM provisioning.
+  `PLATFORM` LLM provisioning. `PLATFORM_OR_CALLER_PROVIDED` is acceptable
+  only for trusted platform/meta agents and must also read `ctx.llm`.
 - Generated/user agents should default to
   `llm_provisioning = LLMProvisioning.CALLER_PROVIDED` and
   `Pricing(..., caller_pays_llm=True, ...)`. If the user explicitly wants

agent_builder/tools.py

@@ -28,7 +28,7 @@ if TYPE_CHECKING:
     from .config import Settings
 
 
-A2A_PACK_MIN_VERSION = "0.1.27"
+A2A_PACK_MIN_VERSION = "0.1.28"
 
 
 @dataclass(frozen=True)

requirements.txt

@@ -1,4 +1,4 @@
-a2a-pack>=0.1.27
+a2a-pack>=0.1.28
 httpx>=0.27
 boto3>=1.34
 deepagents>=0.5.0

tests/test_builder_prompt.py

@@ -24,12 +24,16 @@ class BuilderPromptTests(unittest.TestCase):
         )
         self.assertIn("stream=True", agent_source)
 
-    def test_outer_builder_uses_platform_llm_grants(self) -> None:
+    def test_outer_builder_accepts_user_creds_or_platform_llm_grants(self) -> None:
         agent_source = Path(__file__).resolve().parents[1].joinpath("agent.py").read_text()
 
-        self.assertIn("llm_provisioning = LLMProvisioning.PLATFORM", agent_source)
+        self.assertIn(
+            "llm_provisioning = LLMProvisioning.PLATFORM_OR_CALLER_PROVIDED",
+            agent_source,
+        )
         self.assertIn("caller_pays_llm=False", agent_source)
         self.assertIn("platform-scoped LiteLLM grant", agent_source)
+        self.assertIn("caller-selected LLM credentials", agent_source)
 
     def test_builder_defaults_raise_timeout_budget(self) -> None:
         config_source = Path(__file__).resolve().parents[1].joinpath(