Require user LLM credentials

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` | caller-selected LLM creds when available, otherwise a platform-scoped LiteLLM grant (Card declares `llm_provisioning=platform_or_caller_provided`) | yes |
+| `ctx.llm` | caller's saved LLM credential, routed by the platform (Card declares `llm_provisioning=platform`) | 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
@@ -55,11 +55,9 @@ marketplace — the dashboard already surfaces these on the agent card.
 
 ## Pricing
 
-`$0.10 / call` during the current pricing model. The builder uses your
-selected LLM creds when present and falls back to a scoped platform grant.
-The deployed agent is yours forever; subsequent invocations of a generated LLM
-agent default to scoped platform LLM grants unless you explicitly generate a
-BYOK/caller-paid agent.
+`$0.10 / call` during the current pricing model. The builder uses your saved
+LLM credential. The deployed agent is yours forever; subsequent invocations of
+a generated LLM agent also use the caller's saved LLM credential via `ctx.llm`.
 
 ## 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`` — caller-selected LLM credentials when available, otherwise
-     a platform-scoped LiteLLM grant token for the builder's own DeepAgents loop
+  2. ``ctx.llm`` — the caller's saved LLM credential 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.
 """
@@ -53,17 +53,16 @@ class AgentBuilder(A2AAgent[BuilderConfig, NoAuth]):
     config_model = BuilderConfig
     auth_model = NoAuth
     tools_used = ("deepagents", "langgraph", "a2a-pack", "litellm", "microsandbox")
-    llm_provisioning = LLMProvisioning.PLATFORM_OR_CALLER_PROVIDED
+    llm_provisioning = LLMProvisioning.PLATFORM
     wants_cp_jwt = True
     workspace_access = WorkspaceAccess.dynamic(
         allowed_modes=(WorkspaceMode.READ_ONLY, WorkspaceMode.READ_WRITE_OVERLAY),
     )
     pricing = Pricing(
         price_per_call_usd=0.10,
-        caller_pays_llm=False,
+        caller_pays_llm=True,
         notes=(
-            "Uses caller-selected LLM credentials when available; otherwise "
-            "uses a platform-scoped LiteLLM grant for the builder's own "
+            "Uses the caller's saved LLM credential for the builder's own "
             "reasoning loop. Deployed agents remain yours."
         ),
     )

agent_builder/builder.py

@@ -30,9 +30,9 @@ class BuilderContext:
     sandbox: Any | None = None
     grant_token: str | None = None
     # LLM creds forwarded by the orchestrator according to this agent's Card.
-    # The builder accepts caller-selected creds and falls back to platform
-    # grants. Hosted generated agents should usually use platform grants too.
-    # Falls back to settings for local development.
+    # The builder uses the caller's saved LLM credential, usually proxied
+    # through LiteLLM by the control plane. Falls back to settings only for
+    # local development.
     llm_base_url: str | None = None
     llm_api_key: str | None = None
     llm_model: str | None = None
@@ -49,28 +49,26 @@ user's workspace at ``agents/<name>/`` and then deploy it through the
 control plane.
 
 The default starter is the current a2a-pack DeepAgents scaffold. It declares
-``LLMProvisioning.PLATFORM``, reads the scoped LLM grant from ``ctx.llm``,
-builds a ``ChatOpenAI`` model from those credentials, and wires a small
+``LLMProvisioning.PLATFORM``, reads the caller's saved LLM credential from
+``ctx.llm``, builds a ``ChatOpenAI`` model from those credentials, and wires a small
 tool-calling DeepAgent with ``create_deep_agent`` plus ``wrap_model_call``
 middleware. Do not recreate this from memory: call ``init_agent_template`` first
 and modify the generated files.
 
 Default hosted generated/user agents to ``LLMProvisioning.PLATFORM`` and
-``caller_pays_llm=False`` so main-agent handoffs mint a scoped A2A LiteLLM grant
-for the callee. The generated agent must still read ``ctx.llm`` and must never
+``caller_pays_llm=True`` so main-agent handoffs forward the caller's saved LLM
+credential for the callee. The generated agent must still read ``ctx.llm`` and must never
 read ``A2A_LITELLM_KEY``, ``OPENAI_API_KEY``, LiteLLM master keys, or provider
-keys directly. If the user explicitly asks for BYOK or caller-paid inference,
-use ``LLMProvisioning.CALLER_PROVIDED`` with ``caller_pays_llm=True`` and make
-missing ``ctx.llm.api_key`` a clear setup/config result before constructing
+keys directly. ``LLMProvisioning.CALLER_PROVIDED`` is still acceptable for
+explicit BYOK wording. Always make missing ``ctx.llm.api_key`` a clear setup/config result before constructing
 ``ChatOpenAI``.
 ``Pricing`` accepts only ``price_per_call_usd``, ``caller_pays_llm``, and
 ``notes``. Never set ``runtime.pricing.compute`` or
 ``runtime.pricing.total_usd`` in generated source; those are derived later by
 the control plane/dashboard from declared ``Resources`` and billing policy.
-``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.
+``LLMProvisioning.PLATFORM_OR_CALLER_PROVIDED`` is retained only for backwards
+compatibility; do not use it for ordinary generated agents unless the user
+explicitly asks for that legacy mixed 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

@@ -1,6 +1,6 @@
 ---
 name: a2apack-agent-authoring
-description: Author production a2a-pack agents with clear public @skill schemas, runtime declarations, platform LLM grants, optional caller-provided LLMs, workspace grants, pricing, resources, and secure platform capabilities. Use when writing or reviewing agent.py for an A2A agent.
+description: Author production a2a-pack agents with clear public @skill schemas, runtime declarations, caller-funded LLM credentials, optional agent BYOK, workspace grants, pricing, resources, and secure platform capabilities. Use when writing or reviewing agent.py for an A2A agent.
 ---
 # A2A Pack Agent Authoring
 
@@ -22,12 +22,11 @@ Use `A2AAgent` class attributes for runtime and marketplace behavior:
   caller's control-plane token.
 
 For hosted generated/user-owned agents that call an LLM, default to
-`LLMProvisioning.PLATFORM` with `Pricing(..., caller_pays_llm=False, ...)` so
-main-agent handoffs mint a scoped A2A LiteLLM grant for the callee. Use
-`LLMProvisioning.CALLER_PROVIDED` only when the user explicitly wants BYOK or
-caller-paid inference. Reserve `LLMProvisioning.PLATFORM_OR_CALLER_PROVIDED`
-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`;
+`LLMProvisioning.PLATFORM` with `Pricing(..., caller_pays_llm=True, ...)` so
+main-agent handoffs forward the caller's saved LLM credential for the callee.
+`LLMProvisioning.CALLER_PROVIDED` is still acceptable for explicit BYOK wording.
+Reserve `LLMProvisioning.PLATFORM_OR_CALLER_PROVIDED` for legacy compatibility.
+In every mode, read `ctx.llm`;
 never read `A2A_LITELLM_KEY`, `OPENAI_API_KEY`, provider keys, or platform
 secrets directly. If `ctx.llm.api_key` is empty, return a clear setup/config
 result before constructing `ChatOpenAI`.
@@ -79,8 +78,8 @@ class ResearchAgent(A2AAgent[ResearchConfig, NoAuth]):
     llm_provisioning = LLMProvisioning.PLATFORM
     pricing = Pricing(
         price_per_call_usd=0.05,
-        caller_pays_llm=False,
-        notes="Uses a scoped platform LLM grant through ctx.llm.",
+        caller_pays_llm=True,
+        notes="Uses the caller's saved LLM credential through ctx.llm.",
     )
     resources = Resources(cpu="1", memory="1Gi", max_runtime_seconds=900)
     egress = EgressPolicy(allow_hosts=("api.openai.com",))

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

@@ -35,10 +35,9 @@ Check these items:
   only for trusted platform/meta agents and must also read `ctx.llm`.
 - Hosted generated/user agents should default to
   `llm_provisioning = LLMProvisioning.PLATFORM` and
-  `Pricing(..., caller_pays_llm=False, ...)`. Use
-  `LLMProvisioning.CALLER_PROVIDED` only when the user explicitly wants BYOK or
-  caller-paid inference. In both modes, the code must use `ctx.llm` and never
-  read LiteLLM/provider keys directly.
+  `Pricing(..., caller_pays_llm=True, ...)`. `LLMProvisioning.CALLER_PROVIDED`
+  is acceptable for explicit BYOK wording. In both modes, the code must use
+  `ctx.llm` and never read LiteLLM/provider keys directly.
 - `Pricing` contains only `price_per_call_usd`, `caller_pays_llm`, and `notes`.
   Reject `runtime.pricing.compute`, `runtime.pricing.total_usd`, `compute=`, or
   `total_usd=` in `agent.py`; those are derived platform fields.

tests/test_builder_prompt.py

@@ -32,16 +32,15 @@ class BuilderPromptTests(unittest.TestCase):
         self.assertIn("stream=True", agent_source)
         self.assertIn('grant_write_prefixes=("agents/{name}/",)', agent_source)
 
-    def test_outer_builder_accepts_user_creds_or_platform_llm_grants(self) -> None:
+    def test_outer_builder_uses_user_llm_credentials(self) -> None:
         agent_source = Path(__file__).resolve().parents[1].joinpath("agent.py").read_text()
 
         self.assertIn(
-            "llm_provisioning = LLMProvisioning.PLATFORM_OR_CALLER_PROVIDED",
+            "llm_provisioning = LLMProvisioning.PLATFORM",
             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)
+        self.assertIn("caller_pays_llm=True", agent_source)
+        self.assertIn("caller's saved LLM credential", agent_source)
 
     def test_builder_defaults_raise_timeout_budget(self) -> None:
         config_source = Path(__file__).resolve().parents[1].joinpath(
@@ -60,15 +59,15 @@ class BuilderPromptTests(unittest.TestCase):
         self.assertIn(f"a2a-pack>={A2A_PACK_MIN_VERSION}", requirements)
         self.assertIn("a2a-pack>={A2A_PACK_MIN_VERSION}", tools_source)
 
-    def test_prompt_defaults_generated_agents_to_platform_llm_grants(self) -> None:
+    def test_prompt_defaults_generated_agents_to_user_llm_credentials(self) -> None:
         self.assertIn("LLMProvisioning.PLATFORM", SYSTEM_PROMPT)
-        self.assertIn("caller_pays_llm=False", SYSTEM_PROMPT)
-        self.assertIn("main-agent handoffs mint a scoped A2A LiteLLM grant", SYSTEM_PROMPT)
+        self.assertIn("caller_pays_llm=True", SYSTEM_PROMPT)
+        self.assertIn("main-agent handoffs forward the caller's saved LLM", SYSTEM_PROMPT)
         self.assertIn("ctx.llm.api_key", SYSTEM_PROMPT)
 
         files = _builder_skill_files()
         self.assertIn("LLMProvisioning.PLATFORM", files["a2apack-agent-authoring/SKILL.md"])
-        self.assertIn("caller_pays_llm=False", files["a2apack-agent-authoring/SKILL.md"])
+        self.assertIn("caller_pays_llm=True", files["a2apack-agent-authoring/SKILL.md"])
         self.assertIn("ctx.llm.api_key", files["agent-quality-review/SKILL.md"])
 
     def test_prompt_rejects_platform_derived_pricing_fields(self) -> None: