fix: generate agents with platform LLM grants

README.md

@@ -14,7 +14,7 @@ The single `build(name, prompt)` skill:
    - `a2apack-agent-authoring` — how to shape `agent.py`, public `@skill`
      schemas, runtime metadata, pricing, resources, and auth
    - `deepagents-implementation-patterns` — concrete `create_deep_agent`
-     wiring for caller LLMs, skills, tools, and subagents
+     wiring for `ctx.llm` credentials, skills, tools, and subagents
    - `workspace-artifact-safety` — workspace grants, artifacts, bounded
      subprocesses, and scope expansion
    - `agent-quality-review` — pre-sandbox and pre-deploy checks for generated
@@ -57,8 +57,9 @@ marketplace — the dashboard already surfaces these on the agent card.
 
 `$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 your agent run on
-their own pricing (whatever you set when generating it).
+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.
 
 ## How to call it
 

agent.py

@@ -48,7 +48,7 @@ class AgentBuilder(A2AAgent[BuilderConfig, NoAuth]):
         "Writes the project into the user's workspace, validates it in a "
         "sandbox, then ships it via the control plane."
     )
-    version = "0.1.1"
+    version = "0.1.2"
 
     config_model = BuilderConfig
     auth_model = NoAuth
@@ -278,4 +278,4 @@ def _find_url(value: Any) -> str | None:
         m = _URL_RE.search(value)
         return m.group(0) if m else None
     return None
-# rebuild against a2a-pack 0.1.29 for negotiated write prefixes
+# rebuild against a2a-pack 0.1.30 for generated platform LLM defaults

agent_builder/builder.py

@@ -29,8 +29,8 @@ class BuilderContext:
     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; generated agents usually use caller-provided creds. Falls back
-    # to settings for local development.
+    # grants. Hosted generated agents should usually use platform grants too.
+    # Falls back to settings for local development.
     llm_base_url: str | None = None
     llm_api_key: str | None = None
     llm_model: str | None = None
@@ -47,17 +47,20 @@ 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.CALLER_PROVIDED``, reads caller LLM credentials 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.
+``LLMProvisioning.PLATFORM``, reads the scoped LLM grant 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 generated/user agents to ``LLMProvisioning.CALLER_PROVIDED`` and
-``caller_pays_llm=True`` unless the user explicitly wants platform-paid LLM
-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.
+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
+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
+``ChatOpenAI``.
 ``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

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, 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, 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.
 ---
 # A2A Pack Agent Authoring
 
@@ -21,13 +21,16 @@ Use `A2AAgent` class attributes for runtime and marketplace behavior:
 - `wants_cp_jwt=True` is only for trusted platform agents. It forwards the
   caller's control-plane token.
 
-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. Reserve `LLMProvisioning.PLATFORM_OR_CALLER_PROVIDED`
+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`;
-never read `A2A_LITELLM_KEY`, provider keys, or platform secrets directly.
+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`.
 
 Keep each `@skill` method `async`, put `RunContext[...]` immediately after
 `self`, and annotate every public argument. Do not use `*args` or `**kwargs`;
@@ -68,11 +71,11 @@ class ResearchAgent(A2AAgent[ResearchConfig, NoAuth]):
     config_model = ResearchConfig
     auth_model = NoAuth
 
-    llm_provisioning = LLMProvisioning.CALLER_PROVIDED
+    llm_provisioning = LLMProvisioning.PLATFORM
     pricing = Pricing(
         price_per_call_usd=0.05,
-        caller_pays_llm=True,
-        notes="Caller supplies LLM credentials through the platform.",
+        caller_pays_llm=False,
+        notes="Uses a scoped platform LLM grant through ctx.llm.",
     )
     resources = Resources(cpu="1", memory="1Gi", max_runtime_seconds=900)
     egress = EgressPolicy(allow_hosts=("api.openai.com",))
@@ -97,6 +100,12 @@ class ResearchAgent(A2AAgent[ResearchConfig, NoAuth]):
     ) -> dict[str, Any]:
         creds = ctx.llm
         await ctx.emit_progress(f"researching with {creds.model}")
+        if not creds.api_key:
+            return {
+                "summary": "LLM credentials were not available for this run.",
+                "artifact": None,
+                "path": save_path,
+            }
         graph = self._build_graph(ctx)
         state = await graph.ainvoke(
             {

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

@@ -33,11 +33,14 @@ Check these items:
 - The implementation reads `ctx.llm` for both `CALLER_PROVIDED` and
   `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
-  platform-paid LLM usage, `LLMProvisioning.PLATFORM` is acceptable only if
-  the code still uses `ctx.llm` and never reads LiteLLM/provider keys directly.
+- 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.
+- Any code path that builds `ChatOpenAI` checks `ctx.llm.api_key` first or
+  returns a clear setup/config result before the model constructor runs.
 - Any use of DeepAgents file tools passes `backend=ctx.workspace_backend()`.
 - Any project skills are seeded into the backend and passed with
   `skills=skill_sources or None`.

agent_builder/skills/deepagent-agent-design/SKILL.md

@@ -1,6 +1,6 @@
 ---
 name: deepagent-agent-design
-description: Design generated A2A agents around DeepAgents skills, subagents, durable workspace backends, and caller-provided LLM reasoning. Use whenever building or modifying an agent, deciding whether to add tools, creating skill bundles, wiring create_deep_agent, or avoiding shallow fake tools.
+description: Design generated A2A agents around DeepAgents skills, subagents, durable workspace backends, and ctx.llm-backed reasoning. Use whenever building or modifying an agent, deciding whether to add tools, creating skill bundles, wiring create_deep_agent, or avoiding shallow fake tools.
 ---
 # DeepAgent Agent Design
 
@@ -13,8 +13,8 @@ planning, file work, skill selection, and subagent delegation.
 Use this shape for non-trivial generated agents:
 
 1. Keep one or two public A2A `@skill` methods with clear typed parameters.
-2. Inside each method, read `ctx.llm`, construct `ChatOpenAI`, and build a
-   DeepAgent with `create_deep_agent`.
+2. Inside each method, read `ctx.llm`, verify credentials are available before
+   constructing `ChatOpenAI`, and build a DeepAgent with `create_deep_agent`.
 3. Pass `backend=ctx.workspace_backend()` so DeepAgents file tools write to the
    caller's durable workspace instead of LangGraph state.
 4. Seed project skills into the current grant's write prefix and pass

agent_builder/skills/deepagents-implementation-patterns/SKILL.md

@@ -1,12 +1,14 @@
 ---
 name: deepagents-implementation-patterns
-description: Implement DeepAgents inside a2a-pack agents with real deterministic tools, project skills, custom subagents, workspace-backed files, and caller-provided ChatOpenAI models. Use when editing create_deep_agent calls or deciding what belongs in tools, skills, or subagents.
+description: Implement DeepAgents inside a2a-pack agents with real deterministic tools, project skills, custom subagents, workspace-backed files, and ctx.llm-backed ChatOpenAI models. Use when editing create_deep_agent calls or deciding what belongs in tools, skills, or subagents.
 ---
 # DeepAgents Implementation Patterns
 
 DeepAgents supplies the inner agent loop. In this platform, the reliable shape
-is: caller-provided `ChatOpenAI`, A2A workspace backend, project DeepAgents
-skills, and a small number of exact tools.
+is: `ctx.llm`-backed `ChatOpenAI`, A2A workspace backend, project DeepAgents
+skills, and a small number of exact tools. Hosted generated agents should
+declare `LLMProvisioning.PLATFORM`; explicit BYOK agents can declare
+`CALLER_PROVIDED`, but both modes still read the same `ctx.llm` object.
 
 ## Implementation Shape
 
@@ -32,6 +34,11 @@ stable output paths only when that is one of the grant write prefixes.
 
 def build_graph(ctx: Any) -> Any:
     creds = ctx.llm
+    if not creds.api_key:
+        raise RuntimeError(
+            "LLM credentials were not available; declare PLATFORM for hosted "
+            "generated agents or configure caller-provided credentials."
+        )
     model = ChatOpenAI(
         model=creds.model,
         base_url=creds.base_url,

agent_builder/tools.py

@@ -28,7 +28,7 @@ if TYPE_CHECKING:
     from .config import Settings
 
 
-A2A_PACK_MIN_VERSION = "0.1.29"
+A2A_PACK_MIN_VERSION = "0.1.30"
 
 
 @dataclass(frozen=True)

requirements.txt

@@ -1,4 +1,4 @@
-a2a-pack>=0.1.29
+a2a-pack>=0.1.30
 httpx>=0.27
 boto3>=1.34
 deepagents>=0.5.0

tests/test_builder_prompt.py

@@ -53,6 +53,17 @@ 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:
+        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("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("ctx.llm.api_key", files["agent-quality-review/SKILL.md"])
+
     def test_prompt_requires_resource_mirror_for_heavy_agents(self) -> None:
         self.assertIn("declare resources in BOTH places", SYSTEM_PROMPT)
         self.assertIn("resources = Resources", SYSTEM_PROMPT)

tests/test_template_init.py

@@ -31,7 +31,7 @@ class TemplateInitTests(unittest.TestCase):
         self.assertEqual(set(files), {"agent.py", "a2a.yaml", "requirements.txt"})
         self.assertIn('name = "research-agent"', files["agent.py"])
         self.assertIn('description = "Research helper"', files["agent.py"])
-        self.assertIn("LLMProvisioning.CALLER_PROVIDED", files["agent.py"])
+        self.assertIn("LLMProvisioning.PLATFORM", files["agent.py"])
         self.assertIn("WorkspaceAccess.dynamic", files["agent.py"])
         self.assertIn("RUNTIME_SKILLS_DIR", files["agent.py"])
         self.assertIn("_runtime_skills_root", files["agent.py"])