Handle deterministic agents without LLM creds
All checks were successful
build / build (push) Successful in 11s

This commit is contained in:
robert
2026-06-08 10:39:50 -03:00
parent 8c0dda8405
commit 4cef2ef76f
4 changed files with 42 additions and 16 deletions

View File

@@ -49,20 +49,28 @@ Given a user description, you write a complete agent project under the
user's workspace at ``agents/<name>/`` and then deploy it through the user's workspace at ``agents/<name>/`` and then deploy it through the
control plane. control plane.
The default starter is the current a2a-pack DeepAgents scaffold. It declares The default starter is the current a2a-pack DeepAgents scaffold. For agents
``LLMProvisioning.PLATFORM``, reads the caller's saved LLM credential from that call an LLM, it declares ``LLMProvisioning.PLATFORM``, reads the caller's
``ctx.llm``, resolves the model through ``create_a2a_deep_agent``, and wires a small saved LLM credential from ``ctx.llm``, resolves the model through
tool-calling DeepAgent with the A2A DeepAgents model middleware ``create_a2a_deep_agent``, and wires a small tool-calling DeepAgent with the
middleware. Do not recreate this from memory: call ``init_agent_template`` first A2A DeepAgents model middleware. Do not recreate this from memory: call
``init_agent_template`` first
and modify the generated files. and modify the generated files.
Default hosted generated/user agents to ``LLMProvisioning.PLATFORM`` and For hosted generated/user agents that call an LLM, default to
``caller_pays_llm=True`` so main-agent handoffs forward the caller's saved LLM ``LLMProvisioning.PLATFORM`` and ``caller_pays_llm=True`` so main-agent
credential for the callee. The generated agent must still read ``ctx.llm`` and must never handoffs forward the caller's saved LLM credential for the callee. The
read ``A2A_LITELLM_KEY``, ``OPENAI_API_KEY``, LiteLLM master keys, or provider generated agent must still read ``ctx.llm`` and must never read
keys directly. ``LLMProvisioning.CALLER_PROVIDED`` is still acceptable for ``A2A_LITELLM_KEY``, ``OPENAI_API_KEY``, LiteLLM master keys, or provider keys
explicit BYOK wording. Always make missing ``ctx.llm.api_key`` a clear setup/config result before constructing directly. ``LLMProvisioning.CALLER_PROVIDED`` is still acceptable for explicit
a DeepAgents model. BYOK wording. Always make missing ``ctx.llm.api_key`` a clear setup/config
result before constructing a DeepAgents model.
For explicitly deterministic/local-logic agents that do not call an LLM, remove
the starter's ``LLMProvisioning`` import and class ``llm_provisioning``
declaration, do not read ``ctx.llm``, and set
``Pricing(..., caller_pays_llm=False, ...)``. A no-LLM agent must be callable
without requiring the user's LLM credential.
``Pricing`` accepts only ``price_per_call_usd``, ``caller_pays_llm``, and ``Pricing`` accepts only ``price_per_call_usd``, ``caller_pays_llm``, and
``notes``. Never set ``runtime.pricing.compute`` or ``notes``. Never set ``runtime.pricing.compute`` or
``runtime.pricing.total_usd`` in generated source; those are derived later by ``runtime.pricing.total_usd`` in generated source; those are derived later by

View File

@@ -31,6 +31,12 @@ 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 secrets directly. If `ctx.llm.api_key` is empty, return a clear setup/config
result before constructing a model or calling `create_a2a_deep_agent`. result before constructing a model or calling `create_a2a_deep_agent`.
For explicitly deterministic/local-logic agents that do not call an LLM, omit
the concrete class `llm_provisioning` declaration, remove unused
`LLMProvisioning` imports, do not read `ctx.llm`, and set
`Pricing(..., caller_pays_llm=False, ...)`. A no-LLM agent must be callable
without requiring the user's LLM credential.
`Pricing` only accepts `price_per_call_usd`, `caller_pays_llm`, and `notes`. `Pricing` only accepts `price_per_call_usd`, `caller_pays_llm`, and `notes`.
Do not put `runtime.pricing.compute`, `runtime.pricing.total_usd`, `compute=`, Do not put `runtime.pricing.compute`, `runtime.pricing.total_usd`, `compute=`,
or `total_usd=` in generated `agent.py`; the platform derives those values from or `total_usd=` in generated `agent.py`; the platform derives those values from

View File

@@ -30,14 +30,19 @@ Check these items:
arguments. arguments.
- The public schema is small and stable. No unbounded command strings, hidden - The public schema is small and stable. No unbounded command strings, hidden
prompt fragments, or provider credentials as public inputs. prompt fragments, or provider credentials as public inputs.
- The implementation reads `ctx.llm` for both `CALLER_PROVIDED` and - LLM-backed implementations read `ctx.llm` for both `CALLER_PROVIDED` and
`PLATFORM` LLM provisioning. `PLATFORM_OR_CALLER_PROVIDED` is acceptable `PLATFORM` LLM provisioning. `PLATFORM_OR_CALLER_PROVIDED` is acceptable
only for trusted platform/meta agents and must also read `ctx.llm`. only for trusted platform/meta agents and must also read `ctx.llm`.
- Hosted generated/user agents should default to - Hosted generated/user agents that call an LLM should default to
`llm_provisioning = LLMProvisioning.PLATFORM` and `llm_provisioning = LLMProvisioning.PLATFORM` and
`Pricing(..., caller_pays_llm=True, ...)`. `LLMProvisioning.CALLER_PROVIDED` `Pricing(..., caller_pays_llm=True, ...)`. `LLMProvisioning.CALLER_PROVIDED`
is acceptable for explicit BYOK wording. In both modes, the code must use is acceptable for explicit BYOK wording. In both modes, the code must use
`ctx.llm` and never read LiteLLM/provider keys directly. `ctx.llm` and never read LiteLLM/provider keys directly.
- Deterministic/local-logic agents that do not call an LLM omit the concrete
`llm_provisioning` declaration, remove unused `LLMProvisioning` imports, do
not read `ctx.llm`, and use `Pricing(..., caller_pays_llm=False, ...)`.
Reject no-LLM agents that would require a user's LLM credential before
invoking deterministic skills.
- `Pricing` contains only `price_per_call_usd`, `caller_pays_llm`, and `notes`. - `Pricing` contains only `price_per_call_usd`, `caller_pays_llm`, and `notes`.
Reject `runtime.pricing.compute`, `runtime.pricing.total_usd`, `compute=`, or Reject `runtime.pricing.compute`, `runtime.pricing.total_usd`, `compute=`, or
`total_usd=` in `agent.py`; those are derived platform fields. `total_usd=` in `agent.py`; those are derived platform fields.

View File

@@ -185,16 +185,23 @@ class BuilderPromptTests(unittest.TestCase):
self.assertIn(f"a2a-pack>={A2A_PACK_MIN_VERSION}", requirements) self.assertIn(f"a2a-pack>={A2A_PACK_MIN_VERSION}", requirements)
self.assertIn("a2a-pack>={A2A_PACK_MIN_VERSION}", tools_source) self.assertIn("a2a-pack>={A2A_PACK_MIN_VERSION}", tools_source)
def test_prompt_defaults_generated_agents_to_user_llm_credentials(self) -> None: def test_prompt_routes_llm_and_no_llm_agents_differently(self) -> None:
self.assertIn("LLMProvisioning.PLATFORM", SYSTEM_PROMPT) self.assertIn("LLMProvisioning.PLATFORM", SYSTEM_PROMPT)
self.assertIn("caller_pays_llm=True", 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("handoffs forward the caller's saved LLM credential", SYSTEM_PROMPT)
self.assertIn("ctx.llm.api_key", SYSTEM_PROMPT) self.assertIn("ctx.llm.api_key", SYSTEM_PROMPT)
self.assertIn("deterministic/local-logic agents", SYSTEM_PROMPT)
self.assertIn("caller_pays_llm=False", SYSTEM_PROMPT)
self.assertIn("A no-LLM agent must be callable", SYSTEM_PROMPT)
self.assertIn("without requiring the user's LLM credential", SYSTEM_PROMPT)
files = _builder_skill_files() files = _builder_skill_files()
self.assertIn("LLMProvisioning.PLATFORM", files["a2apack-agent-authoring/SKILL.md"]) self.assertIn("LLMProvisioning.PLATFORM", files["a2apack-agent-authoring/SKILL.md"])
self.assertIn("caller_pays_llm=True", files["a2apack-agent-authoring/SKILL.md"]) self.assertIn("caller_pays_llm=True", 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"]) self.assertIn("ctx.llm.api_key", files["agent-quality-review/SKILL.md"])
self.assertIn("Deterministic/local-logic agents", files["agent-quality-review/SKILL.md"])
self.assertIn("require a user's LLM credential", files["agent-quality-review/SKILL.md"])
def test_prompt_rejects_platform_derived_pricing_fields(self) -> None: def test_prompt_rejects_platform_derived_pricing_fields(self) -> None:
self.assertIn("runtime.pricing.compute", SYSTEM_PROMPT) self.assertIn("runtime.pricing.compute", SYSTEM_PROMPT)