provrun.backends.pydantic_ai

A PydanticAI agent with a filesystem sandbox, in this process.

The agent sees three folders:

/input     the staged inputs, read-only
/output    what the run hands back, writable
/scratch   working space, writable, not part of the bundle

provrun, not the agent, writes the evidence. One PydanticAI capability, evidence_capability, sees every step of the run:

each model request, streamed or not, writes model_called;
each call of a function tool, from any toolset or capability, writes tool_called;
a call that waits for approval, or is deferred, writes tool_called with deferred set, and writes it again when it runs;
the output tool, when the agent answers through one, writes tool_called with output_tool set. In a streamed request it is written before the model_called whose response asked for it, because PydanticAI checks the output while the stream is still open;
each tool that the provider ran itself, and returned in a response, writes tool_called with provider_run set and confidence parsed, because provrun sees only the provider’s account of it.

The backend itself has text output and no provider tools, so it never makes the last two; they matter to other callers of the capability.

After the run, the whole message history is imported as one trace_imported event with confidence parsed.

evidence_capability

source

evidence_capability(ctx: RunContext) -> AbstractCapability

Make the PydanticAI capability that writes provrun’s evidence for one run.

Code other than the backend can pass it to its own agent, streamed or not.

Parameter	Type	Default	Description
`ctx`	RunContext	required	The run whose recorder takes the evidence: model_called, tool_called and the rest the module lists
returns	AbstractCapability		A capability to pass in an agent’s capabilities; it needs provrun[pydantic-ai]

class PydanticAIBackend

source

PydanticAIBackend(model: Any = None, registry: ToolRegistry | None = None)

A PydanticAI agent with a filesystem sandbox over the run’s folders.

The sandbox mounts the input, output and scratch folders. It needs provrun[pydantic-ai].

Parameter	Type	Default	Description
`model`	Any	`None`	A PydanticAI Model to use instead of agent.model, for tests; None infers it from agent.model
`registry`	ToolRegistry \| None	`None`	Where the agent’s tools are found by name; None is a ToolRegistry of the entry points

PydanticAIBackend.execute

source

async PydanticAIBackend.execute(ctx: RunContext, agent: AgentSpec, run: RunSpec) -> RunResult

Run the agent in the sandbox, and import its message history as trace_imported.

It writes the output to response.txt in the output folder unless the agent wrote that file. Raises ValueError for a tool that the registry lacks or that is not in_process.

Parameter	Type	Default	Description
`ctx`	RunContext	required	The run; its recorder takes the evidence the module lists
`agent`	AgentSpec	required	The agent to run; its tools are found in the registry
`run`	RunSpec	required	The staged inputs and the run’s folders, mounted at /input, /output and /scratch
returns	RunResult		The agent’s output as text, and its token and request counts