agentic_tools 0.1.1
agentic_tools: ^0.1.1 copied to clipboard
Tools an AI agent can call: a typed contract, a registry, argument validation and repair, human approval for anything destructive, and execution events.
Changelog #
0.1.1 #
- Shortened the package description to the 60-180 character window pana scores against. Search engines truncate anything longer, so the ten points it withheld were pointing at a real defect: the useful half of the sentence was never being shown.
0.1.0 #
Initial release of the tool layer.
Added #
- Contract —
Tool,ToolSpec,ToolInvocationandToolResult, with specifications validated at construction against the naming rules every major provider enforces. - Authoring —
FunctionToolfor closure-backed tools,FunctionTool.textfor the simplest case,DelegatingToolfor cross-cutting behaviour by composition, andRenamedToolfor resolving naming collisions between packages. - Registry —
ToolRegistrywith eager and lazy registration, andToolSetfor handing each agent only the tools it needs. Specs are readable without constructing implementations. - Execution —
ToolExecutorperforming argument repair and validation, human-approval gating that fails closed, per-tool time budgets enforced even against uncooperative tools, cancellation, bounded-concurrency batches with serialised writes, and tracing. - Events —
ToolCallStarted,ToolCallCompletedandToolApprovalRequested, with aToolFailureKindcategory on every failure.