Module tools 

MCP tool implementations for aprender.

Phase-1 surface (shipped M1–M3):

• M1 scaffold: apr.version.
• M2 subprocess wrappers around apr <cmd> --json: apr.validate, apr.tensors, apr.bench, apr.qa, apr.trace, apr.run, apr.serve.
• M3 streaming slice: apr.finetune (opt-in notifications/progress per non-empty stdout line when params._meta.progressToken is set — see FALSIFY-MCP-PROGRESS-001) + notifications/cancelled → SIGTERM→SIGKILL for apr.run (FALSIFY-MCP-006).

Re-exports

pub use registry::DispatchFn;

pub use registry::McpToolEntry;

pub use registry::ToolIndex;

pub use bench::bench_tool_definition;

pub use finetune::finetune_tool_definition;

pub use qa::qa_tool_definition;

pub use run::run_tool_definition;

pub use serve::serve_tool_definition;

pub use tensors::tensors_tool_definition;

pub use trace::trace_tool_definition;

pub use validate::validate_tool_definition;

pub use version::version_tool_definition;

Modules

bench

apr.bench — M2 tool. Benchmark inference throughput (tok/s, latency percentiles).

finetune

apr.finetune — LoRA/full-finetune subprocess wrapper.

qa

apr.qa — M2 tool. The 8-gate quality checklist; first stop for any model issue.

registry

HELIX-IDEA-002 — link-time tool registration via inventory.

run

apr.run — M2 tool. Synchronous inference via subprocess wrapper.

serve

apr.serve — fire-and-forget subprocess wrapper over apr serve.

subprocess

Shared subprocess wrapper for M2/M3 tools.

tensors

apr.tensors — M2 tool. List tensor names, shapes, and (optionally) stats.

trace

apr.trace — M2 tool. Layer-by-layer tensor trace for debugging a model.

validate

apr.validate — M2 subprocess wrapper over apr validate <model> --json.

version

apr.version — M1 stub tool that reports the aprender-mcp crate version.