PDLt Taskmaster & The System 2 Catalogue¶

PDLt Taskmaster is a protocol harness for verified, auditable model execution: a Python library first, with an extensions SDK, an MCP server, an interactive REPL and a headless CLI on top.

Entry points¶

Set it up on your host¶

Setup differs by operating system, because each has its own sandbox and shell. Each page takes you, or an agent, from nothing to a first verified task, with a check at the end of every step.

Get started explains the steps and how to choose; Reference holds the settings and file locations shared by all three.

Core Premise: Controller-Gated Agentic Execution¶

Most agentic harnesses give large language models uncontrolled execution loops: the model generates code, runs arbitrary shell commands, and loops recursively without deterministic checkpoints or structured semantic bounds.

PDLt fundamentally shifts the paradigm: 1. Semantic Bootstrap Containment: Raw user text and untrusted prompt data remain passive data. The model cannot execute until interpretation and planning phases are complete. 2. Controller-Gated Phases: The model interprets user requests as short, human-readable pseudocode; in review mode, the run waits for confirmation before execution begins. 3. OS-Native Confinement: All generated code executes inside a restricted, session-scoped OS sandbox (Landlock on Linux, Seatbelt on macOS, AppContainers on Windows). 4. Deterministic Delivery Verification: Code and results are checked against formal output contracts, verification suites, and witness projections.

The Two-Plane Separation¶

A foundational design invariant in PDLt is the strict separation between the execution harness and the evaluation benchmark:

graph LR
    subgraph HarnessPlane["Harness Plane (src/pdl_taskmaster/)"]
        H1["REPL Controller"]
        H2["State Machines"]
        H3["OS Confinement"]
        H4["Never knows benchmark questions or solutions"]
    end

    subgraph EvaluationPlane["Evaluation Plane (run_catalogue.py, graders.py)"]
        E1["Catalogue Manifest (112 prompts)"]
        E2["Solution Ground Truths"]
        E3["Automatic Graders"]
        E4["Multi-Arm Parity Sweeps"]
    end

The harness acts strictly as a referee, never a solver: no algorithm hints, no keyword gates, and no fabricated witnesses. A failed prompt is diagnostic data; a gamed pass is an integrity defect.