pmstate 0.1.0

The directory tree IS the process state. A tiny library for filesystem-shaped agent processes.

pmstate

The directory tree IS the process state.

A tiny Python library for agent-driven processes that live on the filesystem. Each node in your process tree owns a slice of state on disk — an append-only event log or a JSON table. The agent navigates, reads, spawns, and prunes the same way a human navigates a project folder.

No DSL. No decorators. No compile step. Plain attrs-shaped values, plain JSON on disk, and a four-tool surface any LLM harness can drive.

Why

Most agent frameworks treat process state as opaque blobs in a vector store or workflow engine. pmstate flips it: the tree of folders and files is the state. That means:

• An agent can ls-style introspect what's happening at any depth.
• Branches spawn and die at runtime — no recompile.
• Every event is a CloudEvents-shaped JSON line. Replayable, auditable, human-readable.
• A human can cat a log file and see exactly what the agent saw.

Install

pip install pmstate              # core
pip install pmstate[claude-sdk]  # with the Claude Agent SDK harness

Example: a procurement flow in 28 lines

from pmstate import Node, Log, Table, Tree
from pmstate import ClaudeHarness

def quote_view(events):
    """Vendor quotes received and pending approval."""
    pending = [e for e in events if e["type"] == "quote.received"
               and not any(a["data"]["quote_id"] == e["id"]
                           for a in events if a["type"] == "quote.approved")]
    return {"pending_count": len(pending), "latest": pending[-1] if pending else None}

def procurement_rollup(children):
    return {
        "open_quotes": children["quotes"]["pending_count"],
        "open_lpos":   children["lpos"]["count"],
        "blocked":     children["quotes"]["pending_count"] > 5,
    }

procurement = Node(
    "procurement",
    description="Vendor quotes, LPOs, approvals.",
    reducer=procurement_rollup,
    children=[
        Node("quotes", state=Log("state/quotes.jsonl"), view=quote_view),
        Node("lpos",   state=Log("state/lpos.jsonl")),
        Node("vendors", state=Table("state/vendors.json")),
    ],
)

tree = Tree("project_alpha", root=Node("active", children=[procurement]))
ClaudeHarness(tree).run()

That is the full procurement flow. One custom view, one reducer, three leaves. The agent gets four tools (list_tree, get_state, find_state, read_log), discovers the structure on its own, and answers questions like "what's blocking us?" by reading the rolled-up state.

Concepts

• Node — a named position in the tree. May own state (a Log or Table), a view (function: events → dict), a reducer (function: children's views → dict), and children.
• Log — append-only JSONL of CloudEvents-shaped events.
• Table — JSON document for slowly-changing reference data.
• Tree — the wrapper that gives you spawn(parent, child) and prune(path) for runtime mutation.
• Harness — adapter that wires the four agent tools into an LLM runtime. v0.1 ships ClaudeHarness; the surface is harness-agnostic.

Quickstart

A runnable end-to-end example lives at examples/procurement/. After installing the claude-sdk extra and setting ANTHROPIC_API_KEY:

git clone https://github.com/matshoppenbrouwers/pmstate
cd pmstate
uv sync --all-extras
uv run python -m examples.procurement.seed_data
uv run python examples/procurement/run.py "what is pending in procurement?"

The agent navigates the procurement subtree, calls the four pmstate tools, and answers from the rolled-up view. Costs a few cents per run.

Status

v0.1 alpha. Open but not supported. One user (Laterite). API will break without warning. PRs not yet accepted. Stars welcome.

License

MIT.