pmstate 0.3.1

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

pmstate

PyPI CI Python

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: build a hiring tracker in three commands

# 1. Edit pmstate.yaml to describe your tree (see docs/spec-authoring.md)
pmstate init                                       # writes pmstate.example.yaml
$EDITOR pmstate.example.yaml && mv pmstate.example.yaml pmstate.yaml

# 2. Generate the project (tree.py, views.py, reducers.py, AGENTS.md, …)
pmstate init --from-spec pmstate.yaml my-process
cd my-process

# 3. Seed deterministic events and ask a question
pmstate seed --n 30
pmstate run "what's pending?"

# 4. Record a real state change with `pmstate append`
pmstate append /pipeline/offers --type candidate.advanced \
  --data '{"from":"interviews","to":"offers","note":"looked strong"}'

# 5. Or hand the writer to the agent (opt-in)
pmstate run --write "advance Ada from interviews to offers"

That's the full loop. The agent gets four tools (list_tree, get_state, find_state, read_log), reads the rolled-up state, and answers questions like "what's blocking us?". One custom view, one reducer, the directory tree on disk is the state.

Driving from Claude Code

pmstate.yaml is the source of truth. To translate a natural-language request into a working tree, point Claude Code (or any orchestrating agent) at [docs/spec-authoring.md](docs/spec-authoring.md). It documents the schema and includes three worked examples (linear pipeline, kanban, log + rollup hierarchy) plus a 5-rule recipe.

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. Ships ClaudeHarness today; the surface is harness-agnostic.

Quickstart

New to pmstate? Walk through [QUICKSTART.md](QUICKSTART.md) — a 10-minute guide that builds a working agent-navigable process tree from scratch using the CLI. Every concept is explained inline; every step has runnable code. The original "write the seven files yourself" path is preserved as an appendix.

The full CLI reference lives at [docs/cli.md](docs/cli.md); the agent-facing spec authoring guide is at [docs/spec-authoring.md](docs/spec-authoring.md).

A larger end-to-end example lives at [examples/procurement/](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?"

Status

v0.3 alpha. Open but not supported. API will break without warning. PRs not yet accepted. Stars welcome.

License

MIT.