markery-langgraph

LangGraph review workflow for Markery patent-trademark research.

Automates the candidate review cycle: loads the project queue, runs historian card --infer for each candidate, routes on the model's recommendation (confirm / reject / defer), writes confirmed pairs via matchmaker confirm, and drafts essays via historian draft. Human review is surfaced via LangGraph's interrupt() mechanism before any confirmation is written.

Requirements

• Python ≥ 3.11
• A working Markery installation with markery on PATH
• MARKERY_ROOT environment variable pointing to the Markery repo root

Setup

This repo runs in its own isolated virtual environment, independent of the Markery venv — nothing here imports markery as a library; it only shells out to the markery CLI. Create the environment with virtualenv, which bundles its own pip (the stdlib python -m venv cannot be used on this host: ensurepip / python3.12-venv is unavailable and installing it needs sudo).

# Clone alongside the Markery repo (the sibling layout lets MARKERY_ROOT auto-resolve)
git clone https://github.com/CosmoGSpacely/markery-langgraph.git
cd markery-langgraph

# Create the isolated env with virtualenv (bundles pip; works without ensurepip)
python3 -m pip install --user virtualenv     # once, if virtualenv is not present
python3 -m virtualenv .venv
.venv/bin/python -m pip install -e '.[dev]'

MARKERY_ROOT no longer has to be exported — config.resolve_markery_root() finds the sibling markery/ directory automatically (see "Markery root resolution" below). An explicit export MARKERY_ROOT=/path/to/markery still works and takes precedence.

Run everything through the env's interpreter (.venv/bin/python, .venv/bin/pytest) or source .venv/bin/activate first.

Contract check

Before running any workflow, verify that the installed Markery version exposes the expected subprocess interface:

python -c "
from langgraph_markery import config
config.check_contract('$MARKERY_ROOT')
print('Contract OK')
"

This reads MANIFEST.json from MARKERY_ROOT and asserts contract_version == "1.1". If the version does not match, update either this package or Markery before proceeding.

Subprocess interface

The workflow calls four Markery CLI commands:

Command	Used by
markery historian card <project> <slug> --infer	run_card_infer()
markery historian digest <project>	run_digest()
markery historian draft <project> <slug>	run_draft()
markery matchmaker confirm <project> <slug>	run_confirm()

Running the graph

export MARKERY_ROOT=/path/to/markery
python -m langgraph_markery.graph <project>

The graph loads all unreviewed candidates for <project>, runs historian card --infer on each, and routes on the model's recommendation:

• confirm → pauses at human_gate for approval, then calls matchmaker confirm and historian draft
• reject → appends the candidate to rejected.jsonl
• defer → logs the slug for later review (no file write)

The graph runs until the candidate queue is empty. Confirmed slugs are printed at the end.

Human gate

When the graph pauses at human_gate, the CLI prompts interactively. To drive the graph programmatically, build a graph with a persistent checkpointer and resume with an override:

from langgraph.checkpoint.memory import MemorySaver
from langgraph_markery.graph import build_graph

graph = build_graph(checkpointer=MemorySaver())
thread = {"configurable": {"thread_id": "radio-pioneers-review"}}

# Run until interrupt
list(graph.stream(initial_state, config=thread))

# Inspect the interrupted state
snapshot = graph.get_state(thread)
print(snapshot.values["infer_result"])

# Resume with a human decision
graph.update_state(thread, {"recommendation_override": "confirm"})
list(graph.stream(None, config=thread))

recommendation_override accepts "confirm" or "reject". Any other value defaults to "reject".

Markery root resolution

config.resolve_markery_root() resolves the Markery repo root in this order (first hit with a MANIFEST.json wins):

1. the MARKERY_ROOT environment variable, if set;
2. a .markery-root pointer file in this repo root (first non-comment line is the path; relative paths resolve against the repo root);
3. a sibling markery/ directory next to this repo.

Running tests

cd markery-langgraph
.venv/bin/pytest        # or: source .venv/bin/activate && pytest

Tests use mocked tool calls — no live Markery CLI or MARKERY_ROOT required. The 30-test suite runs entirely from the isolated .venv.