graph-engineering-loop-workspace

CLI for Graph Engineering Loop. Install the package that already exists on npm:

npm install -g graph-engineering-loop-workspace

Commands after install:

graph-engineering-loop-workspace --help
graph-engineering-loop --help
loopgraph --help

npx graph-engineering-loop-workspace --help
npx graph-engineering-loop-workspace validate .loopgraph/loops.json
npx graph-engineering-loop-workspace run .loopgraph/loops.json --adapter claude --claude-permission-mode acceptEdits
npx graph-engineering-loop-workspace run .loopgraph/loops.json --adapter stdio --adapter-command "node /absolute/path/to/adapter.mjs"
npx graph-engineering-loop-workspace run .loopgraph/loops.json --adapter stdio --adapter-command "node /absolute/path/to/adapter.mjs" --project-graph graphify
npx graph-engineering-loop-workspace cancel

The graph path is resolved from the current directory. The package does not install repository examples into your project, so supply your own graph or run the examples from a clone of the repository.

Every real run generates .loopgraph/status.md and .loopgraph/status.json. Open the Markdown file while the process runs to see the live dependency graph, active objective/tasks, current iterations, and waiting or blocked loops. Normal CLI completion output prints its path.

Graphify project context

Install the optional Graphify CLI, then select it explicitly:

uv tool install graphifyy
npx graph-engineering-loop-workspace run .loopgraph/loops.json \
  --adapter claude \
  --claude-permission-mode acceptEdits \
  --project-graph graphify \
  --project-root "$PWD"

The first run performs local code-only extraction; later runs use Graphify's incremental update command. Full packets are stored in .loopgraph/context/<loop-id>.json. The built-in Claude prompt receives bounded file/node/community scope and omits raw query content. Graphify query logging is disabled for subprocesses launched by LoopGraph.

Options:

• --graphify-path <path> selects a non-default Graphify executable.
• --graphify-graph <path> queries an existing project-local snapshot without refreshing it.

Run --dry-run to validate and preview scheduling; Graphify preflight starts only on a real run. Review graphify-out/ before committing it.

Real Claude smoke test from the repo:

npm run smoke:claude

This creates a fresh temp project, invokes Claude Code with a $1.00 max budget, and verifies that the adapter creates tmp/claude-smoke.txt.

The interactive adapter is the file-backed bridge used by the Claude Code plugin. It waits for structured responses under .loopgraph/bridge/; normal standalone users should select fake, stdio, or claude.

In the plugin, the current interactive packet repeats as a guarded Ralph-style loop until runtime evidence passes or maxIterations is exhausted. The graph scheduler alone advances to dependency-ready work; no additional public command is introduced.

See the repository README for Claude Code plugin usage and loops.json details.