CaseGraph Documentation

CaseGraph is a programmable case engine for AI workflows.

Use it when a workflow should leave behind more than a final model answer: a durable case, workers that record evidence, proposed actions, policy gates, commit receipts, and replayable provenance.

Start Here

1. Use Developer API for the smallest governed agent path with cover.case(...).
2. Read Core Concepts for the mental model.
3. Follow the Quickstart to build a pack-backed workflow from a scaffold.
4. Use Pack Authoring when editing generated pack files.

The browser, CLI, and FastAPI API are still useful. They are reference and operations surfaces, not the first thing a developer needs to understand.

What You Can Build Today

Wrap an existing structured-output agent with cover.case(...). CaseGraph records the decision as a persisted case, proposes the typed local action, previews it, applies automatic local policy, commits, verifies, and replay-checks the result.

Build a tiny workflow from a scaffold when you want a pack-backed CaseGraph project.

For generated pack work, build a tiny workflow by generating a local/fake transactional pack, editing one worker and one action, validating the pack, running it through the top-level Python API, and inspecting replay.

The pack-backed path is:

uv run --all-groups --python 3.13 casegraph packs scaffold --kind transactional \
  --pack-id my_transactional \
  --output .casegraph/my_transactional_pack \
  --json

Then edit the generated files, validate them, and run the generated demo.

Where To Go Next

• Browser Casefile: inspect a case visually.
• Reference Index: compact CLI, API, model, and troubleshooting pointers.
• Status: current capabilities, gaps, and roadmap.

PLAN.md remains the architecture and roadmap record. This docs site is the user workflow guide.