Quickstart

This is the shortest path from a fresh install to a recorded, LLM-judged cycle.

1. Configure an LLM (once)

Pick whichever provider you have. The default is an OpenAI-compatible API key:

pdsa config key <your-api-key>     # or: key-file <path> to keep it out of config
pdsa config model gpt-5.6-terra    # any model your endpoint serves
pdsa check                         # verify with a real round-trip

Other options — keyless local models, OAuth, Codex, or your logged-in Claude Code — are covered in Providers & Auth Modes.

2. Pick a project

Each project gets its own graph DB, so learning stays separated:

pdsa project set my-repo
pdsa project show

3. Run one cycle

pdsa plan  "Add a request timeout to the API client and prove it fires without breaking the happy path"
pdsa do    "Wrapped the call in a linked CancellationTokenSource; added a unit test for the timeout path"
pdsa study "204 tests green; timeout fired at 1.2s with a friendly message; happy path unaffected"
pdsa act   --note "Follow-up: add an OS-level test that no child process survives the kill"

Read the output at each step — plan gives you an expected evaluation to aim at, and study returns a verdict (met / partial / unmet).

4. See what accumulated

pdsa status     # recent cycles + expectation hit-rate (recall)
pdsa eval       # per-cycle expected / verdict / actual
pdsa view       # open the local graph viewer

What just happened

• plan asked the LLM to commit to a verifiable success criterion and started a cycle.
• do organized what you actually did against the plan.
• study compared the result to the expected criterion and recorded a verdict.
• act captured learnings and, if reinforcement is needed, links the next plan as a follow-up cycle.
• Every step became nodes and edges in a per-project Kùzu graph you can query and visualize.

Next: walk through each step in depth in Your First Cycle.