Quick start

1. Give pi a task. Pi picks the entry skill whose description matches: coding for a bug fix or a feature, investigating for a question, inspecting for a review, and so on. Skills lists them all.
2. Read the alignment. The first reply opens with Alignment: <your goal, the problem it solves, and the smallest change expected to meet it>. A misreading caught here costs one message instead of a wrong change, so correct it right away.
3. Watch the loop. The first todo the agent creates opens a loop, and a widget below the editor shows ⚓ <kind> · <last complete phase> ✓ → <next phase>. The agent states its anchor in a Reality anchor: line: the exact check that will observe the result. anchorloop_stage read shows the loop’s reading in full.
4. Read the report. The reply carries what the anchor observed, and anything it did not exercise is marked unknown. Ask for a review, a human check, or a deploy when the report says one is still open.

What you may see

• A one-line warning on an edit: the loop has no anchor yet, or a durable loop is missing a decision, spec, or approval. The edit still goes through.
• A next: advisory at the end of a turn from the second-thought extension, when Jev has a concern about the loop’s evidence.
• A line asking to re-classify when a quick or inline change grew past three source files or two directories.

The extension reference says exactly when each appears.

Next

• Concepts: the loop, the four kinds, and anchors.
• Skills: what each entry skill anchors on.