Two or three agents is where one person loses the thread.

Every session is the agent’s own CLI in a real terminal, so permission prompts, slash commands and resume work exactly as they do in your shell. Give a session its own Git worktree when several share a repository.

Beside the terminal, a timeline shows what the agent did rather than what it said: tool calls, permission waits and compactions on Claude Code; approvals and finished turns on Codex.

Sessions are ranked by what they are blocked on — a permission prompt, an error, a finished turn — not by how much they printed.

When you are the blocker, you hear about it: a macOS banner, a card in the app, or your phone once you pair it and turn its alerts on.

A goal, its tasks, its acceptance checks and the proof sit side by side, so you decide when work is actually finished.

Discarding a worktree with uncommitted changes asks first, and says how many files you would lose.

One SQLite file on your Mac holds what happened. What is left on a plan comes from the provider; what ran comes from Wanigan’s own records, and costs appear only when they were reported.

Session cost comes from the CLI’s own accounting and batch cost from a local price table. Any total that adds them shows both halves. Estimates say so, and missing numbers stay missing.

Projects, goals, evidence, schedules and the policy ledger are there when you reopen, along with archived Claude Code conversations you can resume.

A live agent can’t survive a full quit, and Wanigan won’t pretend otherwise: it asks before stopping one.