login

Do NOT read, open, or inspect any files yourself. Run only the given commands.

The user cannot see tool output (Claude Code collapses it), so any output you are told to show or write verbatim is copied exactly into your reply text — never summarized or paraphrased.

AskUserQuestion, everywhere below: at most 4 questions in one call and at most 4 options in one question; multiSelect: true only with 2 or more options. Make one AskUserQuestion call per message and wait for its answers before the next (two calls in one message show the user only the last). A dismissed question runs nothing.

How this command runs

Every read and check below comes from ONE script. Run it FIRST, before anything else, with the Bash tool's timeout set to 600000 (the sign-in waits for the browser):

node ${CLAUDE_PLUGIN_ROOT}/scripts/login-flow.mjs

It runs the steps below in order and stops at the first one that needs a question. Each step's output is a block: a <name …> line, the output, then a </name> line. Its LAST line says what to do now: next=<step> account=<key>, or just next=stop when it stopped before a link. <key> is the account just linked; every later command passes it as --account <key>.

next=	Go to
stop	Step 0 or Step 1: show what they say, run nothing else
setup	Step 1w, where this machine's analytics go
here	Step 1w, this session's own folder
key	Step 3s-a
payment question=gateway	Step 3a
payment question=tier	Step 3c
joined	Step 4a, new workspaces
routes	Step 4a, repos and folders with no rule
backfill	Step 4b

After a question's answers (or its dismissal) and the commands they call for, carry on with EXACTLY this, with the same timeout, where the step names <step>:

node ${CLAUDE_PLUGIN_ROOT}/scripts/login-flow.mjs --account <key> --from <step>

It prints the blocks of the steps from there on and a new next= line. A listing that fails shows a ✗ line inside its block, and the script goes on: write that line verbatim. Output that ends without any next= line means the script itself failed: show its last ✗ line verbatim and stop (running /beezi:login again resumes safely). Never run the scripts behind the blocks yourself (login-preflight.mjs, login.mjs, the workspace.mjs listings, billing-capture.mjs --from-claude, key-resolve.mjs status): the blocks hold their output. Before doing what next= says, write out what each block printed this time calls for, in order, as its step below says (Step 1's sign-in output, Step 2's summary, Step 3s's report, the no-folder line in Step 4a).

Step 0 — preflight: the <preflight> block.

It checks the three things that make the link fail halfway: plan mode, auto mode (its permission classifier denies this plugin's node scripts), and a state directory this session cannot write to (a sandboxed Bash session only allows writes inside the working directory).

If its output starts with ✗ (the script then stopped with next=stop), STOP: show those lines to the user verbatim and run no other command — not the sign-in, not the plan capture, not the backfill. A flow that dies partway leaves this machine half-linked.

If login-flow.mjs does not run at all — denied by the permission classifier, or blocked by plan mode — that IS the answer, and the same rule applies: STOP. Do not retry it, do not try PowerShell or another shell, do not run any later step. Tell the user their session's permission mode is gating the plugin's scripts, that they should press Shift+Tab to switch to normal mode, and run /beezi:login again.