0201-express-defaults-gate.decision.md

LLP 0201: One question up front accepts every default; declining asks the real questions

Type: Decision Status: Draft Systems: Onboarding, CLI Author: Brendan / Kenny / Claude Date: 2026-08-07 (revised 2026-08-31) Related: LLP 0190 (#pick-gate, #sync-gate: the per-lane defaults gates this retires; the menu semantics there are untouched), LLP 0188 (#never-silent: the floor this must not breach), LLP 0200 (#wizard: the third lane it answers), LLP 0135 (#progress, #orchestration), LLP 0131 (#attended-only), LLP 0129 (#fork), LLP 0191 (#back-edges), LLP 0341 (a dead output stream ends the run as a cancel)

Extends LLP 0190 and retires its per-lane gates. The wizard asks "accept the defaults?" exactly once, right after the fork (and the join, on the team pathway). Yes skips every remaining question; no asks the real questions, linearly, with no further accept-or-customize screens.

Context {#context}

LLP 0190 gave the pick and sync lanes a defaults gate each, and the first revision of this document added an express gate in front of them that could accept all of them at once, while keeping every per-lane gate behind it.

Dogfooding that shape surfaced two failures:

• The gate-of-gates read as the same question twice. Declining "Record and sync all of these" landed immediately on the pick lane's "Record all" over the same rows, then the sync lane's "Sync all", then the folder question: up to three more accept-or-customize screens after the user already said "let me choose". Saying no once must mean "ask me the real questions now".
• The detected list went unread. The gate rendered the detected tools in the prompt's items chrome, above the key-hint line (up/down · enter pick · esc back). Users read the option rows, not the text above the chrome, so "all of these" pointed at a list nobody had looked at.

What must not be lost is the never-silent floor: LLP 0188 #never-silent requires the wizard to say what will ship before anything ships.

Decision {#decision}

One express gate precedes the question lanes and is the wizard's only accept-or-customize screen. On an attended run, after the fork (and the join, on the team pathway) and before the pick lane, the wizard asks one two-row question - on every pathway, local included, whenever seeding yields default rows:

Set up recording
> Record and sync everything
    Configures Claude Code and Codex to record through HypAware.
  Customize
    Choose what to record, what syncs, and how new folders are handled.

The rows are self-explaining; nothing load-bearing rides the items chrome. The first revision's "the list is the explanation" put the detected tools above the prompt as items, where nobody reads. The tool names now live in the accept row's own summary sentence, which is both the disclosure (accepting configures those tools - the one-line side-effect statement LLP 0190 #pick-gate put on the retired pick gate's accept row) and the evidence (what was found), in the one line the happy path is guaranteed to read. Both halves have to stay: an express accept never opens the menu whose per-row summaries carry the specifics, so a sentence that only named the tools would leave "this changes your tools' configuration" stated nowhere on the path most users take. The names are the pick lane's own default rows from one computation (resolvePickSeeding, shared with the lane), joined plainly; fleet-locked rows appear by name like any other, and the fleet detail stays on the later screens and the accept narration, not in this sentence.

sync is claimed only where accepting would in fact sync everything the row names. On a solo machine nothing forwards, so the label drops to "Record everything" and the Customize summary to "Choose what to record." That gloss is shorter for the same reason: the sync and new-folder lanes do not run on a solo machine, so a decline there opens only the pick menu. The same drop applies on an enrolled machine whose client-sync store already withholds one of the named rows: an accept preserves standing opt-outs verbatim rather than clearing them (#narrate, and LLP 0188 #opt-out), so "and sync everything" would be a promise the accept does not keep - the case the retired sync gate handled with its own "Sync all" against "Keep this" split, which this row inherits along with the decision. The gate reads the store the sync lane reads, and a read it cannot complete answers "withheld": a row that cannot prove everything syncs must not say it does. Only the claim narrows; the accept narration still states the split in full, so nothing goes unsaid. The accept summary is the same sentence in every case: what it discloses is the configuring, which happens on a solo machine too. The new-folder policy that rides with an enrolled accept is deliberately not in the row copy - it is stated by the accept narration (#narrate) - so the row stays one readable sentence about the tools.

Declining asks the real questions, linearly. The per-lane defaults gates (LLP 0190 #pick-gate, #sync-gate) are removed. Customize opens the pick multiselect directly, then - on enrolled runs - the sync multiselect and the new-folder question, in that order, then the finale. No later screen ever again asks "defaults or customize".

The decline row names every question it opens. The gate is the one screen a user decides on before anything is configured, so its second row has to say what saying no leads to, in full: on an enrolled run that is three questions, not two, and a gloss naming the two menus left the new-folder question - a standing privacy preference (LLP 0200) - unnamed on the only screen that could have named it. The clauses track the counted lanes' own labels ("Choose what to collect", "Choose what syncs", "Choose how new folders are handled"), so the row and the position lines it opens read as the same list. Unenrolled the row names only the pick menu, which is the only lane a decline opens there.

A two-answer gate splits the prose that documents it, and what applies to both answers may not land under one of them. docs/TEAM_SETUP.md describes the gate and then, below it, the screens Customize opens. That second block is a subsection, so everything after it is nested inside it until the next heading - and the first thing after it is the upload-hold callout, which is the doc's one pointer to the review window, the hypaware-privacy skill, and PRIVACY.md. Nesting it there hides it from exactly the reader who took the one-keypress accept and has no reason to read a section headed "If you choose Customize". The rule is structural, not editorial: a disclosure that holds for every reader of a section sits at that section's level, never under a heading that states a condition. Text unchanged and structure moved is the same regression as text removed, and it does not show up in a diff of sentences.

The menus still open on the same defaults the retired gates stated: the pick menu's boxes are seeded by detection and the locked set, the sync menu's boxes are checked for everything that syncs with locked rows leading read-only, and a bare enter keeps the checked state (LLP 0274). So "inspect everything and change nothing" is still one enter per screen; what is gone is the extra screen per lane that asked whether the user wanted to see the screen.

The menus are the never-silent statement on this path: each one names the rows it records or ships before anything ships, which is the same statement the retired gates carried (LLP 0188 #never-silent binds the statement, not the prompt shape). The sync lane's no-candidate narrations (LLP 0276 #no-candidates, LLP 0289 #ask-the-store) are unchanged: a lane with nothing to ask still states its facts and asks nothing.

Accepting skips the prompts, never the statements. Each lane still runs; it prints the title and items it would have shown and proceeds with the default instead of waiting for an answer. The pick lane narrates "HypAware will record:" with its rows (locked ones fleet-suffixed), the sync lane narrates "These will sync to your server:" with the split it would have shown, and the new-folder lane narrates its question and records its standing answer. This is what keeps the fast path inside LLP 0188 #never-silent.

Accepting takes each lane's standing answer, which is its default only where the user has no standing one. The two stores behind the gate are read, never reset: the sync lane keeps its local-only entries (LLP 0188 #opt-out) and the new-folder lane keeps its recorded mode (LLP 0200 #wizard). "Accept the defaults" is an answer to the questions this run asks, not a licence to discard answers a previous one recorded - and each of these lanes narrates the value it took, so a lane that reset one would announce the new value as though the user had just chosen it.

A narrated statement is finished on stdout even when the write behind it fails. The new-folder lane's narration is a sentence split across two writes: the title is a lead-in ending in a comma and the indented clause under it completes it. A failed preference write warns and leaves the previous mode standing, which is the right outcome, but returning at that point left the lead-in on screen with the next phase's output under it - a question the run asked and never finished. The failed arm therefore prints the same clause, naming the mode that actually stands (the store was not written, so the standing answer is still true of the machine), before the warning rather than after it: stdout and stderr are one terminal on an attended run, so a warning written first sits inside the sentence instead of under it. The "could not record" half stays on stderr and the recorded form's "change later with ..." tail is not repeated, so an unrecorded answer never reads like a recorded one. The failure this arm cannot absorb - the output stream itself dying - is settled run-wide by LLP 0341: lane writes never throw, and a dead stdout cancels the run at the next boundary.

The narration is blocked, not streamed. Each statement is led by a blank line and its detail is indented under a title, because on this path there are no prompts to separate the statements: without the spacing the run reads as one paragraph of mixed subjects. A lane whose answer is a single fact prints it as one more indented line of its own block.

Nothing found means no gate. With nothing detected and nothing locked there is no default worth naming and nothing to accept: "record everything" over nothing would mean "record nothing". The orchestrator skips the gate entirely and the pick lane opens its menu, exactly as it does on a decline.

The gate carries no position line, and an express run states none. The gate is what decides how many questions remain, so it can no more state a total than the fork can (LLP 0135 #progress: the denominator resolves once and never moves). After an accept, no lane states a position either: a "Step 3 of 5" above a narration would count screens nobody is answering. A declining run keeps every position exactly as before.

It behaves like any other wizard question. Escape steps back to the fork (LLP 0191 #back-edges); ctrl+c cancels the run at 130, narrating the enrolled state first when this run has already joined. It is asked once per pass through the lanes, so a back to the fork and forward again asks it again and may be answered differently. With the per-lane gates gone each lane is one screen, so the back chain is: folder question to sync menu (or past a sync lane that asked nothing to the pick menu), sync menu to pick menu, pick menu to this gate whenever it was shown, and this gate to the fork. On a pass where there is nothing to accept and no gate is shown (#no-default-no-accept), the pick menu's back edge reaches the fork directly - including the pass a back into the gate opens, since the rows are the picker's own confirmed defaults and a confirmed empty selection leaves none. Non-interactive runs (--yes, presets, --from-file) never see the gate: they already take every default without prompting (LLP 0131 #attended-only).

Rejected: making express the only path (the menus are how a user with an unusual machine says no, and LLP 0011 #autodetect-vs-default holds that detection seeds and never forces). Rejected: keeping the per-lane gates behind the express gate (the first revision's shape; declining then asked "accept defaults?" up to three more times). Rejected: showing the express gate only on enrolled runs (the first revision's #one-lane-no-gate; with the pick gate gone, the solo pathway needs the accept question too, and one rule on every pathway beats two). Rejected: rendering the detected list in the items chrome (text above the key-hint line goes unread; the row summary is where the eyes are). Rejected: silently auto-accepting a lane whose gate had exactly one obvious answer (the never-silent objection LLP 0190 raised against skipping the sync step).

Consequences {#consequences}

• The attended happy path on every pathway is: fork (plus join on team), express gate, narration, finale. Two questions, and the second names the tools it is about to configure in the row the user actually reads.
• The decline path is linear: pick menu, then on enrolled runs the sync menu and the folder question, then the finale. Each screen is a real question; none of them is a gate.
• Every lane keeps its autoAccept seam. A lane that later adds a question must decide what accepting it means.
• A user who accepts express and later wants to change any of it uses the same standing controls as everyone else: hyp privacy client, hyp privacy folders, and re-running hyp init.
• The step counter describes only the decline path. That is honest, not a regression: an express run has no steps to count.
• The accept summary grows with the number of detected tools, and is not capped. With the shipped picker labels and the four-space indent it measures 54 columns at one row, 64 at two, 75 at three, and 91 at four, so from four detected tools up it soft-wraps on an 80-column terminal with the continuation starting at column 0. Left as it is: the redraw is correct either way (countPhysicalRows counts wrapped physical rows, so frames do not smear), the one and two row cases that cover almost every machine never wrap, and both repairs cost more than the wrap does. Capping the list ("Claude Code, Codex, and 3 more") would drop names from the sentence that is the evidence half of this row, and wrapping in the renderer would change every prompt in the CLI to fix one summary.

References

• LLP 0190, LLP 0188, LLP 0200, LLP 0135, LLP 0131, LLP 0129, LLP 0191, LLP 0274, LLP 0276, LLP 0289
• src/core/cli/wizard/express.js (the gate and the shared narration), src/core/cli/wizard/index.js (orchestration, the run-once detector), src/core/cli/wizard/pick.js (resolvePickSeeding, the menu-only lane), src/core/cli/wizard/sync_scope.js (the menu-only lane), src/core/cli/wizard/folder_ask.js (the autoAccept arms)