Repository navigation
Groom: TUI question component UX — no navigation between questions, long pause after final question, awkward up/down arrow behavior #830
Description
Summary
The question component (the TUI surface where the agent asks the person questions during a task) has several UX problems that need grooming before any implementation work starts.
Observed problems
- No way to move between questions — once a question is shown there is no way to go back to a previous question or jump forward. Answering wrong means living with it or restarting the whole flow.
- Long pause after the final question — after answering the last question, the TUI sits for a noticeable period with no visible feedback before anything happens. Feels frozen; we should show what is happening (or fix the latency itself).
- Up/down arrow behavior is weird — arrow-key navigation doesn't behave the way a person expects (selection highlight, wrap-around, or focus semantics are off). Needs a defined key contract.
What this issue is
A grooming item, not a task: the expected behavior for each of the three problems has to be decided (navigation model, post-final-question feedback, arrow-key contract) before anyone breaks it into tasks.
Grooming questions to settle
- Should people be able to navigate back and edit earlier answers? If yes, what happens to downstream answers?
- What should happen between the final answer and the next visible state — spinner, progress, or eliminating the pause?
- What is the intended key contract for the question list (up/down, tab, vim keys)? Should it match the rest of the TUI conventions?
Notes
- Reported by hand during use of the TUI; no replication steps beyond "answer questions in a chat/task flow."
- Area:
internal/tui3question component.
Drafted with CodeAF · reviewed and owned by the author
Activity
santoshkumarradha commentedon Sep 11, 2026
Recon findings (2026-09-11, dev @ 39039b9)
Four read-only passes: engine, rendering + keys, freeze/latency, design contract. Nothing changed yet. This issue tracks the redesign + fixes; close it with the PR.
Defects — cause first
B1. Enter → ~10 s freeze (likely a deadlock, not slowness). Enter calls
ResolveQuestionover the wire synchronously inside bubbletea's Update (internal/tui3/question.go:2368). The engine wakes the turn and emits phase frames (session/loop.go:2508) that often beat the reply. The client reader hands each phase frame to a handler that callsp.Send(internal/remote/client.go:824-830→internal/tui3/tui3.go:1325-1331, twin at 1287-1314); bubbletea v2'sp.Sendblocks until Update is free — and Update is waiting for the reply that same reader can no longer read. OnlycallDeadline = 10s(remote/client.go:79) breaks it → "the engine did not answer in time". The comment atquestion.go:2352-2366already records "ten seconds to the millisecond". Dates from #747 (phase frames on the wire). Same exposure:AttachReplayon a conversation switch (observe.go:128-160),SetAutonomy, the room.Fix at the cause: non-blocking coalescing wake instead of
p.Sendin the reader (+ test: reader returns ≤50 ms while Update is busy); answer from atea.Cmdand close the question at once (markQuestionSentexists); AST law: no wire door called directly from Update.B2. Every recorded answer busts the prompt cache.
ResolveQuestion→recordDecisionrewrites the system prompt (session/question.go:1571→memory.go:1055), so the next request pays a cold prefix. Tool consent ("allow once") takes this road too. Also re-readsdecisions.jsonlundera.muand may callRememberScoped(a model call) synchronously.B3. Other post-answer waits. Landing accept runs the git merge inside the answer call (
task_audit.go:2627-2687). An answer on a checkpoint round is followed by the markreader (~8 s "taking stock",checkpoint.go:2697). The reflex gate is not on this path.A1. Arrows that don't move (ranked):
app.go:3066): a new question only notifies + bells (questiondelivery.go:158-168), is never pinned, andalt+arefuses because the count is 0 (question.go:3373). Nothing re-delivers it on return. Matches "I come back to a chat and the question has been asked".questionroom.go:1349/1709,app.go:3471).questionsheet.go:612-633).question.go:1308-1320, 2968).alwaysswallows every arrow (question.go:1684); 250 ms settle guard; text in the box → arrows go to the caret (:2677); phone width <60 moves the pointer but never draws it; no wrap at the list ends.Other correctness:
?ask-back on a modelaskis queued as a steer that is read only after the tool returns, i.e. after the question is answered. "Make it a rule" writes a hard-coded reason into memory. Five lanes (connect, harness, subharness, subharness-ask, fuel) never emit answered/withdrawn on the questions stream → stale on other windows. Schema offers confidencelow/medium/high, code readssure/fairly/unsure→ confidence never shown. Schemainput.pairshas no field behind it. Click row math likely off by the receipt rows. Receipt offersc changethat nothing handles. Ratify still counts in? 1 question. A question page stays open after its question is answered or withdrawn. The first key typed into an empty box answers the question (a digit picks,dhands the call back).UX debt (owner's complaints, confirmed)
hueAsk #C08FE8painted on whole rows: head, every label, every key, every·(styles.go:338,palette.ask/askBold,question.go:1021/1374/1476/1839). Breaks the design language's own "colour is stroke, never fill".?is amber on the block and violet in the room; the manual says labels are "amber".questionOffer,question.go:1422). The rule above the block names the wrong keys for a modelask(esc interrupt;escactually folds). Card key-row bolding is inverted (question.go:1843).escfolds straight to the status chip; no in-place collapsed state.· suggestedon the card,my pick · fairly surein the room, nothing on the line, phone sheet, batch sheet, home chips, task record.cor typing.Architecture (for the redesign)
session.Question(session/question.go:627) for 11 lanes: modelask, consent, task proposal, standing, landing/conflict, connect, harness, subharness, subharness-ask, fuel, recovery (borrows consent). Rich options: body, consequence, safe, widening, blocks (text/diagram/table/diff/image/layout), dimensions.Pick{key, reason, confidence, wouldChange}. Input: text/blanks/checklist/dial/secret.line/card/room/sheet) is a hint; the surface decides inapp.questionForm(tui3/question.go:849). Every model ask with no form is a card (the gate requires a reason).ask= one question; a multi-question panel must be gathered from concurrent asks (batch sheet) or needs a schema change (prompt-diet caps apply, The prompt profile is a row,askpays less for it, and the job footer is said once per change #844).esc= later (never cancel; test-pinned), 250 ms settle, never move the box under typing, first window wins, pointer opens on the safe answer for irreversible gates.docs/DESIGN-LANGUAGE.md("What we refused: Borders"). Three dim rounded boxes already ship (/folder,ctrl+kcard, onboarding) — should become one frame.internal/manual/chat/questions.md(1043 lines, 55 probes), ~40 pinned strings ininternal/e2e/tuiwords_test.go, the one key tableinternal/tui3/questionkeys.go, icon slots (pointer vs folded both▸today; checkbox/radio/recommended missing).Next: the view primitives + design prototypes, then the owner picks.
santoshkumarradha commentedon Sep 11, 2026
Owner's picks (2026-09-11) → work started
Pick sheet: https://claude.ai/code/artifact/24190a18-caa4-4a4c-9298-6917583944a4
?▸◆), like home; violet leaves the question surface; words in ink◆ recommendedat the right, reason/confidence under it only while focused (safe-answer law still wins on hands-only questions)↑↓ choose · enter take it · esc laterin the frame; secondary keys on one dim row belowsomething else…becomes an input when focused←→between questions, answers editable until sentAlso in scope (owner): every "waiting on you" surface — status chip, home "needs you" panel/strip, task rows
your call, task record, tab?, terminal title, phone sheet — one grammar, cleared the instant a question settles. Pulls in #677 (question behind the start page eats typed letters; sev:critical) and #840 (waiting chip displaces the task row's door hint).Lanes (Claude Opus, interactive, on the Spark; integration branch
feat/question-viewsoff dev 333acc6; lanes PR against it, then one PR to dev closes this issue):af-qv-F,feat/qv-F): p.Send deadlock + no wire call from Update (+ AST law, ≤50 ms loopback test), decisions out of message[0], away-rule pin + re-deliver, one off-screen key guard (chat: a question behind the New chat start page takes the letters somebody is typing into it, and one of them grants always-allow #677), first-key-answers rule, room scroll-to-focus, checklist single pointer, every lane's answered/withdrawn event, ask-back reaching the model, rule memory reason, schema confidence/pairs.af-qv-P,feat/qv-P): frame primitive + doc amendments, amber, chooser as one function on Question properties, panel/row/consent/fold/receipt, waiting surfaces (Task room: a waiting question's chip displaces the task row's door hint (e2e space_in_the_task_room_pages_the_card) #840), then evidence split + two-pane page, then tabs + review; manual + e2e needles + gallery rewritten.santoshkumarradha commentedon Sep 11, 2026
Owner approved five more (2026-09-11) — one mechanism, not five
Answers are messages. Today an answer is only the return value of the tool call that asked. It becomes a message to the conversation that also releases a waiting call when there is one. Three approved ideas are cases of that one door:
askwithBlocking.Turn=falsereturns at once; the answer arrives at the next step boundary (or starts a turn). Builds the deferred D8 seam./autonomy, not a new constant; any key pauses it.alwaysrevokes the rule; irreversible = not offered.Two more are views over what already exists:
Question.Scope(once/task/project/always).allow all N · one by one · deny all, fanned out as one answer per question; irreversible never grouped.Lane A (
af-qv-A,feat/qv-A) builds the engine half, design first (rewrites D8 indocs/design/questions/DESIGN.md); P draws all five; F gives the answer door a list form and a step/batch property for grouping.santoshkumarradha commentedon Sep 11, 2026
Owner-reported 2026-09-11 15:35 — an answered question keeps coming back. Standing proposal answered
3 just onceat 14:55:59 (recorded once); the card reappears on every tab/conversation switch, and pressing it again records nothing. Cause:internal/remote's held set drops a card only when its old per-lane resolve method is called (server.go:2455-2525); the chat now answers every lane throughMethodQuestionResolve(server.go:2102), which never clears it, so each attach re-hands the answered card andoutstanding()keeps the chatwaiting · your call. Fix at the shape (held set cleared by the question lifecycle events, not by which door was called) assigned to lane F in #910.santoshkumarradha commentedon Sep 11, 2026
Plan change (owner, 2026-09-11 16:05): merge to
devpiece by piece. The integration branchfeat/question-viewsis retired; every PR retargets todevand must stand alone there (manual true for exactly that PR). Order by readiness: #910 (engine: one lifecycle door, cache-stable answers, held-set fix for the answered question that came back) → #919 (answer road: Enter freeze, key ownership, #677) → #914 (answers are messages; rebases onto #910) → P1 (frame, amber, chooser, panel, receipt, chip #840, phone sheet, scope row, /autonomy settings row, pointer by stakes). Two new parallel lanes: R (evidence split + two-pane full page) and T (tabs + review, grouped permissions), stacked on P1. Side runs: #690 (host road drops self-started turns), #926 (keyboard-handover test red under load). This issue closes with the last piece.santoshkumarradha commentedon Sep 11, 2026
First piece merged: #910 → dev
154a0be65. The engine half: oneraiseQuestion/letGodoor for every lane;Question.Waiting()as the single reading of "is anything stopped on this"; the held set cleared by the question's own lifecycle (heldSet.keepOnly), which closes the owner-reported defect where an answered standing question came back on every attach;withoutSecretWords— a secret answer's words never reach the decision record, the questions lane, a--hostframe or message[0] (an OpenRouter key typed into the connect box was written to decisions.jsonl before this); the decision record out of message[0] so answering is byte-stable for the cache; the override memory write registered withstartMemoryJobsoClosewaits for it; ask-back keeps the question open;askschema enums generated from the Go constants.bin/aforgerebuilt on the owner's laptop at154a0be65.Next: #919 (rebasing), #933 (safety fix — owner ruled deny-first until the engine grades a call), #914 (rebasing onto #910), then #946 and R's page work.
santoshkumarradha commentedon Sep 11, 2026
#919 merged → dev
9e803c9ec(the answer road). The ~10 s Enter freeze is gone at its cause: a non-blocking coalescing wake door (doorbell.go) so the wire reader can never be parked by the UI, and every engine door asked off the update loop (offLoop), both held by law tests. Also the one off-screen key guard that closes #677 (a question behind the New chat page ate typed letters, one of which granted always-allow), questions that arrive while away now pinned and re-delivered, the full page scrolling to the focused answer, and one answering road that takes a list.Merged so far: #910, #919.
santoshkumarradha commentedon Sep 11, 2026
#954 and #960 merged. dev now has: #910 (engine lifecycle, secret answers, cache-stable), #919 (Enter freeze, key ownership, #677), #954 (an answer is a message: keep working while asking, opt-in countdown you can stop, change an answer), #960 (ask-back answered with the question still open; the chip counts only what waits on you). Left: #933 (views P1 — two row-count tests and a rebase), then #955 (evidence beside the answers, two-pane page) and #946 (tabs + review, grouped permissions) which are stacked on it, plus #963 (clocks let go on close).
santoshkumarradha commentedon Sep 11, 2026
#933 merged → dev
23f5da1eb. The drawing you picked is live: one frame primitive (the ctrl+k card and onboarding moved onto it), amber on the marks alone with the violet retired, one chooser keyed on the question's properties, the panel with the pointer on the pick and◆ recommendedat the right edge, two tiers of keys,escfolding in place, the✓receipt, the chip carrying the question's words (#840), the phone sheet, the permission lifetimes row, and the/autonomysettings row (plus the wire fix that made every door onto those rules work at all — it reported "no project" on every machine).The permission pointer ships deny-first per the owner's ruling, with
TestTheGateGradesNoCallAsIrreversibleAndMarksDenyTheSafeAnswerfailing the day #953 grades a call, so the allow-once half cannot be forgotten.Merged: #910, #919, #954, #960, #963, #933. Left: #955 (evidence beside the answers, two-pane page) and #946 (tabs + review, grouped permissions), both rebasing onto dev now.
santoshkumarradha commentedon Sep 12, 2026
Closing this. Every complaint in the opening report is fixed and on
devat27506e80b, across eleven pull requests.The four things you reported
"press enter, escape, all in the same line without any hierarchy." The keys are tiered now: the primary verbs sit on the frame's bottom edge, the rest on a second dim row under it, and a question drawn narrow drops to the rules-and-bands sheet rather than cramming. #933, #946.
"when I press enter, there is a long gap where nothing happens. It becomes completely unresponsive." That was a deadlock, not slowness. The tui3 wire reader called a blocking
p.SendwhileUpdatewaited synchronously onResolveQuestion's reply — which the same reader had to deliver. Only the 10-second call deadline broke it, which is exactly the pause you felt. Reproduced deterministically at 10.000362601s; after the fix, 181µs. Fixed with a non-blocking coalescing door (doorbell.go,offloop.go) and two law tests that fail the build if any wire door is ever called fromUpdateagain. #919."the pink text… not ideal. Maybe with the border we kind of show the question there hanging." The question is a framed panel above the box now, foldable to one titled rule with
escand reopened withspace. The violet is retired from both ladders: colour is stroke, never fill — amber on the three marks (?, the pointer, the recommendation) and ordinary ink on every word. A law test walks the renderers to hold it. #933."turning it into a full page is also not ideal. Maybe left right part." The page is two panes: answers on the left, and on the right the case for whichever answer the pointer is on — what it means,
then ·,why this one ·,would switch if,confidence ·, and anything it drew. Under ~100 columns the same lines unfold under the row instead. The panel never takes more than half the screen. #972."we need to see what is the recommended option… it can say recommended or be highlighted cleverly."
◆ recommendedon the row, andenterstarts there — except on a permission, where it starts on the answer that loses nothing.The five you approved afterwards
All five landed: the scope row (
tcycles how long an answer holds), permissions from one batch as one frame, the undo window, the visible countdown (opt-in through/autonomy, per your ruling), and the engine continuing to work while a question waits.Two things found on the way that you did not ask for
A key you typed into the connect box was being written to disk. An OpenRouter key reached
decisions.jsonl, the questions lane,--hostframes andmessage[0]. Fixed at one place,withoutSecretWords. #910.The answered question that kept coming back — the one you sent the screenshot of — was
internal/remote's held set being cleared only by the old per-lane doors, so the unified resolve never cleared it and every attach re-handed the answered card. Now driven by the question's own lifecycle (heldSet.keepOnly). #910.What the wave left behind
approval.AlwaysAsks, soentermeansallow onceon an ordinary call anddenyon a grave one — the second half of your pointer ruling, which Questions: one frame, one chooser, one key grammar — and the amber on the marks alone #933 could not ship because nothing upstream graded anything.scripts/one-suite.shonly guards runs that go through the Makefile; a hand-typedgo test ./internal/session/bypasses it entirely, so a "locked" run is not alone. Evidence and a suggested fix are on that issue.PRs: #910, #919, #933, #946, #954, #960, #963, #972, #973, #976, #985.