Skip to content

Node-based terminal manager for AI coding agents — tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition.

Quality-of-life improvements for Chat question card (#979)

Paul Robertsonc8c0cca7cfb0
diff --git a/docs/hook-reply-approvals.md b/docs/hook-reply-approvals.md
index 51f26fc09..3ee77e9b5 100644
--- a/docs/hook-reply-approvals.md
+++ b/docs/hook-reply-approvals.md
@@ -158,7 +158,11 @@ same order. That is why a held question also carries its texts (`held.questions`
before edits", "Revise…" (feedback → `plan-revise`). Question: radio (single) / checkbox (multiSelect),
an "Other" text field (the only input on a question with no options), Submit once every question is
answered; a multi-select "Other" joins the ticked labels and the text with `", "` as one free-text
-answer. The ticket is re-checked against the store at send time; `false` (or a rejection) shows
+answer, the text quoted (inner quotes escaped) when it contains a comma or a quote — Claude Code's own
+picker format, MEASURED on 2.1.283 (`Red, Blue, "teal, sort of"`, `Red, teal`), so the model can tell
+the user's words from the labels. "Chat about this" (`question-clarify`) declines the question: a deny
+whose message is the native clarify text, built by core from the pending file's questions; no
+`interrupt`, so Claude asks what to clarify and stops at the prompt. The ticket is re-checked against the store at send time; `false` (or a rejection) shows
"Couldn't send — answer in the terminal (⌘M)" and leaves the controls usable — never a stuck "Sent".
While controls are up, the composer placeholder and the status row point at the card
(`chatComposerPlaceholder({answerOnCard})`). The kanban card modal mounts the same `ChatPanel`, so the
diff --git a/src/core/agents/permission-decision.test.ts b/src/core/agents/permission-decision.test.ts
index 4820d2139..3101eae7f 100644
--- a/src/core/agents/permission-decision.test.ts
+++ b/src/core/agents/permission-decision.test.ts
@@ -6,6 +6,7 @@ import {
labelHeldForRevision,
MIN_STRUCTURED_ANSWER_REVISION,
buildPermissionDecision,
+ clarifyMessage,
isBoundedAnswerContent,
parsePermissionAnswer,
parsePendingRequest,
@@ -183,6 +184,31 @@ const SURFACES = 'Which surfaces should get it?'
const ANSWER_1 = { [QUESTION_TEXT]: 'Kim kimi okuyabilir (context)' }
const ANSWER_2 = { [SURFACES]: ['Desktop'] }
+describe('buildPermissionDecision — question-clarify ("Chat about this")', () => {
+ it('parses from untrusted input', () => {
+ expect(parsePermissionAnswer({ kind: 'question-clarify' })).toEqual({ kind: 'question-clarify' })
+ })
+
+ it('is a deny carrying Claude Code\'s own clarify message, listing the HELD request\'s questions', () => {
+ const r = buildPermissionDecision(QUESTION_REQ, { kind: 'question-clarify' })
+ if (!r.ok) throw new Error(r.reason)
+
+ const d = decisionOf(r.content)
+ expect(r.decision).toBe('deny')
+ expect(d.behavior).toBe('deny')
+ expect(d.message).toBe(clarifyMessage([QUESTION_TEXT, 'Which surfaces should get it?']))
+ expect(d.message).toContain('Start by asking them what they would like to clarify.')
+ expect(d.message).toContain(`- "${QUESTION_TEXT}"\n (No answer provided)`)
+ // The turn continues: Claude asks what to clarify, then stops at the prompt.
+ expect(d).not.toHaveProperty('interrupt')
+ })
+
+ it('refuses a hold that is not a question', () => {
+ expect(buildPermissionDecision(PLAN_REQ, { kind: 'question-clarify' }).ok).toBe(false)
+ expect(buildPermissionDecision(null, { kind: 'question-clarify' }).ok).toBe(false)
+ })
+})
+
describe('buildPermissionDecision — question', () => {
it('echoes the PENDING request\'s questions verbatim and adds answers keyed by the exact question text', () => {
const r = buildPermissionDecision(QUESTION_REQ, {
diff --git a/src/core/agents/permission-decision.ts b/src/core/agents/permission-decision.ts
index 4951ceab8..4c4edc03a 100644
--- a/src/core/agents/permission-decision.ts
+++ b/src/core/agents/permission-decision.ts
@@ -109,6 +109,8 @@ export function parsePermissionAnswer(v: unknown): PermissionAnswer | null {
: null
case 'plan-revise':
return typeof v.message === 'string' ? { kind: 'plan-revise', message: v.message } : null
+ case 'question-clarify':
+ return { kind: 'question-clarify' }
case 'question': {
if (!isPlainObject(v.answers)) return null
const answers: Record<string, string | string[]> = Object.create(null)
@@ -160,6 +162,23 @@ function questionIndex(toolInput: Record<string, unknown>): Map<string, Question
return index
}
+/**
+ * What Claude Code tells the model when the user picks "Chat about this" in its own question picker
+ * — MEASURED verbatim on 2.1.283 (the tool result after the native choice), minus the template's
+ * stray indentation. The question texts come from the PENDING FILE, never from the renderer.
+ */
+export function clarifyMessage(questions: readonly string[]): string {
+ return [
+ 'The user wants to clarify these questions.',
+ 'This means they may have additional information, context or questions for you.',
+ 'Take their response into account and then reformulate the questions if appropriate.',
+ 'Start by asking them what they would like to clarify.',
+ '',
+ 'Questions asked:',
+ ...questions.map((q) => `- "${q}"\n (No answer provided)`)
+ ].join('\n')
+}
+
function buildQuestionDecision(
pending: PendingRequest,
answers: Record<string, string | string[]>,
@@ -252,6 +271,14 @@ export function buildPermissionDecision(pending: PendingRequest | null, answer:
case 'question':
if (pending?.toolName !== ASK_USER_QUESTION_TOOL) return { ok: false, reason: 'the held request is not a question' }
return buildQuestionDecision(pending, answer.answers, answer.freeText)
+ case 'question-clarify': {
+ if (pending?.toolName !== ASK_USER_QUESTION_TOOL) return { ok: false, reason: 'the held request is not a question' }
+ const index = questionIndex(pending.toolInput)
+ if (!index) return { ok: false, reason: 'the held question has no readable questions' }
+ // No `interrupt`: the turn continues and Claude asks what the user wants to clarify, then
+ // stops at the prompt — exactly where the native "Chat about this" leaves it.
+ return capped(decisionJson({ behavior: 'deny', message: clarifyMessage([...index.keys()]) }), 'deny')
+ }
default:
return { ok: false, reason: 'unknown answer' }
}
diff --git a/src/renderer/lib/chatAnswer.test.ts b/src/renderer/lib/chatAnswer.test.ts
index 9802a3ce2..e99f73e70 100644
--- a/src/renderer/lib/chatAnswer.test.ts
+++ b/src/renderer/lib/chatAnswer.test.ts
@@ -16,6 +16,7 @@ import {
answerTooLong,
planReviseAnswer,
questionAnswerFrom,
+ quoteCustomItem,
toggleLabel
} from './chatAnswer'
@@ -116,6 +117,31 @@ describe('questionAnswerFrom — UI selection → the structured answer core val
freeText: ['Which surfaces?']
})
})
+ it('quotes a typed "Other" containing a comma, so the model can tell it from the labels', () => {
+ const sel = emptySelection([MULTI])
+ sel[0].labels = ['Desktop', 'Phone']
+ sel[0].other = true
+ sel[0].otherText = 'teal, sort of'
+
+ expect(questionAnswerFrom([MULTI], sel)?.answers['Which surfaces?']).toBe('Desktop, Phone, "teal, sort of"')
+ })
+
+ it('quotes a lone typed "Other" on a multi choice by the same rule', () => {
+ const sel = emptySelection([MULTI])
+ sel[0].other = true
+ sel[0].otherText = 'a, b'
+
+ expect(questionAnswerFrom([MULTI], sel)?.answers['Which surfaces?']).toBe('"a, b"')
+ })
+
+ it('never quotes a single choice\'s "Other": it is the whole answer', () => {
+ const sel = emptySelection([SINGLE])
+ sel[0].other = true
+ sel[0].otherText = 'neither, really'
+
+ expect(questionAnswerFrom([SINGLE], sel)?.answers['Pick one?']).toBe('neither, really')
+ })
+
it('a question with no options is answered by its text field alone', () => {
const sel = emptySelection([FREE])
sel[0].otherText = 'Ada'
@@ -271,3 +297,11 @@ describe('rebindRetryDelay', () => {
])
})
})
+
+describe('quoteCustomItem — Claude Code\'s own multi-select quoting (measured, 2.1.283)', () => {
+ it('matches the native picker on every measured answer', () => {
+ expect(quoteCustomItem('teal, sort of')).toBe('"teal, sort of"')
+ expect(quoteCustomItem('say "hi"')).toBe('"say \\"hi\\""')
+ expect(quoteCustomItem('teal')).toBe('teal')
+ })
+})
diff --git a/src/renderer/lib/chatAnswer.ts b/src/renderer/lib/chatAnswer.ts
index 995efaf98..57db02157 100644
--- a/src/renderer/lib/chatAnswer.ts
+++ b/src/renderer/lib/chatAnswer.ts
@@ -167,18 +167,30 @@ export function toggleLabel(labels: readonly string[], label: string, on: boolea
return on ? [...rest, label] : rest
}
+/**
+ * The typed "Other" as one item of a multi-select answer, quoted the way Claude Code's own picker
+ * quotes it — MEASURED on 2.1.283: `Red, Blue, "teal, sort of"`, `Green, "say \"hi\""`, but
+ * `Red, teal`. Quoted (with the inner quotes escaped) only when it contains a comma or a quote, so
+ * the model can always tell where the user's own words start and end in the joined list.
+ */
+export function quoteCustomItem(text: string): string {
+ return /[,"]/.test(text) ? JSON.stringify(text) : text
+}
+
/**
* The free text a question's answer would carry, or null when it is not a free-text answer. ONE
* definition, so the text that is capped is the text that is sent: on a multi choice with "Other" it
- * is the ticked labels (in option order) plus the typed text, joined with ", " — the TUI's own
- * multi-select format, and one string because core carries one free-text answer per question.
- * Core caps THAT string, so capping only the typed part let a long answer through to a refusal.
+ * is the ticked labels (in option order) plus the typed text (`quoteCustomItem`), joined with ", " —
+ * the TUI's own multi-select format, and one string because core carries one free-text answer per
+ * question. Core caps THAT string, so capping only the typed part let a long answer through to a
+ * refusal. A single choice's "Other" is the typed text as-is: it is the whole answer.
*/
function freeTextOf(q: ChatQuestion, sel: QuestionSelection): string | null {
if (!sel.other && q.options.length > 0) return null
const typed = sel.otherText.trim()
+ if (!q.multiSelect || typed === '') return typed
const labels = q.options.map((o) => o.label).filter((l) => sel.labels.includes(l))
- return q.multiSelect && labels.length && typed ? [...labels, typed].join(', ') : typed
+ return [...labels, quoteCustomItem(typed)].join(', ')
}
/** The first question whose free-text answer is over the shared cap (core would refuse it), or null.
diff --git a/src/renderer/nodes/ChatAnswerControls.tsx b/src/renderer/nodes/ChatAnswerControls.tsx
index d19500cd3..e82072005 100644
--- a/src/renderer/nodes/ChatAnswerControls.tsx
+++ b/src/renderer/nodes/ChatAnswerControls.tsx
@@ -256,6 +256,19 @@ export function QuestionAnswerControls({ questions, onSubmit, agentLabel, chip }
>
Submit
</button>
+ {/* Claude Code's own escape hatch: decline the question and talk it over instead — Claude
+ asks what you want to clarify and waits at the prompt, where the composer takes over. */}
+ <button
+ type="button"
+ className="term-chat__answer-btn"
+ aria-disabled={busy}
+ title={`Decline the question — ${agentLabel} will ask what you'd like to clarify`}
+ onClick={() => {
+ if (!busy) void submit({ kind: 'question-clarify' })
+ }}
+ >
+ Chat about this
+ </button>
</div>
{tooLong && (
// The only incomplete state that is not visible on its own: the ticked labels plus the typed
diff --git a/src/renderer/nodes/ChatPanel.answers.test.tsx b/src/renderer/nodes/ChatPanel.answers.test.tsx
index 8c1ca6ddc..7040a41fc 100644
--- a/src/renderer/nodes/ChatPanel.answers.test.tsx
+++ b/src/renderer/nodes/ChatPanel.answers.test.tsx
@@ -240,6 +240,17 @@ describe('question card', () => {
expect(sent().answer).toEqual({ kind: 'question', answers: { 'Pick one?': 'neither' }, freeText: ['Pick one?'] })
})
+ it('"Chat about this" declines the question without picking anything — available at once', async () => {
+ await hold('waiting', held(['Pick one?']))
+ await mount([askMsg(0, [SINGLE])])
+ const chat = button(cards()[0], 'Chat about this')
+ expect(off(chat)).toBe(false)
+
+ await click(chat)
+
+ expect(sent()).toEqual({ nodeId: NODE, pendingId: 'n-q-1', answer: { kind: 'question-clarify' } })
+ })
+
it('several questions: Submit waits for all of them', async () => {
await hold('waiting', held(['Pick one?', 'Which surfaces?']))
await mount([askMsg(0, [SINGLE, MULTI])])
diff --git a/src/shared/agents/permission-answer.ts b/src/shared/agents/permission-answer.ts
index 136e53ec2..133295e74 100644
--- a/src/shared/agents/permission-answer.ts
+++ b/src/shared/agents/permission-answer.ts
@@ -33,6 +33,8 @@ export const INTERACTIVE_HOLD_TOOLS: readonly string[] = [EXIT_PLAN_MODE_TOOL, A
* option label, a string[] is several labels of a multiSelect question. A key listed in
* `freeText` carries typed text (the picker's "Other") instead of a label — that is the ONLY way
* an answer may be something other than an existing label.
+ * - `question-clarify` — "Chat about this": decline the question so the user can talk it over. A
+ * deny whose message is Claude Code's own for that choice, built by core from the held request.
*/
export type PermissionAnswer =
| { kind: 'allow' }
@@ -40,6 +42,7 @@ export type PermissionAnswer =
| { kind: 'plan'; mode: 'restore' | 'acceptEdits' | 'manual' }
| { kind: 'plan-revise'; message: string }
| { kind: 'question'; answers: Record<string, string | string[]>; freeText?: string[] }
+ | { kind: 'question-clarify' }
/** Typed text in an answer ("Revise…" feedback, a question's "Other"), in UTF-16 code units. ONE
* definition: core refuses past it and the answer fields stop at it, so a field can never accept