commandmate 0.34.1 → 0.36.0

package/dist/server/src/lib/detection/selection-shape.js

@@ -66,6 +66,8 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.DISMISSABLE_PANEL_FOOTER_PATTERN = exports.COMMAND_CODE_SELECTION_LIST_FOOTER = exports.SET_AS_DEFAULT_FOOTER_PATTERN = exports.SESSION_SCOPE_FOOTER_PATTERN = exports.MAX_OPTION_NUMBER = exports.SELECTION_SHAPE_TAIL_LINE_COUNT = void 0;
 exports.hasDismissablePanelFooter = hasDismissablePanelFooter;
+exports.readCommandCodeQuestionRegion = readCommandCodeQuestionRegion;
+exports.hasCommandCodeQuestionChrome = hasCommandCodeQuestionChrome;
 exports.extractCommandCodeSelectionListFrame = extractCommandCodeSelectionListFrame;
 exports.readSelectionListShape = readSelectionListShape;
 exports.shouldOfferOptionNumbers = shouldOfferOptionNumbers;
@@ -249,6 +251,344 @@ const COMMAND_CODE_RULE_ROW_PATTERN = /^\u2500+$/;
  * `tests/fixtures/`.
  */
 const COMMAND_CODE_RULE_MIN_COLUMNS = 40;
+/**
+ * The cursor glyph Command Code draws against the highlighted option.
+ *
+ * U+276F, measured on the capture Issue #2521 was raised from. Only this one,
+ * and not codex's `›` or gemini's `●`: {@link readCommandCodeQuestionRegion}
+ * counts cursors to decide whether the region holds ONE live dialog, and a
+ * wider alphabet would start counting glyphs that are not cursors at all.
+ */
+const COMMAND_CODE_CURSOR_GLYPH = '❯';
+/**
+ * The markers Command Code paints on a question screen's tab strip.
+ *
+ * Measured row: `● Dispatch | ◯ Review` — a filled disc on the tab that has the
+ * screen and a hollow ring on the one that does not. The names are NOT part of
+ * the reading (a different question draws different tabs); what is read is the
+ * strip's structure, which is what {@link isCommandCodeQuestionTabRow} states.
+ *
+ * Two families rather than one alphabet, because "one of these is filled and one
+ * of these is not" is the whole signal. A list of bullets (`● item one`) carries
+ * the first family and nothing else, and a bullet list is the thing this
+ * reading must never mistake for a dialog.
+ */
+const COMMAND_CODE_TAB_SELECTED_MARKERS = '●◉⦿';
+/** The hollow half of {@link COMMAND_CODE_TAB_SELECTED_MARKERS}'s pair. */
+const COMMAND_CODE_TAB_UNSELECTED_MARKERS = '◯○◌⚪';
+/** One `<marker> <label>` cell of the tab strip. */
+const COMMAND_CODE_TAB_SEGMENT_PATTERN = new RegExp(`^\\s*([${COMMAND_CODE_TAB_SELECTED_MARKERS}${COMMAND_CODE_TAB_UNSELECTED_MARKERS}])\\s+\\S`);
+/**
+ * Is this row Command Code's question-screen tab strip?
+ *
+ * Three conditions, and each one is a bullet list this predicate has to refuse:
+ *
+ *  - **two or more `|`-separated cells.** `● Dispatch` on its own is a bullet;
+ *    `● Dispatch | ◯ Review` is a strip. The separator is the cheapest part of
+ *    the structure and the one prose never has in this position;
+ *  - **every cell is `<marker> <label>`.** A row that is a strip for its first
+ *    cell and prose for its second is prose;
+ *  - **at least one filled marker and at least one hollow one.** A tab strip
+ *    says which tab has the screen. A row of identical bullets does not, and
+ *    `● one | ● two` is exactly the list-with-a-pipe this rules out.
+ */
+function isCommandCodeQuestionTabRow(line) {
+    const segments = line.split('|');
+    if (segments.length < 2)
+        return false;
+    let selected = 0;
+    let unselected = 0;
+    for (const segment of segments) {
+        const match = COMMAND_CODE_TAB_SEGMENT_PATTERN.exec(segment);
+        if (match === null)
+            return false;
+        if (COMMAND_CODE_TAB_SELECTED_MARKERS.includes(match[1]))
+            selected += 1;
+        else
+            unselected += 1;
+    }
+    return selected > 0 && unselected > 0;
+}
+/**
+ * Rows that mean the region spans more than the one dialog on screen now.
+ *
+ * Issue #2521's condition 4: the tab strip and its options have to be the
+ * SAME screen, with no other turn's generating or completion UI between them.
+ * Structurally that is mostly settled by taking the LAST rule row as the top
+ * edge — a new turn repaints the composer, and the composer's own lower rule
+ * then becomes that edge — so this pattern is the belt to that braces, for a
+ * frame captured mid-repaint.
+ *
+ * Two rows, both restated here rather than imported: `cli-patterns` owns
+ * Command Code's busy and marker vocabularies and pulls the logger and the tool
+ * registry in with them, and this module is the browser-safe leaf the chat
+ * surface imports (see the module docblock).
+ *
+ *  - `esc to interrupt`, the tail Command Code appends to every status row
+ *    (`COMMAND_CODE_INTERRUPT_HINT_PATTERN`). A turn is in flight;
+ *  - a row that OPENS with U+273B and a word — `✻ Thought for 1 second`,
+ *    `✻ Worked for 4s`. Matched by shape rather than by the two English verbs,
+ *    so a reworded or localised marker still ends the region.
+ */
+const COMMAND_CODE_QUESTION_REGION_REJECT_PATTERN = /\besc\s+to\s+interrupt\b|^[^\S\n]*✻[^\S\n]+\S/im;
+/**
+ * Where a footer-less Command Code question screen sits on a pane (Issue #2521).
+ *
+ * ## The frame this exists for
+ *
+ * Command Code's `AskUserQuestion` draws no hint-bar footer. It draws a rule,
+ * a tab strip, the question, and the options — and when an option's description
+ * WRAPS, the continuation row begins with a single space:
+ *
+ *     ────────────────────────────… (200 columns of U+2500)
+ *     ● Dispatch | ◯ Review
+ *
+ *     Approve proceeding from the plan into worktree creation and dispatch?
+ *
+ *     ❯ 1. Prepare worktrees + dispatch (Recommended)
+ *          I create the worktrees and pause for you to
+ *      answer).
+ *       2. Worktrees only, then pause
+ *     …
+ *
+ * Measured 2026-09-12 on a pane reported as Command Code 1.53.0 at the
+ * production 200x1000; the anonymised capture is
+ * `tests/fixtures/command-code-askuserquestion-2521/`. Nothing read it:
+ * `detectPrompt` stops its scan at ` answer).` and never reaches option 1, and
+ * the selection-list branch below wants a footer this screen does not draw. So
+ * the frame fell through to the generic composer check — which matches, because
+ * `COMMAND_CODE_PROMPT_PATTERN` is `^❯(\s*$|\s+\S)` and `❯ 1. Prepare …` is a
+ * `❯` followed by a space and a glyph — and the pane was published as
+ * `ready` / `input_prompt`. `wait` read that as a finished turn and exited 0 on
+ * a session that was asking a question (Issue #2521's "偽完了").
+ *
+ * ## What this function is NOT
+ *
+ * It is not a general reading of footer-less dialogs, and it deliberately does
+ * not make `optionCount` into a dialog test. {@link readSelectionListShape}
+ * counts the numbers in the tail of ANY frame, an ordinary numbered answer
+ * included, and the whole point of the conditions below is that a number list
+ * on its own proves nothing. Five things have to line up:
+ *
+ *  1. the LAST rule row that qualifies as the seam ({@link
+ *     COMMAND_CODE_RULE_ROW_PATTERN} at {@link COMMAND_CODE_RULE_MIN_COLUMNS} or
+ *     wider) is the top edge, and the last row carrying content is the bottom.
+ *     Taking the LAST one is what keeps an older dialog higher up the pane out:
+ *     a new turn repaints the composer, whose own rules sit below the old
+ *     dialog, and the region is then the composer's hint row;
+ *  2. the first row under that rule is a tab strip
+ *     ({@link isCommandCodeQuestionTabRow}) and a non-blank question body
+ *     follows it before the first option;
+ *  3. the options are a STRICT run `1.` … `N.`, 2 ≤ N ≤ {@link
+ *     MAX_OPTION_NUMBER}. A gap, a repeat or a list that starts at `2` is not
+ *     this screen;
+ *  4. exactly one {@link COMMAND_CODE_CURSOR_GLYPH} in the region, and it is on
+ *     an option row. Two cursors means two screens — typically a composer's own
+ *     `❯` under the dialog — and none means the highlight is elsewhere;
+ *  5. no filter box, no `/model` footer, no dismiss-only footer and no other
+ *     turn's UI in the region ({@link FILTER_INPUT_PATTERN},
+ *     {@link COMMAND_CODE_SELECTION_LIST_FOOTER},
+ *     {@link DISMISSABLE_PANEL_FOOTER_PATTERN},
+ *     {@link COMMAND_CODE_QUESTION_REGION_REJECT_PATTERN}). `Type something...`
+ *     is an OPTION on this screen, not a filter, and the filter pattern is
+ *     narrow enough to tell them apart.
+ *
+ * `null` for anything else, which leaves every existing verdict and every
+ * existing crop exactly as they were.
+ *
+ * ## Tool-id-free and browser-safe, on purpose
+ *
+ * Three call sites read the same answer: the command-code detector's
+ * `afterPrompt`, {@link extractCommandCodeSelectionListFrame} (which runs for
+ * every CLI, with no tool id in hand) and `ChatSurface`, a client component.
+ * So this takes a raw frame and nothing else, and imports nothing but
+ * `./ansi` — the same trade the rest of this module makes.
+ *
+ * ## The one spelling it cannot answer on
+ *
+ * A capture that has been through `stripBoxDrawing` — which blanks a pure-U+2500
+ * row, so condition 1 finds no seam and this returns `null`. Same limitation
+ * `extractOpenCodeModalOverlayFrame` records for the same reason. Issue #2522
+ * did not weaken the reading to work around it: the three consumers that clean
+ * their frames (the response poller, Auto-Yes, `/prompt-response`) now keep the
+ * same tick's RAW capture and hand that here, so both spellings of one capture
+ * exist and the right one is passed.
+ *
+ * ## Issue #2522: the structural half is shared
+ *
+ * Conditions 1, 2, 3-minus-strictness and 5 moved into
+ * {@link scanCommandCodeQuestionScreen}, which {@link hasCommandCodeQuestionChrome}
+ * also reads. This function is exactly that scan plus the two conditions that
+ * are about the LIST — a strict run and a single cursor — so every verdict it
+ * gave before is the verdict it gives now.
+ *
+ * @param frame - a raw `capture-pane -p -e` frame, ANSI intact (CRLF tolerated)
+ * @returns the region's line indices into the LF-normalised frame, or `null`
+ */
+function readCommandCodeQuestionRegion(frame) {
+    const screen = scanCommandCodeQuestionScreen(frame);
+    if (screen === null)
+        return null;
+    // 3 (continued). A STRICT run. The numbers were collected in the order they
+    // are drawn and are compared against the run, rather than restarted on every
+    // fresh `1` the way `countTrailingOptionRun` does: that function is reading a
+    // tail that may hold two lists, and this one is asserting that the region
+    // holds exactly one.
+    if (!screen.numbers.every((value, index) => value === index + 1))
+        return null;
+    // 4 (continued). Exactly ONE cursor. Two means two screens — typically a
+    // composer's own `❯` under the dialog, or a frame caught mid-repaint.
+    if (screen.cursorCount !== 1)
+        return null;
+    return {
+        ruleLineIndex: screen.ruleLineIndex,
+        firstLineIndex: screen.firstLineIndex,
+        lastLineIndex: screen.lastLineIndex,
+        tabLineIndex: screen.tabLineIndex,
+        cursorLineIndex: screen.cursorLineIndex,
+        optionCount: screen.numbers.length,
+    };
+}
+/**
+ * Is Command Code's question CHROME on this frame, whatever its numbering says?
+ * (Issue #2522)
+ *
+ * The same five conditions {@link readCommandCodeQuestionRegion} applies, minus
+ * the two that are about the LIST rather than about the screen: the run need not
+ * be a strict `1.`…`N.`, and there may be more than one cursor on it.
+ *
+ * ## Why the weaker reading is worth having
+ *
+ * Issue #2522 確定仕様 B needs three answers where #2521 had two. "Not this
+ * screen" keeps every existing verdict; "read in full" produces an answerable
+ * prompt; and **"this screen, unreadable"** — a gap in the numbering, a repeated
+ * row, a list that starts at `2`, a second cursor from a half-finished repaint —
+ * has to reach the manual-operation fallback rather than the `ready` the generic
+ * composer check would otherwise publish off the dialog's own `❯` row. That is
+ * the very failure both Issues exist for: `wait` exiting 0 on a pane that is
+ * asking a human a question.
+ *
+ * It is deliberately NOT a relaxation of the strict reading. Nothing that was
+ * `null` becomes a REGION, so no partial option list is produced anywhere and no
+ * existing crop, card or prompt payload changes. What this adds is the ability
+ * to say "declined for a numbering reason" out loud.
+ *
+ * ## Why it is still narrow
+ *
+ * Every structural condition stays: the last qualifying rule row, a genuine tab
+ * strip as the first row under it ({@link isCommandCodeQuestionTabRow} — one
+ * filled marker, one hollow, `|`-separated), a non-blank question body before
+ * the first option, at least two numbered rows, at least one `❯` ON one of them,
+ * and none of the pickers, panels or other-turn UI the other branches own. An
+ * assistant answering in a numbered list carries none of that, and a frame with
+ * one option or with the `❯` on a composer row is still `false` — the two cases
+ * #2521 recorded as "the highlight is elsewhere".
+ *
+ * @param frame - a raw `capture-pane -p -e` frame, ANSI intact (CRLF tolerated)
+ */
+function hasCommandCodeQuestionChrome(frame) {
+    return scanCommandCodeQuestionScreen(frame) !== null;
+}
+/**
+ * The structural half of the question-screen reading, shared by the strict
+ * {@link readCommandCodeQuestionRegion} and the weaker
+ * {@link hasCommandCodeQuestionChrome} (Issue #2522 split them apart).
+ *
+ * Written once rather than twice on purpose: the conditions below are what make
+ * the reading safe for every other CLI's frames, and a second copy of them is a
+ * second thing to forget to narrow.
+ */
+function scanCommandCodeQuestionScreen(frame) {
+    if (!frame)
+        return null;
+    // Stripped per line off the same split, so every index below is valid against
+    // the ANSI-bearing rows the cropper slices.
+    const lines = frame.replace(/\r\n/g, '\n').split('\n').map(ansi_1.stripAnsi);
+    let last = lines.length - 1;
+    while (last >= 0 && lines[last].trim() === '')
+        last -= 1;
+    // A region needs a rule above it, so the last content row cannot be row 0.
+    if (last < 1)
+        return null;
+    let ruleLineIndex = -1;
+    for (let i = last - 1; i >= 0; i -= 1) {
+        const row = lines[i].trim();
+        if (row.length < COMMAND_CODE_RULE_MIN_COLUMNS)
+            continue;
+        if (!COMMAND_CODE_RULE_ROW_PATTERN.test(row))
+            continue;
+        ruleLineIndex = i;
+        break;
+    }
+    if (ruleLineIndex < 0)
+        return null;
+    const region = lines.slice(ruleLineIndex + 1, last + 1);
+    const body = region.join('\n');
+    // 5. The screens this reading must not take over.
+    if (FILTER_INPUT_PATTERN.test(body))
+        return null;
+    if (exports.COMMAND_CODE_SELECTION_LIST_FOOTER.test(body))
+        return null;
+    if (exports.DISMISSABLE_PANEL_FOOTER_PATTERN.test(body))
+        return null;
+    if (COMMAND_CODE_QUESTION_REGION_REJECT_PATTERN.test(body))
+        return null;
+    // 2. The tab strip opens the region.
+    const tabOffset = region.findIndex((line) => line.trim() !== '');
+    if (tabOffset < 0)
+        return null;
+    if (!isCommandCodeQuestionTabRow(region[tabOffset]))
+        return null;
+    // 3. Numbered rows under it.
+    const numbers = [];
+    let firstOptionOffset = -1;
+    for (let i = tabOffset + 1; i < region.length; i += 1) {
+        const match = OPTION_ROW_PATTERN.exec(region[i]);
+        if (match === null)
+            continue;
+        if (firstOptionOffset < 0)
+            firstOptionOffset = i;
+        numbers.push(Number(match[1]));
+    }
+    // Two is the floor: one numbered row under a tab strip is not a choice. The
+    // ceiling is {@link MAX_OPTION_NUMBER} and it is {@link OPTION_ROW_PATTERN}'s
+    // rather than a length check — that pattern captures a SINGLE digit, so a
+    // `10.` row is not read as an option at all and an eleven-option screen
+    // reports its first nine. Recognising such a screen and under-counting it is
+    // the right failure: declining it would put the pane back on the `ready` that
+    // made `wait` exit 0. (Issue #2522's reader cross-checks the count against its
+    // own block reading and declines to ANSWER such a screen; what it must not do
+    // is hand the pane back to the composer check.)
+    if (numbers.length < 2)
+        return null;
+    // 2 (continued). The question itself. A tab strip sitting straight on top of
+    // its options is some other screen.
+    if (!region.slice(tabOffset + 1, firstOptionOffset).some((line) => line.trim() !== '')) {
+        return null;
+    }
+    // 4. A cursor, on an option row.
+    let cursorCount = 0;
+    let cursorOffset = -1;
+    for (let i = 0; i < region.length; i += 1) {
+        const inRow = region[i].split(COMMAND_CODE_CURSOR_GLYPH).length - 1;
+        if (inRow === 0)
+            continue;
+        cursorCount += inRow;
+        if (OPTION_ROW_PATTERN.test(region[i]))
+            cursorOffset = i;
+    }
+    if (cursorCount === 0 || cursorOffset < 0)
+        return null;
+    return {
+        ruleLineIndex,
+        firstLineIndex: ruleLineIndex + 1,
+        lastLineIndex: last,
+        tabLineIndex: ruleLineIndex + 1 + tabOffset,
+        numbers,
+        cursorCount,
+        cursorLineIndex: ruleLineIndex + 1 + cursorOffset,
+    };
+}
 /**
  * The rows of a Command Code dialog, cut out of the pane it is painted on
  * (Issue #2326).
@@ -290,7 +630,11 @@ const COMMAND_CODE_RULE_MIN_COLUMNS = 40;
  *  - **no footer** — the picker was closed between the flags being read and the
  *    pane being captured, or the frame is some other dialog. Cropping on a
  *    guess would blank the card, which Issue #2326 calls out as worse than
- *    showing too much;
+ *    showing too much. Issue #2521 carved ONE screen out of this: a frame that
+ *    {@link readCommandCodeQuestionRegion} positively recognises is cropped to
+ *    that region, because `AskUserQuestion` draws no footer at all and its card
+ *    was otherwise the whole pane. The carve-out is a positive reading, not a
+ *    relaxation — everything it declines still lands here;
  *  - **a footer but no rule above it** — copilot's `/model` footer is
  *    `↑/↓ to navigate · … · enter to select · esc to cancel` and matches the
  *    same pattern, but copilot draws its picker in a corner-bordered box, so no
@@ -314,8 +658,22 @@ function extractCommandCodeSelectionListFrame(frame) {
             break;
         }
     }
-    if (footer < 0)
-        return null;
+    // Issue #2521: no footer is no longer unconditionally "do not crop". One
+    // screen draws none — `AskUserQuestion` — and it is the screen whose
+    // uncropped card was 409 rows of transcript with the question below the fold.
+    // The reading is the narrow one, so every other footer-less frame (a closed
+    // picker, a permission dialog, another CLI's overlay) still takes the `null`
+    // path and is returned untouched. Checked only when the footer is absent, so
+    // a picker's own crop is decided exactly as #2326 left it.
+    if (footer < 0) {
+        const region = readCommandCodeQuestionRegion(frame);
+        if (region === null)
+            return null;
+        // The rule row itself is the transcript's boundary and not the dialog's
+        // first row — the same edge the footer path takes — and `lastLineIndex` is
+        // the last row with content, so the pane's trailing padding is left out.
+        return lines.slice(region.firstLineIndex, region.lastLineIndex + 1).join('\n');
+    }
     for (let i = footer - 1; i >= 0; i -= 1) {
         const row = (0, ansi_1.stripAnsi)(lines[i]).trim();
         if (row.length < COMMAND_CODE_RULE_MIN_COLUMNS)

package/dist/server/src/lib/detection/tools/command-code/detect.js

@@ -72,6 +72,24 @@
  * box. Measured live on v1.40.1 at the production 200x1000 geometry; the capture
  * is `tests/fixtures/chat-dialog-card-2254/command-code-model-1-40-1.txt`.
  *
+ * `beforePrompt` arrived with Issue #2521 as a third `afterPrompt` branch and
+ * was moved ahead of the shared parser by Issue #2522. It is the only branch
+ * here not read off a footer: `AskUserQuestion` draws none. The screen is a
+ * rule, a tab strip (`● Dispatch | ◯ Review`), the question and a strict
+ * `1.`…`N.` run with one `❯` on it, and `readCommandCodeQuestionRegion` is what
+ * recognises that shape. Before #2521, a wrapped option description took the
+ * frame away from the shared parser and the generic composer check answered
+ * `ready` off the dialog's own cursor row — so `wait` exited 0 on a pane that
+ * was asking a human a question.
+ *
+ * #2521 answered that with a manual-operation fallback (the selection-list
+ * reason, no `promptData`). #2522 added `tools/command-code/dialog.ts`, which
+ * READS the same region, and the branch now publishes one of two verdicts:
+ * `PROMPT_DETECTED` with the parsed question, options and `❯` default where the
+ * reading succeeds, and #2521's fallback where the screen is plainly up and the
+ * reading declines. The move to `beforePrompt` is part of that: see the branch
+ * for why the generic parser must not get this frame either before or after.
+ *
  * ## What #2304 re-measured, and what it did not change
  *
  * Nothing here changed, and that is the finding. Seven frames were captured
@@ -104,6 +122,7 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.commandCodeStatusDetector = exports.VERIFIED_AGAINST = void 0;
 const cli_patterns_1 = require("../../cli-patterns");
 const selection_shape_1 = require("../../selection-shape");
+const dialog_1 = require("./dialog");
 const status_reason_1 = require("../../status-reason");
 const run_detection_1 = require("../run-detection");
 const verified_against_1 = require("../verified-against");
@@ -112,6 +131,83 @@ exports.VERIFIED_AGAINST = verified_against_1.COMMAND_CODE_VERIFIED_AGAINST;
 exports.commandCodeStatusDetector = (0, run_detection_1.createToolStatusDetector)({
     tool: 'command-code',
     verifiedAgainst: exports.VERIFIED_AGAINST,
+    beforePrompt(frame) {
+        // Issue #2521 recognised this screen; Issue #2522 reads it.
+        //
+        // `AskUserQuestion` draws no footer, so neither `afterPrompt` branch below
+        // reads it — and the shared `detectPrompt` at step 1 does not either, because
+        // a wrapped option description continues on a row that begins with a SINGLE
+        // space (` answer).`) and the generic multiple-choice parser ends its scan
+        // there, one row short of option 1. The frame therefore reached step 3, where
+        // `COMMAND_CODE_PROMPT_PATTERN` (`^❯(\s*$|\s+\S)`) matched the dialog's own
+        // cursor row, `❯ 1. Prepare worktrees + dispatch (Recommended)`, and the pane
+        // was published as `ready` / `input_prompt`. A human was being asked a
+        // question and `wait` exited 0 on it.
+        //
+        // ## Why `beforePrompt` and not `afterPrompt` (Issue #2522 確定仕様 A / B)
+        //
+        // #2521's branch sat at 1.5, after the shared parser, because it only had to
+        // catch the frames that parser could not read. This one has to run FIRST, and
+        // for two separate reasons:
+        //
+        //  - **the short spelling.** With one-line descriptions the generic parser
+        //    SUCCEEDS on this screen, and succeeds badly: `extractQuestionText` walks
+        //    up to five rows above option 1 and takes the tab strip, the rule and the
+        //    transcript's TODOS row into the question. Placing the reader after it
+        //    would leave exactly the frames that already look fine — and are wrong;
+        //  - **the unreadable spelling.** A question screen with a gap in its
+        //    numbering, or one taller than the reader's cap, must reach #2521's
+        //    manual-operation fallback and NOT the generic parser's partial list.
+        //    That is only expressible ahead of the parser.
+        //
+        // `/usage`, `/model` and the four permission dialogs are unaffected, and not
+        // by ordering: `readCommandCodeQuestionRegion` positively declines every one
+        // of them (the pickers by their own footers, the permission dialogs by having
+        // no tab strip), which is what the fixture sweep in
+        // `command-code-askuserquestion-2521.test.ts` pins.
+        //
+        // Read against `frame.raw` rather than `frame.lastLines` or `frame.clean`:
+        // the region runs from the last rule row to the last content row, which is
+        // more rows than the 15-row window holds, and the rule row itself is what
+        // `stripBoxDrawing` would erase. Every other consumer of this reading keeps
+        // the same tick's raw capture for the same reason (確定仕様 C).
+        const reading = (0, dialog_1.readCommandCodeQuestionDialog)(frame.raw);
+        // Read in full: the question, its options, the `❯` default and the free-text
+        // row are all in hand, so the pane is published exactly as an ordinary
+        // answerable dialog is. `hasActivePrompt: true` is what puts `promptData` on
+        // `/current-output`, raises the answer buttons, lets `respond <id> 2` reach
+        // the pane and lets Auto-Yes consider it — every one of them through the
+        // SAME payload, which is the point of reading it once here.
+        if (reading.kind === 'prompt') {
+            return {
+                status: 'waiting',
+                confidence: 'high',
+                reason: status_reason_1.STATUS_REASON.PROMPT_DETECTED,
+                hasActivePrompt: true,
+                evidence: 'positive',
+                promptDetection: reading.prompt,
+            };
+        }
+        // The question UI is plainly up and this reader could not read it. #2521's
+        // verdict, unchanged and deliberately kept: `COMMAND_CODE_SELECTION_LIST` is
+        // in `SELECTION_LIST_REASONS`, which publishes `isSelectionListActive`,
+        // raises the arrow-driven card and makes `wait` exit 10 instead of polling a
+        // blocked agent to its `--timeout`.
+        //
+        // `hasActivePrompt: false` and no `promptDetection`: nothing parsed the
+        // options, so there is no payload to answer with and none is invented — and
+        // the generic parser is not given a second chance to invent a partial one.
+        if (reading.kind === 'unsupported') {
+            return {
+                status: 'waiting',
+                confidence: 'high',
+                reason: status_reason_1.STATUS_REASON.COMMAND_CODE_SELECTION_LIST,
+                hasActivePrompt: false,
+                evidence: 'positive',
+            };
+        }
+        return null;
+    },
     afterPrompt(frame) {
         // Issue #2369. `/usage` opens a READ-ONLY panel: a plan header, two usage
         // meters, a breakdown URL, and `Press Esc to close` as its last row. It is