commandmate 0.44.1 → 0.44.3

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

@@ -0,0 +1,461 @@
+"use strict";
+/**
+ * Command Code's selection-list / plan-review / question-page patterns and
+ * readers, moved out of `selection-shape.ts` (Issue #3218). `selection-shape.ts`
+ * re-exports every name it used to export from here.
+ *
+ * Imports only `ansi` and `selection-shape-patterns`: the screen-side modules
+ * reach this through `selection-shape.ts`, so nothing server-only may come in,
+ * and importing `selection-shape.ts` back would be a cycle.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.COMMAND_CODE_QUESTION_REGION_REJECT_PATTERN = exports.COMMAND_CODE_CURSOR_GLYPH = exports.COMMAND_CODE_RULE_MIN_COLUMNS = exports.COMMAND_CODE_RULE_ROW_PATTERN = exports.COMMAND_CODE_PLAN_APPROVE_CHOICE_FOOTER = exports.COMMAND_CODE_PLAN_APPROVE_CHOICE_ROW = exports.DISMISSABLE_PANEL_FOOTER_PATTERN = exports.COMMAND_CODE_SELECTION_LIST_FOOTER = void 0;
+exports.isCommandCodePlanApproveChoice = isCommandCodePlanApproveChoice;
+exports.hasCommandCodeSelectionListFooterAtBottom = hasCommandCodeSelectionListFooterAtBottom;
+exports.hasCommandCodeDismissablePanelFooterAtBottom = hasCommandCodeDismissablePanelFooterAtBottom;
+exports.hasCommandCodePlanReviewFooterAtBottom = hasCommandCodePlanReviewFooterAtBottom;
+exports.hasDismissablePanelFooter = hasDismissablePanelFooter;
+exports.isCommandCodeQuestionTabRow = isCommandCodeQuestionTabRow;
+exports.readCommandCodeReviewPage = readCommandCodeReviewPage;
+const ansi_1 = require("../../ansi");
+const selection_shape_patterns_1 = require("../../selection-shape-patterns");
+/**
+ * Command Code's picker footer (v1.40.1, measured at 200x1000 for Issue #2297).
+ *
+ *     type to search · ↑/↓ navigate · shift+↑/↓ jump provider · enter to select · esc to cancel
+ *
+ * Before this Issue no detector branch matched it, so the overlay fell through
+ * to the `default` floor and the chat surface raised its `unclassified` card —
+ * whose controls are the `1`–`9` / `y` / `n` answer keys, every one of which
+ * would have been typed into that `Type to search models...` box. The
+ * command-code detector imports this and answers
+ * `STATUS_REASON.COMMAND_CODE_SELECTION_LIST` instead, which puts the arrow pad
+ * on the card.
+ *
+ * Lower case is what Command Code draws, and the `·` separator keeps the pattern
+ * off ordinary prose — the same two narrowings `COPILOT_SELECTION_FOOTER_PATTERN`
+ * settled on for the same reason.
+ */
+exports.COMMAND_CODE_SELECTION_LIST_FOOTER = /\benter\s+to\s+select\b\s*[·•]\s*esc\s+to\s+cancel\b/i;
+/**
+ * The footer of an overlay whose ONLY exit is the dismiss (Issue #2369).
+ *
+ * Measured on Command Code 1.49's `/usage`, whose last row is verbatim
+ * `Press Esc to close`. The panel above it is read-only — a plan header, two
+ * meters, a breakdown URL — so there is no highlight to move and no key to
+ * confirm with. Before this pattern existed nothing matched the screen, it
+ * reached the `default` floor, and the chat surface answered an
+ * eighteen-button `unclassified` card to a panel with one working key.
+ *
+ * ## Why it is this narrow
+ *
+ * Three narrowings, and each one is there because this predicate SUPPRESSES
+ * controls — a false positive takes the arrow pad and the answer keys away from
+ * a screen that needed them:
+ *
+ *  - **a verb before the key.** `Press`/`Hit`/`Type`, or nothing at all, but not
+ *    an arbitrary run of prose. `esc to cancel` on its own is the tail of every
+ *    picker footer in this file, including
+ *    {@link COMMAND_CODE_SELECTION_LIST_FOOTER}, and those screens have a
+ *    highlight;
+ *  - **`close`/`dismiss`/`exit`, never `cancel` or `go back`.** The distinction
+ *    is measured rather than stylistic: `cancel` is what a screen with a
+ *    COMMITTABLE choice calls its escape (claude's `/model`, codex's picker,
+ *    Command Code's own), while `close` is what a panel that decided nothing
+ *    calls it;
+ *  - **the whole row.** Anchored at both ends (leading/trailing space aside), so
+ *    a hint bar that offers a dismiss ALONGSIDE something else — `↑/↓ navigate ·
+ *    enter to select · esc to close` — does not match, and keeps the arrows the
+ *    other half of its sentence promises.
+ *
+ * Case-insensitive because the measured row is `Press Esc to close` and Command
+ * Code writes its picker hints in lower case; nothing rests on which it uses.
+ */
+exports.DISMISSABLE_PANEL_FOOTER_PATTERN = /^\s*(?:press|hit|type)?\s*(?:<)?esc(?:ape)?(?:>)?\s+to\s+(?:close|dismiss|exit)\b[\s.·•]*$/im;
+/**
+ * The radio row `ctrl+a` puts in place of the action list when comments are
+ * pending (Issue #2793).
+ *
+ * Measured on 1.58.0 (`tests/fixtures/command-code-plan-review-2763/`): with
+ * zero comments `ctrl+a` approves at once, and with one or more it first asks
+ * what to do with them. The three action rows are replaced by ONE row, and the
+ * selected side is `(•)`, the other `( )` — readable without the ANSI:
+ *
+ *     Approve (•) with 4 comments as notes ( ) original plan · discard comments
+ *
+ * `comments?` because the badge above it is measured in both numbers
+ * (`1 pending comment` / `4 pending comments`); the radio itself was captured
+ * with four.
+ */
+exports.COMMAND_CODE_PLAN_APPROVE_CHOICE_ROW = /^\s*Approve\s+\([• ]\)\s+with\s+\d+\s+comments?\s+as\s+notes\s+\([• ]\)\s+original\s+plan\b/im;
+/**
+ * The hint bar under {@link COMMAND_CODE_PLAN_APPROVE_CHOICE_ROW}, verbatim
+ * `←/→ choose · enter confirm · esc back`.
+ *
+ * Anchored to the LAST row of the text it is tested against (no `m` flag: `$` is
+ * the end of the input), because that is where the overlay draws every one of
+ * its five measured hint bars, and because it is what keeps a transcript that
+ * QUOTES this screen from matching — the composer is always drawn below a
+ * transcript, so there the quote is never the last row.
+ */
+exports.COMMAND_CODE_PLAN_APPROVE_CHOICE_FOOTER = /(?:^|\n)[ \t]*←\/→\s+choose\s*·\s*enter\s+confirm\s*·\s*esc\s+back\s*$/i;
+/**
+ * Whether this tail is the approve-with-comments confirmation (Issue #2793).
+ *
+ * BOTH rows, for the reason {@link COMMAND_CODE_PLAN_REVIEW_FOOTER} gives for
+ * its two: either half alone is a sentence something else could print. Not
+ * adjacent — `Editor exited with code 127` and a blank row sit between them on
+ * both captures — so each is tested on its own.
+ *
+ * Deliberately NOT folded into {@link COMMAND_CODE_PLAN_REVIEW_FOOTER}, and
+ * therefore not {@link SelectionListShape.offersPlanApprove}: on this screen
+ * `enter` is the documented CONFIRM of an approval the human already started
+ * with `ctrl+a`, so the chat surface must keep its `Enter`, and what `ctrl+a`
+ * does here was not measured.
+ *
+ * @param tail - the frame's last rows, ANSI stripped, last content row last
+ */
+function isCommandCodePlanApproveChoice(tail) {
+    return (exports.COMMAND_CODE_PLAN_APPROVE_CHOICE_ROW.test(tail) &&
+        exports.COMMAND_CODE_PLAN_APPROVE_CHOICE_FOOTER.test(tail));
+}
+/**
+ * How many content rows the plan review footer's LAST row (`Cancel esc`) may
+ * have under it and still be the overlay's own footer (Issue #2846): the footer's
+ * two rows plus three.
+ *
+ * Measured over the sixteen plan review captures that carry the footer
+ * (`tests/fixtures/command-code-plan-review-2761/`, `-2763/`): fourteen draw ONE
+ * row below it, the hint bar (`type + enter to comment · …`); one draws two, an
+ * `Editor exited with code 127` row and then the hint bar; and the garbled
+ * `editor-killed-staircase.txt` draws three, because the pane wrapped the hint
+ * bar. So the plan review footer is NOT the last row the way the picker's is,
+ * and a "last two rows" rule misses every one of them.
+ *
+ * What bounds it from above is the composer. A transcript is always drawn above
+ * one, and its shortest measured form is four rows — rule, `❯ Ask your
+ * question...`, rule, status row — so a footer QUOTED in a reply has at least
+ * four content rows under it and its `Approve` row falls outside this window.
+ * The margin between the widest overlay (3) and the narrowest composer (4) is
+ * one row; a capture that narrows it is the reason to re-measure, not to widen
+ * this.
+ */
+const COMMAND_CODE_PLAN_REVIEW_TAIL_ROW_COUNT = 5;
+/**
+ * The last `count` rows of `text` that carry content, joined with `\n`.
+ *
+ * Blank rows are dropped rather than counted: the pane is padded with them, and
+ * a real footer is set off from its hint bar by one.
+ */
+function lastContentRows(text, count) {
+    return text
+        .split('\n')
+        .filter((row) => row.trim() !== '')
+        .slice(-count)
+        .join('\n');
+}
+/**
+ * Whether the LAST content row of this tail is Command Code's picker footer
+ * (Issue #2846).
+ *
+ * {@link COMMAND_CODE_SELECTION_LIST_FOOTER} matches the sentence anywhere, and
+ * a reply that quotes the `/model` footer is that sentence in the middle of the
+ * transcript. The picker draws the footer as the last row of the pane (all four
+ * captures that carry it do, `tests/fixtures/chat-dialog-card-2254/`), and a
+ * transcript always has the composer under it — the same reasoning
+ * {@link COMMAND_CODE_PLAN_APPROVE_CHOICE_FOOTER} anchors with `$`.
+ *
+ * @param tail - the frame's last rows, ANSI stripped, last content row last
+ */
+function hasCommandCodeSelectionListFooterAtBottom(tail) {
+    return exports.COMMAND_CODE_SELECTION_LIST_FOOTER.test(lastContentRows(tail, 1));
+}
+/**
+ * Whether the LAST content row of this tail is the dismiss-only panel's footer
+ * (Issue #2846). `Press Esc to close` is the panel's last row
+ * ({@link DISMISSABLE_PANEL_FOOTER_PATTERN}); the same row inside a reply has
+ * the composer under it.
+ *
+ * @param tail - the frame's last rows, ANSI stripped, last content row last
+ */
+function hasCommandCodeDismissablePanelFooterAtBottom(tail) {
+    return exports.DISMISSABLE_PANEL_FOOTER_PATTERN.test(lastContentRows(tail, 1));
+}
+/**
+ * Whether the plan review footer ({@link COMMAND_CODE_PLAN_REVIEW_FOOTER}) sits
+ * at the bottom of this tail (Issue #2846): both of its rows are within the last
+ * {@link COMMAND_CODE_PLAN_REVIEW_TAIL_ROW_COUNT} content rows, which is the
+ * overlay's own hint bar and nothing as tall as a composer.
+ *
+ * @param tail - the frame's last rows, ANSI stripped, last content row last
+ */
+function hasCommandCodePlanReviewFooterAtBottom(tail) {
+    return selection_shape_patterns_1.COMMAND_CODE_PLAN_REVIEW_FOOTER.test(lastContentRows(tail, COMMAND_CODE_PLAN_REVIEW_TAIL_ROW_COUNT));
+}
+/**
+ * How many rows from the end of the content the dismiss footer is looked for.
+ *
+ * The same 15-row tail the detection chain hands a tool module as
+ * `NormalizedFrame.lastLines` (`STATUS_CHECK_LINE_COUNT`), restated as a number
+ * here because this module is a browser-safe leaf and `tools/frame.ts` pulls in
+ * `cli-patterns`. `detect.ts` reads `frame.lastLines` and never this constant,
+ * so the two cannot disagree about the detector's own window; this one bounds
+ * the CLIENT-side reading in {@link hasDismissablePanelFooter}.
+ */
+const DISMISSABLE_PANEL_TAIL_LINE_COUNT = 15;
+/**
+ * Whether the tail of this frame is a dismiss-only panel (Issue #2369).
+ *
+ * The same question `commandCodeStatusDetector.afterPrompt` asks of
+ * `frame.lastLines`, asked of a raw capture, so the chat surface can answer for
+ * a frame whose server verdict has not reached it. The pattern is shared rather
+ * than restated — one expression, two call sites, exactly as
+ * {@link COMMAND_CODE_SELECTION_LIST_FOOTER} is shared between the detector and
+ * the dialog card.
+ *
+ * @param frame - a raw `capture-pane -p -e` frame, ANSI intact
+ */
+function hasDismissablePanelFooter(frame) {
+    if (!frame)
+        return false;
+    const lines = (0, ansi_1.stripAnsi)(frame.replace(/\r\n/g, '\n')).split('\n');
+    let last = lines.length - 1;
+    while (last >= 0 && lines[last].trim() === '')
+        last -= 1;
+    if (last < 0)
+        return false;
+    const tail = lines.slice(Math.max(0, last + 1 - DISMISSABLE_PANEL_TAIL_LINE_COUNT), last + 1);
+    return tail.some((line) => exports.DISMISSABLE_PANEL_FOOTER_PATTERN.test(line));
+}
+/**
+ * The horizontal rule Command Code draws directly above a full-screen dialog.
+ *
+ * Command Code is an INLINE tool (`alternate_on=0`): opening `/model` does not
+ * clear the pane, it paints the picker under whatever the session has already
+ * printed. On the capture taken for Issue #2326 that is 256 rows of banner and
+ * transcript followed by a 77-row picker, and this rule row is the seam between
+ * them — 200 columns of U+2500 at the production `TUI_PANE_WIDTH`, the only
+ * such row anywhere on the frame while the picker is open (the composer's own
+ * two rules are not drawn while a dialog has the screen).
+ *
+ * Matched as "nothing but the rule glyph", not as "contains one", because a box
+ * border (`╭──…──╮`, which is how copilot draws ITS pickers) carries corners and
+ * must NOT be read as this seam — see
+ * {@link extractCommandCodeSelectionListFrame} for why that non-match is the
+ * safe answer rather than a missed case.
+ */
+exports.COMMAND_CODE_RULE_ROW_PATTERN = /^\u2500+$/;
+/**
+ * How wide a rule row must be before it counts as the dialog seam.
+ *
+ * The measured row is the full pane width (200 columns), and a pure function
+ * cannot know that width, so this is a floor rather than an equality: it only
+ * has to reject a short dash run inside a reply. Forty columns is a quarter of
+ * the production pane and wider than any decorative rule measured in
+ * `tests/fixtures/`.
+ */
+exports.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.
+ */
+exports.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 = '◯○◌⚪';
+/**
+ * The ANSWERED half, measured on 1.54.1 (Issue #2753).
+ *
+ * A multi-question `AskUserQuestion` keeps one tab per question and marks the
+ * ones already answered: `✔ Party size | ✔ Rental car | ● Update scope | ◯ Review`.
+ * Before this family existed the whole strip failed the cell test, the reader answered
+ * `none`, and the generic parser read the checkbox list under it as a single
+ * select — see the Issue for the capture.
+ *
+ * U+2714 and nothing else. `✓` (U+2713) and `☑` (U+2611) are the shapes a check
+ * mark could take; neither has been seen on a pane, and this reading is shared
+ * with `ChatSurface` and `extractCommandCodeSelectionListFrame`, which run for
+ * EVERY tool — so an unmeasured glyph here widens what other tools' frames can
+ * be claimed as this screen.
+ */
+const COMMAND_CODE_TAB_ANSWERED_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}${COMMAND_CODE_TAB_ANSWERED_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 other cell.** 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. The other
+ *    cell may be hollow (not answered yet) or the `✔` of an answered tab
+ *    (measured on 1.54.1 — Issue #2753); what may not be missing is the filled
+ *    one, so `✔ one | ✔ two` is still refused.
+ */
+function isCommandCodeQuestionTabRow(line) {
+    const segments = line.split('|');
+    if (segments.length < 2)
+        return false;
+    let selected = 0;
+    let others = 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
+            others += 1;
+    }
+    return selected > 0 && others > 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.
+ */
+exports.COMMAND_CODE_QUESTION_REGION_REJECT_PATTERN = /\besc\s+to\s+interrupt\b|^[^\S\n]*✻[^\S\n]+\S/im;
+/**
+ * The Review page's own three rows, measured on 1.54.1 (Issue #2755 §7).
+ *
+ * `Enter` on the `Submit` row does not send the answers — it opens a SECOND
+ * screen:
+ *
+ *     ✔ Update scope | ● Review
+ *
+ *     1. Which files should I update?
+ *        calc.js
+ *
+ *     ❯ 1. Submit
+ *       2. Cancel
+ *
+ *     ← to go back and edit
+ *
+ * ## Why this needs a reading of its own
+ *
+ * Because it is a real numbered list under a real tab strip, and the GENERIC
+ * parser answers it: a two-option `multiple_choice` whose default is `Submit`.
+ * The default answer on that payload does not pick anything — it COMMITS
+ * whatever the human has ticked so far, from a payload nobody meant to expose,
+ * which is #1928's shape one screen later. Auto-Yes fires on defaults.
+ *
+ * Words rather than structure, and deliberately: the two numbered rows are
+ * `Submit` and `Cancel`, which is the same shape as any other two-option
+ * question. What separates this page from a question is only what it SAYS, so
+ * all three rows are required together — the confirm pair AND the back hint —
+ * and the whole thing is anchored inside a recognised question region. A screen
+ * that carries two of the three keeps every verdict it had.
+ *
+ * The cursor glyph is optional on both numbered rows: it rests on `1. Submit`
+ * when the page is reached by `Enter`, and the Issue is explicit that a pattern
+ * decided on the bare word would miss it.
+ */
+const COMMAND_CODE_REVIEW_SUBMIT_ROW_PATTERN = /^[^\S\n]*(?:\u276F[^\S\n]*)?1[.)][^\S\n]+Submit[^\S\n]*$/im;
+/** The `Cancel` half of {@link COMMAND_CODE_REVIEW_SUBMIT_ROW_PATTERN}'s pair. */
+const COMMAND_CODE_REVIEW_CANCEL_ROW_PATTERN = /^[^\S\n]*(?:\u276F[^\S\n]*)?2[.)][^\S\n]+Cancel[^\S\n]*$/im;
+/** The footer only the Review page draws. */
+const COMMAND_CODE_REVIEW_BACK_HINT_PATTERN = /^[^\S\n]*\u2190[^\S\n]+to[^\S\n]+go[^\S\n]+back[^\S\n]+and[^\S\n]+edit[^\S\n]*$/im;
+/**
+ * The warning the Review page carries when it was reached with answers missing.
+ *
+ * Measured from the undocumented `d`, which ends a multi-select from any row:
+ * `⚠ You have not answered all questions` over a `No answer`. Issue #2755 §7
+ * requires that such a page is never read as "answered", which is why the
+ * predicate below reports it rather than hiding it.
+ */
+const COMMAND_CODE_REVIEW_UNANSWERED_PATTERN = /have\s+not\s+answered\s+all\s+questions/i;
+/**
+ * Read Command Code's `AskUserQuestion` Review page, if this frame is one
+ * (Issue #2755 §7).
+ *
+ * The same region {@link readCommandCodeQuestionRegion} works on — last rule
+ * row, tab strip as the first row under it — plus the three rows above. The
+ * question-shaped conditions are deliberately NOT applied: this page has no
+ * question body of its own on the `Submit`/`Cancel` spelling (the answers sit
+ * where the question would), and it draws `1.` twice, so the strict reading
+ * declines it for a numbering reason that says nothing about what it is.
+ *
+ * `null` for everything else, so no existing verdict moves except the two
+ * captures this is about.
+ *
+ * @param frame - a raw `capture-pane -p -e` frame, ANSI intact (CRLF tolerated)
+ */
+function readCommandCodeReviewPage(frame) {
+    if (!frame)
+        return null;
+    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;
+    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 < exports.COMMAND_CODE_RULE_MIN_COLUMNS)
+            continue;
+        if (!exports.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 tabOffset = region.findIndex((line) => line.trim() !== '');
+    if (tabOffset < 0)
+        return null;
+    if (!isCommandCodeQuestionTabRow(region[tabOffset]))
+        return null;
+    const body = region.join('\n');
+    if (!COMMAND_CODE_REVIEW_SUBMIT_ROW_PATTERN.test(body))
+        return null;
+    if (!COMMAND_CODE_REVIEW_CANCEL_ROW_PATTERN.test(body))
+        return null;
+    if (!COMMAND_CODE_REVIEW_BACK_HINT_PATTERN.test(body))
+        return null;
+    const submitRow = region.find((row) => COMMAND_CODE_REVIEW_SUBMIT_ROW_PATTERN.test(row));
+    return {
+        hasUnansweredWarning: COMMAND_CODE_REVIEW_UNANSWERED_PATTERN.test(body),
+        cursorOnSubmit: submitRow !== undefined && submitRow.includes(exports.COMMAND_CODE_CURSOR_GLYPH),
+    };
+}

package/dist/server/src/lib/detection/tools/context-above.js

@@ -0,0 +1,20 @@
+"use strict";
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.readContextAbove = readContextAbove;
+/** How many non-blank rows above the options may hold the question. */
+const QUESTION_SCAN_ROWS = 4;
+/**
+ * The non-blank rows above `firstRow`, nearest first, joined by newlines.
+ * Rows at or above `floorRow` are not read: pass -1 to read to the top of the
+ * frame.
+ */
+function readContextAbove(lines, firstRow, floorRow) {
+    const rows = [];
+    for (let i = firstRow - 1; i > floorRow && rows.length < QUESTION_SCAN_ROWS; i--) {
+        const row = lines[i].trim();
+        if (row === '')
+            continue;
+        rows.push(row);
+    }
+    return rows.join('\n');
+}

package/dist/server/src/lib/detection/tools/copilot/detect.js

@@ -13,10 +13,15 @@ exports.copilotStatusDetector = exports.VERIFIED_AGAINST = void 0;
 exports.readIdleEvidence = readIdleEvidence;
 const cli_patterns_1 = require("../../cli-patterns");
 const prompt_detector_1 = require("../../prompt-detector");
+const verdicts_1 = require("../verdicts");
 const status_reason_1 = require("../../status-reason");
 const prompt_1 = require("./prompt");
+const live_region_1 = require("./live-region");
 const run_detection_1 = require("../run-detection");
 const verified_against_1 = require("../verified-against");
+function detectCopilotPrompt(clean) {
+    return (0, prompt_detector_1.detectPrompt)((0, cli_patterns_1.stripBoxDrawing)(clean), (0, cli_patterns_1.buildDetectPromptOptions)('copilot'));
+}
 /** copilot-cli build these rules were read off (#1885 / #1895; value in ../verified-against, #1929). */
 exports.VERIFIED_AGAINST = verified_against_1.COPILOT_VERIFIED_AGAINST;
 /**
@@ -39,6 +44,8 @@ function readIdleEvidence(frame) {
 exports.copilotStatusDetector = (0, run_detection_1.createToolStatusDetector)({
     tool: 'copilot',
     verifiedAgainst: exports.VERIFIED_AGAINST,
+    // Issue #3183: where the live part of the frame begins (`./live-region.ts`).
+    liveRegion: live_region_1.COPILOT_LIVE_REGION,
     beforePrompt(frame) {
         // 0. Copilot: picker detection BEFORE thinking detection
         // COPILOT_THINKING_PATTERN includes "Reasoning\s+[■▪▮]" which matches the
@@ -63,33 +70,18 @@ exports.copilotStatusDetector = (0, run_detection_1.createToolStatusDetector)({
         // never the transcript. Ordering against them is therefore settled inside the
         // helper: it declines any frame that still has a status bar.
         if ((0, cli_patterns_1.isCopilotSelectionFrame)(frame.contentLines)) {
-            const promptOptions = (0, cli_patterns_1.buildDetectPromptOptions)('copilot');
-            const promptDetection = (0, prompt_detector_1.detectPrompt)((0, cli_patterns_1.stripBoxDrawing)(frame.clean), promptOptions);
+            const promptDetection = detectCopilotPrompt(frame.clean);
             if (promptDetection.isPrompt) {
                 // Distinguish yes/no prompts (2-3 options, e.g., "Do you want to run this command?")
                 // from ask_user multi-select prompts (4+ options). Yes/no prompts should show
                 // PromptPanel with buttons; ask_user prompts need NavigationButtons for ↑↓ selection.
                 const optionsCount = promptDetection.promptData?.options?.length ?? 0;
                 if (optionsCount <= 3) {
-                    return {
-                        status: 'waiting',
-                        confidence: 'high',
-                        reason: status_reason_1.STATUS_REASON.PROMPT_DETECTED,
-                        hasActivePrompt: true,
-                        evidence: 'positive',
-                        promptDetection,
-                    };
+                    return (0, verdicts_1.activePromptVerdict)(status_reason_1.STATUS_REASON.PROMPT_DETECTED, promptDetection);
                 }
                 // 4+ options: treat as selection list (NavigationButtons)
             }
-            return {
-                status: 'waiting',
-                confidence: 'high',
-                reason: status_reason_1.STATUS_REASON.COPILOT_SELECTION_LIST,
-                hasActivePrompt: false,
-                evidence: 'positive',
-                promptDetection,
-            };
+            return (0, verdicts_1.positiveVerdictWithPrompt)('waiting', status_reason_1.STATUS_REASON.COPILOT_SELECTION_LIST, promptDetection);
         }
         // 0.5. Copilot: the bottom status bar carries the running half of the turn
         // (Issue #547 put a thinking check here; Issue #1885 replaced what it reads).
@@ -113,16 +105,8 @@ exports.copilotStatusDetector = (0, run_detection_1.createToolStatusDetector)({
         // window form would also have matched copilot's own response text -- see
         // `status-vocabulary-in-response.txt`.
         if ((0, cli_patterns_1.readCopilotStatusBar)(frame.contentLines) === 'working') {
-            const promptOptions = (0, cli_patterns_1.buildDetectPromptOptions)('copilot');
-            const promptDetection = (0, prompt_detector_1.detectPrompt)((0, cli_patterns_1.stripBoxDrawing)(frame.clean), promptOptions);
-            return {
-                status: 'running',
-                confidence: 'high',
-                reason: status_reason_1.STATUS_REASON.THINKING_INDICATOR,
-                hasActivePrompt: false,
-                evidence: 'positive',
-                promptDetection,
-            };
+            const promptDetection = detectCopilotPrompt(frame.clean);
+            return (0, verdicts_1.positiveVerdictWithPrompt)('running', status_reason_1.STATUS_REASON.THINKING_INDICATOR, promptDetection);
         }
         return null;
     },
@@ -147,13 +131,7 @@ exports.copilotStatusDetector = (0, run_detection_1.createToolStatusDetector)({
         // verdict claude's `❯` row and codex's `›` row publish, so nothing downstream
         // has to learn a new reason code.
         if ((0, cli_patterns_1.readCopilotStatusBar)(frame.contentLines) === 'idle') {
-            return {
-                status: 'ready',
-                confidence: 'high',
-                reason: status_reason_1.STATUS_REASON.INPUT_PROMPT,
-                hasActivePrompt: false,
-                evidence: 'positive',
-            };
+            return (0, verdicts_1.positiveVerdict)('ready', status_reason_1.STATUS_REASON.INPUT_PROMPT);
         }
         return null;
     },

package/dist/server/src/lib/detection/tools/copilot/live-region.js

@@ -0,0 +1,52 @@
+"use strict";
+/**
+ * Copilot's live region (Issue #3183; design doc §7 "copilot").
+ *
+ * - **composer** — the bottom chrome `findCopilotChromeStart` locates: the cwd
+ *   row, the fenced composer (`─` rules on 1.0.80, `╻▄` / `╹▀` frames on
+ *   1.0.82) and the status bar under it. Measured: `copilot-live-1885/
+ *   turn-complete.txt` L997-L1001 (1.0.80) and `copilot-live-2269/
+ *   turn-complete.txt` L996-L1000 (1.0.82).
+ * - **dialog top** — the `╭` corner the bottom `╰` row closes. copilot draws its
+ *   dialogs in a box over the whole bottom of the pane, chrome gone. Measured:
+ *   `copilot-live-1885/permission-dialog.txt` L987 / L1001 (1.0.80; 1.0.82's
+ *   dialog has not been captured — design doc §6 item 3).
+ *
+ * `composerHidesDialogs: false`: the `/model` picker is drawn with the fenced
+ * composer still on screen (`copilot-live-1885/model-picker.txt`; the comment
+ * on `findCopilotChromeStart` records the blank header row it leaves), so only
+ * numbered readings are vetoed by it. The numbered permission dialog replaces
+ * the chrome (`permission-dialog.txt`).
+ *
+ * @module lib/detection/tools/copilot/live-region
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.COPILOT_LIVE_REGION = void 0;
+const cli_patterns_1 = require("../../cli-patterns");
+const live_region_1 = require("../live-region");
+/** How far above the bottom row the box's closing corner may sit. */
+const BOX_BOTTOM_MAX_ROWS = 2;
+/** How tall a dialog box may be before it stops looking like one. */
+const BOX_MAX_ROWS = 120;
+function findCopilotDialogBoxTop(lines) {
+    const last = (0, live_region_1.lastContentRow)(lines);
+    let bottom = -1;
+    for (let i = last; i >= Math.max(0, last - BOX_BOTTOM_MAX_ROWS); i--) {
+        if (lines[i].trimStart().startsWith('╰')) {
+            bottom = i;
+            break;
+        }
+    }
+    if (bottom < 0)
+        return -1;
+    for (let i = bottom - 1; i >= Math.max(0, bottom - BOX_MAX_ROWS); i--) {
+        if (lines[i].trimStart().startsWith('╭'))
+            return i;
+    }
+    return -1;
+}
+exports.COPILOT_LIVE_REGION = {
+    composer: (0, live_region_1.fencedComposer)(lines => (0, cli_patterns_1.findCopilotChromeStart)(lines)),
+    dialogTop: (0, live_region_1.dialogTopFrom)(findCopilotDialogBoxTop),
+    composerHidesDialogs: false,
+};