commandmate 0.27.1 → 0.29.0

package/dist/server/src/lib/detection/opencode-modal-overlay.js

@@ -0,0 +1,400 @@
+"use strict";
+/**
+ * opencode's modal overlays, detected structurally from a captured frame
+ * (Issue #2112).
+ *
+ * ## What this is about
+ *
+ * opencode 1.18 opens its session list (`ctrl+x l`), agent list (`ctrl+x a`),
+ * timeline (`ctrl+x g`), command palette (`ctrl+p`) and model picker
+ * (`ctrl+x m`) as an overlay painted OVER the transcript. The overlay hides
+ * nothing the detector reads: the finished-turn marker `▣ Build · <model> ·
+ * 2.8s` from the previous turn is still on the pane above it, and the input box
+ * is still drawn below it. So branch D of `tools/opencode/detect.ts` matched the
+ * old marker and published a pane that is blocked on a human as
+ * `ready` / `opencode_response_complete` — measured on the committed #2046
+ * fixtures, `dialog-{session-list,agent-list,timeline}.txt`.
+ *
+ * That is worse than it sounds, and it is why this module exists rather than a
+ * widened allowlist. `ready` is POSITIVE evidence, so the frame never reaches
+ * the "nothing could read this" path #1017 / #1494 built for unknown overlays:
+ * `isUnclassifiedActive` stays false, the 60-second escape hatch never opens,
+ * and `commandmate wait` reports the blocked pane as **exit 0, completed**.
+ * `claude`'s `/help` overlay is the benign version of the same shape — it lands
+ * on `running` / `default`, i.e. on the no-evidence side, where the hatch does
+ * open.
+ *
+ * ## The signature, and why it is not a word list
+ *
+ * `Sessions`, `Commands`, `Timeline` and `Select agent` are the overlay
+ * headings. They are also ordinary English an agent writes in a reply, and
+ * Issue #1883 already deleted one "the word is somewhere on the pane" inference
+ * for exactly that reason (#1896 measured the harm: an answer that said
+ * `Select model to continue:` parked the session on `waiting` for the rest of
+ * the session). `OPENCODE_SELECTION_LIST_PATTERN` reacted by narrowing to three
+ * allowlisted headings, which is why `Select agent` does not match it today.
+ *
+ * This module reads the LAYOUT instead. opencode paints the overlay as a
+ * background-coloured rectangle, and `capture-pane -e` re-emits that background
+ * verbatim:
+ *
+ * ```
+ *   ␛[48;2;20;20;20m    Commands                     ␛[38;2;128;128;128mesc␛[0m    ␛[48;2;4;4;4mlogical
+ *   └ column 10 ─────────────────────────────────────────────────────────┘ column 70
+ * ```
+ *
+ * Every row of the overlay switches the background on at the SAME column and
+ * off at the SAME column, whatever is written on it — and one of those rows
+ * carries the overlay's own dismiss affordance, `esc`, right-aligned INSIDE the
+ * rectangle. Both halves are needed:
+ *
+ *  - the rectangle alone is also what opencode's input box and its echoed user
+ *    prompts are (measured: the composer is a painted rectangle at columns
+ *    3–78 on every 80-column frame in `opencode-live-2046`);
+ *  - the `esc` hatch alone is also what the composer footer says while the agent
+ *    is generating (`⬝⬝⬝⬝⬝⬝  esc interrupt`, `esc again to interrupt`) — which
+ *    is why the hatch is required to be the LAST thing inside the rectangle, as
+ *    `OPENCODE_SELECTION_LIST_PATTERN` requires it to be the last thing on the
+ *    line.
+ *
+ * Reading the rectangle rather than the row's total width is what makes this
+ * work over a busy transcript. At 120 and 200 columns the same palette is
+ * centred with response text visible on BOTH sides of it
+ * (`opencode-live-2047/w200/command-palette.txt`), so the rows are not equal in
+ * length and nothing outside the rectangle is constant — only its two edges are.
+ *
+ * ## What a match means, and what it does not
+ *
+ * A match means a background-painted rectangle offering an `esc` hatch was on
+ * the pane. A **non**-match means no such rectangle could be read, and NEVER
+ * "no overlay is open": the rule needs the SGR the capture was taken with, so a
+ * caller that hands over an already-stripped frame — Auto-Yes's
+ * `captureAndCleanOutput` is the one in this repository — always gets `null`.
+ * That is fail-open by construction and matches how `tools/opencode/prompt.ts`
+ * already describes its own gutter anchor on that path. Every production caller
+ * of `detectSessionStatus` passes the raw `capture-pane -e` output
+ * (`lib/tmux/tmux.ts` always passes `-e`).
+ *
+ * A leaf module by design: it imports the ANSI primitives, the column rule and
+ * the excerpt bound and nothing else.
+ *
+ * @module lib/detection/opencode-modal-overlay
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.OPENCODE_OVERLAY_TRUNCATION_MARKER = exports.OPENCODE_OVERLAY_EXCERPT_MAX_BYTES = exports.OPENCODE_OVERLAY_HEADER_ROW_LIMIT = exports.OPENCODE_OVERLAY_MIN_ROWS = exports.OPENCODE_OVERLAY_MIN_WIDTH = exports.OPENCODE_OVERLAY_MIN_LEFT = exports.OPENCODE_MODAL_OVERLAY_RECOVERY_KEY = exports.OPENCODE_MODAL_OVERLAY_ID = void 0;
+exports.detectOpenCodeModalOverlay = detectOpenCodeModalOverlay;
+const excerpt_1 = require("./excerpt");
+const terminal_columns_1 = require("./terminal-columns");
+/**
+ * Published `sessionStatusReason` for this signature.
+ *
+ * Snake case, like every other reason, and distinct from
+ * `opencode_selection_list` on purpose: an operator reading `capture --json`
+ * can then tell "the allowlisted picker heading matched" from "a painted
+ * rectangle with an esc hatch was on the pane", which are two different
+ * readings of the same screen and fail in different ways.
+ */
+exports.OPENCODE_MODAL_OVERLAY_ID = 'opencode_modal_overlay';
+/**
+ * The key that closes it, in opencode's own notation.
+ *
+ * Taken from the overlay's own chrome — the hatch this module anchors on IS the
+ * affordance, printed by opencode at the right edge of the header row. Exported
+ * for the same reason {@link OPENCODE_SIDEBAR_RECOVERY_CHORD} is: no surface
+ * should hold a bare key literal of its own.
+ */
+exports.OPENCODE_MODAL_OVERLAY_RECOVERY_KEY = 'esc';
+/**
+ * Leftmost column an overlay rectangle may start at.
+ *
+ * One, not zero. A rectangle that starts at column 0 is not an overlay drawn
+ * OVER the pane — it is the pane's own background, which some themes paint
+ * explicitly (measured: `opencode-live-2047/w200/*` paint every row
+ * `48;2;4;4;4` from column 0), and treating that as a rectangle makes every row
+ * of every frame "match" whatever else is asserted. The measured overlays start
+ * at column 10 (80 and 120 columns), 70 (200 columns) and 1 (the 80-column
+ * session list and timeline, which are nearly pane-wide), so the constraint has
+ * never been the thing that decided a verdict.
+ */
+exports.OPENCODE_OVERLAY_MIN_LEFT = 1;
+/**
+ * Narrowest rectangle that can be an overlay, in columns.
+ *
+ * The measured overlays are 60 columns wide (palette and pickers at every
+ * width) and 78 (the session list and timeline at 80 columns), so this is a
+ * floor with four times the headroom rather than a tuned threshold. Its job is
+ * to reject a single highlighted CELL — opencode paints the selected row of a
+ * list in its own background, and a one-word highlight elsewhere on the pane
+ * should not be able to pose as a dialog.
+ */
+exports.OPENCODE_OVERLAY_MIN_WIDTH = 16;
+/**
+ * How many rows must carry the rectangle before this reports a match.
+ *
+ * Three: the measured overlays carry 8 (the timeline, the shortest) to 72 (the
+ * command palette), so there is an order of magnitude of headroom, and three is
+ * what a real dialog cannot be under — opencode draws a padding row, the header
+ * row with the hatch, and a padding row before any content at all.
+ */
+exports.OPENCODE_OVERLAY_MIN_ROWS = 3;
+/**
+ * How far down the rectangle the `esc` hatch may sit, in rows.
+ *
+ * A dialog puts its dismiss affordance in its TITLE BAR, and the measurement is
+ * unusually clean: on all eleven overlays in this repository — four #2046
+ * dialogs, the palette at 80/120/200 columns, the #2049 and #1896 palettes, the
+ * #1896 model picker and the canary picker — the hatch is on the SECOND row that
+ * carries the rectangle, the first being the overlay's top padding. Three is
+ * that measurement plus one row of headroom.
+ *
+ * What it buys is the case the other guards cannot reach: opencode paints code
+ * blocks in an agent's reply with their own background, so a fenced keybinding
+ * table (`quit    esc`) is a painted rectangle with a hatch-shaped row in it.
+ * Requiring the hatch at the top makes that a table in the middle of a block
+ * rather than a dialog's title bar. The 120- and 200-column captures show the
+ * other half of the same problem — transcript painting produces up to fourteen
+ * incidental rectangles per frame — and not one of them carries a hatch at all.
+ */
+exports.OPENCODE_OVERLAY_HEADER_ROW_LIMIT = 3;
+/** Bound on {@link OpenCodeModalOverlay.headerText}, in UTF-8 bytes. */
+exports.OPENCODE_OVERLAY_EXCERPT_MAX_BYTES = 200;
+/** Appended when {@link OPENCODE_OVERLAY_EXCERPT_MAX_BYTES} cut the excerpt. */
+exports.OPENCODE_OVERLAY_TRUNCATION_MARKER = '…[truncated]';
+/**
+ * The overlay's dismiss affordance, right-aligned inside the rectangle.
+ *
+ * Deliberately the same shape as `OPENCODE_SELECTION_LIST_PATTERN`'s tail
+ * (Issue #1896): text, then two or more spaces, then `esc`, then nothing but
+ * padding. The `$` is what separates a dialog's hatch from the composer footer's
+ * `esc interrupt` / `esc again to interrupt`, which carry a word AFTER the key —
+ * both of those are live frames in this repository
+ * (`opencode-live-1883/turn-running.txt`,
+ * `opencode-live-1894/esc-again-after-marker.txt`) and both are painted
+ * rectangles, so without the anchor a generating pane would be read as a dialog.
+ *
+ * Applied to the rectangle's own columns, never to the whole row: at 120 and 200
+ * columns the transcript resumes to the right of the overlay on the same line,
+ * so a row-anchored `$` would never match there.
+ *
+ * The leading `\S` keeps a rectangle whose only content is the hatch from
+ * counting: a dialog has a heading.
+ *
+ * Linear pattern, no nested quantifiers — ReDoS safe (S4-001).
+ */
+const OVERLAY_ESCAPE_HATCH = /\S[^\S\n]{2,}esc[^\S\n]*$/;
+/** Cheap pre-filter for the rows worth measuring at all. */
+const ESCAPE_HATCH_WORD = 'esc';
+/**
+ * One of opencode's own box gutters: the composer, the permission dialog, or an
+ * echoed user prompt.
+ *
+ * Same character class as `OPENCODE_GUTTER_ROW_PATTERN` (Issue #1893) plus the
+ * bottom border's `╹`. Those boxes are painted rectangles too — the composer is
+ * one at columns 3–78 on every 80-column frame here — and a user whose prompt
+ * happened to end in `…  esc` would otherwise be drawing an overlay in the
+ * transcript. That is not hypothetical: it is `words-in-response.txt`, this
+ * Issue's negative control, whose echoed prompt says all four dialog headings
+ * and ends in a hatch.
+ *
+ * Tested at the column immediately LEFT of the rectangle, because that is where
+ * opencode puts the glyph: the gutter is drawn first and the background is
+ * switched on after it, so the box's painted interior starts one column to its
+ * right (measured — `[(null, 0, 3), ('48;2;30;30;30', 3, 78)]` on every gutter
+ * row of `sidebar-off.txt`). The rectangle's own first column is tested too, so
+ * a theme that paints the gutter itself is covered by the same rule.
+ */
+const BOX_GUTTER = /^[│┃╹]/;
+/**
+ * Every ANSI sequence, in the repository's own spelling.
+ *
+ * Restated from `ANSI_PATTERN` (`ansi.ts`) rather than imported because this
+ * module needs to walk the sequences IN ORDER while counting columns, and that
+ * one is a shared `g`-flagged instance whose `lastIndex` a caller must not
+ * disturb. `tests/unit/detection-opencode-modal-overlay-2112.test.ts` asserts
+ * the text this scan produces is byte-identical to `stripAnsi`'s, so the
+ * restatement cannot drift silently.
+ */
+const ANSI_SEQUENCE = /\x1b\[[0-9;]*[a-zA-Z]|\x1b\][^\x07\x1b]*(?:\x07|\x1b\\)|\[[0-9;]*m/g;
+/** An SGR sequence, with its parameters. Only these can change the background. */
+const SGR_PARAMETERS = /^(?:\x1b\[|\[)([0-9;]*)m$/;
+/**
+ * Fold one SGR sequence's parameters into the current background.
+ *
+ * Handles the three forms tmux emits: the 16-colour codes (`40`–`47`,
+ * `100`–`107`), the 256-colour and 24-bit extended forms (`48;5;n`,
+ * `48;2;r;g;b`), and the two resets (`0`, `49`). Foreground and attribute codes
+ * leave the background alone, which is what keeps a syntax-highlighted
+ * transcript row from producing dozens of spurious edges.
+ */
+function applySgr(current, parameters) {
+    const parts = (parameters === '' ? '0' : parameters).split(';');
+    let bg = current;
+    for (let i = 0; i < parts.length; i++) {
+        const code = Number(parts[i] === '' ? '0' : parts[i]);
+        if (!Number.isFinite(code))
+            continue;
+        if (code === 0 || code === 49) {
+            bg = null;
+        }
+        else if ((code >= 40 && code <= 47) || (code >= 100 && code <= 107)) {
+            bg = parts[i];
+        }
+        else if (code === 48) {
+            if (parts[i + 1] === '2') {
+                bg = parts.slice(i, i + 5).join(';');
+                i += 4;
+            }
+            else if (parts[i + 1] === '5') {
+                bg = parts.slice(i, i + 3).join(';');
+                i += 2;
+            }
+        }
+    }
+    return bg;
+}
+/** Measure one captured row: its text, its background runs and their edges. */
+function scanRow(line) {
+    ANSI_SEQUENCE.lastIndex = 0;
+    const segments = [];
+    let text = '';
+    let bg = null;
+    let columns = 0;
+    let segmentStart = 0;
+    let cursor = 0;
+    let match;
+    while ((match = ANSI_SEQUENCE.exec(line)) !== null) {
+        const chunk = line.slice(cursor, match.index);
+        text += chunk;
+        columns += (0, terminal_columns_1.visibleWidth)(chunk);
+        cursor = match.index + match[0].length;
+        const sgr = SGR_PARAMETERS.exec(match[0]);
+        if (sgr === null)
+            continue;
+        const next = applySgr(bg, sgr[1]);
+        if (next === bg)
+            continue;
+        if (columns > segmentStart)
+            segments.push({ bg, start: segmentStart, end: columns });
+        bg = next;
+        segmentStart = columns;
+    }
+    const tail = line.slice(cursor);
+    text += tail;
+    columns += (0, terminal_columns_1.visibleWidth)(tail);
+    if (columns > segmentStart)
+        segments.push({ bg, start: segmentStart, end: columns });
+    const edges = new Set(segments.map(segment => segment.start));
+    if (segments.length > 0)
+        edges.add(segments[segments.length - 1].end);
+    return { text, segments, edges, width: columns };
+}
+/** The background painted at one column, or `null` for the terminal default. */
+function backgroundAt(row, column) {
+    for (const segment of row.segments) {
+        if (segment.start <= column && column < segment.end)
+            return segment.bg;
+    }
+    return null;
+}
+/**
+ * Does this row carry the rectangle `[left, right)`?
+ *
+ * Four conditions, and every one of them is about the LAYOUT:
+ *
+ *  1. the row has a background boundary at both edges;
+ *  2. every column between them is painted — an unpainted gap means the two
+ *    boundaries belong to two different things that happen to line up;
+ *  3. the boundaries are real transitions, i.e. what is painted just outside
+ *    differs from what is painted just inside. This is what stops a theme that
+ *    paints the whole pane from turning every column pair into a rectangle;
+ *  4. neither the rectangle's first column nor the column just left of it holds
+ *    one of opencode's own box gutters.
+ */
+function carriesRectangle(row, left, right) {
+    if (!row.edges.has(left) || !row.edges.has(right))
+        return false;
+    for (const segment of row.segments) {
+        if (segment.end <= left || segment.start >= right)
+            continue;
+        if (segment.bg === null)
+            return false;
+    }
+    if (backgroundAt(row, left - 1) === backgroundAt(row, left))
+        return false;
+    if (right < row.width && backgroundAt(row, right) === backgroundAt(row, right - 1))
+        return false;
+    if (BOX_GUTTER.test((0, terminal_columns_1.sliceColumns)(row.text, left - 1, left)))
+        return false;
+    return !BOX_GUTTER.test((0, terminal_columns_1.sliceColumns)(row.text, left, right));
+}
+/**
+ * The modal overlay painted over an opencode transcript, or null.
+ *
+ * Call it only for opencode: the rule is about opencode's own dialog chrome, and
+ * every other tool is left exactly as it was (Issue #2112 changes no other
+ * tool's detection).
+ *
+ * The search is seeded from the rows that contain the literal `esc` and never
+ * from the frame at large, so the cost is `rows × (background boundaries on a
+ * hatch row)²` rather than anything quadratic in the pane. Background boundaries
+ * are rare — a foreground-only SGR does not make one — so a 200-row opencode
+ * capture typically seeds from a single row with fewer than ten of them.
+ *
+ * When more than one rectangle qualifies the widest-supported one wins, measured
+ * by how many rows carry it: an overlay's own highlighted list item is a
+ * rectangle too, and it is drawn on exactly one row.
+ *
+ * A caveat the seeding makes easy to miss: two identical rows are two entries in
+ * `carrying`, so the title-bar check below compares object identity rather than
+ * value.
+ *
+ * @param frame - The capture with its ANSI intact (`capture-pane -e`)
+ */
+function detectOpenCodeModalOverlay(frame) {
+    const rows = frame.split('\n').map(scanRow);
+    const examined = new Set();
+    let best = null;
+    for (const row of rows) {
+        if (!row.text.includes(ESCAPE_HATCH_WORD))
+            continue;
+        const edges = [...row.edges].sort((a, b) => a - b);
+        for (const left of edges) {
+            if (left < exports.OPENCODE_OVERLAY_MIN_LEFT)
+                continue;
+            for (const right of edges) {
+                if (right - left < exports.OPENCODE_OVERLAY_MIN_WIDTH)
+                    continue;
+                const key = `${left}:${right}`;
+                if (examined.has(key))
+                    continue;
+                if (!carriesRectangle(row, left, right))
+                    continue;
+                const header = (0, terminal_columns_1.sliceColumns)(row.text, left, right);
+                if (!OVERLAY_ESCAPE_HATCH.test(header))
+                    continue;
+                examined.add(key);
+                const carrying = rows.filter(candidate => carriesRectangle(candidate, left, right));
+                if (carrying.length < exports.OPENCODE_OVERLAY_MIN_ROWS)
+                    continue;
+                // Identity, not a re-test: `row` is the row whose slice just matched, and
+                // asking which position it holds in the rectangle is the whole of the
+                // title-bar rule. A second search by text would find the FIRST hatch-
+                // shaped row instead, which is a different question.
+                const headerRow = carrying.indexOf(row);
+                if (headerRow < 0 || headerRow >= exports.OPENCODE_OVERLAY_HEADER_ROW_LIMIT)
+                    continue;
+                if (best === null || carrying.length > best.rows) {
+                    best = {
+                        id: exports.OPENCODE_MODAL_OVERLAY_ID,
+                        headerText: (0, excerpt_1.truncateToByteBudget)(header.trim(), exports.OPENCODE_OVERLAY_EXCERPT_MAX_BYTES, exports.OPENCODE_OVERLAY_TRUNCATION_MARKER),
+                        left,
+                        right,
+                        rows: carrying.length,
+                        headerRow,
+                    };
+                }
+            }
+        }
+    }
+    return best;
+}

package/dist/server/src/lib/detection/opencode-pane-obstruction.js

@@ -0,0 +1,181 @@
+"use strict";
+/**
+ * opencode's sidebar, detected structurally from a captured frame (Issue #2095).
+ *
+ * ## What this is about
+ *
+ * opencode 1.18.22 paints a right-hand sidebar (session title, `Context` /
+ * token counters, `LSP` status) that does **not** get its own region of a
+ * `capture-pane`. It shares ROWS with the transcript and with the input box, so
+ * every captured line becomes `<transcript text> … <sidebar text>` and every
+ * reader in this repo sees one row (Issue #2047 measured this at ≥121 columns).
+ *
+ * Issue #2046 then measured the consequence at the width production actually
+ * runs opencode at. `ctrl+x b` (`sidebar_toggle`) turns the sidebar on at **80
+ * columns**, below the width #2047 found it appearing at on its own, and the
+ * same finished turn flips from `ready` / `opencode_response_complete` to
+ * `running` / `unknown_frame` with the sidebar's own text saved as the
+ * assistant's reply. Escape does not close it; pressing `ctrl+x b` again does.
+ *
+ * #2046 removed `b` from {@link OPENCODE_LEADER_CHORD_VALUES} and the
+ * special-keys route refuses it, but that only closes CommandMate's own door:
+ * opencode's `ctrl+p` command palette lists `Show sidebar   ctrl+x b` and a user
+ * can run it from there. **This module does not stop the sidebar. It makes the
+ * sidebar visible to the people and the tools that are waiting on the pane.**
+ *
+ * ## The signature, and why it is not a word list
+ *
+ * `Context`, `LSP` and `tokens` are on the sidebar, and they are also ordinary
+ * English an agent writes in a reply — a word list would flag any turn that
+ * happens to discuss LSPs. The signature this module uses is the geometry
+ * instead:
+ *
+ * ```
+ *   ┃                                     Context          ← gutter row, second column
+ *      OK2046                             1% used
+ *   ┃  Build · Claude Sonnet 4.6 GitHub   maenokota-share-work-…
+ *   ╹▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀▀  commandmate-issue-2046/…
+ *                                       ↑ the box ends here
+ * ```
+ *
+ * opencode draws the input box's bottom edge as `╹` + a run of `▀` spanning the
+ * box's full width, and pads every gutter (`┃`) row out to that same width. So
+ * the box's right edge is measurable from the frame itself, and **content to the
+ * right of it on a row that belongs to the box is a second column** — something
+ * a box that owns the full pane width can never produce, because its own text
+ * wraps at that edge.
+ *
+ * Deliberately NOT gated on the pane being wide. #2046's measurement is that the
+ * explicit toggle ignores the 121-column threshold, so a width branch would miss
+ * the case this Issue exists for.
+ *
+ * ## What a match means, and what it does not
+ *
+ * A match means the frame was laid out in two columns. A **non**-match means the
+ * layout could not be read — most often because no bottom border was on the
+ * frame at all, which is what a permission dialog produces (measured:
+ * `permission-bash.txt` / `permission-edit.txt` under every width of
+ * `tests/fixtures/opencode-live-2047` carry no border row) — and NEVER "the
+ * sidebar is off". Nothing reading this may report
+ * `null` as an all-clear, for the same reason `upstream-faults.ts` says so.
+ *
+ * A leaf module by design: it imports {@link stripAnsi}, the excerpt bound and
+ * the column rule and nothing else, so the payload builder, the CLI types and the
+ * client components that derive this from the frame they already hold can all
+ * share one rule.
+ *
+ * @module lib/detection/opencode-pane-obstruction
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.OPENCODE_SIDEBAR_TRUNCATION_MARKER = exports.OPENCODE_SIDEBAR_EXCERPT_MAX_BYTES = exports.OPENCODE_SIDEBAR_MIN_ROWS = exports.OPENCODE_SIDEBAR_RECOVERY_CHORD = exports.OPENCODE_SIDEBAR_OBSTRUCTION_ID = void 0;
+exports.detectOpenCodePaneObstruction = detectOpenCodePaneObstruction;
+const ansi_1 = require("./ansi");
+const excerpt_1 = require("./excerpt");
+const terminal_columns_1 = require("./terminal-columns");
+/**
+ * Published `paneObstruction.id` for this signature.
+ *
+ * Snake case, like every `sessionStatusReason` value, because the same operators
+ * read the two next to each other in `capture --json`.
+ */
+exports.OPENCODE_SIDEBAR_OBSTRUCTION_ID = 'opencode_sidebar';
+/**
+ * The keystroke that closes it, in opencode's own notation.
+ *
+ * Taken verbatim from opencode's keybind table, which its `ctrl+p` palette
+ * prints as `Show sidebar   ctrl+x b` — see
+ * `tests/fixtures/opencode-live-2046/w80/dialog-command-palette.txt:70`.
+ * Exported rather than written into each surface so the UI banner, the history
+ * row and `wait`'s stderr cannot drift apart, and so no surface has to hold a
+ * bare key literal of its own.
+ */
+exports.OPENCODE_SIDEBAR_RECOVERY_CHORD = 'ctrl+x b';
+/**
+ * How many box rows must carry a second column before this reports a match.
+ *
+ * Two, as cheap insurance against a single stray cell — a wide character
+ * straddling the edge, a repaint caught mid-write. Not a tuned threshold: the
+ * measured frames carry 3 (in the 100-row window this is judged on) to 11, so
+ * the guard has an order of magnitude of headroom and has never been the thing
+ * that decided a verdict.
+ */
+exports.OPENCODE_SIDEBAR_MIN_ROWS = 2;
+/** Bound on {@link OpenCodePaneObstruction.matchedText}, in UTF-8 bytes. */
+exports.OPENCODE_SIDEBAR_EXCERPT_MAX_BYTES = 200;
+/** Appended when {@link OPENCODE_SIDEBAR_EXCERPT_MAX_BYTES} cut the excerpt. */
+exports.OPENCODE_SIDEBAR_TRUNCATION_MARKER = '…[truncated]';
+/**
+ * The input box's bottom edge: `  ╹▀▀▀▀▀…`.
+ *
+ * The same anchor `OPENCODE_COMPOSER_BOTTOM_BORDER` (`cli-patterns.ts`, Issue
+ * #1911) uses, restated here with capture groups because this module needs the
+ * edge's COLUMN and that one only needs to know the row exists.
+ * `tests/unit/detection-opencode-pane-obstruction-2095.test.ts` asserts the two
+ * still select the same rows, so the restatement cannot drift silently.
+ */
+const COMPOSER_BOTTOM_BORDER_ROW = /^([^\S\n]*)╹(▀{4,})/;
+/**
+ * A row of the input box: leading space, then the gutter.
+ *
+ * Character class shared with `OPENCODE_GUTTER_ROW_PATTERN` (Issue #1893):
+ * opencode draws the same box with `┃` and, in some themes, `│`.
+ */
+const COMPOSER_GUTTER_ROW = /^([^\S\n]*)[│┃]/;
+/**
+ * The second column sharing rows with an opencode transcript, or null.
+ *
+ * Call it only for opencode: the anchors are opencode's own box drawing, and
+ * every other tool is left exactly as it was (Issue #2095 changes no other
+ * tool's detection).
+ *
+ * ANSI is stripped first — the sidebar is painted with an SGR background, so on
+ * a raw `capture-pane -e` frame the column boundary is buried in escape
+ * sequences.
+ */
+function detectOpenCodePaneObstruction(frame) {
+    const lines = (0, ansi_1.stripAnsi)(frame).split('\n');
+    // The LAST border row, not the first: a long capture holds the borders of
+    // every repaint still in scrollback, and the only one that describes the box
+    // on screen now is the last one. Reading an older, wider edge would report
+    // "no second column" on a pane that has one.
+    let borderIndex = -1;
+    let boxLeft = -1;
+    let boxRight = -1;
+    for (let i = 0; i < lines.length; i++) {
+        const border = COMPOSER_BOTTOM_BORDER_ROW.exec(lines[i]);
+        if (!border)
+            continue;
+        borderIndex = i;
+        boxLeft = border[1].length;
+        boxRight = boxLeft + 1 + border[2].length;
+    }
+    if (borderIndex === -1)
+        return null;
+    let rows = 0;
+    let matchedText = '';
+    for (let i = 0; i < lines.length; i++) {
+        const line = lines[i];
+        if (i !== borderIndex) {
+            const gutter = COMPOSER_GUTTER_ROW.exec(line);
+            // Same indent as the border, or it is not this box's row. At 200 columns
+            // opencode centres the home screen's box (measured: `boxLeft` 63), so an
+            // indent-agnostic test would count transcript rows as box rows.
+            if (!gutter || gutter[1].length !== boxLeft)
+                continue;
+        }
+        // Cut at the COLUMN, not at the UTF-16 unit — see {@link indexAtColumn} for
+        // which of the two ways those disagree can invent a column that is not
+        // there. A box that owns the full pane width slices to '' whatever is
+        // written in it.
+        const tail = line.slice((0, terminal_columns_1.indexAtColumn)(line, boxRight)).trim();
+        if (tail === '')
+            continue;
+        rows += 1;
+        if (matchedText === '') {
+            matchedText = (0, excerpt_1.truncateToByteBudget)(tail, exports.OPENCODE_SIDEBAR_EXCERPT_MAX_BYTES, exports.OPENCODE_SIDEBAR_TRUNCATION_MARKER);
+        }
+    }
+    if (rows < exports.OPENCODE_SIDEBAR_MIN_ROWS)
+        return null;
+    return { id: exports.OPENCODE_SIDEBAR_OBSTRUCTION_ID, matchedText, boxRight, rows };
+}

package/dist/server/src/lib/detection/status-detector.js

@@ -42,6 +42,13 @@ Object.defineProperty(exports, "STATUS_REASON", { enumerable: true, get: functio
  */
 exports.SELECTION_LIST_REASONS = new Set([
     status_reason_1.STATUS_REASON.OPENCODE_SELECTION_LIST,
+    // Issue #2112: opencode's session list / agent list / timeline / command
+    // palette, recognised from the rectangle they are painted as rather than from
+    // their headings. Like every other opencode overlay they are driven by ↑/↓ +
+    // Enter and dismissed with `esc`, which is what NavigationButtons sends —
+    // and, critically, they must never be read as a finished turn: the marker
+    // from the PREVIOUS turn is still on the pane behind them.
+    status_reason_1.STATUS_REASON.OPENCODE_MODAL_OVERLAY,
     // Issue #1893: opencode's permission dialog is a horizontal button strip that
     // only ←/→ + Enter drive — typing an option number does nothing (measured on
     // 1.18.21), so it is a `menu`, not something `respond <id> N` can answer.

package/dist/server/src/lib/detection/status-reason.js

@@ -23,6 +23,26 @@ exports.STATUS_REASON = {
     THINKING_INDICATOR: 'thinking_indicator',
     OPENCODE_PROCESSING_INDICATOR: 'opencode_processing_indicator',
     OPENCODE_SELECTION_LIST: 'opencode_selection_list',
+    /**
+     * Issue #2112: a background-painted rectangle with an `esc` hatch was drawn
+     * over the transcript — opencode's session list, agent list, timeline,
+     * command palette or picker.
+     *
+     * Distinct from {@link STATUS_REASON.OPENCODE_SELECTION_LIST} because the two
+     * are different READINGS of the same screen: that one is the allowlisted
+     * heading (`Select model` / `Select provider` / `Connect a provider`, #1896)
+     * found as text and survives ANSI stripping; this one is the layout and needs
+     * the SGR the capture was taken with. An operator reading `capture --json`
+     * can tell which one answered, which is the difference between "the heading is
+     * one we know" and "something modal is on the pane".
+     *
+     * The literal is restated from `OPENCODE_MODAL_OVERLAY_ID`
+     * (`lib/detection/opencode-modal-overlay.ts`) rather than imported, so this
+     * module keeps the no-imports property its docblock above depends on;
+     * `tests/unit/detection-opencode-modal-overlay-2112.test.ts` pins the two
+     * equal so the restatement cannot drift.
+     */
+    OPENCODE_MODAL_OVERLAY: 'opencode_modal_overlay',
     /** Issue #1893: opencode's permission dialog — an arrow-key button strip. */
     OPENCODE_PERMISSION_PROMPT: 'opencode_permission_prompt',
     CLAUDE_SELECTION_LIST: 'claude_selection_list',