[Bug]: Cmd+C copies the entire conversation instead of the selection

[ggllzz04]

GUI version

0.1.40

Operating system

macOS (Apple silicon)

Reproduction steps

1. Open a conversation in which a single chat row can fill the mounted window — e.g. a turn whose work trace is expanded, a turn containing a large inline diff, or a very long answer.
2. Press Ctrl+O ("Expand or collapse all traces") to expand every trace, or click a turn's "Worked for … · N steps" line to expand just that turn.
3. Select a few characters inside that turn. A small selection is enough; it does not need to span the conversation.
4. Press Cmd+C and paste into any text editor.

Expected behavior

Cmd+C copies exactly the current selection. A whole-conversation transcript should only be produced by an explicit select-all action.

Actual behavior

The clipboard receives a plain-text transcript of the entire session — every message, each prefixed with User: / Assistant: — including messages that were never rendered on screen and were never part of the selection. The result is indistinguishable from the app's own select-all. Once it happens it repeats on every subsequent Cmd+C until the selection is cleared, because nothing clears the feed selection after the override.

Contributing factors, from reading the bundled renderer:

• The feed installs a copy listener that builds a transcript of the whole message list and replaces the clipboard when a geometric predicate passes: the current DOM selection must partially contain (containsNode(row, true)) both the first and the last currently mounted [data-feed-row] element.
• Because the feed is virtualized (increaseViewportBy: {top: 800, bottom: 800}), the mounted rows are only the viewport ± 800 px window. When one tall row fills that window, the first and last row are the same element, so any selection inside that row satisfies the predicate — even three characters. Tall rows are easy to produce: expanding a trace renders the full reasoning text, tool calls and output cards inside a single row.
• The predicate is evaluated against the rendered window, but the payload is always the entire session. So the clipboard can contain messages that were never rendered, let alone selected.

Related, same root cause: Mod+C is bound to run.interrupt ("Stops the agent when nothing is selected; copies otherwise"), and the Cmd+A handler only skips INPUT / TEXTAREA / contentEditable targets. Key events originating inside a shadow root are retargeted to the host element, so pressing Cmd+A while focus is inside the diff viewer (diffs-container, an open shadow root) is not excluded — it selects the entire feed, which then trips the transcript override on the next Cmd+C.

Screenshots, recordings, or sanitized logs

Sanitized reconstruction of the predicate (minified identifiers omitted), verified in a stock Chromium 153 harness using the same DOM shape ([data-feed-row] rows inside a scroll container):

const sel = feedEl.ownerDocument.getSelection();
if (!sel || sel.rangeCount === 0 || sel.isCollapsed) return false;
const rows = feedEl.querySelectorAll("[data-feed-row]");
const first = rows[0], last = rows[rows.length - 1];
return sel.containsNode(first, true) && sel.containsNode(last, true);

setup	predicate result
6 rows mounted, a few characters selected in a middle row	false
6 rows mounted, selection spanning the first to the last mounted row	true
1 row mounted (tall row), 3 characters selected inside it	true
selection made inside an open shadow root	false

Also verified: when the row containing the selection is unmounted, Chromium collapses the selection — so detached/stale selections are not the cause; the geometry alone is.

Fix direction: gate the transcript override on an explicit flag set by the app's own select-all handler instead of inferring intent from selection geometry; if geometry is kept, require full containment (containsNode(row, false)) and emit only the rows actually covered by the selection; clear the feed selection after an override; skip the override when the selection's range root is not the document.

Safety check

• I removed credentials and private project information from this report.

[ggllzz04]

Adding the diagnostics for the session this was observed in, in the same shape /feedback appends them, in case the traces help:

Session: 854effe6-7c36-4cf5-b8e1-6c7a17453ae9
Traces: 086c5e7c507a908488f9aa2dfa8173ca, d66fd11bac23eccdafb281447a7e0c21, 9be2bbd58c46ac0a1d81ca73931293cb, 7585b9ad84a7be813804ef8f029e60a0, 6802618225025feaf9fa905fdb9a28e0

For context on the repro: in this session the misbehaviour shows up once traces are expanded (Ctrl+O) and the feed is scrolled to the bottom of the last turn — i.e. exactly the state where the mounted window shrinks to a single [data-feed-row], making the predicate in the report trivially true for any selection.

[natenjo]

Same on Windows v.0.1.40 ..really anoying.