Skip to content

feat: edit queued chat messages - #29555

Merged
mafredri merged 32 commits into
mainfrom
mathias/codagt-971-fe-3-edit-queued
Nov 2, 2026
Merged

mafredri merged 32 commits into
mainfrom
mathias/codagt-971-fe-3-edit-queued

Conversation

@mafredri

@mafredri mafredri commented Sep 18, 2026 •

Copy link
Copy Markdown
Member

Queued rows gain Edit and Cancel edit actions. Edit opens the composer at once and marks the row on the server (editing_since). Save patches the content and clears the marker; Cancel only clears the marker. On a paused chat, Save sends the edited head and Cancel sends it unchanged. A chat that loads with a row under edit reopens it in the composer. While the chat is paused, Edit on other rows is unavailable, and hovering or focusing it shows why. The row the composer edits shows Cancel edit instead of Edit. While a turn streams, a history edit keeps Stop in the send slot and a queued edit shows Save Edit.

The composer target is derived on every render from the row the store marks, the composer mode the user chose in this session ("follow", "draft", or a history or queued row), and the state of the latest marker request. In "follow", the composer edits the marked row, so a reload or chat switch restores edit mode in the same render the store gets the row, and every tab shows the same edit. A queued mode holds while the server marks that row or the request to begin its edit is in flight, and closes when the server no longer marks it. The page shows edit mode only for the row the editor holds, and never for an edit the server does not have.

When the marked row changes, a modified input (text or file-reference chips that differ from what was last loaded into the editor) stays as a new-message draft and no edit opens. An unmodified input, including text typed and deleted again, gets back the draft from before the edit, captured once per edit session, and the composer follows the marked row.

Text typed in a queued edit is saved in local storage per row, like a new-message draft, so a reload or chat switch brings it back. It is saved with the row content it was typed against and loads only while the row still has that content. Text a tab loaded from the saved edit and did not change counts as unchanged when the edit ends, so a tab that followed another tab's edit gets its own draft back.

Editing a sent message also keeps its skill and file-reference chips now.

For review: deriveComposerTarget and deriveQueuedMessageUnderEditID in useQueuedMessageEdit.ts are pure functions with table tests. The hook holds one piece of state, unappliedEdit, which keeps a refused Edit's row as the target until the refusal restores the earlier composer mode. useConversationEditingState owns the editor text and the pre-edit draft and loads the editor in one layout effect. The page owns composerMode and the saveQueuedMessage and setQueuedMessageEditing requests, whose store updates are in runEditQueuedMessage (chatQueueReconciliation.ts). A successful request that sets editing writes the marker into the store without waiting for queue_update, and a 404 drops the row. Marker requests share a mutation scope with save, so an end or save never runs before its begin, but use a separate mutation instance, so they do not mark the composer pending.

Limitations:

  • A 409 (another client began an edit on a different row of the paused chat) is shown as a toast with the server's message.
  • Cases that drop user content: switching to another row during a modified edit replaces the modified text; when the server closes a modified edit, the draft from before the edit is lost; a queued save that fails after another edit opened rolls the composer back to the saved row. Attachments added during an edit are reset when it closes.
  • Discarding a modified edit does not ask for confirmation, and Escape on a modified edit cancels it without asking (CODAGT-1513).
  • Edit can race with a pending send, a 404 refusal, or Send now on the same row (CODAGT-1519).
  • queue_update has no version, so the client cannot order it against its own writes. A queue snapshot the stream read before an edit began can arrive after the begin returned; the composer treats it as the edit ending and keeps typed text as a new-message draft until Edit reopens the row (CODAGT-1503).

Linear: https://linear.app/codercom/issue/CODAGT-971

Part of the CODAGT-971 frontend stack: 3 of 3.

🤖 This PR was created with the help of Coder Agents on behalf of @mafredri, and will be reviewed by a human. 🏂🏻

@linear-code

linear-code Bot commented Sep 18, 2026

Copy link
Copy Markdown

CODAGT-971

@mafredri

This comment has been minimized.

@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Sep 18, 2026 •

Copy link
Copy Markdown

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-10-30T15:21:35.883723Z be28da4 Manual request
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 60bb342e91

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment on lines +941 to +944
const request: TypesGen.EditChatQueuedMessageRequest = {
content,
model_config_id: editSelectedModelConfigID,
reasoning_effort: editReasoningEffort,

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Preserve the queued row's model on content-only edits

When a queued row was created with a different model from the chat's current picker value, saving a text-only edit passes that picker value to resolveEditModelConfigID, which treats the mismatch as a user-requested model change. This silently rewrites model_config_id even though the user never touched the selector, which is especially likely when multiple queued messages were created with different models. Track whether the selector changed during this edit, as is already done for reasoning effort, or initialize it from the queued row before including an override.

Useful? React with 👍 / 👎.

@mafredri mafredri Sep 18, 2026 •

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Partially valid: the rewrite is real, but it is the existing history-edit behavior on main (resolveEditModelConfigID is a verbatim extraction, and no model dirty flag exists anywhere). Fixing only queued rows would create exactly the divergence this warns about, so it is tracked for both paths in https://linear.app/codercom/issue/CODAGT-1075 rather than in this stack.

🤖 Reply by Coder Agents on behalf of @mafredri.

Comment on lines +668 to +670
queuedEditRestoredRef.current = true;
const row = chatQueuedMessages.find((message) => message.editing_since);
if (!row || isViewerNotOwner) {

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Wait for fresh queue data before consuming restoration

When React Query has cached messages from before another client began editing, this effect runs immediately with that stale queue, permanently sets queuedEditRestoredRef, and then ignores the authoritative background-refetch result containing editing_since. Revisiting such a chat therefore fails to reopen its server-side edit as intended. Only mark restoration complete after the current mount has received authoritative queue data, or reconsider the marker when the query data timestamp advances.

Useful? React with 👍 / 👎.

@mafredri mafredri Sep 18, 2026 •

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Valid for the code this was posted on. The head has no one-time restore: deriveComposerTarget in useQueuedMessageEdit.ts derives the composer target on every render, and a composer in "follow" (untouched) targets the row the store marks. When a refetch or queue_update brings editing_since after cached data, the composer opens that row.

🤖 Reply by Coder Agents on behalf of @mafredri.

Comment on lines +116 to +120
// Leaves edit mode with the composer text kept as the draft.
const leaveEdit = () => {
setEditingTarget(null);
setEditingFileBlocks([]);
};

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Persist text when an external update ends editing

When another client ends or moves the queued edit, leaveEdit makes the current editor text look like the new draft but never writes its text or serialized state to draft storage. If the user reloads or navigates away before typing another character, the retained text is lost and the pre-edit draft is restored instead. Persist the current editor state when leaving, and clear the obsolete draftBeforeEdit, so the visually retained content is actually treated as the draft.

Useful? React with 👍 / 👎.

@mafredri mafredri Sep 18, 2026 •

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Valid. When the server stops marking the row and the edit was modified, loadTargetIntoEditor in useConversationEditingState.ts keeps the text as a new-message draft and writes it to draft storage. A change counts as modified when the text or the editor state, such as a file-reference chip, differs from what was loaded. The "when the server ends the edit" table in useConversationEditingState.test.ts asserts the stored draft. Keeping the draft from before the edit as well is still open (CRF-8).

🤖 Reply by Coder Agents on behalf of @mafredri.

Comment on lines +345 to +347
{onEdit && (
<QueuedMessageActionButton
label="Edit"

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Reveal queued actions when they receive keyboard focus

For every non-head row, this new Edit control is rendered inside the existing opacity-0 action container until pointer hover, but tabbing to the button does not update hoveredID; keyboard users therefore focus and activate an invisible edit control. Reveal the action group with focus-within or remove hidden controls from the tab order until they are displayed.

AGENTS.md reference: site/AGENTS.md:L61-L65

Useful? React with 👍 / 👎.

@mafredri mafredri Sep 18, 2026 •

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Valid; Send now and Remove had the same gap on main. The actions container in QueuedMessagesList.tsx has focus-within:opacity-100, as AttachmentPreview does, so an action that receives keyboard focus is visible.

🤖 Reply by Coder Agents on behalf of @mafredri.

Comment on lines +235 to +239
expect(editBehind).toBeDisabled();
await user.hover(editBehind);
expect(
await screen.findByText("Finish the current edit first."),
).toBeInTheDocument();

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Move tooltip rendering assertions to Storybook

This behavior test asserts the tooltip's rendered presence with toBeInTheDocument, although FE1 explicitly reserves rendered visual states for Pixel-backed stories and requires Vitest assertions to target non-visual outcomes. Drive the disabled-tooltip state in the existing paused story and keep this test focused on whether the edit callback is or is not invoked.

AGENTS.md reference: site/AGENTS.md:L5-L7

Useful? React with 👍 / 👎.

@mafredri mafredri Sep 18, 2026 •

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Valid. Edit on a row behind a paused head is aria-disabled and stays focusable, and its tooltip gives the reason. QueuedMessagesList.test.tsx focuses it, checks toHaveAccessibleDescription("Finish the current edit first."), and checks that a click does not call onEdit; the PausedEditBehindHead story hovers it for the screenshot.

🤖 Reply by Coder Agents on behalf of @mafredri.

Comment on lines +12 to +16
export const trackQueuedEditTarget = (
targetID: number | null,
row: TypesGen.ChatQueuedMessage | undefined,
seenID: number | null,
): { seenID: number | null; lost: boolean } => {

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do we need to be able to handle | null and | undefined on every single parameter?

@mafredri mafredri Sep 18, 2026 •

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The helper this was posted on is not in the current head. The functions this stack adds to chatQueueReconciliation.ts, runDeleteQueuedMessage and runEditQueuedMessage, take required parameters. The composer target is derived in deriveComposerTarget (useQueuedMessageEdit.ts) from the composer mode, the marked row ID (number | null, null when no row is marked), and the latest marker request.

🤖 Reply by Coder Agents on behalf of @mafredri.

Comment on lines +659 to +676
const queuedEditRestoredRef = useRef(false);
useEffect(() => {
if (
queuedEditRestoredRef.current ||
chat === undefined ||
chatQueuedMessages === undefined
) {
return;
}
queuedEditRestoredRef.current = true;
const row = chatQueuedMessages.find((message) => message.editing_since);
if (!row || isViewerNotOwner) {
return;
}
const { text, fileBlocks } = getEditableContentPayload(row.content);
isEditReasoningEffortDirtyRef.current = false;
editing.handleBeginEdit({ kind: "queued", id: row.id }, text, fileBlocks);
}, [chat, chatQueuedMessages, isViewerNotOwner, editing]);

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have no idea what is happening here. My immediate thought is to say something is wrong here (although I could be wrong). Could you have an agent confirm that this is definitely valid?

@mafredri mafredri Sep 18, 2026 •

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That effect is not in the current head. Edit mode is restored by derivation: while the composer is in "follow" (the user has not typed or chosen anything), deriveComposerTarget in useQueuedMessageEdit.ts targets the row the store marks for the chat owner, and the layout effect in useConversationEditingState.ts loads that row into the editor when the target changes. That covers a reload, a chat switch, and a row marked from another tab. The first keystroke leaves "follow", so a later marker does not replace typed text.

🤖 Reply by Coder Agents on behalf of @mafredri.

// under edit opens the composer on it. A marker that arrives later
// from another client does not, as that would replace what the user
// is typing here.
const queuedEditRestoredRef = useRef(false);

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

do we need this as a ref?

@mafredri mafredri Sep 18, 2026 •

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The ref is gone. The composer target is derived on every render from the store marker, the composer mode, and the marker request (deriveComposerTarget in useQueuedMessageEdit.ts), so no once-per-mount flag is needed. The first keystroke moves the composer out of "follow" (handleContentChange in useConversationEditingState.ts), so a marker that arrives later from another tab does not replace typed text.

🤖 Reply by Coder Agents on behalf of @mafredri.

@mafredri
mafredri force-pushed the mathias/codagt-971-fe-3-edit-queued branch 2 times, most recently from 4c2cc30 to 1b8fceb Compare September 18, 2026 13:12
@mafredri
mafredri force-pushed the mathias/codagt-971-fe-3-edit-queued branch from 1b8fceb to d4de206 Compare September 18, 2026 13:27
@mafredri
mafredri force-pushed the mathias/codagt-971-fe-3-edit-queued branch from d4de206 to 17cd798 Compare September 18, 2026 14:05
@mafredri
mafredri force-pushed the mathias/codagt-971-fe-3-edit-queued branch from 17cd798 to b608b0b Compare September 18, 2026 15:48
@mafredri
mafredri force-pushed the mathias/codagt-971-fe-3-edit-queued branch from b608b0b to f938d82 Compare September 18, 2026 18:25
@mafredri
mafredri force-pushed the mathias/codagt-971-fe-3-edit-queued branch from f938d82 to c14a860 Compare September 18, 2026 18:36
@mafredri
mafredri force-pushed the mathias/codagt-971-fe-3-edit-queued branch from c14a860 to 471bf40 Compare September 21, 2026 23:07
@mafredri
mafredri force-pushed the mathias/codagt-971-fe-3-edit-queued branch 3 times, most recently from 29272f9 to fca5a0c Compare September 22, 2026 11:52
Loading a row into the composer remounts the Lexical editor, and two
of its updates reached handleContentChange after the hook had reset
its refs for the new editor. The outgoing editor commits once more
when its root detaches, before the listener is removed, and an editor
superseded within the same commit chain seeded from a passive effect
after the parent's next load. Either one overwrote inputValueRef and
the serialized state, so the new editor's seed looked like user input.
An untouched edit then counted as modified: it kept the row's text as
a draft instead of following another client's edit or giving the
pre-edit draft back.

ContentChangePlugin now ignores updates of an editor without a root
and registers in a layout effect; ValueSyncPlugin seeds synchronously
in a layout effect, so each editor reports its seed in the commit that
mounts it. The queued edit tests for a server-side end or move now run
against the real editor.
An edit counted as modified from the first change after it loaded, and
a composer in follow mode switched to draft on its first keystroke.
Neither flag went back, so typing and deleting the text again left the
composer holding the old row as a draft when another client moved the
edit, or never following another client's edit at all.

The composer now compares the editor with the content last loaded into
it, by text and file-reference chips, when the marked row changes. A
changed editor keeps its content as before; an unchanged one follows.
Follow mode no longer switches on keystrokes, so the composer only
leaves it when a row change finds the editor changed.

The file-reference cases now run against the real editor, which also
covers a stored draft that loses one of its chips.
The move tests reached the first row through Edit, so nothing checked
that a composer following the marked row keeps its changed text when
the marker moves. Two rows now open the first row by following the
marker: typed text is kept, and text deleted again follows the next
row. Every row starts from a stored draft so the composer is unchanged
when the first marker arrives.
A send clears the editor, and the composer must then count as
unchanged. The test starts from a stored draft, so without the reset
the empty editor would differ from the draft and not follow.
When a marker arrived at a changed composer in follow mode, one commit
reported the marked row as the editing target before the load effect
declined it. That commit unmounted the row's focused Edit button,
dropping focus to the page, and dropped the slash commands, which can
close an open "/" menu for good.

The hook now reports the target the editor holds. A load still runs
in a layout effect, so the edit chrome appears before paint.
…ory edits

Opening an edit loaded only the row's text, so a queued row's or past
message's skill chips were missing from the composer and a save dropped
them. The editor now loads the serialized state getEditableContentPayload
builds from the content, and the modified check compares skill chips in
document order alongside file references.

ValueSyncPlugin keeps restoring a serialized state with a deferred update,
as main does, because a discrete commit there makes chip decorators
flushSync inside the effect; only the plain-text seed is discrete.
Opening an edit of a history message or a queued row built the editor
state from text and skill chips only, so file-reference parts never
reached the composer and saving sent the content without them.
serializeMessageDraft now rebuilds FileReferenceNodes in document order,
and getEditableContentPayload builds the state whenever the content has
a skill or file-reference part. The composer already sends file
references from the editor, so a loaded chip round-trips and a removed
one is dropped.
QueuedMessagesList keys the row under edit on queuedMessageUnderEditID
and never reads editing_since, so the Cancel edit test now uses an
unmarked row and absorbs the separate pending-begin case.

buildEditModelOverrides tests repeated resolveEditModelConfigID and the
submitChatTurn edit test. The edit test now covers both dirty states,
so the function no longer needs an export for testing.

The file-reference-only reopen case is the only one without a skill
chip, which is what makes getEditableContentPayload build editor state
for file references. Its name now says so.
…remounts

Text typed into a queued edit lived only in the editor, so a reload or a
switch to another chat and back loaded the row's server content over
it. The edit's editor state is now stored per row next to the
new-message draft and loaded instead of the row's content. It still
counts as typed, so a marker that leaves keeps it as a new-message
draft. The stored state goes away when the edit is saved, cancelled or
loses its row.
…emoving its row

Remove on the row the composer edited deleted the row while the editor
still held its text, so the typed text stayed behind as a new-message
draft and the draft from before the edit was lost. Remove now closes
the edit first, as Cancel edit does.
…ending its row

Send now on the row the composer edited promoted the row while the
editor still held its unsaved text, which then stayed behind as a
new-message draft. Send now now closes the edit first, like Remove, so
the server sends the row's saved content and the draft from before the
edit comes back.
The model picker shows the chat's model, not the queued row's, so a
queued save sent the picker's model whenever it differed from the
row's and changed the row's model without the user picking one. A
queued save now sends model_config_id only when the picker changed
during this edit, tracked like the reasoning effort.

History edits keep their rules: they also fall back to a usable model
when the message's model is no longer available.
…used Edit

When the server refused to begin an edit (409 or another non-404
error), the target went away and the editor switched the composer to
follow mode, so it opened the row as soon as another client began
editing it. The composer now returns to the mode it had before Edit,
unless a later choice replaced it, and the editor keeps a draft mode
it finds when the target goes away.
…s the row

On a paused chat, cancelling the queued edit sends the row unchanged,
which the row action already says. The composer's cancel button now
says the same. History edits keep "Cancel editing", since cancelling
one sends nothing.
react-query reports a settled request to the page on its notify
scheduler, separately from rejecting the request's promise. The
refused-Edit restore only worked because the default scheduler runs
after the promise's handlers. When the page saw the failure first, the
editor closed the target, switched the composer to follow, and later
opened another client's edit of the row.

The Edit's choice now stays the composer's target until the begin's own
handlers have applied the outcome, so the editor sees the restored mode
whichever order react-query uses.

@DanielleMaywood DanielleMaywood left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This review was performed by ai-review, an automated review bot.

10 inline findings and 1 below.

whole PR

SCOPE-1: This PR changes 4772 lines in 33 files (limit 1000). Split it into smaller PRs.

Remedy: split

Coverage
Coverage: 32 of 34 files reviewed, 1 unsupported, 1 deleted; 35 rules.
Candidates: 880 found, 743 outside changed lines, 6 already discussed, 131 evaluated.
Findings: 11 (40 sites, 8 pre-existing sites), 0 over budget, 0 failed.
Model: deepseek/deepseek-v4.1-flash; 103 questions, 78 answered, 25 abstained (truncated 25), 0 cached; 0 uncalibrated findings held back; 90900 input and 98508 output tokens.
Experiments: 4 run, 0 passed, 4 vetoed, 0 unverified, 0 skipped for time.
  vetoed: TYPE-1 site/src/pages/AgentsPage/components/QueuedMessagesList.tsx:128: 3 new type errors, first src/pages/AgentsPage/components/AgentChatInput.tsx: TS2345 Argument of type 'number 
  vetoed: TYPE-1 site/src/pages/AgentsPage/hooks/useQueuedMessageEdit.ts:33: 1 new type error, first src/pages/AgentsPage/hooks/useQueuedMessageEdit.ts: TS2345 Argument of type 'number 
  vetoed: TYPE-1 site/src/pages/AgentsPage/hooks/useQueuedMessageEdit.ts:56: 4 new type errors, first src/pages/AgentsPage/hooks/useConversationEditingState.test.tsx: TS2345 Argument of type 'number 
  vetoed: TYPE-1 site/src/pages/AgentsPage/utils/draftStorage.ts:116: 6 new type errors, first src/pages/AgentsPage/hooks/useConversationEditingState.test.tsx: TS2345 Argument of type 'string 

store={store}
initialMessages={chatMessagesList ?? []}
editing={{ ...editing, handleBeginHistoryEdit }}
editing={{ ...editing, handleBeginHistoryEdit, handleCancelEdit }}

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

READ-1: Nested ternary. Use early returns, an if/else, or a lookup instead.

Remedy: split

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changing it here: this ternary is on main unchanged; the PR only touches a prop inside it.

🤖 Reply by Coder Agents on behalf of @mafredri.


// Behavioral: clicking edit fires onEditUserMessage with the
// correct message ID and text.
// message ID.

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

TEST-1 (pre-existing, non-blocking): Story UserMessageCopyButton asserts in its play function. Stories are covered by visual regression; keep play to state setup and move behaviour assertions to a Vitest test.

Remedy: move

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pre-existing story, outside this PR's change. Leaving it.

🤖 Reply by Coder Agents on behalf of @mafredri.

"cursor-not-allowed text-content-disabled hover:bg-transparent hover:text-content-disabled",
)}
>
<Spinner className="h-3.5 w-3.5" loading={busy}>

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

STYLE-1 (pre-existing, non-blocking): Use size-3.5 instead of w-3.5 h-3.5.

Remedy: simplify

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pre-existing line, outside this PR's change. Leaving it.

🤖 Reply by Coder Agents on behalf of @mafredri.

const [selectedModel, setSelectedModel] = useState("");
const [selectedReasoningEffort, setSelectedReasoningEffort] = useState("");
const isEditReasoningEffortDirtyRef = useRef(false);
const isEditModelDirtyRef = useRef(false);

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

READ-2 (pre-existing, non-blocking): 21 lines and 12 statements with no blank line between them. Separate the logical steps with blank lines.

Remedy: split

Same at 5 other sites:

  • site/src/pages/AgentsPage/AgentChatPage/AgentChatPage.tsx:336: 59 lines and 15 statements with no blank line between them. Separate the logical steps with blank lines.
  • site/src/pages/AgentsPage/AgentChatPage/AgentChatPage.tsx:525: 38 lines and 8 statements with no blank line between them. Separate the logical steps with blank lines.
  • site/src/pages/AgentsPage/components/AgentChatInput.tsx:1159: 81 lines and 11 statements with no blank line between them. Separate the logical steps with blank lines.
  • site/src/pages/AgentsPage/components/AgentChatInput.tsx:1358: 45 lines and 10 statements with no blank line between them. Separate the logical steps with blank lines.
  • site/src/pages/AgentsPage/components/ChatConversation/submitChatTurn.ts:385: 53 lines and 12 statements with no blank line between them. Separate the logical steps with blank lines.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

All six sites are pre-existing code this PR does not change. Leaving them.

🤖 Reply by Coder Agents on behalf of @mafredri.

@mafredri

Copy link
Copy Markdown
Member Author

SCOPE-1: keeping this as one PR. It is the last of the CODAGT-971 frontend stack; #29553 and #29554 already split out the refactors and the paused display, and the remaining edit flow does not work in parts.

🤖 Reply by Coder Agents on behalf of @mafredri.

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm concerned with how we're adding an extra 135 lines to this already bloated page. Not going to block with this but thoughts on trying to pull some logic into children components?

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed the page is too big. The queued-edit code here can't move into children on its own: the composer target, save pending state and queued save all feed the editing state and send path that #31476 keeps in the page. Shrinking the page means moving the whole composer, which is your #30638's "composer as a node" step. I'd rather do that after this lands than add it to a 4.7k-line PR; happy to take it on or build on #30638.

🤖 Reply by Coder Agents on behalf of @mafredri.

Comment on lines +126 to +129
export const isQueuedMessageUnderEdit = (
message: ChatQueuedMessage,
queuedMessageUnderEditID: number | null,
): boolean => message.id === queuedMessageUnderEditID;

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

uhh?

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I mean.. it gave a descriptive name to the check 😂, definitely couldn't have been a variable!

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It was a one-line ID comparison left over from when it also read editing_since. Inlined at its three call sites and removed.

🤖 Reply by Coder Agents on behalf of @mafredri.

@mafredri

Copy link
Copy Markdown
Member Author

@codex review
/coder-agents-review

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: be28da4661

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment on lines +204 to +208
store.setQueuedMessages(
store.getSnapshot().queuedMessages.map((row) => {
if (row.id === id) {
return { ...row, editing_since: editingSince };
}

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Skip stale marker writes after queue updates

When another client moves the edit after this request commits but before editQueuedMessage resolves locally, its authoritative queue_update can already have marked a different row in the store. This unconditional post-await write then marks the older requested row again and clears the newer marker, leaving this client following a row the server no longer considers under edit with no later event guaranteed to repair it. Capture and check the queue convergence fence before applying this optimistic marker.

Useful? React with 👍 / 👎.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Real, but not changing it here. queue_update has no version, so the client cannot tell whether a snapshot that arrived during the request is newer or older than its own write. Skipping the write-back on any such update would fix this case but drop the write-back that corrects a stale snapshot read before the begin committed. Both need a queue version; I added this sequence to CODAGT-1503, which the PR's Limitations already link.

🤖 Reply by Coder Agents on behalf of @mafredri.

Comment on lines +706 to +710
closeQueuedEditOf(id);
return requestDeleteQueuedMessage(id);
};
const handlePromoteQueuedMessage = (id: number) => {
closeQueuedEditOf(id);

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Restore the queued edit when an action fails

When the user has unsaved changes in a queued edit and Delete or Send now fails due to a network, conflict, or server error, closeQueuedEditOf has already restored the old draft and removed the persisted queued-edit state. Although the reconciliation path restores the queued row, nothing restores the user's edited content, so the failed action causes permanent input loss. Close the edit only after success or retain a rollback snapshot.

AGENTS.md reference: site/AGENTS.md:L11-L11

Useful? React with 👍 / 👎.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changing this. Remove and Send now discard the typed text on purpose: the user chose to drop the edit. If the request fails, the row comes back with its saved content and the user can retry the action.

🤖 Reply by Coder Agents on behalf of @mafredri.

@coder-agents-review coder-agents-review Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This round re-reviews queued-message editing at be28da4 (rebased, 30 fix commits): 1 P1, 4 P2, 10 P3, 11 P4 and 1 Nit are open, six of them re-raised from round 1.

The fixes the panel checked hold, including cancel routing, chips on edit, the transcript refetch and the dirty-ref reset; the disputes on silent draft conversion, the 404 toast, the delete refetch and the two naming nits are closed. Three PR Limitations have no ticket: the content-loss cases, the confirmation dialog, and Escape on a modified edit; this needs a human decision.

The chip fix for edits (898dd11) changes editing of sent messages on main and does not depend on the queued-edit flow, so it can ship as its own PR; if it stays, the description needs the addition the messageParsing.ts comment asks for.

Out of scope (needs a ticket or explicit acceptance by a human):

  • site/src/api/queries/chats.ts:2767: deleteChatQueuedMessage.onSuccess awaits invalidateChatMessages, so each queued-row delete refetches every loaded transcript page in sequence and stays pending until they all return.
  • coderd/exp_chats.go:3166: patchChatQueuedMessage returns 404 for a failed authorization check, a missing chat and a missing row with no machine-readable code, so a client can tell them apart only by matching message text.
  • site/src/pages/AgentsPage/hooks/useConversationEditingState.test.tsx:576: the comment moved from the deleted test file describes a past change ("now") instead of the behavior.

Notes:

  • A refused Edit on another row returns the composer to the row it was editing, but that row's typed text was deleted at the switch, so the composer shows the row's stored content; this is the row-switch case in Limitations.

🤖 This review was automatically generated with Coder Agents.

isOwner &&
serverMarkedID !== null
) {
editing.handleCancelEdit({ kind: "queued", id: serverMarkedID });

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P3 [CRF-46] The page's handleCancelEdit branch that switches a cancelled history edit to the server-marked queued row has no test. (Netero)

Replacing this line with editing.handleCancelEdit(); leaves AgentChatPage.test.tsx, useConversationEditingState.test.tsx, useQueuedMessageEdit.test.ts and submitChatTurn.test.ts green (252 passed), because the hook test at useConversationEditingState.test.tsx:1617 passes the target to the hook directly. If the branch regresses, the marked row does not reopen in this tab after a history cancel. Add an AgentChatPage.test.tsx case that begins a history edit while a queued row is marked, cancels, and expects the composer to hold the queued row's text.

🤖

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changing this in this PR; it is approved, and the remaining changes are limited to user-facing defects.

🤖 Reply by Coder Agents on behalf of @mafredri.

model_config_id: isEditModelDirtyRef.current
? effectiveSelectedModel || undefined
: undefined,
reasoning_effort: isEditReasoningEffortDirtyRef.current

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P4 [CRF-47] The queued save's dirty reasoning-effort branch has no test. (Netero)

Replacing lines 375 to 377 with reasoning_effort: undefined leaves the same 252 tests green; submitChatTurn.test.ts:311 covers only the clean case, and the dirty-effort table at line 367 runs only for history edits. Add the queued target to that table, or add a { current: true } case to the test at line 311.

🤖

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changing this in this PR; it is approved, and the remaining changes are limited to user-facing defects.

🤖 Reply by Coder Agents on behalf of @mafredri.

* resolveEditModelConfigID picks one, the effort unless the user changed
* it for this edit.
*/
const buildEditModelOverrides = ({

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P4 [CRF-48] buildEditModelOverrides has one caller since ac386e9 moved the queued branch off it. (Netero)

7321374 extracted it so both branches shared the model and effort logic; it now wraps one resolveEditModelConfigID call and one ternary for the history branch at line 391. Inline it back into the history branch, which also removes its doc comment at line 93.

🤖

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changing this in this PR; it is approved, and the remaining changes are limited to user-facing defects.

🤖 Reply by Coder Agents on behalf of @mafredri.


// Closes the composer's edit of the row and discards its typed text; the
// server acts on the row's saved content.
const closeQueuedEditOf = (id: number) => {

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P3 [CRF-49] closeQueuedEditOf discards the composer's edit before Send now or Remove runs, so when that request fails the row comes back still under edit and the rewrite is gone. (Hisoka, Kite)

On failure, runPromoteQueuedMessage and runDeleteQueuedMessage restore the queue snapshot, which still has the row's editing_since, so the server edit and the pause remain while the composer is in "draft". Reproduced with a temporary page test (type, Edit, type " edit", Send now with the promote rejecting): {"composer":"draft","editing":1,"edit":1,"cancelEditing":1,"storage":[]}. Fix: close the composer's edit only after the request resolves, as runEditQueuedMessage applies its store write only after success.

🤖

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Intended, same as Send now: the user gave up the edit. If the request fails, the row comes back with its saved content and the action can be retried.

🤖 Reply by Coder Agents on behalf of @mafredri.

finalizeSuccessfulSend(target);
if (sendTarget?.kind === "queued") {
removeQueuedEdit(sendTarget);
// Saving puts the row back in the queue, and on a paused chat's

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P4 [CRF-61] The save-tail comment says "Saving puts the row back in the queue", but the row never leaves the queue; a save clears editing_since. (Gon)

No code removes the row from queuedMessages during an edit, so the comment describes a step that does not happen. Rewrite: "Saving ends the edit, and on a paused chat's head it sends the row. The saved text is not a new message, so the pre-edit draft comes back unless another edit opened during the save."

🤖

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changing this in this PR; it is approved, and the remaining changes are limited to user-facing defects.

🤖 Reply by Coder Agents on behalf of @mafredri.

readOnlyNotice?: string;
isChatPaused?: boolean;
// Editing state, owned by the parent.
// Editing state, owned by the parent. The kind selects the banner text.

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P4 [CRF-62] The editingKind prop comment says the kind selects the banner text, but it also selects the cancel label and whether the edit holds the Stop slot. (Gon)

editingHoldsStop (line 1369) is true only for history edits, so a queued edit gives up Stop while a turn streams, which the comment does not tell the reader. Rewrite: "The kind selects the banner text, the cancel label, and whether the edit holds the Stop slot."

🤖

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changing this in this PR; it is approved, and the remaining changes are limited to user-facing defects.

🤖 Reply by Coder Agents on behalf of @mafredri.

// layout.
Object.defineProperty(Range.prototype, "getBoundingClientRect", {
configurable: true,
value: () => new DOMRect(0, 0, 1, 16),

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P4 [CRF-63] The Lexical jsdom shim new DOMRect(0, 0, 1, 16) and the type paste helper are copied between useConversationEditingState.test.tsx and AgentChatPage.test.tsx. (Gon)

Both copies are new in this PR (AgentChatPage.test.tsx:56, :119; here and at line 384), so a change to one leaves the two Lexical setups different. Move the shim and the helper into one shared test helper module, with 16 as a named line height.

🤖

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changing this in this PR; it is approved, and the remaining changes are limited to user-facing defects.

🤖 Reply by Coder Agents on behalf of @mafredri.


const loaded = loadedTarget?.target;
if (loaded?.kind === "queued") {
if (isEditorChanged() && (content.trim() || hasFileReferences)) {

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P4 [CRF-64] The queued-edit localStorage write in handleContentChange repeats writeDraft with a different key. (Zoro)

Both do the same non-empty check, setItem in a try with a quota catch, and removeItem otherwise (writeDraft at lines 142 to 156). Pass the key to writeDraft, then call writeDraft(queuedEditStorageKey(loaded.id), ...) when isEditorChanged() and removeQueuedEdit(loaded) otherwise.

🤖

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changing this in this PR; it is approved, and the remaining changes are limited to user-facing defects.

🤖 Reply by Coder Agents on behalf of @mafredri.


type SerializedNode = { type?: string; children?: SerializedNode[] };

const chipNodeTypes: ReadonlySet<string> = new Set([

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P4 [CRF-65] chipNodeTypes is a second list of the chip node types; isChipNode in fileReferenceDisplay.ts:9-12 already lists the same two. (Zoro)

FileReferenceNode.getType() returns the same "file-reference" string the serialized state uses, so both lists describe one set, and a third chip type added to only one of them makes isEditorChanged miss changes to it. Export chipNodeTypes from one module and make isChipNode (node) => chipNodeTypes.has(node.getType()).

🤖

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not changing this in this PR; it is approved, and the remaining changes are limited to user-facing defects.

🤖 Reply by Coder Agents on behalf of @mafredri.

… was typed against

A tab following another tab's queued edit loaded that tab's saved typed
text and counted it as its own change, so when the edit ended it kept
the text as a new-message draft and overwrote the shared draft. Saved
text could also load over a row another client had saved since.

The saved edit now records the row content it was typed against and
loads only while the row still has it. Text loaded from the saved edit
and not changed in this tab counts as unchanged when the edit ends.
@mafredri

mafredri commented Nov 2, 2026

Copy link
Copy Markdown
Member Author

Reply to the review summary:

  • The history chip fix stays here: queued edits load content through the same path, so the description now says editing a sent message keeps its chips.
  • The content-loss cases, the confirmation dialog and Escape on a modified edit are tracked in CODAGT-1513.
  • Out of scope: the delete refetch is on main and tracked in CODAGT-1512; the untyped 404s and the test comment are not changing in this PR.

🤖 Reply by Coder Agents on behalf of @mafredri.

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants