Skip to content

fix(chat): show the answer as it streams, and keep a model override across /clear#49

Merged
dnviti merged 1 commit into
chore/version-bump-5.1.2from
fix/5-1-2-regressions
Jul 26, 2026
Merged

fix(chat): show the answer as it streams, and keep a model override across /clear#49
dnviti merged 1 commit into
chore/version-bump-5.1.2from
fix/5-1-2-regressions

Conversation

@dnviti

@dnviti dnviti commented Jul 26, 2026

Copy link
Copy Markdown
Owner

Five defects introduced by the four PRs already merged onto chore/version-bump-5.1.2 (#39, #40, #41, #47). Each PR's own tests passed; all but one lived where two of them met — which is why every gate was green while the chat was visibly broken.

Closes #48.

The chat stopped rendering

User-reported, and the worst of the five. MessageList looked messages up in a messageById map memoised on [messages] — but the transcript appends to its array in place, so that array's identity never changes. The map was built once at mount and every message added afterwards resolved to undefined and rendered nothing.

The turn strips kept moving (they come from a memo that does depend on version), so the header, tool counts and cost ticked away above a body that stayed empty, and a reload made everything appear at once — because reloading replaces the array.

Keyed on version instead. This does not make the list re-render per token: apply() bumps version on every event but only notifies the list listeners on structural ones, so the map is rebuilt at exactly the cadence the neighbouring messages memo already used.

A model override no longer survives /clear

/clear restarts the agent in place from the options the session was launched with, and the model is the one thing in them that can change while the session is alive. So a conversation quietly went back to the model it opened with — after the browser had already been told the switch was applied, and while the record still held the override.

The choice now reaches those options by either door: the picker, and a /model typed straight into the composer, which was forwarded untouched and hit the same reversion.

A collapsed turn's title escaped its bar

It set neither a font size nor nowrap, while every sibling in that fixed 28px row sets both — so it inherited the page default and wrapped, and without nowrap the ellipsis was inert. Measured at 924px: 68px of content in a 28px bar.

An override could not be undone

The server had a cleared path, user-facing message and all, that nothing in the UI could reach — the text field refuses to submit empty and every listed model carries a name. A typo therefore stayed in force for every later launch of that conversation. There is now a way back to the runtime's default.

Model names are also capped and stripped of control characters, which would otherwise ride into the best-effort /model turn as extra lines.

The browser checks gated nothing

npm test runs only the unit suite; the browser checks were a separate script no workflow invoked, and the runner exited 0 when Chrome was absent. With the rendering bug restored, npm test passes 1357/1357 while the browser check fails. They now run in CI and refuse to skip there.

Their fixture also grew a second turn — without one, no turn is ever collapsed, so the broken title was never rendered during the check at all.

Verification

Typecheck clean · 1359 unit tests · 54/54 browser checks · zero failures.

Every fix has a test that was confirmed to fail against the unfixed code:

FAIL :: the answer is on screen before it has finished :: the streaming turn body stayed empty
FAIL :: the turn strip does not wrap at 924px :: scrollHeight=68 clientHeight=28
1) carries the new model into the options a /clear restart replays
1) records a /model typed straight into the composer, and still forwards it

The live-rendering check also asserts the transcript holds the streamed message, and that passes in both states — so a failure can only mean the view did not render, never that the test fed it bad events.

Notes

  • The changelog describes 5.1.2's four features rather than these fixes: the defects were introduced by PRs that never shipped, so listing them would document bugs against a release that never existed.
  • Left open deliberately: after clearing an override the model chip falls back to the model reported at session start-up, which is stale if the switch had been applied live. What the chip should claim while a process runs a model the conversation no longer asks for is a product decision, not a silent patch.

🤖 Generated with Claude Code

…cross /clear

Five defects introduced by the four PRs already merged onto the 5.1.2 branch.
Each one's own tests passed; all but one lived where two of them met.

The chat stopped rendering. MessageList looked messages up in a map memoised
on the messages array, but the transcript appends to that array in place, so
its identity never changes: the map was built once at mount and every message
after it resolved to nothing. The strips above kept ticking while the bodies
stayed empty, and reloading "fixed" it because reloading replaces the array.
Keyed on the transcript version instead, which is the cadence the neighbouring
memo already used — the list still does not re-render per token.

A model override no longer survives /clear. The restart replays the options
the session was launched with, and the model is the one thing in them that can
change while the session is alive, so the conversation quietly went back to
the model it opened with after the browser had been told the switch applied.
The choice now reaches those options too, by either door: the picker, and a
/model typed straight into the composer, which was forwarded untouched and hit
the same reversion.

A collapsed turn's title escaped its bar. It set neither a size nor nowrap
while every sibling in that fixed-height row sets both, so it inherited the
page default and wrapped — and without nowrap the ellipsis was inert.

An override could not be undone. The server had a "cleared" path, message and
all, that nothing in the UI could reach: the text field refuses to submit
empty and every listed model carries a name. A typo therefore stayed in force
for every later launch of that conversation.

Model names are now capped and stripped of control characters, which would
otherwise ride into the best-effort /model turn as extra lines.

The browser checks gated nothing: npm test runs only the unit suite, and the
runner exited 0 when Chrome was absent. With the rendering bug restored, npm
test passes 1357/1357 while the browser check fails. They now run in CI and
refuse to skip there. Their fixture also grew a second turn, without which no
turn is ever collapsed and the broken title was never rendered at all.

Every fix has a test confirmed to fail against the unfixed code.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Copilot AI review requested due to automatic review settings July 26, 2026 16:22
@dnviti
dnviti merged commit dc6ddf0 into chore/version-bump-5.1.2 Jul 26, 2026
5 checks passed
@dnviti
dnviti deleted the fix/5-1-2-regressions branch July 26, 2026 16:25

Copilot AI 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.

Pull request overview

Fixes several regressions in the chat UI/session behavior on chore/version-bump-5.1.2, including restoring streaming rendering, persisting model overrides across in-place /clear restarts, tightening model-name sanitization, and ensuring browser-based layout checks actually run in CI.

Changes:

  • Fix message rendering during streaming by rebuilding the message-id lookup map based on transcript version changes.
  • Persist model override changes into the restart options used by /clear, including a UI path to clear an override, with server-side model-name normalization.
  • Run browser layout checks in CI (and fail CI if Chrome is missing), plus expand fixtures/checks to cover the newly-guarded cases.

Reviewed changes

Copilot reviewed 12 out of 12 changed files in this pull request and generated 2 comments.

Show a summary per file
File Description
test/chat-wiring.test.js Adds wiring-level tests for remembering model overrides (including clearing + /model typed path) and model-name sanitization.
test/chat-clear-reset.test.js Verifies /clear restarts with the remembered model (override and cleared-to-default cases).
test/browser/run.js Makes missing Chrome a hard failure under CI instead of a silent skip.
test/browser/checks.ts Adds a live streaming rendering check and expands fixed-bar wrap coverage with a second turn.
src/server/websocket/messages.ts Normalizes typed model names and persists the effective model into /clear restart options.
src/server/chat/session.ts Introduces rememberModel() to update stored restart options for in-place restarts.
src/server/chat/manager.ts Plumbs rememberModel() through the chat session manager.
src/client/shell/chat/TurnStrip.tsx Ensures collapsed turn titles don’t wrap/overflow and match bar typography.
src/client/shell/chat/MessageList.tsx Fixes message lookup memoization so messages render as they stream.
src/client/shell/chat/Composer.tsx Adds an explicit UI option to clear a model override back to default.
CHANGELOG.md Updates 5.1.2 notes to describe shipped features/fixes for that release line.
.github/workflows/ci.yml Adds npm run test:browser to CI to gate layout/streaming regressions.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Comment on lines 260 to +262
return map;
}, [messages]);
// eslint-disable-next-line react-hooks/exhaustive-deps
}, [messages, version]);
session.agent as AgentKind,
session.workingDir,
);
this.deps.chatManager?.rememberModel(session.id, model || profile?.model);
dnviti added a commit that referenced this pull request Jul 26, 2026
…l /clear (v5.1.2) (#37)

* chore: bump version to 5.1.2

* feat(chat): let users override a conversation's model, independent of profile default (#39)

Adds a per-conversation model override so any chat session can switch or
type a custom model at any time, regardless of whether the underlying
runtime already supports it. Persisted on the session (mirrors the
chatBypassPermissions pattern), applied via a three-tier fallback
(live adapter.setModel -> best-effort /model slash command -> saved for
next launch), and always beats the runtime profile's default for that
conversation only. Composer's model chip becomes an always-enabled
combobox with free-text entry and honest per-outcome feedback, replacing
the old dead-end disabled state.

Closes #38

Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>

* feat(chat): collapsible turn sections, auto-folded on a new turn (#34) (#40)

A long conversation used to be one unbroken wall of fully-expanded turns.
Each turn's strip now discloses/hides its own body; only the newest turn
opens by default, a turn the user has explicitly opened or closed stays
that way across new turns, and jumping to a turn via the rail or search
force-opens it. Expand-all/collapse-all lives in the turn index header.

Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>

* fix(chat): show Claude's slash commands from the moment a session opens (#41)

Claude's CLI does not report `slash_commands` until it has processed a
first turn's `system/init`, which only arrives after the first message
is written to its stdin. The chat session spawns the process as soon as
a chat is opened, well before any message is sent, so the command menu
and its composer button stayed empty — looking broken — until a
throwaway message unlocked them.

The adapter now advertises a static baseline built from this app's own
table of Claude's built-in commands as soon as it starts, so the menu
and button are populated immediately. The real `init` list — including
any project or plugin commands — still arrives with the first turn and
replaces this baseline outright, exactly as before.

ACP runtimes (kimi, omp) already report commands during their handshake,
before any message is sent, so they were never affected. codex, grok and
pi have no command support at all; the composer already hides the button
entirely rather than showing an empty menu, which is the honest behavior
for a runtime that offers none.

Closes #30

Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>

* fix(chat): make /clear and /new actually reset the conversation (#47)

`/clear` and `/new` forwarded their text to the still-running agent
process like any other message, then only added a cosmetic marker
that hid prior messages in the UI. The next message sent it right
back into the same process, so the agent kept whatever context it
already had — the reset never really happened, only its display did.

Route clearing commands through the same restart path a manual
"start fresh" relaunch already uses: stop the live adapter and start
a new one with no resume id, so the next turn talks to a process that
was never handed the prior conversation. This also fixes the id
persisted for later reconnects, so a rejoin can't resurrect it either.

Fixes #43

Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>

* fix(chat): show the answer as it streams, and keep a model override across /clear (#49)

Five defects introduced by the four PRs already merged onto the 5.1.2 branch.
Each one's own tests passed; all but one lived where two of them met.

The chat stopped rendering. MessageList looked messages up in a map memoised
on the messages array, but the transcript appends to that array in place, so
its identity never changes: the map was built once at mount and every message
after it resolved to nothing. The strips above kept ticking while the bodies
stayed empty, and reloading "fixed" it because reloading replaces the array.
Keyed on the transcript version instead, which is the cadence the neighbouring
memo already used — the list still does not re-render per token.

A model override no longer survives /clear. The restart replays the options
the session was launched with, and the model is the one thing in them that can
change while the session is alive, so the conversation quietly went back to
the model it opened with after the browser had been told the switch applied.
The choice now reaches those options too, by either door: the picker, and a
/model typed straight into the composer, which was forwarded untouched and hit
the same reversion.

A collapsed turn's title escaped its bar. It set neither a size nor nowrap
while every sibling in that fixed-height row sets both, so it inherited the
page default and wrapped — and without nowrap the ellipsis was inert.

An override could not be undone. The server had a "cleared" path, message and
all, that nothing in the UI could reach: the text field refuses to submit
empty and every listed model carries a name. A typo therefore stayed in force
for every later launch of that conversation.

Model names are now capped and stripped of control characters, which would
otherwise ride into the best-effort /model turn as extra lines.

The browser checks gated nothing: npm test runs only the unit suite, and the
runner exited 0 when Chrome was absent. With the rendering bug restored, npm
test passes 1357/1357 while the browser check fails. They now run in CI and
refuse to skip there. Their fixture also grew a second turn, without which no
turn is ever collapsed and the broken title was never rendered at all.

Every fix has a test confirmed to fail against the unfixed code.

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants