feat(desktop): space picker in the composer, and composer chrome cleanup

[adamleithp]

Problem

Two unrelated itches in the spaces composer, plus a few papercuts found on the way.

A new task always filed into whichever space you happened to open the composer from. There was no way to retarget it without backing out and starting again. The picker for this already existed in the static-spaces-sidebar prototype (#76355), but that PR is a large flag-gated rebuild of the whole sidebar and is not close to landing. The picker does not depend on any of it, so it is pulled out here on its own.

The rest is chrome. The model trigger sat a pixel below its neighbours, the channel CONTEXT.md was disguised as a file attachment, and the slash-command surfaces had drifted off the design system.

Changes

Space picker. A quill Combobox chip in the composer's selector row. Starred spaces (with #me leading) sit above the rest, searchable, since a project can carry hundreds. Picking one navigates to that space's own new-task route; the draft survives because it lives in the shared draft store.

Decoupled from #76355: that version sorted starred spaces by the prototype sidebar's drag order and matched the personal space by name. This one derives everything from useChannels and keys off channelType, so it carries none of the prototype's state.

Composer chrome.

• Model trigger alignment. The model and reasoning selectors were the only toolbar items wrapped in a bare span, which builds a line box whose leading pushed them below Attach, Mode and the history button. Measured 481.07 against 479.73; all four now share a top. The fix is in shared PromptInput, so every composer with that toolbar gets it.
• CONTEXT.md moved out of the attachments row and pinned beside the send button as a quill Chip. It is not a file you attached, it comes with the space and rides along with every send, so it no longer collapses to an extension square with its name hidden in a tooltip. The editor's right inset is now measured from that cluster rather than hardcoded, otherwise a long line runs under the chip.
• Space and workspace-mode chips move to variant="default" and drop their carets.

Slash commands.

• A Slash commands item in the attach menu, which writes / at the start of the composer. It inserts at position 1 and drops the caret at 2, leaving the editor in exactly the state typing / produces, so the existing suggestion plugin opens the list itself.
• The suggestion popup's rows were hand-rolled buttons with raw Radix vars inside a role="listbox". They are now ItemMenuItem with role="option" and aria-selected, per the guidance in packages/quill/packages/primitives/AGENTS.md for custom menu-like lists.
• Skill chips in the editor render as plain text in the primary color instead of a chip, going destructive when selected.
• Attach-menu icons dropped from weight="bold" to regular, which GithubLogo needed most.

Note

Removing the chip treatment from skill chips also removes their hover ×. A skill is now deleted by selecting it and pressing backspace, which is what the destructive state signals.

How did you test this code?

Automated, run by me (Claude):

• pnpm --filter @posthog/ui typecheck — clean
• pnpm --filter @posthog/ui test — 306 files, 2545 tests, all passing
• biome check on every changed file
• hogli ci:preflight --strict — clean

I also drove the running dev app over CDP and verified: the picker renders, filters (bill to two results), marks the current space selected and navigates on pick; the measured editor padding tracks the context chip and reflows when it is removed; the attach menu's new item inserts / and opens the command list; the suggestion rows render as quill items with exactly one aria-selected.

Two things I could not verify that way and would like a human to confirm:

• Arrow-key navigation in the suggestion popup. My synthetic key events never reached the tiptap plugin. I A/B tested by reverting the file and re-running: the original behaves identically, so this is not a regression, but it is the primary interaction and deserves a real keypress.
• The skill-chip rendering and the context chip's Enter activation. The dev app was in active use and I stopped driving it rather than fight for the composer.

No tests added. The changes are presentational or one-line wiring, and the parts worth testing (whether a tiptap suggestion plugin fires, whether Base UI binds Enter on a non-native button) are not reachable from jsdom, so a test would assert the wiring rather than the behavior.

Automatic notifications

• Publish to changelog?

Docs update

No docs update needed.

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

I drove this; Claude (Claude Code, Opus 5) wrote most of it across an interactive session. Repo skills invoked: /writing-user-facing-copy before changing the context chip's label, /test-electron-app for the CDP verification against the running app. The quill component choices came from packages/quill/packages/primitives/AGENTS.md.

A few decisions worth surfacing for review:

• The space picker started as a DropdownMenu and became a Combobox once it was clear a project can have hundreds of spaces. It then had a real bug: Channel objects were used as combobox items with a controlled value, which selects correctly once and then silently stops, because the channels query repolls and rebuilds its objects so identity matching fails. Items are now ids.
• The suggestion popup was asked to become a quill Autocomplete and deliberately is not one. Base UI's Autocomplete owns a focused text input, its own filtering, its own highlight state and its own positioner. Here the tiptap plugin owns the query and keyboard and tippy owns positioning, and critically the caret has to stay in the editor, so an Autocomplete input holding focus would break typing in place. It was rebuilt with the menu-list primitives quill documents for exactly this case instead.
• nativeButton={false} on the context chip is not cosmetic. quill's Chip renders a div so ChipClose can nest legally, and without that prop Base UI never binds Enter or Space, so the chip took focus and did nothing.

The branch is named posthog-code/composer-space-select rather than desktop/... because a remote branch named exactly desktop blocks desktop/* as a ref path.

This holds several separable changes. I kept them together because they are all small and all in the same composer, but I am happy to split if a reviewer would rather see them apart.

[trunk-io]

😎 Merged successfully - details.

[github-actions]

React Doctor found 1 issue in 1 file · 1 error.

Errors

• ❌ packages/ui/src/features/message-editor/components/PromptInput.tsx:207 Effect subscription or timer never cleaned up effect-needs-cleanup

_{Reviewed by React Doctor for commit 6fa635f.}

[trunk-io]

   

View Full Report ↗︎ ⋅ Docs

[greptile-apps]

Prompt To Fix All With AI

### Issue 1
products/desktop/packages/ui/src/features/canvas/components/WebsiteNewTask.tsx:135-137
**Space retargeting races task creation**

When a user selects another space while task creation is in flight, the picker navigates immediately but the pending submission retains the previous `onTaskCreated` callback, causing the completed task to be filed in the old space and redirecting the user away from the newly selected composer. Disable space selection while the task mutation is active.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (1): Last reviewed commit: "fix(desktop): correct space grouping and..." | Re-trigger Greptile}

[stamphog]

Note

🤖 stamphog reviewed f769fe804687195a29da9e1054c77d84f889033f — verdict: REFUSED

Greptile flagged a real, unresolved P1 race in the new space picker: WebsiteNewTask's onTaskCreated closes over the channel active at submit time and nothing disables SpaceSelect while a task creation is in flight, so switching spaces mid-submission files the finished task into the old space and navigates the user away from the space they just picked. Verified by reading the current file — the bug is present and unaddressed on the current head.

• greptile-apps[bot] reviewed the current head.
• Unresolved inline review comment (greptile, P1) on WebsiteNewTask.tsx:135-137: selecting another space while a task creation is in flight files the task into the old space and redirects the user away from the newly selected composer — the SpaceSelect trigger has no disabled state tied to isCreatingTask/the mutation's in-flight status.

Gate mechanics and policy version

Gate		Result
prerequisites	✓	all clear
deny-list	✓	no deny categories matched
size	✓	481L, 12F substantive — within ceiling
tier	✓	T1-agent / T1d-complex (481L, 12F, single-area, feat)
stamphog 2.0.0b4		.stamphog/policy.yml @ fceb6cf · reviewed head f769fe8

[deployment-status-posthog]

Deploy status

Environment	Status	Deployed At	Workflow
dev	✅ Deployed	2026-08-04 22:37 UTC	Run
prod-us	✅ Deployed	2026-08-04 22:52 UTC	Run
prod-eu	✅ Deployed	2026-08-04 22:55 UTC	Run