mirror of https://github.com/windmill-labs/windmill.git synced 2026-08-25 00:01:55 +00:00

Files

T

Guilhem 64b089cd23 feat(frontend): use unified drill picker for AI chat @-mention dropdown (#9159 )

* feat(frontend): use unified drill picker for AI chat @-mention dropdown

* fix(frontend): chat picker review followups + overlay alignment

- AIChatDisplay: migrate @-badge popover to ChatContextPicker (was still
  importing the deleted AvailableContextList after the rebase onto #9034,
  causing a build break).
- DrillPicker: handle Tab as Enter so the inline @<word> mention completes
  without losing focus. Tweak leaf-row weight to font-normal; secondary
  text uses text-hint.
- ContextTextarea: drop px-0.5 from the highlight span — extra horizontal
  padding made every glyph typed after a mention drift right of the
  invisible textarea below. box-decoration-clone keeps the rounded corners.
- ContextElementBadge: explicit font-normal label, hoist label into a
  {@const} and pass to title= so the truncated badge shows the full title
  on hover.
- workspaceTree: drop orphaned doc-comment left dangling by the rebase.
- Add unit tests for drillPicker.ts and workspaceTree.ts (51 tests cover
  resolveScope/scopeChain/collectLeavesGrouped/leafHaystack, buildWorkspaceTree
  shape + loading + dir forest + leaf shape, withCurrent rename suppression,
  extraItemsByKind dedup, legacyScopeToPath, relativizeWorkspacePath).

* fix(flow-editor): ignore keyboard shortcuts when focus is outside the flow root

Menus, modals, drawers etc. live outside the flow root and capture focus
explicitly. Flow nodes aren't focusable, so the unfocused default
(activeElement === body) means "flow is the canvas" and we should react;
anything else means another surface has the user's attention and our
shortcuts would steal it.

* fix(frontend): inline @ mention picker + chat layout polish

- ContextTextarea: swap manual Portal+caret-math positioning for
  svelte-floating-ui anchored at the `@` character (virtual reference,
  middleware [offset, flip(crossAxis:false), shift]). Picker stays
  pinned to `@` while the user types the query, slides leftward when
  hitting the right edge instead of flipping alignment, and floating-ui
  handles above-vs-below + edge clamping automatically. Drops the
  60vh-worst-case reservation that left a big gap above the caret in
  sessions, and the now-unused isFirstMessage prop is marked deprecated.
- AIChatDisplay: the `@`-button Popover now opens with placement
  bottom-start (was the default `bottom`), aligning its left edge with
  the button instead of centering under it.
- ChatContextPicker: when no Diffs/Modules/Databases branches are
  present (e.g. global chat), return the Workspace tree's children at
  the root instead of wrapping them under a redundant "Workspace" row.
  handleScopeChange handles both the wrapped and unwrapped layouts and
  the single-kind `dir:` top segment.

* chore(frontend): address review suggestions on chat picker PR

- DrillPicker: clamp width to viewport on narrow screens —
  w-[420px] → w-[min(420px,calc(100vw-20px))].
- workspaceTree.buildWorkspaceTree: make loadingKind optional (defaults
  to {}). Chat picker still passes it; callers that don't track loading
  no longer need to thread an empty object.
- ChatContextPicker.handleScopeChange: name the WRAPPED vs UNWRAPPED
  layouts in a comment block so the dir:/kind: branches are obvious.
- ContextTextarea: drop deprecated isFirstMessage prop (floating-ui
  handles direction); drop defensive Math.max on the @ index now that
  the invariant is documented; comment the floatingRef(anchorRef) call
  as the supported virtual-reference path in svelte-floating-ui.
- AIChatInput: stop forwarding isFirstMessage to ContextTextarea.

* feat(frontend): sync selectedContext with @-mentions in textarea

Both picker entry points now insert a visible `@title` token in the
textarea, and deleting that token drops the matching entry from
selectedContext.

- AIChatInput: new insertMention(title) export. Appends `@title ` to
  instructions, prefixing a space only if the existing text doesn't
  already end in whitespace.
- AIChatDisplay: the `@`-button popover calls insertMention after
  addContextToSelection so its picks match the inline-mention path's
  textarea state.
- ContextTextarea: new onRemoveContext callback. A $effect compares the
  set of `@title` tokens in `value` (derived) against the previous
  snapshot; titles that disappeared trigger onRemoveContext for any
  selectedContext entry with `deletable !== false`. The diff lives in
  an effect (not handleInput) so it catches both keystroke deletions
  AND programmatic value updates from updateInstructionsWithContext.
- AIChatInput: passes onRemoveContext that filters selectedContext by
  type+title — mirrors the existing badge X-button handler.

* chore(frontend): narrow ChatContextPicker `inner` from `any` to `DrillPicker | undefined`

The previous `let inner: any` worked around svelte-check rejecting
`DrillPicker<ChatLeafData>` (the imported component is seen as the
non-generic `Comp`). Dropping the type parameter keeps the workaround
without `any`, so handleKeydown / pickHighlighted are at least typed
at the call site.

Addresses May-14 PR review.

* fix(frontend): address PR #9159 bot-review findings (eager preload, focus, dedup, icon types)

- [P1] ChatContextPicker.handleScopeChange: stop preloading workspace
  kinds at the wrapped picker root. New `isWorkspaceOnly` $derived (true
  when no Diffs/Modules/Databases branches are present) gates the at-
  root preload, so the chat root no longer fires two list requests
  before the user enters Workspace. Reported by Codex.
- [P2] AIChatDisplay @-button popover: call aiChatInput.focusInput()
  after close() so the textarea is focused for immediate typing — mirrors
  the inline-mention path's setTimeout(textarea.focus, 0). Reported by
  Claude.
- [P2] AIChatInput.insertMention: no-op when the `@title` token is
  already present in instructions, so re-picking a workspace item
  doesn't leave duplicate visible tokens for a single selectedContext
  entry. Reported by Codex.
- [P2] drillPicker.ts: introduce `DrillIcon = ComponentType |
  Component<any, {}, ''>` and replace `icon: any` on DrillLeaf,
  DrillBranch, and ChatContextPicker.buildContextBranch. Mirrors the
  ComponentType | Component pattern used in TriggersBadge.svelte for
  the same Svelte 4/5 compatibility window. Reported by Pi.

* fix(frontend): preserve workspace context on refresh + load all kinds for internal search

- [P1, Codex] ContextManager.updateAvailableContextForScript/Flow:
  preserve workspace_script and workspace_flow entries through the
  selectedContext filter on editor refresh. They're user-picked refs
  that don't appear in availableContext, so the previous filter was
  silently dropping them whenever the script/flow editor refreshed
  options (e.g. on any code change).
- [P2, cubic-dev-ai] WorkspaceItemDrillPicker: in internal-search mode
  (externalFilter === undefined, DrillPicker renders its own search
  box), preload all kinds on mount. Without this, typing in the
  picker's search before clicking a kind branch produced incomplete
  results since DrillPicker can't reach back through the adapter to
  trigger fetches on internalFilter change. Cached items keep the
  effective cost near-zero on warm sessions.

* fix(frontend): preserve workspace refs through script-mode context refresh

The script-mode updateAvailableContext overwrites newSelectedContext
with a fresh [code] entry, defeating the workspace_script /
workspace_flow preservation in the later filter — the entries are
already gone by the time the filter runs.

Seed newSelectedContext with the refreshed code block AND the user-
picked workspace_script / workspace_flow / code_piece entries from
currentlySelectedContext, so editor refreshes don't wipe @-mention
badges in script chat. The existing line-271 filter still validates
each entry against newAvailableContext + the per-type allowlist.

Reported by Codex on PR #9159 — completes the prior workspace-context-
on-refresh fix (b02d1f2d35) which only patched the filter, not the
rebuild step that runs before it.

* fix(frontend): preserve all previously-selected contexts on script refresh

The prior c2775fe0c5 fix only carried over workspace_script /
workspace_flow / code_piece entries from currentlySelectedContext.
That preserved the workspace P1 path but still dropped previously-
selected diff / error / db / runtime-context badges, which cubic
flagged in its 16:55 review.

Spread the full currentlySelectedContext (minus `code`, which we just
rebuilt). The downstream filter validates each entry against
newAvailableContext + the per-type allowlist, so auto-derived types
like diff / error / db survive when still applicable, and unrelated
items are dropped automatically.

Reported by cubic-dev-ai on PR #9159.

* fix(frontend): rehydrate auto-derived context + sync badge X with textarea

- [P2, cubic] ContextManager.updateAvailableContext: when the rebuild
  carries over previously-selected diff/error/db entries, swap each one
  for the matching freshly-built entry from newAvailableContext in the
  final .map() step. Preserves the user's `deletable` override on top
  of the fresh content/diff/schema, so refreshes don't keep stale
  payloads while still surviving the badge across edits.
- [P2, Pi/Codex] AIChatInput: new `removeMention(title)` export that
  strips `@title` tokens from `instructions` (whitespace-bounded so
  substring matches don't bleed). The badge X-button now calls it
  after filtering selectedContext, mirroring the inverse textarea-to-
  badge sync. No double-remove: ContextTextarea's $effect-driven
  onRemoveContext is a no-op once selectedContext no longer holds the
  entry.

* fix(frontend): retype ChatContextPicker.inner to DrillPicker<ChatLeafData>

`npm run check:fast` (TypeScript-only) and `npm run check` (svelte-check)
disagree on whether the imported DrillPicker is generic — `check:fast`
sees it as `Comp` and rejects the type parameter, while `svelte-check`
sees the real generic component and requires it. CI runs `check`, so
follow that: `DrillPicker<ChatLeafData> | undefined`.

This also fully replaces the prior `inner: any` workaround called out in
multiple bot reviews — handleKeydown / pickHighlighted now type-check at
the call site against the correct component instance.

* fix(frontend): scope removeMention's whitespace collapse to the mention site

The trailing `.replace(/  +/g, ' ')` in `removeMention` was global,
collapsing any pre-existing double-spaces in the prompt — e.g. a user
typing `"hello  world  @foo  bar"` lost their intentional formatting
when they deleted the `@foo` badge.

Rework the regex to match `(^|\s)@title(\s|$)` and decide per-match:
- Mention at a boundary (no lead or no trail): drop entirely.
- Mention in the middle: keep ONE bordering whitespace char (the
  leading one verbatim, so newlines/tabs aren't downgraded to spaces).

No global pass over `instructions`. Unrelated whitespace stays intact.

Reported by cubic-dev-ai on PR #9159 (07:27 review of 9e07eac4).

* fix(frontend): expose DrillPicker.onFilterChange + lazy-load workspace kinds

Both Codex P1s came from over-eager preload heuristics on my prior fixes:
the workspace picker cold-loaded every configured kind on mount in
internal-filter mode, and the chat badge popover never observed its own
internal filter so workspace results were missing from search until the
user drilled into Workspace.

Replace both ad-hoc effects with a single `onFilterChange` callback on
DrillPicker that fires whenever the EFFECTIVE filter (external or
internal) changes:

- [P1] WorkspaceItemDrillPicker: drop the "cold-load on mount when
  externalFilter === undefined" effect. Workspace kinds now load only
  once the user actually types something — closer to the pre-refactor
  behavior where the breadcrumb / "Open editor" pickers only fetched
  the drilled-into kind plus all kinds on search.
- [P1] ChatContextPicker: handleFilterChange replaces the prior
  externalFilter-only effect. Badge-popover search (internal filter)
  now triggers the same preload as inline-mention search (external
  filter), so workspace results appear without needing to drill first.

Both fixes reported by Codex on PR #9159.

* fix(frontend): skip mention-removal sync when textarea is programmatically cleared

sendRequest() sets `instructions = ''` immediately after dispatching to
AIChatManager. The mention-removal effect treated this as user-initiated
deletion and cleared selectedContext BEFORE AIChatManager.beforeSend
snapshotted it — selected `@` contexts disappeared from the outgoing
request. Skip the sync when value is empty; user-initiated mention
deletes happen in-place against non-empty content.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* fix(frontend): scope post-send wipe protection to the send path only

Replace the blanket `if (value !== '')` guard on the mention-removal
effect with an explicit `clearForSend()` export. `sendRequest()` now
calls it instead of `instructions = ''`, so a user manually clearing
the whole textarea still drops the corresponding context badges while
the post-dispatch programmatic wipe is silent.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* refactor(frontend): extract useWorkspaceItemsLoader composable shared by both drill picker adapters

WorkspaceItemDrillPicker and ChatContextPicker each duplicated the same
machinery: loaded/loadingKind state seeded from the module cache, a
stale-while-revalidate ensureLoaded coroutine with an untrack guard, a
kind:/dir: scope-segment decoder, and the "load every kind once the user
starts searching" filter callback.

Move that to a single useWorkspaceItemsLoader() returning
{loaded, loadingKind, ensureLoaded, ensureAll, ensureForScopeSegment,
onFilterChange}. Adapters keep their own scope-walking policy (chat
collapses an optional 'workspace' wrapper, workspace handles single-kind
mode) but delegate kind decoding and lazy fetch to the composable.

Net: -135 +28 LOC in the two adapters; +109 LOC in the new composable.
The cache-version race, untrack discipline, and stale-while-revalidate
semantics now live in one place.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* fix(frontend): address Codex P1+P2s — non-context clear, same-title cross-removal, single-kind cold load

P1: sendRequest() now clears `instructions` unconditionally after the
optional `clearForSend()` so APP/NAVIGATOR/ASK/API modes (which don't
mount ContextTextarea) still reset the input after send.

P2: removeMention() now calls a new `unsyncMention(title)` on the
textarea before stripping `@title` from `value`, so the mention-removal
effect doesn't fire a second onRemoveContext on a same-title sibling
(e.g. workspace_script + workspace_flow sharing a path).

P2: single-kind WorkspaceItemDrillPicker loads its kind at mount even
when scope is empty — buildWorkspaceTree collapses to the kind's
children, so there's no kind row to drill into to trigger the load.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>