mirror of
https://github.com/simstudioai/sim.git
synced 2026-09-24 15:45:35 +08:00
35a7bf61d2e68f4c78d6c76d394dfdce2631d66b
5013
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Vikhyath Mondreti
|
35a7bf61d2 |
fix(executor): stop HITL error edges from firing on successful resume (#5152)
* fix(executor): stop HITL error edges from firing on successful resume * add comments |
||
|
Waleed
|
3ebb9a5029 |
feat(connectors): add Google Meet knowledge base connector (#5149)
* feat(connectors): add Google Meet knowledge base connector Syncs Google Meet meeting transcripts into a knowledge base via the Meet REST API v2. Lists conference records, fetches transcript entries lazily per meeting (contentDeferred), resolves speaker display names, and maps participants/duration/meeting-date tags. OAuth via the existing google-meet provider (meetings.space.readonly). * fix(connectors): finalize Google Meet transcripts before indexing - Only index a meeting once every transcript is FILE_GENERATED, so a partial transcript is never stored under an endTime-keyed hash that would never refresh - Sort merged transcript entries by start time to preserve chronology across multiple transcripts in one conference * refactor(connectors): dedicated TRANSCRIPTS_PAGE_SIZE constant for Meet transcripts * fix(connectors): only flag Meet listing capped when cap truncates source Previously listingCapped was set whenever the fetched count reached maxMeetings, even when the API returned every record and no next page existed. That suppressed the sync engine's deletion reconciliation when the cap happened to equal the true source size. Now flag only when more pages remain or records were dropped from the page. |
||
|
Waleed
|
ce283fa1c9 |
feat(scheduled-tasks): expose Google Calendar-style recurrence options (#5146)
* feat(scheduled-tasks): expose Google Calendar-style recurrence options Add a per-day weekly toggle (repeat on arbitrary weekdays), monthly nth-/last-weekday anchoring, and a yearly frequency to the scheduled task modal, closing the gap with a calendar app's recurrence picker. The recurrence UI compiles to cron, so this is front-end only: croner already speaks the nth/last-weekday (#/#L) syntax, and the display path normalizes #L to cronstrue's L so labels read "last Monday" not "null". * fix(scheduled-tasks): preserve monthly anchor when reselecting Monthly Selecting Monthly from the frequency dropdown hard-reset monthlyMode to day-of-month, silently dropping a previously chosen nth-/last-weekday anchor when switching cadence away and back. Preserve the existing mode on reselect, mirroring how the last recurring cadence is restored across the recurring toggle. * fix(scheduled-tasks): fold 5th-occurrence monthly into last-weekday; align weekday-digit parsing Address review edge cases in the monthly recurrence anchors: - The picker no longer offers a fifth weekday (a 5th occurrence is always the month's last), and recurrenceToCron clamps any nth-weekday that resolves to a 5th occurrence to #L — so a launch date drifting to day 29-31 can never emit #5 and silently skip months without one. - cronToRecurrence accepts croner's alternate Sunday digit (7) for the #/#L monthly anchors, matching parseCronToHumanReadable's normalizer; externally-authored 7#L crons now round-trip (canonicalized to 0#L). - #5 crons are left as custom pass-through so their month-skipping behavior is preserved verbatim rather than rewritten. |
||
|
Theodore Li
|
1248f8eef4 | fix(files): only show Share in context menu for files, not folders (#5147) | ||
|
Theodore Li
|
3b78436ae3 |
feat(pii): gate data retention PII redaction behind feature flag (#5144)
* feat(pii): gate data retention PII redaction behind feature flag * fix(pii): evaluate pii-redaction flag globally with no org/user context |
||
|
Waleed
|
ecbe1919d8 |
feat(files): inline rich markdown editor (#5133)
* feat(files): inline rich markdown editor
Replace the raw/preview split for markdown files with a Linear-style inline WYSIWYG editor (TipTap/ProseMirror): bubble + slash menus, code-block language picker with Prism highlighting and line-wrap, resizable images (HTML <img>), GFM tables, and frontmatter held byte-exact out of band.
A round-trip preflight gate (decided once per open) falls back to the raw Monaco editor for any file that can't be edited losslessly, so the rich editor never silently corrupts a file.
* fix(files): chain autosave unmount flush after in-flight save
The unmount flush no longer fires a concurrent PUT alongside an in-flight save; it awaits the in-flight save and then writes the latest content sequentially, so an out-of-order completion can't clobber newer edits with a stale snapshot (addresses Cursor Bugbot).
* fix(files): read pasted images from clipboard items, not just files
Some browsers expose a pasted or copied image only via DataTransfer.items (with an empty files list), so screenshot paste was silently ignored. extractImageFiles now falls back to items; moved to a testable module with unit tests (addresses Cursor Bugbot).
* fix(files): destroy round-trip probe editor on serialization error
Wrap the probe serialize() in try/finally so the throwaway Editor is always destroyed even if setContent/getMarkdown throws (addresses Greptile). Adds a test proving PipeSafeTable escapes only interior cell pipes, not structural delimiters.
* fix(resource): hold breadcrumb nav latch across the route swap
scheduleClose fired on the pointer/focus exit that immediately follows a click-to-navigate and was clearing the reopen latch before the route swapped, letting the popover flash back open. The latch is now released by a short timer instead (addresses Cursor Bugbot).
* chore(files): drop platform references and non-essential inline comments
* fix(files): scope inline markdown editor to the files view
The mothership preview was routing streaming markdown through the inline editor path: it showed Monaco during streaming (previewMode fell back to 'editor') and lost the streamed content on the TextEditor→MarkdownFileEditor swap (the TextEditor unmounted before it could reconcile + autosave). The inline rich editor is now opt-in via a FileViewer prop that only the files view sets, so the mothership keeps its raw/preview streaming editor and persists as before.
* fix(mothership): use the inline markdown editor in the chat resource view
Idle markdown in the chat resource view now renders the single-surface inline editor (no raw/split/preview pencil toggle), matching the files view. While the agent streams, FileViewer forces the rendered preview instead of Monaco, and the streamed file persists via the agent's server write + the existing content-query invalidation on tool completion — so the idle editor refetches the persisted content.
* refactor(files): collapse the duplicate raw-editor fallback branch in the markdown gate
* fix(mothership): swap to the inline editor once a file preview finishes streaming
The preview session keeps status='complete' and previewText after streaming ends, so streamingContent stayed defined and the file stuck on the read-only rendered preview. Treat content as streaming only while status==='streaming'; once complete the EmbeddedFile sees no streamingContent and mounts the editable inline editor (which refetches the persisted content). The synthetic streaming-file stays a pure preview.
* Revert "fix(mothership): swap to the inline editor once a file preview finishes streaming"
This reverts commit 25b12e4caa389109c2d5ed7f7d122e35d441f980.
* Revert "fix(mothership): use the inline markdown editor in the chat resource view"
This reverts commit 9430aa7fdc5050d002f2312d31dcf4a255e2de18.
* feat(files): rich markdown editor across files + chat, read-only for unsafe, robust load/save
- chat resource view streams into the rich editor (streamdown while streaming → editable on completion); agent persists server-side, editor never saves mid-stream
- round-trip-unsafe / >128KB markdown renders read-only in the rich editor (no Monaco, no corruption)
- markdown always uses the rich editor (dropped the inline-markdown opt-in flag)
- editor loads content as TipTap's initial content keyed by file id — strict-mode/SSR-safe, no content-sync effect
- fix autosave "Saving…" status suppression under React strict mode
- lock the streamed-file persistence handoff with a state-machine lifecycle test
* chore(files): remove dead code (unused FileViewer logger + EmbeddedWorkflowActions router)
* fix(files): derive markdown round-trip verdict from live content, not a locked stale snapshot
The gate locked isRoundTripSafe on the first post-stream snapshot, which is often the empty create_file buffer before the agent's server write lands — wrongly leaving an unsafe document editable. Derive the verdict from the current content (memoized on the bytes) so canEdit tracks the real payload.
* test(files): guard the rich editor dirty signal — open is never dirty, edits emit
* fix(files): lock the markdown round-trip verdict on opened content, never strand dirty edits
The round-trip-safety verdict now gates editability only at open time — computed once, on the exact
content the editor mounts with, and locked for its lifetime. A dirty document is round-trip-safe by
construction (the editor only emits safe markdown), so the verdict must never flip off mid-edit:
doing so disabled autosave, ⌘S, the toolbar Save and the unmount flush, stranding unsaved edits.
Locking on the opened (reconciled) content also fixes the stale post-stream empty-buffer snapshot,
and lets the redundant MarkdownFileEditor gate (plus its duplicate content fetch) be deleted.
* improvement(file-viewer): reuse shared copy hook, lazy frontmatter split
- code-block: replace hand-rolled copy-with-timeout with shared useCopyToClipboard
- rich-markdown-editor: compute frontmatter split once via lazy ref, drop redundant frontmatterRef
- round-trip-safety: correct stale comments (read-only, not raw editor fallback)
* feat(file-viewer): linked images, typed-link input rule, drag-to-reorder, churn fixes
- image: round-trip linked images/badges via an href attr + custom markdown tokenizer; make
the image a drag handle so it can be grabbed and reordered
- link-input-rule: convert typed [text](url) to a link on the closing paren (normalized href)
- markdown-paste: render pasted markdown as rich content, guarded against code blocks
- round-trip-safety: behavioral link-count check replaces the static linked-image rejection
- extensions: trim the table serializer's blank lines to stop interior-table whitespace churn
* improvement(file-viewer): Backspace at start of a heading reverts it to a paragraph
Notion-style: ProseMirror's default joins or no-ops at a heading boundary, stranding the
heading style. A second Backspace then merges as usual.
* fix(file-viewer): don't upload pasted/dropped images into a read-only editor
handlePaste/handleDrop ran the workspace image upload without checking editability, so a
read-only doc (canEdit=false or a round-trip-unsafe file) could still trigger an upload.
Guard both on view.editable.
* fix(file-viewer): sanitize linked-image href; drop global leading-newline strip
- image: run the linked-image (badge) anchor target through normalizeLinkHref so a
javascript:/data: href in a file can't execute on click; the markdown still preserves the
raw target (file content unchanged)
- markdown-fidelity: the table serializer now trims its own surrounding blank lines, so the
global leading-newline strip in postProcessSerializedMarkdown is redundant — removing it
stops clobbering content that legitimately begins with whitespace
* feat(file-viewer): stream agent output directly into the rich editor; add more code languages
- rich-markdown-editor: the TipTap editor is now the only markdown surface. Agent output streams
into it read-only (synced per chunk, autoscrolled), then the same instance hands off to an
editable editor on settle — no separate streamdown preview, so no stream→edit flash. The
round-trip verdict + frontmatter lock when the content settles.
- code-block/code-highlight/detect-language: register Go, Rust, Java, C, C++, C#, Ruby, PHP grammars
and add detectors, so those blocks highlight and the picker offers them.
- css: style h5/h6 in the prose stylesheet.
* fix(sidebar): hydrate collapse state before paint to stop refresh flash
The collapsed sidebar swaps entire subtrees (collapsed flyout vs expanded
lists), but isCollapsed only resolved after the first paint via auto
rehydration, so a collapsed reload rendered the expanded tree into the 51px
rail and then reflowed — the misplaced/flashing content on refresh.
Adopt zustand's documented SSR pattern: skipHydration on the persist config
(first render keeps the default false, matching SSR HTML) and flush
persist.rehydrate() from a useLayoutEffect so the correct structure commits
in the same pre-paint frame. Removes the old race where onRehydrateStorage
lifted the data-sidebar-collapsed mask before React committed the rail.
* refactor(file-viewer): audit fixes — stale docs, DRY settle-lock, language detection
- rich-markdown-editor: rewrite the now-stale single-surface docstring (no PreviewPanel); extract a
shared lockSettled() helper used by both the mount and stream-settle paths; guard the settle
re-seed so it only setContent's when the body actually changed (no redundant doc rebuild)
- detect-language: stop misreading generics (List<String>) as HTML markup; detect Go type/struct
- code-block: export LANGUAGE_OPTIONS + add a test asserting every picker language has a registered
Prism grammar (prevents picker/highlighter drift)
* refactor(file-viewer): remove dead markdown-preview renderer now superseded by the rich editor
Markdown files route exclusively to RichMarkdownEditor on both the read-only
and editable paths, so PreviewPanel's markdown branch and its Streamdown-based
renderer were unreachable. Delete MarkdownPreview and its renderers, callout/
frontmatter/checkbox machinery, and the now-unused remark/rehype/prism/streamdown
imports; drop the dead toggleMarkdownCheckbox/onCheckboxToggle plumbing in
text-editor. Keep the html/csv/svg/mermaid branches intact.
* refactor(file-viewer): drop dead streamingMode/append path, align naming, cover autosave
The streaming engine only ever runs in 'replace' mode (the only runtime callers
pass it); the 'append' branch of resolveStreamingEditorContent was unreachable.
Remove streamingMode + the StreamingMode type and thread it out of the 6
components that forwarded it — nextContent is now simply the streamed snapshot,
behavior-identical on the live path. Rename for codebase semantics: the boolean
prop streaming -> isStreaming, EditorKeymap -> RichMarkdownKeymap, the highlight
PluginKey KEY -> HIGHLIGHT_PLUGIN_KEY. Add a defensive isEditable guard to the
markdown paste handler (parity with the image handler; read-only must never
mutate). Add a dependency-free useAutosave test suite (debounce, min-display
window, no-data-loss when an edit lands mid-save, error/no-retry, Cmd+S flush,
streaming-disabled lock, unmount flush).
* fix(file-viewer): re-lock round-trip verdict + frontmatter on each stream settle
LoadedRichMarkdownEditor stays mounted across multiple agent edits to the same
file within a chat (previewContextKey is the chat id), but the settle effect only
locked settledRef when it was null — so a second stream into the same instance
kept editability and frontmatter tied to the first settled snapshot. A repeat
edit that is round-trip-unsafe would stay editable, and saves would re-attach the
stale frontmatter. Track wasStreaming and re-derive the verdict + frontmatter on
every stream->settle transition (user edits never re-derive, preserving the
don't-strand-edits rule). Verified red/green in the e2e streaming harness.
* test(file-viewer): lock link href sanitization for dangerous schemes from file content
Greptile flagged a possible javascript: link XSS. Verified TipTap 3.26.1 already
neutralizes javascript:/data:/vbscript: (and mixed-case/whitespace variants) from
file-loaded markdown to an empty href. Add a committed regression test that asserts
this against the real headless editor, so a future TipTap bump can't silently
reintroduce the issue.
* perf(file-viewer): cap the round-trip probe at 24KB and coalesce streaming syncs
@tiptap/markdown's parse is superlinear (~O(n2)) in document size — measured ~170ms
at 11KB, ~875ms at 23KB, multiple seconds past ~35KB — and it runs synchronously at
mount inside the round-trip-safety probe (twice) and the editor's own setContent. The
128KB cap allowed multi-second main-thread freezes; lower it to 24KB so the worst-case
mount stays near a second while still covering the vast majority of real markdown files
(larger files open read-only). Separately, coalesce streaming chunk-syncs to one re-parse
per animation frame so a fast-streaming agent doesn't re-parse the whole accumulating
doc per token. Typing latency was measured to be already excellent (sub-ms median, no
change needed); the only hot cost was the mount parse.
* perf(file-viewer): chunked markdown parsing to remove the O(n2) mount cost
@tiptap/markdown's whole-document setContent(md,'markdown') is superlinear in size,
freezing the main thread at mount for large files (~2.5s at 34KB, ~11s at 65KB) and
forcing a restrictive read-only cap. Parse block-by-block instead: a conservative
blank-line/fence-aware splitter (merges list/quote runs and indented continuations so
ambiguous structures stay atomic; reference-link/footnote/raw-HTML docs fall back to a
whole parse), each block parsed with the editor's own lexer via one reused headless
parser, assembled into a doc. This is linear and byte-identical to the one-shot parse —
measured ~15ms vs multiple seconds at 124KB+ — so the editor mount, streaming sync, and
round-trip probe are all linear, and the editable-size cap goes 24KB -> 256KB (covers
the p99 of real files). Fidelity + idempotency are pinned by unit tests, a 400-document
property/fuzz test, and adversarial edge cases (nested/loose lists, blockquotes, setext,
indented code, lazy continuation, HTML, reference links).
* fix(sidebar): render collapse state from a cookie so SSR matches
The server couldn't read localStorage, so a collapsed user's first paint
rendered the *expanded* tree at 51px — prefetched chat/workflow lists,
pinned-chat pin icons, and loading skeletons all crammed into the rail and
then reflowed once the store hydrated.
Mirror the collapse state into a sidebar_collapsed cookie (the shadcn/ui
sidebar pattern), read it in the workspace server layout, and seed the
sidebar's first render with it: structure is now correct on the server, so
the first paint is the real rail with no skeleton/pin/shift. The store
remains the post-hydration source of truth; the blocking script honors the
cookie for width when localStorage is absent so width and structure agree.
* refactor(sidebar): make the cookie the single source of truth for collapse
Consolidates the collapse machinery onto one source of truth instead of
layering the cookie on top of the legacy localStorage + CSS-mask system:
- Collapse persists only in the sidebar_collapsed cookie; the store seeds
isCollapsed from it and drops it from localStorage (partialize + merge),
removing the dual-write and the cross-tab desync it caused.
- Retire the redundant html[data-sidebar-collapsed] attribute + CSS mask now
that the server emits the correct data-collapsed structure; also delete the
dead sidebar-collapse-show/-remove/-btn rules.
- Blocking script reads the cookie for collapse (width stays in localStorage)
and seeds the cookie once from the legacy flag so existing collapsed users
keep their preference.
- Keep skipHydration + a pre-paint rehydrate for width only — the documented
zustand SSR pattern, so _hasHydrated is deterministically false during SSR.
Width stays in localStorage; each field now has exactly one home.
* refactor(file-viewer): simplify + cleanup chunked-parse (linear merge, parse-once seed)
From the /simplify + /cleanup passes:
- splitMarkdownBlocks: build continuation runs and join each once instead of
concatenating onto the growing previous block per group, which was O(n2) for a
pathological single long loose list (now linear: 208KB loose list splits in ~3ms).
- rich-markdown-editor: seed the editor's initial content via a lazy useState
initializer instead of useRef(parseMarkdownToDoc(...)), whose argument re-parsed the
whole document on every render (i.e. every keystroke). Parses exactly once at mount.
- Document that the indent-merge rule is load-bearing for nested fenced code, and
tighten the verbose inline comment blocks.
* refactor(sidebar): drop orphaned sidebar-collapse-btn class
Its CSS rule was removed with the data-sidebar-collapsed mask; the button's
collapse behavior is fully driven by the React isCollapsed ternary, leaving the
class name pointing at nothing.
* test(file-viewer): consolidate split test files into one per module
Match the dir's one-test-per-module convention: fold the markdown-parse property/fuzz
suite into markdown-parse.test.ts and the editability corpus into round-trip-safety.test.ts
(both already tested the same module from a separate-concern file). No coverage change —
same assertions, fewer files (12 -> 10).
* fix(file-viewer): make all editor controls respect read-only permissions
Every interactive control that calls updateAttributes/dispatches a command mutates
the doc even when read-only (ProseMirror commands run regardless of editable), so
gate them on editor.isEditable:
- bubble menu: the Cmd/Ctrl+K shortcut and shouldShow now bail when not editable, so
a read-only doc can't open the link bar and setLink into it (Cursor finding).
- code block: the language picker renders as a static label when read-only (its
onSelect mutates); copy + view-only wrap stay.
- image: no drag-to-reorder (draggable=false, no drag handle) and no resize handle
when read-only; the image still renders and follows its link.
- links: a plain click now follows the link in read-only (reader) mode, while edit
mode still requires a modifier so a plain click can place the cursor (Cursor finding).
Verified with new read-only permission e2e tests.
* fix(sidebar): honor collapsed cookie even when localStorage is corrupt
The blocking script read the collapse cookie inside the same try as
JSON.parse(localStorage); invalid persisted JSON fell through to the 248px
fallback and ignored a collapsed cookie, painting an expanded-width rail on
first load. Read collapse from the cookie first and parse the persisted width
in its own try so the two are independent.
* docs(sidebar): convert inline comments to TSDoc
* fix(file-viewer): resolve in-app workspace image URLs in the rich editor
The removed MarkdownPreview rewrote /workspace/{id}/files/{fileId} image src to the
serving endpoint /api/files/view/{fileId}; without it, in-app image URLs 404 in the
rich editor (Cursor finding). Re-add the rewrite as a display-only transform on the
rendered <img src> — the node's stored src attribute keeps the original path so markdown
round-trips unchanged. Absolute/non-workspace URLs pass through. Unit tested.
* fix(files): restore same-page anchor links in the rich markdown editor
Headings rendered by the TipTap editor had no slug ids (the old MarkdownPreview
got them from rehype-slug), so in-document table-of-contents links like
[section](#section) had no targets. Resolve the slug to its heading on click
(GitHub-style, duplicate-disambiguated) and scroll to it, with zero per-keystroke
cost.
* feat(files): render mermaid diagrams in the rich markdown editor
A code block renders as a Mermaid diagram when it is fenced ```mermaid or
auto-detected (an untagged fence whose first line opens with a diagram keyword,
the Linear/GitHub heuristic). Detection is display-only — the node stays an
ordinary code block and the markdown round-trips unchanged.
- Source while the caret is inside the block, diagram on blur; a Show source /
Show diagram control plus copy, matching the code block's hover chrome.
- Clicking the diagram selects the node (same ring as an image), not flips source.
- Theme-aware (light/dark) via next-themes; the diagram frame shares the code
block's chrome (one CSS source of truth).
- Extracted MermaidDiagram into a shared module so the editor reuses it without
pulling preview-panel's heavy deps; rendered SVGs are memoized so toggling the
source view and back is instant.
Covered by mermaid-diagram unit tests and the editor e2e harness.
* fix(files): harden the markdown editor (CRLF chunking, href allowlist, image escaping)
Final-audit follow-ups:
- splitMarkdownBlocks normalizes CRLF/CR first — a closing fence ending in \r
no longer fails to match, which had collapsed Windows-authored files with
fenced code into one block and defeated the linear chunker (perf regression).
- normalizeLinkHref rejects file://, blob:, and other non-network schemes
(script/data schemes already rejected); network scheme:// (http/ftp/…) and
bare host:port still pass.
- Image markdown serialization escapes alt/title delimiters and angle-brackets
a src with spaces/parens, so they round-trip losslessly; linked-image anchors
open in a new tab (target=_blank).
- Markdown paste routes through the chunker so a large pasted blob can't freeze
the main thread.
* test(files): cover the code-highlight incremental re-tokenization gate
Export and unit-test changeTouchesCodeBlock: prose-only edits map decorations
(false), edits inside a code block or a setNodeMarkup language change re-tokenize
(true) — the perf-correctness path that keeps highlighting off the keystroke path.
* fix(files): keep relative links relative, navigate in-app links within the SPA
- normalizeLinkHref no longer prefixes `./`/`../` relative paths into `https://./…`
(they round-trip and resolve correctly).
- Following a same-origin in-app link (e.g. /workspace/…) routes through the
Next router (same tab) instead of always opening a new tab; modifier-click and
external URLs still open a new tab.
* fix(files): linked images don't open a tab on a plain click in the editor
The linked-image anchor's native navigation was firing on a plain click in edit
mode (where handleClick intentionally returns false for caret placement). Prevent
the anchor's default so the editor's handleClick — gated on editable/modifier,
matching text links via openOnClick:false — is the sole navigator.
* fix(sidebar): match the collapse cookie value strictly (not a substring)
A substring search for 'sidebar_collapsed=1' also matched 'sidebar_collapsed=10',
desyncing the pre-paint sidebar rail and client store from the strict server read.
Parse the cookie value and compare it to '1' exactly, in both the pre-paint inline
script and readCollapsedCookie. Added a store test.
* fix(sidebar): reconcile migrated-legacy collapse before paint
A user whose collapse lived only in localStorage has no sidebar_collapsed cookie
at SSR (initialCollapsed=false), but the pre-paint script migrates them to a
cookie. The store's persist.rehydrate() is async (flips _hasHydrated after paint),
so the first paint showed expanded labels in the collapsed 51px rail. Reconcile to
the cookie synchronously in a useLayoutEffect (first render still matches the
server, so no hydration mismatch) — no narrow-rail flash.
|
||
|
Theodore Li
|
7349bf403f |
feat(files): password, email-OTP, and SSO auth for public file shares (#5140)
* feat(files): password, email-OTP, and SSO auth for public file shares * fix(files): suppress filename in share previews for email/sso, not just password * fix(files): normalize allow-list emails to lowercase; genericize shared SSO denial message * fix(security): make isEmailAllowed case-insensitive; normalize email at client gates * test(security): cover isEmailAllowed case-insensitive matching * fix(security): bind auth cookie to auth type; password endpoint rejects non-password shares * chore(db): format generated migration meta * fix(files): share upsert validation returns 400 not 500; disabling always succeeds * feat(access-control): org admins can restrict allowed file-share auth types |
||
|
Siddharth Ganesan
|
5925651cbc |
feat(vfs): add lazy vfs + remove dynamic fields for prompt caching hits (#5138)
* feat(vfs): add lazy vfs + remove dynamic fields for prompt caching hits * feat(vfs): send typed workspace snapshot for append-only deltas Build the workspace inventory from the primary db (fixes replica-lag staleness) and emit it as a typed VfsSnapshotV1 `vfs` payload alongside the markdown, so the mothership can diff it into append-only baseline/delta messages. Generate the TS contract mirror from the Go-owned JSON schema (sync-vfs-snapshot-contract) and sort connector types so diffs stay byte-stable. * fix(lint): fix lint * fix(vfs): forward the typed snapshot through the branch payload builder The branch buildPayload implementations hand-list the params they pass to buildCopilotRequestPayload and forwarded workspaceContext but dropped vfs, so the typed snapshot never reached the Go request (req.Vfs was always nil and the append-only delta path never engaged). Forward vfs in both the workflow and workspace branches, and add a regression guard asserting the branch threads it through (the bug slipped past tests because post.test mocked the payload builder and payload.test called it directly, bypassing the branch). * improvement(contracts): update vfs contracts |
||
|
Theodore Li
|
208d135dac |
feat(enrichment): add enrichment details sidebar with cost + provider cascade (#5139)
* feat(enrichment): add enrichment details sidebar with cost + provider cascade * fix(enrichment): address review — persist detail on cancel/skip, exclude not_run from ran count, refetch on panel open * fix(enrichment): keep cascade detail sticky on upsert; mark unattempted providers not_run on abort * fix(enrichment): show Cancelled in details panel for aborted runs |
||
|
Theodore Li
|
9d2a6ef043 |
feat(logs): redact PII from workflow logs via configurable rules (#5136)
* feat(logs): redact PII from workflow logs via configurable rules Enterprise PII redaction for workflow execution logs, configured under Data Retention as org-scoped rules (each rule picks entity types + which workspaces it applies to). Reuses the guardrails Presidio engine in mask mode at the log-persist choke point, with a check-digit-validated VIN recognizer. Also adds per-workspace data-retention-hours overrides. * fix(logs): widen PII entity visibleValues to string[] for strict build typecheck * fix(logs): redact error/trigger/executionState; keep guardrails import lazy - Extend PII redaction to span error/errorMessage/toolCalls and top-level error/completionFailure/trigger/executionState (Bugbot: PII in execution metadata). executionState is safe to redact — resume reads from the separate pausedExecutions table, not the log copy. - Lazy-import validate_pii in pii-redaction so the Python/child_process guardrails module stays out of the static middleware/RSC graph. - Type the org retention mutation to the contract body (optional, non-null). * refactor(logs): drop per-workspace retention override; PII redaction stays org-scoped - Remove the unused per-workspace data-retention-hours override (no UI; superseded by workspace-scoped PII rules). Reverts cleanup-dispatcher to org-only retention, drops resolveEffectiveRetentionHours, the workspace.dataRetentionSettings column + migration, and the workspace data-retention route/contract/hooks. Fixes Bugbot's null-as-unset finding by removing the buggy path entirely; org retention behavior is unchanged. - Stop re-checking isWorkspaceOnEnterprisePlan at persist time (it returns false on transient errors, which would fail-open and leak PII). Enabled rules already imply entitlement; redact whenever rules apply (fail-safe). * fix(logs): redact oversized strings and executionData.environment - Drop the per-string size cap in PII redaction: oversized strings were left unmasked (leak). Nothing is skipped now; large payloads still fail-safe via the total-bytes ceiling + per-chunk timeout (scrub, never leak). - Add executionData.environment (incl. variables) to the redaction set. * refactor(logs): single-scope PII rules with most-specific-wins resolution Each rule now targets one scope — all workspaces (workspaceId: null) or a single workspace — with workspaceId unique across rules. Resolution is most-specific-wins (a workspace's own rule overrides the all rule), not union; an empty specific rule exempts that workspace. Matches Access Control's resolveWorkspaceGroup precedence. UI 'Applies to' becomes a single-select; Add rule disables when all scopes are taken. * feat(logs): default + workspace-overrides UI for PII redaction Reshape the PII redaction settings into a 'Default (all workspaces)' block plus a 'Workspace overrides' list, making the most-specific-wins precedence explicit (overrides replace the default; unlisted workspaces use it). Same data model (workspaceId null = default), UI only. * improvement(logs): clearer default/overrides PII UI Drop the uppercase section labels and the overrides description; gate the Workspace overrides section behind a configured default; use a single Delete action; 'Add redaction' creates the all-workspaces default and disappears once set. * fix(guardrails): handle stdin EPIPE in PII python spawns Attach an 'error' listener to the child's stdin in both runPythonScript (the batch masking hot path) and executePythonPIIDetection. A 256KB chunk can exceed the OS pipe buffer, so if the Python process exits mid-read (OOM/kill) the EPIPE emitted on stdin was unhandled and would crash the Node process. Funnel it into the promise rejection so the fail-safe scrub path handles it gracefully. * fix(logs): redact executionData.correlation The top-level correlation field is copied from pre-redaction trigger data, so webhook/schedule correlation values could persist unredacted. Add it to the redaction set alongside trigger/environment. * fix(logs): enforce unique PII rule scope server-side The contract accepted multiple rules with the same workspaceId (or several null all-rules); resolution is first-match, so duplicates could disagree with the UI. Add a schema refine rejecting duplicate scopes. * fix(logs): re-hydrate data-retention form on org switch The form hydrated once via a boolean ref, so switching the active org left stale retention days + PII rules and saves targeted the new org with old config. Key hydration on orgId so it re-loads per org. |
||
|
Vikhyath Mondreti
|
13b5d215e9 |
improvement(access-controls): docs, terminology, fix delete bug (#5141)
* improvement(access-controls): dedup independent block names * improvement(access-controls): fix delete button, naming in perm group modal |
||
|
Vikhyath Mondreti
|
91f9dfdaec |
improvement(governance): derived access (#5134)
* improvement(governance): org-ws-credential roles clarity * revert isHosted * improvement(credentials): code cleanup * address comments * make kb cascade delete on user hard delete * revert env flags * chore(db): drop local 0242 migration to regenerate after merging staging Our 0242 collides with staging's 0242. Remove it (and its snapshot + journal entry) so the KB-cascade migration can be regenerated with the correct number on top of the merged staging migrations. * chore(db): regenerate kb→workspace cascade migration as 0243 Regenerated via drizzle-kit generate on top of the merged staging migrations (staging took 0242). Re-applied the safety edits: NOT VALID + separate VALIDATE on the FK re-add, and the -- migration-safe note on the DROP. check:migrations passes. * improve copy * update docs |
||
|
Theodore LiandClaude Opus 4.8
|
c419a34317 |
feat(tables): raise per-plan table limits (free 5/50k, pro 100/100k, max 1k/500k) (#5135)
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> |
||
|
Theodore Li
|
f0b3550729 |
feat(files): public share links for workspace files (#5130)
* feat(files): public share links for workspace files * improvement(files): drop reserved public_share columns until used; sync audit mock * fix(files): share modal tracks authoritative saved state until toggled * feat(files): per-IP rate limit on public share endpoints * fix(files): address PR review — public CSV OOM, content cache, share FK, soft-delete filter, download anchor * fix(files): disable CSV import action in read-only preview (public share) * refactor(files): drive CSV preview import affordance off readOnly, not disableImport * fix(files): version public viewer caches by file updatedAt so edits aren't stale * fix(files): 409 (not corrupt source) when a shared generated doc has no compiled artifact * feat(files): gate public sharing behind an access-control permission |
||
|
Vikhyath Mondreti
|
267e49c69c |
improvement(workspaces): auto-add without invite if part of organization (#5132)
* feat(workspaces): auto-add without invite if part of organization * reverse feature flag hardcoding * address comments * improve ux for org invite modal |
||
|
Theodore LiandClaude Fable 5
|
63fdc472c1 |
improvement(block): table empty-state filter/sort builders + upsert conflict-column selection (#5123)
* ci(migrations): fail dev schema push with an actionable error on rename/drop prompt
`drizzle-kit push --force` only suppresses the data-loss confirm, not the
rename-vs-drop disambiguation prompt. That prompt fires whenever a diff both
adds and drops tables/columns at once (e.g. migration 0231 created
sim_trigger_state while dropping the workspace_notification_* tables), and in
CI it crashes with a bare "Interactive prompts require a TTY" stack trace.
Catch that specific failure in the dev push step and emit a GitHub error
annotation explaining the cause and the fix (drop the stale objects on the dev
DB to match schema.ts — the same DROPs the versioned migration already applied
to staging/prod), instead of leaving an opaque trace. Exit status is preserved
either way.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* improvement(tables): empty-state filter/sort builders + upsert conflict-column selection
* improvement(tables): throw on ambiguous upsert instead of guessing the conflict column
* Revert "ci(migrations): fail dev schema push with an actionable error on rename/drop prompt"
This reverts commit
|
||
|
Vikhyath Mondreti
|
58312a10e5 | improvement(misc): add more sportmonks tools, improvestreaming ux (#5129) | ||
|
Siddharth Ganesan
|
ea83be3bbe |
fix(mship): add folder rename tools and locked workflow status (#5126)
* fix(mship): add folder rename tools and locked workflow status * fix(mship): manage_folder bug fixes * improvement(mship): clean up deprecated fields from contracts * test(mship): update tool-call display title tests for client-derived titles Display titles now come from the Sim-side name resolver, not the stream's ui.title/phaseLabel. Update the read lifecycle test to expect the name-derived title and drop the obsolete phaseLabel-fallback test. * fix(contracts): lint |
||
|
Siddharth GanesanandVikhyath Mondreti
|
e5f3965ed1 |
feat(mship): add parallel subagents, improve streaming performance (#5122)
* feat(subagents): add support for parallel subagents * fix(subagents): address parallel-subagent bugs * progress on streaming refactor * improvement(subagents): update comment to reflect new go feature flag * debug mode progress * remove debug logs * fix(validation): add escape annotation * improvement(code): remove dead fallbacks * fix subagent lane fallback issue * fix(mothership): increase default redis event limit to 100k from 5k * fix(mothership): streaming invariant projection enforcement --------- Co-authored-by: Vikhyath Mondreti <vikhyath@simstudio.ai> |
||
|
Theodore Li
|
597d7eafb5 |
fix(tables): enforce row limits against the current plan, not a frozen per-table cap (#5120)
* fix(tables): enforce row limits against the current plan, not a frozen per-table cap * fix(tables): gate multi-batch CSV create + initial rows against the plan, harden limits cache bound * fix(tables): thread running row count through copilot batchInsertAll capacity check * chore(tables): align tx-variant capacity docstrings * fix(tables): map row-limit errors to 400 in create-from-CSV import * feat(tables): add Upgrade action to the row-limit toast * fix(tables): keep CreateTableData.maxRows so staging callers type-check after merge * improvement(tables): route row-limit Upgrade action to the explore-plans page |
||
|
Theodore Li
|
badfbc3bdf |
fix(resource): left-align table filter/sort when there's no search (#5128)
The unconditional ml-auto from #5117 right-aligned the embedded table editor's filter/sort cluster, which has no search bar. Only push the aside + filter/sort group right when a search occupies the left; without a search it stays left-aligned as before. |
||
|
Theodore Li
|
63a3e6d2cb |
feat(files): stream large CSV previews and add import-as-table (#5125)
* feat(files): stream large CSV previews and add import-as-table * fix(files): validate fileId in csv-preview route, guard double-import, fix sniff perf and toggle flash * fix(files): scope mothership preview-toggle loading guard to CSV files only |
||
|
Theodore LiandClaude Fable 5
|
a028d07e7b |
improvement(mothership): user_table speed parity — limit bounds, background import/delete/update jobs (#5012)
* improvement(mothership): user_table speed parity — limit bounds, async import/delete/update jobs
- query_rows / filter ops clamp limit to the contract maxes; query_rows
skips execution metadata.
- import_file / create_from_file (large CSV/TSV) and delete_rows_by_filter
(>1000 unbounded matches) dispatch background table jobs, claiming the
per-table job slot; inline paths claim the slot too.
- update_rows_by_filter now escalates the same way: >1000 unbounded matches
run as a background table job (new 'update' job type + runTableUpdate worker
+ tableUpdateTask), so a broad update on a huge table no longer loads every
row into the request. Best-effort/non-atomic and skips workflow recompute
(documented); unique-column patches stay inline. Pagination is limit/offset.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs(mothership): trim user_table catalog copy to the essentials
Drop the verbose doomedCount/affectedCount, delete-mask, workflow-recompute,
and unique-column asides from the bulk-op descriptions. The model only needs:
large ops return { jobId }, limit maxes at 1000, one job per table.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* improvement(mothership): make user_table limit cap internal, not model-facing
The model can now pass any limit — no "cannot exceed 1000" rejection. 1000
becomes an internal threshold: query_rows clamps the page to MAX_QUERY_LIMIT
(totalCount signals truncation; the model pages with offset), and bulk filter
ops above the cap run as background jobs.
update_rows_by_filter loads full row data inline, so an explicit limit above
the cap escalates to the background worker with a new maxRows budget (the worker
stops after maxRows; update has no read mask so the cap is exact). delete only
loads ids inline, so an explicit limit (any size) stays inline — only unbounded
deletes use the masked background path, which would over-hide a bounded delete.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* improvement(mothership): bounded delete above the cap runs async, not inline
An explicit delete limit now mirrors update: ≤1000 runs inline, above the cap it
escalates to the background worker honoring the limit via maxRows — instead of
always staying inline. The worker stops after maxRows (per-page fetch capped to
the remaining budget).
Bounded background deletes skip pendingDeleteMask: the filter-based mask hides
every match, which would over-hide the rows beyond the cap the job never deletes.
Unmasked, a bounded delete is eventually consistent like a bounded update (rows
disappear as deleted), and doomedCount is omitted from the payload so the count
isn't double-subtracted.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs(mothership): tidy user_table limit/offset param copy
Drop "Any value is allowed" from the limit description and restore the original
offset description.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(tables): skip pendingDeleteMask for bounded background deletes
The bounded-delete commit (
|
||
|
Waleed
|
7d46103d09 |
chore(deps): remove unused dependencies and harden CI supply chain (#5119)
* chore(deps): remove unused dependencies and harden CI supply chain Dependency cleanup: - Remove unused deps: papaparse, unified, and 6 unused Radix primitives (alert-dialog, radio-group, scroll-area, separator, toggle, visually-hidden) plus @tanstack/react-query-devtools (all verified zero imports repo-wide) - Consolidate jwt-decode into the existing jose dependency (decodeJwt) - Migrate react-window to @tanstack/react-virtual to drop a redundant virtualization library (terminal, structured-output, code viewer) - Remove the better-auth-harmony plugin and its gating env flag Supply-chain hardening: - SHA-pin every GitHub Action to a full commit SHA with a version comment - Pin CI bun-version to 1.3.13 (was "latest" in the release job) - Raise bun minimumReleaseAge cooldown from 3 to 7 days - Add a non-blocking `bun audit` step in CI - Add a CODEOWNERS gate routing dependency-manifest changes to @simstudioai/deps * chore(deps): remove unused apps/docs dependencies (@tabler/icons-react, dotenv-cli) * style(search-modal): use Send icon for Invite teammates action * feat(search-modal): surface New chat as the top action above Create workflow * feat(search-modal): add Secrets to the pages list |
||
|
Theodore Li
|
08bcacd74e | fix(copilot): mount input tables with display-name CSV headers, not column IDs (#5121) | ||
|
Waleed
|
fcfa41cd42 | fix(azure): replace Azure DevOps icon with Azure icon and remove AzureDevOpsIcon (#5118) | ||
|
Waleed
|
cae1769117 |
improvement(knowledge): align connected-sources rows and move source chip left of filter/sort (#5117)
* improvement(knowledge): align connected-sources rows and move source chip left of filter/sort - Drop the -mx-2 on the connectors list so rows respect the ChipModalBody gutter: the row hover no longer bleeds to the modal edges and row content lines up with the px-4 header. - Add a 'leading' slot to ResourceOptions (left of the filter/sort cluster) and render the knowledge connected-source chip there instead of the far-right 'aside', so it reads as part of the control row. 'aside' stays right-aligned for the table editor's run/stop control. * improvement(resource): render options aside left of filter/sort The options-bar aside has a single other consumer (the table editor's embedded run/stop control), so instead of adding a separate slot, render aside itself to the left of the filter/sort cluster. Drops the extra slot and keeps one canonical control position; the run/stop control moves left too, which is fine for a status widget. * fix(resource): keep options aside grouped with filter/sort without a search bar Group aside + the filter/sort cluster in one ml-auto right-aligned container instead of relying on the search's flex-1 to anchor them. Without this, an options bar with no search (the embedded mothership table editor) split aside to the far left and filter/sort to the far right via justify-between. * docs(resource): clarify aside groups with filter/sort regardless of search |
||
|
Waleed
|
4d39b0cbf8 |
feat(connectors): use resource selectors for KB connector config (#5116)
* feat(connectors): use resource selectors for KB connector config Replace raw ID text inputs with selector pickers (canonical selector + manual-input pairs) across Google Drive/Docs/Forms/Sheets, Notion, Monday, and Webflow KB connectors, so users pick folders/spreadsheets/pages/boards/ collections instead of pasting IDs — matching the workflow blocks. - Add multi-select where the sync handler supports it (Drive/Docs/Forms folders, Monday boards, Webflow collections) via parseMultiValue - Add shared escapeDriveQueryValue/buildDriveParentsClause helpers for safe multi-folder Drive queries - Add ConnectorConfigField.mimeType, plumbed into the selector context - Fix Webflow listingCapped not set on maxItems truncation (deletion- reconciliation data-loss safety) Fully backward compatible: legacy single-string IDs and CSV both normalize via parseMultiValue; resolved canonical keys are unchanged. * fix(webflow): set listingCapped on within-page maxItems truncation When a collection's items fit in a single API page but maxItems cuts the list within that page, neither hasMoreInCollection nor hasMoreCollections is true, so listingCapped was not set and the sync engine could hard-delete still-existing documents. Add the within-page drop signal to the guard. |
||
|
Theodore Li
|
ea505f0388 |
improvement(tables): versioned CSV snapshot cache for table mounts + parallel multipart uploader (#5108)
* improvement(tables): versioned CSV snapshot cache for table mounts + parallel multipart uploader * chore(db): drop colliding 0239 migration (renumber pending) * chore(db): renumber rows_version migration to 0240 (off staging's 0239) * improvement(tables): mount snapshots by presigned URL so the sandbox fetches directly (raise cap to 500MB) * fix(tables): allow url sandbox entries in the function-execute contract; key snapshot by column shape so schema edits invalidate it * chore(e2b): log sandbox inputs split by url-fetch vs inline write * improvement(tables): order export + snapshot rows by order_key so the CSV matches the grid under fractional ordering |
||
|
Waleed
|
c907b1194b |
improvement(supabase): add Edge Functions tool; correct storage output shapes + harden tools (#5112)
* improvement(supabase): add Edge Functions tool; correct storage output shapes + harden tools
- Add supabase_invoke_function tool (POST/GET/PUT/PATCH/DELETE /functions/v1/{name})
- upsert: support on_conflict; storage upload: support cache-control
- Fix storage copy/move/upload/delete-bucket output properties to match live API
- get_public_url: build URL via directExecution (no spurious network call)
- text_search: validate column identifier
- Strip non-TSDoc section-label comments
* fix(supabase): harden rpc/text_search identifiers; drop unused get_public_url apiKey
- rpc: validate + encode functionName (SSRF/injection parity with vector_search)
- text_search: validate language config interpolated into the PostgREST operator
- get_public_url: remove unused apiKey param + dead auth headers (public endpoint needs no auth)
- create_bucket: tighten output description to match the {name}-only response
* improvement(tavily): mark optional params advanced; fix empty content output
- Mark 29 optional search/extract/crawl/map subBlocks as mode: advanced (keep query/urls/url/apiKey basic)
- Fix search transformResponse: populate content from result.content (was result.snippet, always empty)
- Guard data.results with ?? []; correct country placeholder to a lowercase name
- Rewrite stale TavilySearch/Extract response interfaces; drop dead duplicate interfaces
* fix(supabase): valid Cache-Control directive on upload; clarify functionName description
- storage upload: expand a bare numeric cache-control to `max-age=<n>` (a raw number is not a valid Cache-Control header)
- block: functionName input description now covers RPC, vector search, and Edge Function invoke
* fix(supabase): edge-function error handling + reject array headers
- invoke_function: drop unreachable !response.ok branch (executor throws on non-OK before transformResponse runs and surfaces the error body); document the success-only contract
- invoke_function: ignore non-object (array) headers so JSON arrays can't produce numeric-index header names
- block: reject array/non-object Edge Function headers with a clear error in config.params
* fix(supabase): scope Edge Function method/body/headers to invoke_function
Prevents a stale `method` value (e.g. from the Edge Function field) from leaking
into other operations' params. The tool executor lets `params.method` override a
tool's static verb (tools/utils.ts), so an unscoped value could turn a read into
DELETE/POST against PostgREST. Now method/body/headers are only passed for the
invoke_function operation. Adds a block-level regression test.
* fix(supabase): only parse Edge Function body/headers for invoke_function
Stale or invalid functionBody/functionHeaders left in the block (common when
switching operations) were parsed and validated for every operation, so they
could throw before unrelated tools ran even while hidden. Moved parsing and
validation inside the invoke_function guard; added a regression test.
* fix(supabase): include last_accessed_at as a storage list sort option
The Storage list API accepts last_accessed_at for sortBy; add it to the tool
description and the block dropdown so the surfaced options match the API.
|
||
|
Waleed
|
11e23131fe |
feat(google): Maps Pollen/Solar, Custom Search expansion, and live-API fixes across Google integrations (#5113)
* feat(google): add Maps Pollen/Solar, expand Custom Search, fix Ads/Groups/Contacts/Slides New capability: - Google Maps: add Pollen Forecast and Solar Potential tools (API-key, google_cloud BYOK) - Google Custom Search: add start/dateRestrict/fileType/safe/searchType/siteSearch/ siteSearchFilter/lr/gl/sort params, htmlTitle/htmlSnippet/formattedUrl/mime/fileFormat/ cacheId/image result fields, and nextPageStartIndex pagination Fixes (validated against live API docs): - Google Ads: bump all tools from sunset v19 to v24 - Google Groups: forward OAuth credential under oauthCredential (was dropping token in 11 ops), forward all update_settings fields, JSON.stringify update_settings/add_alias bodies - Google Contacts: include required metadata.sources[].etag in updateContact body (fixed 400) - Google Slides: remove unsupported GIF thumbnail mimeType (API only allows PNG) - Google Sheets: wire delete_rows/delete_sheet/delete_spreadsheet into the V2 block - Google Custom Search: throw on API error responses instead of returning empty success; num optional + Number-coerced; pagemap typed unknown * docs(google): regenerate integration docs for new and updated operations * fix(google_maps): correct Solar requiredQuality enum to BASE The Solar API ImageryQuality enum is HIGH/MEDIUM/BASE (+ UNSPECIFIED) per the live docs; there is no LOW. Selecting "Low" sent requiredQuality=LOW which the API rejects as INVALID_ARGUMENT, and the valid BASE tier was unreachable. Replace LOW with BASE in the tool param/output descriptions, the type union, and the block dropdown. * fix(google_maps): guard !response.ok in Pollen/Solar; use ?? for color channels Address Greptile review: - Pollen and Solar transformResponse now check !response.ok || data.error (matches the Custom Search fix); a gateway error without an error key in the body no longer returns empty/zeroed output silently. - Pollen color channels use ?? instead of || so a legitimate 0 isn't treated as missing (consistent with the other numeric fields in the file). * fix(google_maps): guard against NaN days in Pollen forecast Address Cursor Bugbot: a non-numeric `days` input parsed to NaN and was forwarded as `days=NaN` (the tool's `?? 1` only catches undefined, not NaN), breaking the forecast call. The block now coerces invalid input to undefined, and the tool defaults to 1 unless `days` is a finite number. * fix(google): clamp Pollen days to 1-5; stop forwarding stale group settings fields Address Cursor Bugbot: - Pollen: clamp days to the documented 1-5 range (truncating fractionals) so 0, negatives, or >5 can't be sent to the API. - Google Groups update_settings: the block has no dedicated settings subblocks, so forwarding name/description from params could leak stale values from create_group/update_group and unintentionally rename the group. Forward only oauthCredential + groupEmail from the block (the tool's own param schema still exposes the settings fields for the agent path). * fix(google_sheets): fail fast on non-numeric delete indices Address Cursor Bugbot: delete_sheet/delete_rows parsed deleteSheetId/startIndex/ endIndex with Number.parseInt but didn't validate, so non-numeric UI input became NaN and was forwarded (the v2 delete tools only reject null/undefined), breaking the batchUpdate. The block now throws a clear error when any of these is not a valid number. * fix(google_search): clamp num to 1-10 and normalize start Address Cursor Bugbot: num was coerced with Number() but not bounded, so values like 11 or fractionals reached the API and failed. The tool now truncates and clamps num to the documented 1-10 range and only sends a positive integer start, ignoring non-numeric/out-of-range input. |
||
|
Waleed
|
d7fd0405f9 |
improvement(search): align cmd+k action icons + highlight with the design system (#5114)
* improvement(search): align cmd+k action icons + highlight with the design system - Each Actions verb now uses the exact icon from its real location: Fit to view -> Scan (workflow-controls), Copy workflow link -> Duplicate (nav context menu), Invite teammates -> User (settings teammates nav). Run/Create/Import already matched. - Remove the Toggle theme action and its now-dead useTheme wiring. - Matched-text highlight now uses the design-system search tokens (--highlight-match-bg / --highlight-match-text), matching the SearchHighlight component used in knowledge-base and code search, instead of an ad-hoc font-semibold. * improvement(search): use font-medium for matched-text emphasis in cmd+k Drop the colored background highlight in favor of the design system's standard emphasis weight (font-medium, used by Button/Label/Input/Table). Lighter than the previous semibold and avoids a background, keeping the palette's clean, undecorated text style. |
||
|
Waleed
|
9e9f2b9e1f |
fix(realtime): debounce the reconnecting toast to stop transient-blip flashes (#5111)
* fix(realtime): debounce the reconnecting toast to stop transient-blip flashes The "Reconnecting..." persistent toast fired the instant isReconnecting flipped true, so sub-second transport blips that self-heal on the first retry flashed a scary alert. Add useStableFlag, an anti-flicker boolean that delays the rising edge (2s, so brief blips never surface) and holds the falling edge (1.5s min visible, so a drop just past the delay does not flash-and-vanish). The socket flag stays accurate; only the user-facing alarm is smoothed. State machine extracted into a framework-agnostic controller with unit coverage for both flicker modes. * fix(realtime): reset stable-flag React state on options change; de-vacuous blip test Address Greptile review: - useStableFlag: reset React state to the fresh controller's baseline when the controller is recreated on an options change, so a dynamic consumer changing delayMs/minVisibleMs while active with value already false can no longer strand the flag at true. - test: read the live probe.active getter in the blip test instead of a destructured snapshot, which was bound to false at destructure time and made the assertion vacuous. |
||
|
Waleed
|
05cd7d92a8 |
feat(search): actions, fuzzy matching, and highlighting in cmd+k palette (#5110)
* feat(search): actions, fuzzy matching, and highlighting in cmd+k palette Add a context-aware actions layer to the cmd+k search palette (Run workflow, Create workflow/folder, Import workflow, Fit to view, Copy link, Invite teammates, Toggle theme), replace the substring matcher with a boundary-anchored fuzzy matcher (initialisms, typos, multi-word) that is a strict superset of the old behavior, highlight matched characters, and rank against clean human text instead of structural id/uuid tokens. Expose invoke() on the global commands provider so the palette runs real registered commands. * fix(search): highlight the matched substring, not an earlier scattered occurrence Contiguous substring matches (exact/prefix/contains) now report the substring's own indices instead of the greedy subsequence scan positions, so HighlightedText bolds the characters the user actually matched. Restructures fuzzyMatch to handle the substring tier first; scores are unchanged for these cases. * fix(search): log clipboard copy failures and make fuzzy positions read-only - Copy workflow link now logs on clipboard write failure instead of silently swallowing the error, matching the sidebar's copy-link convention. - FuzzyResult.positions is now readonly and the NO_MATCH singleton's array is frozen, so the shared instance can never be mutated by a caller. |
||
|
Waleed
|
8b93e43037 |
improvement(integrations): validate BigQuery/Forms/PageSpeed + regenerate integration docs (#5109)
* improvement(integrations): validate BigQuery/Forms/PageSpeed + regenerate integration docs - BigQuery: mark null-defaulted outputs optional (get_table type/numRows/numBytes/creationTime/lastModifiedTime/location, list_datasets location, list_tables type, query totalBytesProcessed) - Google Forms: add response pagination (pageToken + filter params, nextPageToken output), fix pageSize visibility, advanced-mode pagination subBlocks + filter wandConfig - PageSpeed: add a 7th BlockMeta template (competitor benchmark) - Regenerate integration docs; add manual intro sections to new datagma/dropcontact/enrow/icypeas/leadmagic pages * fix(docs-gen): preserve apostrophes in tool descriptions when generating docs The doc generator extracted tool descriptions with a character class that excluded both quote types (['"]([^'"]...)['"]), so a double-quoted description containing an apostrophe (e.g. "Find someone's email") was truncated at the apostrophe — the generated docs/catalog showed stubs like "Find someone". Anchor extraction on the actual opening quote (single/double/backtick), matching the existing extractDescription helper, in both buildToolDescriptionMap and extractToolInfo. Regenerated docs restore full descriptions across all affected integrations (Apollo, Ahrefs, LeadMagic, Findymail, OpenAI, Slack, etc.). * fix(docs-gen): resolve tools defined in a sibling file + scope params per tool The doc generator located a tool's definition only by filename convention (decompress.ts / index.ts), so file_decompress — which lives in compress.ts alongside file_compress — fell back to index.ts and rendered an empty Input table. It also read the params block from the first tool in a multi-tool file, so every tool in such a file inherited the first tool's inputs/outputs. - getToolInfo: when no candidate file declares the exact tool ID, scan the whole tool-prefix directory for the file that does. - extractToolInfo: read the params block scoped to the specific tool, falling back to the full file for tools that inherit params via spread. Regenerated docs eliminate ~50 empty/incorrect input tables across integrations (clickhouse, rb2b, reddit, file, etc.); param-less OAuth-only tools correctly keep an empty input table. |
||
|
Waleed
|
80735b424b |
fix(locks): enforce workflow/folder locks on the agent + close manual-UI create gaps (#5107)
* fix(locks): enforce workflow/folder locks on the agent + close manual-UI create gaps The copilot/agent workflow & folder mutation tools and the edit_workflow tool bypassed lock enforcement, so the agent could edit a locked workflow and move/create workflows into a locked folder. Add assertWorkflowMutable/ assertFolderMutable guards (from @sim/workflow-authz) to every agent mutation path, mirroring the REST API. Also close two parity gaps on the manual-UI REST side: creating a workflow into a locked folder and creating a subfolder under a locked parent were previously unguarded. The realtime collaborative canvas already enforced workflow-level locks server-side. * fix(locks): normalize optional folderId to null for assertFolderMutable * refactor(locks): hoist constant folder-lock check out of move loop; scope test mocks with Once * refactor(locks): drop redundant ensureWorkflowAccess fetch in rename |
||
|
Waleed
|
83531452e5 |
fix(sidebar): prefetch chats + workflows so cold loads don't flash skeletons (#5104)
* fix(sidebar): prefetch chats + workflows so cold loads don't flash skeletons On a cold load (e.g. when the browser discards an idle tab and reloads), the persistent sidebar started with an empty React Query cache and client-fetched its chat + workflow lists, flashing loading skeletons. Prefetch both lists server-side in the workspace layout and hydrate them via HydrationBoundary, under the same query keys and mappers the client hooks use, so the sidebar paints populated on the first render. The prefetch runs concurrently with the existing org-settings fetch and never throws, so it adds no blocking work in the common case and falls back to client fetching on error. * refactor(prefetch): call data layer directly instead of internal HTTP self-fetch The sidebar and settings prefetches fetched their data by making internal HTTP requests to our own API routes. Replace those self-fetches with direct calls to shared server-side data functions, so each route handler and its prefetch read from one source with no extra network hop, serialization, or re-auth. - Extract listWorkflowsForUser (lib/workflows/queries) and listMothershipChats (lib/copilot/chat) from their routes; both routes and the sidebar prefetch now call them. - Extract getUserSettings/getUserProfile (lib/users/queries) shared by the settings/profile routes and their prefetches. - Subscription prefetch calls the existing getSimplifiedBillingSummary + getEffectiveBillingStatus directly. - Sidebar prefetch checks workspace access once via checkWorkspaceAccess and skips silently when denied. * refactor(prefetch): share mothership chat list staleTime constant Export MOTHERSHIP_CHAT_LIST_STALE_TIME from the chats hook and use it in both useMothershipChats and the sidebar prefetch, mirroring WORKFLOW_LIST_STALE_TIME so the prefetch and client hook can't drift. * fix(prefetch): keep subscription prefetch on the wire shape via internal billing API The billing summary returns Date fields (and an untyped metadata blob) that the JSON API serializes to strings. Calling the data layer directly would cache Date objects (App Router preserves them through RSC serialization), mismatching the string wire shape the client useSubscriptionData hook caches. Route the subscription prefetch through the internal billing API so server-hydrated and client-fetched data share the exact same shape. The date-free general-settings and profile prefetches keep calling the data layer directly. |
||
|
Waleed
|
a82b44d36d |
perf(db): logs-list index, drop redundant indexes, replica routing, hot-path write cleanups (#5105)
* perf(db): logs-list index, drop redundant indexes, replica routing, hot-path write cleanups * fix(logs): keep /api/v1/logs on primary db — its permissions join is the auth gate, not replica-safe |
||
|
Vikhyath Mondreti
|
8fe090a3a1 |
fix(input-format): field not editable race condition (#5102)
* fix(input-format): field not editable race condition * remove dead code * simplify |
||
|
Waleed
|
2ffc004a0a | improvement(models): add DeepSeek V4 + Mistral Medium 3.5, fix Codestral context window (#5103) | ||
|
Theodore LiandClaude Opus 4.8
|
15a970d805 |
feat(integrations): hosted email-enrichment providers + cascade wiring (#5087)
* feat(integrations): hosted email-enrichment providers + cascade wiring Add Datagma, Dropcontact, LeadMagic, Icypeas, and Enrow integrations — tools, blocks, brand icons, and BYOK + metered hosted-key support — and register each in the tool/block registries and BYOK provider list. Wire the new finders/verifiers into the enrichment cascades: - work-email: Datagma, LeadMagic, Dropcontact, Icypeas, Enrow - phone-number: LeadMagic, Datagma, Dropcontact - email-verification: Icypeas, Enrow - company-info: Datagma, LeadMagic - company-domain: Datagma Add hosting tests for all five providers and cascade tests covering the new providers (incl. new test files for email-verification, company-info, and company-domain). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(enrichment): address PR review on Icypeas success + Datagma billing - Icypeas find_email/verify_email postProcess return success:true for all terminal statuses (NOT_FOUND/DEBITED_NOT_FOUND included) so the cascade runner calls mapOutput and records invalid/not-found verdicts instead of throwing and inflating the error count - Bill Icypeas verify FOUND (not just DEBITED*) per the documented 0.1-credit charge - Datagma enrich_person only applies the 30-credit phone surcharge when a phone lookup (phoneFull) was requested - Note Datagma's URL-param (apiId) auth in the hosted-key doc comment - Update hosting tests to match Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(enrichment): only bill Enrow verify on a completed verification getCost returned a flat 0.25 credits regardless of output, so a job that fell back to the initial submit response (poll never completed, no qualification) was still metered. Charge 0.25 only when a qualification is present; 0 otherwise. Add a no-qualification test case. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * chore(enrichment): peg hosted credit cost to each provider's lowest paid plan Align *_CREDIT_USD to the entry tier Sim will provision: - Datagma: Regular $49/3,000 emails → $0.0163 (was Popular $0.0132) - LeadMagic: Basic $49/2,000 → $0.0245 (was Growth $0.0104) Icypeas (Basic $0.019), Enrow (Starter $0.012), and Dropcontact (Starter ~$0.17) already reflect their lowest plan. Tests derive from the constants, so values stay consistent. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(enrichment): address PR review on mononyms + Icypeas verify email map - work-email LeadMagic: pass full_name + domain so single-token (mononym) names are no longer skipped - work-email Icypeas: firstname/lastname are optional on the API, so run a mononym with firstname alone instead of self-skipping - icypeas_verify_email mapItem reads item.email (verify payload shape) with a fallback to the nested results.emails[0].email Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(enrichment): case-insensitive Enrow find billing getCost compared qualification to exactly 'valid' while the cascade normalizes with toLowerCase(), so a differently-cased API qualifier could zero out billing on a valid email. Lowercase before comparing; add a test. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(enrichment): drop Dropcontact from the phone-number cascade Dropcontact is an email/company-data enrichment service, not a phone-discovery provider — its phone/mobile_phone fields are unreliable and were surfacing firmographic data (an employee-count range like "5000-20077") as the phone. Keep the two purpose-built phone finders (LeadMagic find_mobile, Datagma find_phone); Dropcontact stays in the work-email and company cascades where its data is reliable. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(enrichment): only accept valid-qualified Enrow emails in work-email Enrow's finder qualifies each email valid/invalid. The work-email mapOutput accepted any non-empty email, so an invalid-qualified address could fill the cell while hosted billing (which only charges on valid) charged zero. Gate the cell on qualification === 'valid', consistent with billing. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> |
||
|
Vikhyath Mondreti
|
feca5fa6f6 |
improvement(execution, connectors): offload large function inputs, increase connector limits + better error propagation (#5089)
* fix(execution,connectors): offload large function inputs; harden KB connector size limits Addresses a class of 10 MB limit failures: - executor/variables: offload over-budget function block-output context values to durable large-value refs (lazy `sim.values.read`) so JS function blocks can merge medium files without exceeding the 10 MB inter-block request-body cap. - connectors: stream downloads via `readBodyWithLimit` (memory-safe), and surface oversized files as visible `failed` KB documents instead of silently dropping them — listing-time for github/s3/dropbox/onedrive/sharepoint, fetch-time for gitlab/azure/google-drive via a shared `ConnectorFileTooLargeError`. Raise the per-file cap from a hardcoded 10 MB to the canonical 100 MB KB document limit (`CONNECTOR_MAX_FILE_BYTES`), except Google Drive's export path (Google's hard 10 MB export-API limit). - sync-engine: `classifyExternalDoc` + bulk `skipDocuments` (failed rows with a reason, excluded from retry), byte-bounded batch concurrency to cap peak worker memory at the raised cap, and a `metadata.fileSize ?? size` fallback. * fix zoom * update skill * address comments + fix terminal event in sse stream * fix accounting issue |
||
|
Waleed
|
cc56408be3 |
perf(execution): parallelize preflight gates, cache deployed state, memoize Anthropic client (#5098)
* perf(execution): parallelize preflight gates, cache deployed state, memoize Anthropic client
- Memoize Anthropic + Azure-Anthropic SDK clients (new client-cache.ts) keyed
by apiKey (+beta header; +baseURL/version/pinnedIP for Azure) so HTTP
keep-alive connections are reused instead of a fresh TLS handshake per call.
apiKey is the tenant boundary.
- Parallelize the read-only preflight gates in preprocessing.ts (ban +
subscription, then usage + org-member + rate-limit) while preserving exact
error precedence (ban 403 -> usage 402 -> rate 429) and keeping the sole
write (admission reservation) last.
- Parallelize the independent workflow-state and env-var loads in execution-core.
- Cache deployed workflow state by immutable deploymentVersionId with
deep-clone-on-read, oldest-first eviction, and a 5-min TTL bounding the
credential-mapping edge across ECS tasks.
- Parallelize the independent personal-subscription + membership queries in
getHighestPrioritySubscription.
- BYOK: drop the redundant getWorkspaceById existence check (auth already
validates the workspace); read the key list fresh every call for zero
cross-instance staleness.
Billing/usage/ban/permission reads stay fresh on the primary (no cache, no
replica). Adds tests for every new mechanism and fixes a pre-existing vitest
class-mock incompatibility that had execution-core.test.ts fully red on staging.
* fix(execution): run rate-limit gate only after ban/usage pass
The rate-limit gate is not read-only — checkRateLimitWithSubscription consumes
a token — so running it in parallel with the read-only gates debited rate-limit
quota for requests that the ban (403) or usage (402) gates reject, which the
original sequential flow never did.
Move the rate-limit gate to run sequentially after the ban and usage gates pass,
preserving the read-only gates' parallelism (ban + subscription + usage) and the
exact ban -> usage -> rate precedence. Add regression tests asserting the rate
limiter is not consumed when an earlier gate rejects, and is consumed once when
they pass.
Caught by Cursor Bugbot review.
* chore(execution): trim redundant preflight comments
Tighten the gate overview to match the sequential rate-limit gate and drop
inline notes that duplicated it or the runRateLimitGate doc.
* refactor(cache): address review — idle TTL for client cache, LRUCache for deployed state
- client-cache: add updateAgeOnGet so the TTL is genuinely idle-based (active
clients keep their warm keep-alive connections; the JSDoc now matches behavior).
- deployed-state: replace the hand-rolled Map + manual FIFO eviction/TTL with
LRUCache (real LRU eviction, built-in TTL), matching the effectiveDecryptedEnv
and integration-tool-schema caches. TTL stays absolute (not reset on read) so
the credential-migration remap still propagates across ECS tasks.
Both per review feedback from Greptile.
* test(execution): isolate rate-limit gate test from STEP 7 reservation
The 'consumes the rate-limit gate once' test reached the STEP 7 admission
reservation, which depends on Redis — it passed locally (reserve throws and is
swallowed) but failed in CI (reserve returns not-reserved -> 429). Pass
skipConcurrencyReservation so the test isolates the rate gate deterministically.
* perf(providers): memoize SDK clients where the pool is per-client (bedrock, vllm)
Generalize the Anthropic client cache into one shared memoizer
(providers/client-cache.ts) and apply it only where each new client owns its own
connection pool — so reuse actually keeps connections warm:
- bedrock: AWS SDK clients hold a per-client connection pool (reuse is the AWS
best practice). Keyed by region + credential identity.
- vllm: a pinned endpoint creates its own undici Agent per call; key by the
resolved IP so DNS re-validation still runs each request.
- anthropic + azure-anthropic: migrated onto the shared memoizer.
Deliberately NOT applied to the OpenAI-compatible providers, groq, cerebras, or
google: their SDKs share a process-global keep-alive pool (Node openai-sdk module
singleton agent; anthropic/global undici), so a fresh client per request already
reuses connections and memoization would add complexity with ~no benefit. litellm
uses a plain shared-agent client (no pinning) and is likewise skipped.
Bounded LRU (max 1000, 30m idle TTL) with no close-on-eviction, avoiding the
unbounded-growth and eviction-closes-in-use-client failure modes seen in similar
client caches.
* chore(perf): trim verbose comments to terse why-notes
* chore(perf): drop obvious inline comments, keep nuance as TSDoc
* fix(bedrock): key client cache on full credential, not just access key id
A corrected secret under the same access key id would otherwise keep serving the
stale cached client until TTL/eviction. Caught by Cursor Bugbot.
* test(execution,providers): fix preflight mock reset + isolate provider client cache in tests
- preprocessing.test: re-establish the checkOrgMemberUsageLimit mock in beforeEach
(the only gate mock not re-set). In the full suite its implementation was reset
so the success-path test got undefined -> threw -> 500 -> success:false. Mirrors
how checkServerSideUsageLimits is handled.
- client-cache: add clearProviderClientCacheForTests; call it in the bedrock and
vllm test beforeEach so construction assertions always start from a cache miss
now that those providers memoize their client.
* test(execution): make RateLimiter mock constructable under vitest 4.x
The RateLimiter mock used an arrow factory (vi.fn(() => ({...}))). vitest 4.x
(CI) rejects `new` on an arrow-implemented mock ("not a constructor"); 3.2.4
allowed it. The new rate-gate test is the first to actually `new RateLimiter()`,
so it surfaced the failure only in CI. Switch the mock to a regular function and
drop the speculative beforeEach re-establishments that didn't address it.
|
||
|
Waleed
|
f238184fa0 |
feat(file): add Compress and Decompress operations to the File block (#5100)
* feat(file): add Compress operation to bundle files into a .zip archive * feat(file): add Decompress operation to extract .zip archives Adds the inbound half of the archive pair: extracts a .zip back into the workspace with zip-slip path sanitization, symlink skipping, and entry/ size caps to bound zip-bomb expansion. Extracted files are returned in the files output, ready to chain downstream. * fix(file): align archive ops with v5 output surface and zip mime - Drop the single 'file' output reintroduced for compress/decompress; v5 intentionally exposes only 'files' (plus id/name/size/url scalars), so compress/decompress reuse the existing surface with no new block output - Add zip/gz to EXTENSION_TO_MIME (previously only in the reverse map), so archive extensions resolve to a real mime instead of octet-stream - Update File v5 block test for the two new operations * fix(file): harden compress naming per review - Flatten zip entry names to a safe basename so untrusted fileInput names with .. or / cannot produce zip-slip entry paths (cursor) - Treat archiveName as a flat name landing at the workspace root instead of passing it through splitWorkspaceFilePath, which silently created folders for names with separators (greptile) - Add the upfront empty-input guard before any DB calls, matching the read and content operations (greptile) * fix(file): make decompress extraction atomic and bound per-entry size - Read and validate every entry before writing any file, so hitting a size cap no longer leaves partially-extracted files in the workspace (cursor) - Enforce the per-entry cap on the materialized buffer in addition to the declared size, covering entries that omit an uncompressed size (cursor) - Pre-check declared sizes up front to reject standard zip bombs before materializing, and return 422 when no files could be extracted (cursor) * fix(file): exclude skipped entries from caps and reject multi-archive decompress - Resolve safe (sanitized) zip entries up front so unsafe/skipped entries no longer count toward the per-entry and total uncompressed-size caps (cursor) - Reject decompress input that resolves to more than one archive with a clear error instead of silently extracting only the first (cursor) * fix(file): enforce single-archive decompress at the API boundary The block already rejects multiple archives, but the manage route is the real boundary (callable directly and by the LLM tool) and still took the first of multiple resolved inputs. Add the empty-input and >1-archive guards in the route so extra archives are rejected with a clear error rather than silently ignored (cursor). * docs(file): correct compress description and stale file-output references - Drop the misleading 'under provider upload limits' claim from the compress tool description (models cannot read zip archives) - Fix bestPractices to reference the 'files' output, not a non-existent 'file' - Remove the stale 'file' property from the compress test fixture so it matches the real API response (greptile) |
||
|
Waleed
|
c864a928bf |
improvement(models): sort model dropdown by latest release date within each provider (#5099)
* improvement(models): sort model dropdown by latest release date within each provider * fix(models): preserve input provider order and build catalog index once |
||
|
Vikhyath Mondreti
|
73c73ff0f2 |
fix(kb): canonicalize knowledge-base upload keys (#5096)
* fix(kb): canonicalize knowledge-base upload keys * fix tests |
||
|
Vikhyath Mondreti
|
d14bc78d03 |
fix(realtime): re-check workspace role on mutating socket events (#5080)
* fix(realtime): re-check workspace role on mutating socket events * address comments |
||
|
Waleed
|
2b7f57a788 |
improvement(providers): tighten Gemini and vLLM agent-attachment ceilings (#5095)
A live-doc audit of the merged large-file feature found two ceilings that were higher than the provider actually accepts: - Gemini: 100MB -> 50MB. Gemini hard-caps PDFs at 50MB, so a 50-100MB PDF passed our gate, got uploaded + polled, then failed at generateContent. 50MB respects the documented limit and is more memory-safe. - vLLM: 50MB -> 25MB. vLLM's default image-fetch timeout is 5s; a 50MB remote fetch routinely exceeds it. 25MB aligns with that reality and matches Baseten (the other vLLM-backed provider). |
||
|
Waleed
|
2c1392e3a0 |
fix(chat): autoscroll follow-ups — re-engage threshold + keep end-of-turn options in view (#5094)
* fix(chat): align scrollbar/keyboard detach with wheel/touch re-engage threshold The onScroll detach branch set only stickyRef.current = false, leaving userDetachedRef false, so a scrollbar-drag or keyboard detach kept the lenient 30px (STICK_THRESHOLD) re-engage threshold instead of the strict 5px (REATTACH_THRESHOLD) used after wheel/touch. A programmatic virtualizer re-pin landing within 30px could then snap autoscroll back on right after the user deliberately scrolled away. Reuse the detach() helper so all detach paths set userDetachedRef consistently. * fix(chat): keep end-of-turn options in view after streaming When a stream ends, the suggested-follow-up options and the actions row (gated on !isStreaming) mount, but the virtualizer's getTotalSize — which drives the scroll container's scrollHeight — only catches up a frame or two later via its ResizeObserver. The single scrollToBottom() on effect teardown therefore landed on a stale, too-short bottom and the options were clipped behind the input. (Pre-virtualization this worked because scrollHeight reflected the new rows immediately.) Extract the rAF follow loop already used for CSS height animations into a shared followToBottom(window) helper and run it for a short settle window on teardown, so the bottom is chased until the virtualizer re-measures. The follow is self-interrupting — height growth leaves scrollTop where we put it, while a user scroll moves it up, so it bails the instant the user scrolls and never fights a real gesture even with listeners torn down. |
||
|
Waleed
|
e1f22bdac9 |
feat(providers): support large agent-block attachments via Files APIs and remote URLs (#5092)
* feat(providers): support large agent-block attachments via Files APIs and remote URLs
Agent-block file uploads were inlined as base64 with a hard 10MB cap. Files
above the threshold now use each provider's native large-file path:
- OpenAI / Gemini: upload to the provider Files API, reference by file_id/uri
- Anthropic: GA url content-block source (no Files API beta, no upload)
- OpenRouter/Groq/Together/Baseten/xAI/vLLM: remote signed URL in image_url/file
- Limits live per-provider in models.ts; the agent block + /models page reflect them
Files <=10MB keep the identical base64 path (zero regression). Server-only file
handles are stripped from untrusted input to prevent SSRF.
* fix(providers): clear forged file handles for inline providers too
attachLargeFileRemoteUrls early-returned for inline-strategy providers before
clearing server-only handle fields, so a forged remoteUrl on an inline-provider
file could still reach a builder (e.g. buildOpenAICompatibleChatContent for
mistral/ollama). Clear the handles for every provider before the strategy check.
* fix(providers): correct OpenAI expiry serialization and Anthropic large-text-doc handling
- OpenAI upload now uses the SDK (client.files.create) so expires_after is
serialized as a real nested object; the prior expires_after[anchor] bracket
FormData keys were ignored by OpenAI's server, leaving files un-expiring.
- Anthropic url document source only supports PDFs/images; large non-PDF text
docs now throw a clear error instead of emitting an unsupported url source.
- Warn when an oversized file can't be sent because cloud storage is unavailable.
* fix(providers): harden large-file path (SSRF fetch, ceiling gate, per-file UI limit)
- Download files for OpenAI/Gemini uploads via validateUrlWithDNS + IP-pinned
fetch so a forged URL can't reach internal addresses (covers all callers).
- Reject files above the provider ceiling before downloading/uploading.
- UI now validates each file against the provider's per-file ceiling instead of
summing all files against it, matching server-side per-file validation.
- Lower Anthropic ceiling to 50MB (documented 32MB request cap / page limits).
* refactor(providers): read files-api upload bytes via storage SDK
Read OpenAI/Gemini upload bytes through downloadFileFromStorage instead of
HTTP-fetching the presigned URL. Removes any server-side URL fetch (no SSRF
vector) and works with internal object storage (e.g. self-hosted MinIO), which
an IP-pinned URL fetch would have blocked.
* docs(providers): clarify files-api bytes are read from storage at upload time
* fix(providers): enforce access checks and strip forged ids in the upload path
uploadLargeFilesToProvider runs on raw request messages for every caller (incl.
the internal providers passthrough), so harden it independently of the agent path:
- verifyFileAccess on each file's storage key before reading its bytes, so a forged
key can't exfiltrate another user's file.
- clear any inbound providerFileId/providerFileUri up front (legit ids are only set
by the upload itself), so a forged id can't reference a file in a hosted account.
* fix(providers): resolve UI attachment limit with the same model->provider helper as execution
The file-upload control imported getProviderFromModel from @/providers/models, but
the execution path and every other consumer use the one in @/providers/utils (runtime
registry + reseller patterns). Align the UI so its size cap can't disagree with
server-side validation for reseller or dynamically-listed models.
* test(providers): add new models.ts exports to provider mocks
attachments.ts now reads getProviderFileAttachment / INLINE_ATTACHMENT_MAX_BYTES
from @/providers/models; the provider unit tests that fully mock that module need
both exports or attachments.ts fails to load.
* fix(providers): guard Gemini upload response name before polling
ai.files.upload returns name as string | undefined; guard it (instead of an
as-string cast) so a missing name surfaces a clear error at the upload site
rather than an opaque files.get failure on the first poll.
* fix(uploads): type the file-handle key list so omit preserves UserFile fields
The 'as const' readonly tuple widened omit's K to all keys, collapsing
Omit<UserFile, K> to {} and failing the production build's type check. Declare
the array as Array<keyof handle fields> so K is the precise literal union.
* refactor(providers): run handle-clear + URL-mint in executeProviderRequest for all callers
Move attachLargeFileRemoteUrls out of the agent handler and into
executeProviderRequest (right before uploadLargeFilesToProvider), so every entry
point — including the internal providers passthrough — clears forged handles and
mints/access-checks large-file URLs uniformly. The agent handler now only hydrates
base64; its missing-file guard exempts large files (resolved downstream).
* fix(azure-openai): guard optional attachment dataUrl in inline image part
PreparedProviderAttachment.dataUrl is now optional (large files carry a handle
instead); azure-openai builds chat content inline and assigned it directly to a
required url field, failing the production build's type check.
* fix(providers): upload OpenAI files via multipart and fix Buffer Blob part
The installed openai SDK (4.104) does not type expires_after on files.create, so
upload via POST /v1/files directly with the documented expires_after[...] form
fields (gives the file an auto-expiry). Also wrap the storage Buffer in a
Uint8Array for the Blob, which the production build's stricter lib types require.
These two type errors were masked locally because tsc was OOMing silently without
the type-check script's --max-old-space-size flag.
* fix(providers): forward userId from the providers API to executeProviderRequest
Large-attachment prep now needs request.userId for presigned URLs and access
checks; the authenticated providers proxy has auth.userId but wasn't passing it,
so oversized attachments failed for logged-in callers. Forwarding it makes large
files work there and keeps the access check (verifyFileAccess) intact.
* fix(providers): fail clearly when a large attachment has no cloud storage
The doc claimed a base64 fallback that doesn't exist — above the inline cap there
is no base64, so without cloud storage the file previously reached the builder and
died with a generic read error. Throw a clear 'requires cloud file storage' error
at the point of detection and correct the doc.
|