)

* fix(landing): fix image rendering and navbar blog/docs navigation

* fix(navbar): remove disclosure aria attributes from navigation links

* chore: fix conflicts

* chore: fix conflicts

* chore: pause marquee

* chore: fix conflicts

* chore: fix conflicts

* chore: address review comments

* chore: fix conflicts

* chore: other bug fixes

* chore: fix conflicts

* chore: fix radius

* chore: fix review changes

* chore: fix conflicts

* chore: revert other

* chore: fix conflicts

* chore: fix review changes

* chore: fix conflicts

* chore: fix review changes

* chore: fix button state

* chore: fix button state

* chore: fix review changes

* chore: fix lint

* chore: fix conflicts

* chore: add metadata

* chore: fix things

* chore: fix overwritten states

* chore: fix warnings

* chore: fix button state

* chore: fix review changes

* chore: fix review changes

* chore: fix hover state

* chore: fix popover

* chore: fix review changes

* chore: fix review changes

* chore: fix conflicts

* chore: update contents

* chore: fix review changes

* chore(config): clean up bun, turbo, and next.js config

* chore(ci): bump bun to 1.3.11 in dockerfiles and workflows

* chore: fix cn

* chore: fix scale

* feat(granola): add Granola meeting notes integration

* fix(granola): use string comparison for includeTranscript to avoid truthy string bug

* fix(granola): add missing get_note output fields to block definition

* regen docs

* feat(ketch): add Ketch privacy consent integration

* fix(ketch): add response.ok guards and fix registry ordering

* fix(ketch): include errorMessage in error response output for all tools

* fix(ketch): wire optional purposes filter for get_consent operation

…es (#3792)

* fix: prevent auth bypass via user-controlled context query param in file serve

The /api/files/serve endpoint trusted a user-supplied `context` query
parameter to skip authentication. An attacker could append
`?context=profile-pictures` to any file URL and download files without
auth. Now the public access gate checks the key prefix instead of the
query param, and `og-images/` is added to `inferContextFromKey`.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: use randomized heredoc delimiter in SSH execute-script route

Prevents accidental heredoc termination if script content contains
the delimiter string on its own line.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: escape workingDirectory in SSH execute-command route

Use escapeShellArg() with single quotes for the workingDirectory
parameter, consistent with all other SSH routes (execute-script,
create-directory, delete-file, move-rename).

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: harden chat/form deployment auth (OTP brute-force, CSPRNG, HMAC tokens)

- Add brute-force protection to OTP verification with attempt tracking (CWE-307)
- Replace Math.random() with crypto.randomInt() for OTP generation (CWE-338)
- Replace unsigned Base64 auth tokens with HMAC-SHA256 signed tokens (CWE-327)
- Use shared isEmailAllowed utility in OTP route instead of inline duplicate
- Simplify Redis OTP update to single KEEPTTL call

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: harden SSRF protections and input validation across API routes

Add DNS-based SSRF validation for MCP server URLs, secure OIDC discovery
with IP-pinned fetch, strengthen OTP/chat/form input validation, sanitize
1Password vault parameters, and tighten deployment security checks.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* lint

* fix(file-serve): remove user-controlled context param from authenticated path

The `?context` query param was still being passed to `handleCloudProxy`
in the authenticated code path, allowing any logged-in user to spoof
context as `profile-pictures` and bypass ownership checks in
`verifyFileAccess`. Now always use `inferContextFromKey` from the
server-controlled key prefix.

* fix: handle legacy OTP format in decodeOTPValue for deploy-time compat

Add guard for OTP values without colon separator (pre-deploy format)
to avoid misparse that would lock out users with in-flight OTPs.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix(mcp): distinguish DNS resolution failures from SSRF policy blocks

DNS lookup failures now throw McpDnsResolutionError (502) instead of
McpSsrfError (403), so transient DNS hiccups surface as retryable
upstream errors rather than confusing permission rejections.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: make OTP attempt counting atomic to prevent TOCTOU race

Redis path: use Lua script for atomic read-increment-conditional-delete.
DB path: use optimistic locking (UPDATE WHERE value = currentValue) with
re-read fallback on conflict. Prevents concurrent wrong guesses from
each counting as a single attempt.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: check attempt count before OTP comparison to prevent bypass

Reject OTPs that have already reached max failed attempts before
comparing the code, closing a race window where a correct guess
could bypass brute-force protection.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: validate OIDC discovered endpoints against SSRF

The discovery URL itself was SSRF-validated, but endpoint URLs returned
in the discovery document (tokenEndpoint, userInfoEndpoint, jwksEndpoint)
were stored without validation. A malicious OIDC issuer on a public IP
could return internal network URLs in the discovery response.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: remove duplicate OIDC endpoint SSRF validation block

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: validate OIDC discovered endpoints and pin DNS for 1Password Connect

- SSRF-validate all endpoint URLs returned by OIDC discovery documents
  before storing them (authorization, token, userinfo, jwks endpoints)
- Pin DNS resolution in 1Password Connect requests using
  secureFetchWithPinnedIP to prevent TOCTOU DNS rebinding attacks

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* lint

* fix: replace KEEPTTL with TTL+EX for Redis <6.0 compat, add DB retry loop

- Lua script now reads TTL and uses SET...EX instead of KEEPTTL
- DB optimistic locking now retries up to 3 times on conflict

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: address review feedback on OTP atomicity and 1Password fetch

- Replace Redis KEEPTTL with TTL+SET EX for Redis <6.0 compatibility
- Add retry loop to DB optimistic lock path so concurrent OTP attempts
  are actually counted instead of silently dropped
- Remove unreachable fallback fetch in 1Password Connect; make
  validateConnectServerUrl return non-nullable string

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: treat Lua nil return as locked when OTP key is missing

When the Redis key is deleted/expired between getOTP and
incrementOTPAttempts, the Lua script returns nil. Handle this
as 'locked' instead of silently treating it as 'incremented'.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: handle Lua nil as locked OTP and add SSRF check to MCP env resolution

- Treat Redis Lua nil return (expired/deleted key) as 'locked' instead
  of silently treating it as a successful increment
- Add validateMcpServerSsrf to MCP service resolveConfigEnvVars so
  env-var URLs are SSRF-validated after resolution at execution time

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: narrow resolvedIP type guard instead of non-null assertion

Replace urlValidation.resolvedIP! with proper type narrowing by adding
!urlValidation.resolvedIP to the guard clause, so TypeScript can infer
the string type without a fragile assertion.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: bind auth tokens to deployment password for immediate revocation

Include a SHA-256 hash of the encrypted password in the HMAC-signed
token payload. Changing the deployment password now immediately
invalidates all existing auth cookies, restoring the pre-HMAC behavior.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: bind auth tokens to deployment password and remove resolvedIP non-null assertion

- Include SHA-256 hash of encryptedPassword in HMAC token payload so
  changing a deployment's password immediately invalidates all sessions
- Pass encryptedPassword through setChatAuthCookie/setFormAuthCookie
  and validateAuthToken at all call sites
- Replace non-null assertion on resolvedIP with proper narrowing guard

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: update test assertions for new encryptedPassword parameter

Tests now expect the encryptedPassword arg passed to validateAuthToken
and setDeploymentAuthCookie after the password-binding change.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: format long lines in chat/form test assertions

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix: pass encryptedPassword through OTP route cookie generation

Select chat.password in PUT handler DB query and pass it to
setChatAuthCookie so OTP-issued tokens include the correct
password slot for subsequent validation.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

---------

Co-authored-by: Claude Opus 4.6 <[email protected]>

* fix(light): tag dropdown, code highlight

* remove more hardcoded vals

…l connectors (#3793)

* fix(knowledge): enqueue connector docs per-batch to survive sync timeouts

* fix(connectors): convert all connectors to contentDeferred pattern and fix validation issues

All 10 connectors now use contentDeferred: true in listDocuments, returning
lightweight metadata stubs instead of downloading content during listing.
Content is fetched lazily via getDocument only for new/changed documents,
preventing Trigger.dev task timeouts on large syncs.

Connector-specific fixes from validation audit:
- Google Drive: metadata-based contentHash, orderBy for deterministic pagination,
  precise maxFiles, byte-length size check with truncation warning
- OneDrive: metadata-based contentHash, orderBy for deterministic pagination
- SharePoint: metadata-based contentHash, byte-length size check
- Dropbox: metadata-based contentHash using content_hash field
- Notion: code/equation block extraction, empty page fallback to title,
  reduced CHILD_PAGE_CONCURRENCY to 5, syncContext parameter
- Confluence: syncContext caching for cloudId, reduced label concurrency to 5
- Gmail: use joinTagArray for label tags
- Obsidian: syncRunId-based stub hash for forced re-fetch, mtime-based hash
  in getDocument, .trim() on vaultUrl, lightweight validateConfig
- Evernote: retryOptions threaded through apiFindNotesMetadata and apiGetNote
- GitHub: added contentDeferred: false to getDocument, syncContext parameter

Infrastructure:
- sync-engine: added syncRunId to syncContext for Obsidian change detection
- confluence/utils: replaced raw fetch with fetchWithRetry, added retryOptions
- oauth: added supportsRefreshTokenRotation: false for Dropbox
- Updated add-connector and validate-connector skills with contentDeferred docs

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix(connectors): address PR review comments - metadata merge, retryOptions, UTF-8 safety

- Sync engine: merge metadata from getDocument during deferred hydration,
  so Gmail/Obsidian/Confluence tags and metadata survive the stub→full transition
- Evernote: pass retryOptions {retries:3, backoff:500} from listDocuments and
  getDocument callers into apiFindNotesMetadata and apiGetNote
- Google Drive + SharePoint: safe UTF-8 truncation that walks back to the last
  complete character boundary instead of splitting multi-byte chars

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix(evernote): use correct RetryOptions property names

maxRetries/initialDelayMs instead of retries/backoff to match the
RetryOptions interface from lib/knowledge/documents/utils.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix(sync-engine): merge title from getDocument and skip unchanged docs after hydration

- Merge title from getDocument during deferred hydration so Gmail
  documents get the email Subject header instead of the snippet text
- After hydration, compare the hydrated contentHash against the stored
  DB hash — if they match, skip the update. This prevents Obsidian
  (and any connector with a force-refresh stub hash) from re-uploading
  and re-processing unchanged documents every sync

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix(sync-engine): dedup externalIds, enable deletion reconciliation, merge sourceUrl

Three sync engine gaps identified during audit:

1. Duplicate externalId guard: if a connector returns the same externalId
   across pages (pagination overlap), skip the second occurrence to prevent
   unique constraint violations on add and double-uploads on update.

2. Deletion reconciliation: previously required explicit fullSync or
   syncMode='full', meaning docs deleted from the source accumulated in
   the KB forever. Now runs on all non-incremental syncs (which return
   ALL docs). Includes a safety threshold: if >50% of existing docs
   (and >5 docs) would be deleted, skip and warn — protects against
   partial listing failures. Explicit fullSync bypasses the threshold.

3. sourceUrl merge: hydration now picks up sourceUrl from getDocument,
   falling back to the stub's sourceUrl if getDocument doesn't set one.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* lint

* fix(connectors): confluence version metadata fallback and google drive maxFiles guard

- Confluence: use `version?.number` directly (undefined) in metadata instead
  of `?? ''` (empty string) to prevent Number('') = 0 passing NaN check in
  mapTags. Hash still uses `?? ''` for string interpolation.
- Google Drive: add early return when previouslyFetched >= maxFiles to prevent
  effectivePageSize <= 0 which violates the API's pageSize requirement (1-1000).

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix(connectors): blogpost labels and capped listing deletion reconciliation

- Confluence: fetchLabelsForPages now tries both /pages/{id}/labels and
  /blogposts/{id}/labels, preventing label loss when getDocument hydrates
  blogpost content (previously returned empty labels on 404).
- Sync engine: skip deletion reconciliation when listing was capped
  (maxFiles/maxThreads). Connectors signal this via syncContext.listingCapped.
  Prevents incorrect deletion of docs beyond the cap that still exist in source.
  fullSync override still forces deletion for explicit cleanup.
- Google Drive & Gmail: set syncContext.listingCapped = true when cap is hit.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix(connectors): set syncContext.listingCapped in all connectors with caps

OneDrive, Dropbox, SharePoint, Confluence (v2 + CQL), and Notion (3 listing
functions) now set syncContext.listingCapped = true when their respective
maxFiles/maxPages limit is hit. Without this, the sync engine's deletion
reconciliation would run against an incomplete listing and incorrectly
hard-delete documents that exist in the source but fell outside the cap window.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

* fix(evernote): thread retryOptions through apiListTags and apiListNotebooks

All calls to apiListTags and apiListNotebooks in both listDocuments and
getDocument now pass retryOptions for consistent retry protection across
all Thrift RPC calls.

Co-Authored-By: Claude Opus 4.6 <[email protected]>

---------

Co-authored-by: Claude Opus 4.6 <[email protected]>

…earch (#3800)

* feat(search): add tables, files, knowledge bases, and jobs to cmd-k search

* fix(search): address PR feedback — drop files/jobs, add onSelect to memo

* fix(search): add files back with per-file deep links, keep jobs out

* fix(search): remove onSelect from memo comparator to match existing pattern

* Change modal field to be company size

* Adjust dropdown options

* Fix lint

---------

Co-authored-by: Theodore Li <[email protected]>

* improvement(terminal): prevent canvas crashes

* checkpoint

* make persistence execution scoped not debounce

* revert feature flags

* address bugbot comments

* fix test

* fix

* fix type

* fix abortion of manual run

* fix type errors

* fix diff comment

* fix chat history query

* fix comment

* Ignore previously executed tool call ids

* fix eval input color

* fix copilot run workflow

* fix tests

---------

Co-authored-by: Theodore Li <[email protected]>