The playbook listed search_queries as a video source in three places, then contradicted itself. Replace with action-oriented guidance: videos via hashtags/URLs, accounts via user_search. Drop 'login-walled' (the agent has no login path).
The profile feed is withheld from flagged IPs and the proxy rotates per request, so re-fetching an empty capture (up to TIKTOK_LISTING_MAX_ATTEMPTS) turns a bad first draw into a hit instead of an ErrorItem.
fetch_item_list goes headful only when the flag promises a display; otherwise it stays headless so the browser launch never fails and the empty feed degrades to an ErrorItem. Verified under xvfb-run: real videos returned on a clean proxy IP, graceful degrade on a flagged one.
For every browser-running role (api/worker/all, not migrate) the entrypoint brings up a virtual display and exports DISPLAY, registering Xvfb for graceful shutdown. Off keeps browsers headless.
The profile feed (/api/post/item_list) returns an empty 200 to headless
sessions but serves data headful on the same proxy IP. Run fetch_item_list
headful and dismiss the login modal that blocks mid-scroll.
The run-reader set gained export_run (CSV export of a stored run), but this
manifest assertion still expected only read_run/search_run. Add export_run so
the agent-door test reflects the shipped readers.
Brings the three newer verbs to parity with tiktok.scrape everywhere it was
wired: MCP tools (+ selfcheck manifest), the chat subagent prompt/description,
the playground catalog, the native docs page, and the SEO marketing page. Also
adds live e2e stages for comments, user search, and trending. Docs/FAQ now
state the real contract: profile metadata is reliable while its video list can
be withheld, and keyword video search is walled (use user search for accounts).
The Explore feed (/api/explore/item_list) is a global trending-video feed
served to anonymous sessions, and it returns the same itemStruct shape as the
other listings — so the verb reuses parse_video, the listing flow, the
TikTokVideoItem output, and the per-video billing meter wholesale. Adds a
browser-capture marker + fetch_trending, a synthetic-target orchestrator entry,
and the tiktok.trending capability, surfaced on the chat subagent.
Comments load over a signed /api/comment/list XHR that TikTok serves to
anonymous sessions once the comments panel is opened (unlike profile-video and
general-search feeds), so this is a reliable verb. Given video URLs it returns
CommentItems (text, author, likes, reply counts; replies carry repliesToId),
deduped per video, capped, and degraded to an ErrorItem for empty/withheld
videos or a bad_url ErrorItem for non-video inputs.
Generalizes the browser capture over a pluggable interaction step so the
comments flow (open panel, scroll the panel to paginate) reuses the same
warm+capture scaffolding as listing/user-search. Billed per comment on a new
TIKTOK_COMMENT meter (TIKTOK_MICROS_PER_COMMENT, matching the per-comment
market), surfaced on the chat subagent alongside tiktok.scrape/user_search.
Video/general search is login-walled for anonymous sessions, but the Users
tab (/api/search/user) returns public account records without a redirect, so
this exposes the one reliably-unblocked search path. A keyword yields
TikTokProfileItems (name, followers, bio, verification), deduped per query,
capped, and degraded to an ErrorItem when a query is empty/withheld.
Reuses the browser capture (generalized over XHR markers + extractor) and the
shared profile item shape. Billed per account on a new TIKTOK_USER meter
(TIKTOK_MICROS_PER_USER), surfaced on the chat subagent alongside tiktok.scrape.
A profile's account data (name, followers, bio, verification) lives in the
page's rehydration blob and loads over plain HTTP without a signed request,
so emit it first and always. The video listing needs a signed item_list XHR
that TikTok withholds from anonymous sessions, so it stays best-effort and
degrades to an ErrorItem. A blocked profile now yields its metadata instead
of only an ErrorItem.
Profile and search feeds are trust-gated: an anonymous headless session
gets an empty item_list (profile) or no results XHR (search), while
hashtag feeds load. A zero-item listing now yields one honest ErrorItem
(errorCode="no_items") instead of vanishing silently, and ErrorItems are
excluded from billing so a blocked target is surfaced but never charged.
- Added a new WorkspaceApiAccessControl component for managing API key access within workspaces.
- Enhanced the PlaygroundApiKeysPage to display API key creation options and access information.
- Updated PlaygroundIndex and PlaygroundSidebar to include links to the new API keys management section.
- Improved user feedback with alerts and descriptions regarding API key usage and requirements.
- Updated the import of FolderInput to FilePlus for improved icon representation in the sidebar.
- Enhanced the button for document import to align with the new iconography, improving visual consistency.
- Adjusted text formatting in the API reference for better readability.
- Simplified import statements in PlaygroundRunner for cleaner code.
- Enhanced button styling and layout in PlaygroundRunner for improved user experience.
- Streamlined output display logic in PlaygroundRunner for better clarity.
