Sergey Shunko
fc09823f80
GET /{platform}/live/status?live_ids=<a,b,...> now answers on Douyin as well as TikTok.
live.douyin.com answers /webcast/room/check_alive/ with a 404, which is why ab25d6e left douyin.live_status unwired. Like reflow/info, the endpoint lives on the webcast host instead: webcast.amemv.com/webcast/room/check_alive/ answered a guest with HTTP 200 and status_code 0 for up to 200 rooms in one call (measured 2026-10-26).
Two differences from TikTok's check_alive answer:
- Each entry in data carries room_id as a 64-bit JSON integer and alive as a boolean, with no room_id_str sibling; Python's JSON reader keeps every digit, and entry.id('room_id') reads it as the string live_id.
- A broadcast that has ended is alive: false in the order asked, while an id that was never a broadcast is omitted from data rather than returned as false (an ask of only nonexistent ids answers data: []).
Replay fixtures
Input for the L2 replay tests in tests/replay/ (see docs/design/13-testing.md).
Each file is one decoded JSON response body, exactly as a platform endpoint returns it,
fed straight into a parser.
These are hand-built. They must be replaced before release.
Every file here was written by hand from V4's parsing and request code
(branch main, commit 8c98fb7), not captured from a live platform. They reproduce the response
shapes that code demonstrably worked against, so the parsers are exercised against a
realistic structure - but a hand-built fixture can only contain the fields whoever wrote
it already knew about. It cannot tell you that the platform renamed a key last week, and
it cannot show you the field nobody thought to look for.
Replacing them with real captured responses is the browser task described in
docs/design/16-salvage-and-debug.md - the first of the five browser tasks listed
there, capturing real responses as fixtures. Until that lands, a green replay suite means
"the parsers do what we think the platform does", not "the parsers match the platform".
The capture task, per file:
- Open the relevant page in a fresh browser context, find the call in the network panel
(
/aweme/v1/web/aweme/detail/,/api/item/detail/, and so on). - Save the complete response body - do not trim fields you think are irrelevant. The whole point of a fixture is being a permanent record of what the platform sent.
- Run it through the redaction pass below.
- Replace the file and re-run
uv run pytest tests/replay -q. Every difference the parsers notice is real information: either the fixture was wrong, or the parser is.
Redaction rules
Captured responses contain real people. Before a file lands in this directory:
- No credentials of any shape. No key whose name resembles
cookie,token,msToken,sessionid,sid_tt,passport,ttwid,odin_tt,verifyFp,s_v_web_id,_signature,authorizationorcredential.tests/replay/test_platform_contract.pyasserts this on every fixture, and CI runs it. A baresecretis not on that list on purpose - TikTok'sitemStruct.secretis a post visibility flag, and a check that fires on a legitimate field gets switched off. - No real personal data. Nicknames, handles, signatures, comment text and avatar
paths are replaced with obviously synthetic values ("Synthetic Creator",
synthetic.viewer.a). Ids are replaced with reserved-looking runs of zeros. - No live CDN hosts. Media URLs point at
*.example-cdn.invalid..invalidis reserved by RFC 2606 and can never resolve, so no test can accidentally reach out to the network and no expired signed URL leaks a real account's media.
Keep those properties when you replace a fixture with a captured one: capture the structure, then substitute the values.
What each file is for
Same set on both platforms, so the cross-platform normalization can be compared
directly (docs/design/11-data-contracts.md, acceptance criterion 1). Douyin adds
video_live_photo.json, a shape TikTok has no counterpart for.
| File | Covers |
|---|---|
video_normal.json |
The happy path: a public video with media, stats, music, hashtags and a location |
video_image_album.json |
Douyin image album / TikTok photo mode - kind=image_album, no video stream, duration_ms=None |
video_live_photo.json |
Douyin only: an album whose slides are live photos, each with a clip under images[i].video, around one still. Structure from a capture of 2026-10-20, values synthetic (#792) |
video_deleted.json |
Withdrawn post: metadata survives, media URLs do not. Must parse, not raise |
video_private.json |
Restricted post, same media-free shape but a different flag |
video_long_desc.json |
Multi-line description with emoji and six hashtags - pins the title/description split |
user_profile.json |
Author profile with counters |
user_posts_page1.json |
Paged post list with has_more and a cursor |
comments_with_replies.json |
Top level comments, one pinned, one carrying inline reply previews |
comment_replies_page1.json |
Reply page, last page (has_more false, so no cursor) |
risk_control_empty.json |
Structurally valid envelope with the payload emptied out - must raise UPSTREAM_RISK_CONTROL, never UPSTREAM_CHANGED |
risk_control_captcha.json |
Captcha / verification decision response - same requirement |
The two risk-control files are the ones most worth replacing with real captures first.
The parsers' risk signatures are inferred from V4's behaviour, and a signature that is
too permissive cools down healthy identities every time somebody looks up a deleted
video - the failure mode docs/design/02-identity-pool.md warns about.
Adding a fixture
Per docs/design/13-testing.md: when an endpoint breaks in production, the first step is
to save that response here and write a failing replay test. Then fix the parser. Every
incident leaves a permanent regression guard behind instead of being forgotten.