antisocial/CLAUDE.md

antisocial — working notes

A self-hosted page that shows a social post without the app. StopTheMadness rewrites links to X, Threads, Instagram, TikTok and Bluesky into /<prefix>/<original path>; this resolves the post by driving a real headless Chromium and renders the media, the text, a platform badge and a copy-the-original button.

Single user, no auth, tailnet only, running on Kallone (lab3). README.md has the rewrite rules and the user-facing description. This file is the working context.

The main job: "this link didn't work"

That is the recurring task. Work it in this order and do not skip step 1.

1. Reproduce against the real browser.

npm run resolve -- 'https://www.instagram.com/user/reel/ABC123/'

Prints the resolved Post as JSON, or the error. This is the ground truth.

2. Decide what kind of failure it is. These are genuinely different and get different fixes:

Symptom	What it usually is
Works now, failed before	Rate limiting. TikTok especially. Not a bug — verify by waiting and retrying before changing anything.
Verification puzzle	Expected. Should redirect to /challenge/<id>. If it reports an error instead, the detection raced the render.
Right media, wrong count	The structured payload was missed and it fell through to the DOM, which only shows the first carousel item.
Poster image where a video belongs	Same fall-through, plus the <video> had not hydrated when the DOM was read.
Nothing at all	The platform refused this specific post logged out, or a key moved.

3. Look at what the page actually served.

npm run probe -- 'https://www.instagram.com/reel/ABC123/embed/captioned/' 5000

Pass the URL the adapter loads, not the post URL — for Instagram that is the /embed/captioned/ form. It reports which payload keys are present, the DOM's <video>/<img>, the OG tags, whether a challenge is up, and the API traffic with bundle noise filtered out. Every adapter fix so far started here.

4. Fix the adapter, then capture a fixture for the case that broke, and assert on it in test/<platform>.test.ts. Fixtures are real captured payloads — see the ones in test/fixtures/. A fix without a fixture will silently regress.

5. Re-run the whole suite (npm test), and re-resolve the link end to end.

Architecture, briefly

Request → src/routes/post.ts → src/platforms/index.ts maps prefix to adapter → src/resolve.ts (cache + timeout + concurrency gate) → withPage from src/browser/pool.ts → the adapter → a Post → src/render/post.ts.

Adding a platform is one file in src/platforms/ plus one row in the table in index.ts. Everything downstream already handles a Post.

Things worth knowing before editing:

• One Chromium, one context, persistent. Cookies and dismissed banners accumulate on purpose — that is what makes traffic look like ordinary browsing. MAX_CONCURRENT page loads, the rest queue.
• Images, video and fonts are aborted at the route layer. The URL is all we need; the bytes get fetched later through our own proxy. Aborting does not remove src attributes, so DOM extraction still works.
• Extraction is layered, most structured first: the platform's own API response caught in flight → an inline payload → the rendered DOM → OG tags. Keep that order when you touch an adapter; each layer is the fallback for the one above.
• src/platforms/scan.ts pulls a balanced JSON object out of a page by key, including when it arrives escaped inside a JS string. Instagram needs this. If a payload looks present but parses to nothing, suspect an extra encoding level.
• Media is never linked straight at a CDN. Assets are registered in src/media/registry.ts and served from /m/<id> with the Referer/Cookie the CDN demands. Range is forwarded — without it the native video scrubber cannot seek. The exception is HLS (direct: true), because proxying would mean rewriting playlists.
• Never a bare error page. A failed resolve renders a card carrying the platform, the original URL and the copy button. A broken adapter must still leave the link one tap away.

Per-platform notes

• Bluesky — asks the public API directly (still through the browser context), so it is the most reliable. The web app calls getPostThreadV2 now; the page fallback deliberately matches only V1.
• X — the platform.twitter.com embed calls the syndication endpoint; we catch that response. A quote post carries no media of its own, so the quoted post's media is used.
• Instagram — the least reliable. It ships the structured payload only some of the time, and refuses some posts from the embed entirely ("the link may be broken"). The DOM is the floor and only ever shows the first carousel item. Reels hydrate their <video> roughly 0.6–1.3s after domcontentloaded, so the DOM read waits when a video is expected.
• TikTok — the post lives under whichever __UNIVERSAL_DATA_FOR_REHYDRATION__ scope carries an itemStruct; do not hardcode the key, photo posts use a different one. Its CDN needs Referer and cookies. Short vm./vt. links arrive as a bare path segment and are rebuilt in buildOriginalUrl.
• Threads — same media schema as Instagram (src/platforms/meta-media.ts). Its payloads are full of empty stub nodes, so the finder only accepts a node with actual candidates in it.

Verification puzzles

TikTok challenges sometimes. The page is parked (src/challenge/registry.ts), kept open but outside the pool holding no concurrency permit, and the viewer is redirected to /challenge/<id>, which streams screenshots of the puzzle and replays their pointer events onto it. Solving it deposits the cookie in the shared context, so the retry is an ordinary request.

Login walls are not passed through — those need an account and cannot be solved this way.

Conventions

• TypeScript 7, ESM, .ts import specifiers (tsc rewrites them to .js on emit). Node's type stripping runs the tree directly for dev, tests and the CLIs.
• No ESLint — typescript-eslint does not support TS 7 yet. npm run typecheck is the lint step.
• Tests are node:test against captured fixtures. No network in the test suite.
• Comments explain why, especially where the code looks odd because a platform is odd. Match that; do not add narration of what the next line does.
• Post text comes from strangers: everything goes through the html tagged template or linkify, both of which escape. Do not build markup by concatenation.

Commands

npm run resolve -- '<url>'   # resolve one post, print JSON       <- start here
npm run probe   -- '<url>'   # what the page actually served      <- then here
npm test                     # fixtures, no network
npm run typecheck
npm run dev                  # http://localhost:8080
container build --tag antisocial:dev .   # Apple container, not Docker

Deployment

CI (.gitea/workflows/publish.yml) builds on main and pushes git.unsupervised.ca/unsupervised/antisocial:latest and :sha-<short>. Manifests live in the infra repo at k8s/kallone/antisocial.unsupervised.studio/; Flux syncs k8s/kallone every 10 minutes. The DNS CNAME is in infra/cloudflare/unsupervised.studio.tf.

Chromium needs the memory-backed /dev/shm volume in the deployment — it crashes against the 64Mi default. Measured 623Mi resident.