diff --git a/AGENTS.md b/AGENTS.md
index 39e5c65..6ec2596 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -78,6 +78,24 @@ uv run pytest app/tests/
 
 All of these run in CI (`.github/workflows/main.yml`) on every push to master and must pass.
 
+Gotchas:
+
+- Backend tests must run **from the repo root**: `main.py` resolves the static-assets
+  path relative to the cwd, and several test modules import `main`. Running from
+  `app/` makes five test modules fail to import.
+- The frontend must be **built before** running backend tests (same reason — the
+  assets at `ui/dist/metube/browser` must exist). The command order above is
+  load-bearing.
+- `app/tests/test_ytdl_utils.py` stubs `yt_dlp` at import time. Run standalone,
+  two tests fail with `AttributeError: <module 'yt_dlp'> does not have the
+  attribute 'YoutubeDL'`; under the full suite the real module is imported first
+  and they pass. This is a known quirk, not a bug to fix in the code under test.
+
+Every non-markdown push to master builds multi-arch Docker images and cuts a dated
+release the same day. **Master is continuously released** — a PR must be
+release-ready exactly as merged; there is no stabilization window for follow-up
+fixes.
+
 ## Code style
 
 Follow `.editorconfig`:
@@ -104,5 +122,36 @@ ui/src/app/          — Angular standalone components (no NgModules)
 - Backend configuration lives in the `Config` class in `app/main.py` with env-var defaults in `_DEFAULTS`. New env vars go there.
 - Real-time communication uses Socket.IO events, not REST polling.
 - Frontend uses standalone Angular components with `inject()` for DI, RxJS Subjects for state, and `takeUntilDestroyed()` for cleanup.
+- Frontend components use OnPush change detection: subscribe callbacks must call `cdr.markForCheck()`.
 - State is persisted as JSON files via `AtomicJsonStore` in `app/state_store.py`.
+- Persisted state stays compact: the completed queue deliberately drops bulky entry data (see `_compact_persisted_entry` in `app/ytdl.py`). Don't expand what gets persisted without discussion.
+- Custom yt-dlp postprocessors added to `ytdl_params['postprocessors']` run in **list order** within a stage. When combining postprocessors, mirror the ordering the yt-dlp CLI would produce (e.g. sponsor-segment removal before chapter splitting).
 - No pre-commit hooks — linting and tests are enforced in CI only.
+
+## Checklist: adding a per-download option
+
+New options on the download form (the `split_by_chapters` pattern) need **all** of
+these pieces — the last three are the ones commonly missed:
+
+1. `parse_download_options` in `app/main.py`.
+2. A field on `DownloadInfo` in `app/ytdl.py`.
+3. A `hasattr` backfill in `DownloadInfo.__setstate__` for old persisted records.
+4. The safe-deserialization field list in `app/ytdl.py`.
+5. UI form control + cookie persistence in `ui/src/app/app.ts` / `app.html`, and
+   the payload in `downloads.service.ts` (plus its spec).
+6. The redownload path in `app.ts`, so retries carry the option.
+7. If the option makes sense for unattended downloads: threading through
+   `app/subscriptions.py` (`SubscriptionInfo` field, serializer, add/update
+   routes, the enqueue call) — or a note in the PR that it's deliberately
+   direct-downloads-only.
+
+## Security invariants
+
+User input and extractor-provided metadata (titles, playlist names, URLs) are
+untrusted. Use the existing guards instead of hand-rolling:
+
+- User-submitted URLs go through the SSRF guard (see `test_url_guard.py` for the
+  expected behavior).
+- Anything that becomes a filesystem path goes through `_is_within_directory` and
+  `_sanitize_path_component` in `app/ytdl.py` — including values that arrive via
+  yt-dlp metadata, which sites can influence.