Roadmap

Ordered by scope, smallest first.

Small

A config flag, one function, a CI tweak, or docs.

• Rewrite git history to remove leaked identifiers: the working tree was scrubbed, but older commits still carry real usernames, captions and places (the list is in PROFILES.md’s leak scan). git filter-repo --replace-text with a replacements file, then a force-push and a fresh clone everywhere. The repository owner’s call; not done yet. The maintainer has an offline runbook for it, kept outside the repository.
• Report a blocked upstream as drift: for a row it cannot read, check_tool_versions.sh prints (could not read upstream releases) and counts no problem, while tool-versions.yml blocks egress to a fixed host list, so a pin whose upstream lives on a host missing from that list reads as fine forever. Count a problem when every row one host serves went unread (a blocked host, not a one-off rate limit), naming the host in the tracking issue.
• Mask a story’s header: a story is saved whole, so its top carries the viewer’s username and age text over reel_viewer_top_shadow. Paint over the header’s own nodes (avatar, username, timestamp), not the whole gradient, or hold a long press while the screenshot is taken, which hides the overlay but is one more gesture inside the viewer.

Done: Markdown lint and format checks, a link check (relative links on pull requests, external links after merge and weekly), spell check (typos), container image scanning (Trivy: advisory in CI, blocking at release), dependency review on pull requests, shell formatting (shfmt), import boundaries (import-linter), hashed dependency locks (pip-compile), a no_media_node debug dump on the “media node not found” path, whole story frames, story frames matched across accounts, a retake of a dark story frame, the twin-post merge migration and /opml?aggregate=0. Earlier: credentials from a file, caption hashtag/mention links, optional feed auth, the compatibility table (COMPATIBILITY.md) and the committed OpenAPI spec.

Medium

A feature across several parts of the scraper, compose or CI, or repeated real-device work.

• Confirm the double-capture fix on a device: a card used to be captured a second time once capture.expand_caption() had expanded its caption, since the collapsed and expanded captions hash differently. A post is now stored under both hashes, a card with no date is matched on its caption, and a Copy link that repeats the last shortcode is checked against the stored row; the test suite covers each, no device run has. A clean run shows no link_clipboard_failures on a feed with nothing new. Also unconfirmed: whether the dumpsys clipboard fallback fixes Copy link leaving the clipboard empty (6 of 8 attempts in the 445 baseline, run log).
• Stories that reshare a stored post: a feed post shared to a story, by its own account or another followed one, repeats an entry the posts feed already has. Find the reshared-post sticker’s node in the story viewer (no dump of one is recorded yet), hash that rectangle, and compare it with the covers and slides of recently stored posts; then drop the story or store it as a link to the post, which is still to decide. Matching the whole frame against stored images also works without the node, but needs template matching, which the app has no dependency for.
• Retry Instagram 446: 446.0.0.49.77 is in v424.validated but has crashed on launch since 2026-09-15 (a native SIGSEGV in RenderThread; run log), so igprofiles.DEFAULT_BUILD is pinned to 445. Retry it; if it still crashes, drop it from v424.validated and update its COMPATIBILITY.md row. If it works, promote its fixtures too: none have been recorded for 446.
• Baseline what hasn’t been checked yet: no capture-mode baseline has covered the own profile, the Following list (--following) or the login form, so their required selector keys are unchecked on every build. And 425-439 run with the “hasn’t been validated” warning until each gets a new-profile baseline/validate.
• Instagram update path: install-on-missing is automatic (see README.md’s “First-time setup”), but an outdated install isn’t handled yet — detect the forced “update Instagram” screen (as a challenge-style stop) and reuse install.install_instagram() (app/instadroid/install.py) with a newer validated build (new-profile baseline/validate, or fork if it drifted), plus a scraper.py dump smoke check, keeping the previous xapk in APK_CACHE_DIR for rollback.
• Resource-id check for new Instagram builds in CI: .github/workflows/new-builds.yml already opens an issue weekly when APKPure lists a major version newer than every validated build (check-new-builds, app/devtools/check_new_builds.py). Still to add: a static resource-id report in that issue. aapt2 dump resources on the new build’s base APK (about a second) lists every selector resource id missing from it, and the ids added or removed since the newest validated build.
  • A research pass on 2026-09-14 ran this on the cached 443-446 builds. All 18 Instagram resource ids the selectors use were present in every one, matching the 445 selectors working unchanged on 446.
  • 12 of the 17 required keys in igprofiles/screens.py are resource ids. A removed id is near-certain drift, but text and content-desc keys can’t be checked this way (those strings aren’t in the APK), and layout changes don’t show up either.
  • It means downloading the ~140MB xapk in CI; confirm APKPure serves GitHub’s IP addresses. No account, session, dump or APK may end up in a cache or artifact: in a public repo, anyone can read both.

Large

Open investigations, new capture mechanisms, or changes to the container/process topology.

• Reach the real Following feed without the switcher: under gpu_mode=guest the switcher’s bottom sheet may not open, and the scraper then falls back to Home — algorithmic, with suggested posts mixed in. Investigate a deep link or activity intent that opens Following directly; unconfirmed whether one exists. (A followed-accounts allowlist now filters the fallback’s suggested posts after the fact — see README.md’s “Followed-accounts allowlist” — but reaching the real feed directly would still be cheaper than the extra Following-list navigation that costs.)
• Detect username changes automatically: today a rename has to be noticed and reconciled by hand (scraper.py rename <old> <new>). Instagram’s numeric user id never appears in the feed’s accessibility tree, so detecting a rename would mean visiting each account’s profile — extra in-app navigation and detection surface per run, which is why it wasn’t done automatically here.
• Full story-reel capture: only a story’s current frame is captured (see README.md’s “Stories”) — a deliberate scope decision, not a gap left for later, given that tapping to advance a story has been observed to eject the app to the OS launcher on this host once its queue is exhausted. Worth revisiting only with a materially different navigation approach (e.g. reading the tray’s own total count to know exactly how many frames to expect, so the loop never has to discover exhaustion by tapping past the end).
• Real video capture: still a poster-frame still — Reels/videos never get the actual video. Likely needs screen recording rather than a screenshot, plus somewhere to store and serve a video file per post, and meaningfully longer dwell time per video post (see README.md’s “Staying under the radar”) — a real cost/benefit call, not just effort.
• Replay whole navigation sequences: the scraper is now split into app/instadroid/ modules, and single recorded screens replay through the parsers (promote-dump scrubs a DEBUG_DIR dump into igprofiles/vXYZ/fixtures/, tests/test_replay.py checks it). The device flows still run against hand-written tests/fakedevice.py screens; recording a real run’s sequence of dumps and taps, and replaying it through fakedevice, would catch navigation drift (a moved tab, a new interstitial) the same way.
• arm64 host support: on an arm64 host, official redroid/redroid images run Instagram’s arm64 code natively — no NDK translation, sidestepping the whole compatibility matrix (COMPATIBILITY.md). Needs a multi-arch app image (platforms: in publish.yml) and host docs (binder in the kernel).