feat(drugs_com): add drugs.com mirror site (port 40016)

[boyugou]

TL;DR

Adds a fully functional drugs.com mirror site to WebHarbor at port 40016, with 21 benchmark tasks covering drug lookup, interaction checking, pill identification, condition browsing, news, and authenticated account actions. Real (bcrypt-checked) authentication and CSRF protection throughout; search results and browse pages are click-through-only (no answer visible before opening the relevant detail page).

Motivation

drugs.com is the most-visited drug information portal (24,000+ drugs). It covers pharmaceutical/medical information — a domain not represented in the existing 16 sites — making it a high-signal addition for web-agent benchmarks.

Design

Flask application (sites/drugs_com/app.py)

Ten SQLAlchemy models: User, DrugClass, Drug, DrugImage, DrugInteraction, DrugReview, Condition, DrugCondition, NewsArticle, SavedDrug. All seeded idempotently — each seed_*() function gates on an already-populated table so seed_database() is safe to call repeatedly without breaking the byte-identical reset invariant.

Route coverage mirrors the real site's navigation:

Route	Description
/<slug>.html	Drug detail with tabbed sections
/drug_information.html	Drugs A-Z index
/drug-interactions	Drug interaction checker
/pill-identifier	Pill identifier by imprint/shape/color
/drug-classes/, /drug-classes/<slug>/	Drug class browser
/conditions/, /condition/<slug>/	Conditions A-Z
/news/	News with category filter
/search	Drug/class/condition search
/my-med-list	Saved drugs (auth required)

app.url_map.strict_slashes = False is set globally so all routes accept trailing-slash variants. The last @app.route decorator on each view is the canonical URL used by url_for(); alias routes are placed before it.

Auth and CSRF

/login and /register perform real bcrypt-checked authentication, matching the pattern used by every other site in the repo. CSRFProtect is enabled repo-wide-consistently: every <form method="post"> carries a token; the one route that's genuinely JSON-only and unreferenced by any template (/api/interaction-check) is deliberately exempt, while the two routes reachable from both a real form and JavaScript carry their token through both paths (an X-CSRFToken header for the JSON call, an appended FormData field for the other) rather than being exempted.

Answer-leak prevention

Search results (.drug-card / "Best match" cards) show only the drug name and a link — no drug class, brand names, CSA schedule, rating, or description snippet, all of which require opening the detail page. The homepage, the /drug-classes index, and /search's condition/class-match sidebar do not pre-list category members by name, so "name at least N drugs in category X"-style tasks require actually opening that category's page rather than reading a summary widget.

Seed content integrity

Drug detail text (uses/warnings/dosage/side_effects) is sourced from the FDA's public drug-label API at seed time where available, falling back to a generated description otherwise. Because that API does fuzzy/tokenized matching rather than exact matching, a fetched label is verified to actually describe the queried drug (by generic-name match) before being accepted, rather than trusting the top search hit — this prevents one drug's content from silently containing a different, unrelated drug's real label text.

Seed database

• 246 drugs across diverse Rx and OTC drug classes, 104 drug classes
• 716 reviews, 76 drug interactions, 103 pill images, 69 conditions, 80 news articles, 12 users
• Seed DB MD5: d3d228f3fc0b3b880e149ac35550e3c5
• Benchmark users: alice.j@test.com / TestPass123! (primary), plus bob_c, carol_d, david_k
• Alice's med list is pre-seeded with ibuprofen, metformin, atorvastatin (lisinopril is not in it)

Pill images

Rendered as inline SVGs via the _pill_svg.html macro (no external files required for the benchmark). Real pill photos at static/images/pills/<slug>.jpg are HuggingFace assets; a pill_image_exists Jinja filter falls back to the SVG macro when absent.

Benchmark tasks (tasks.jsonl)

21 tasks (Drugs.com--0 through Drugs.com--20) covering:

• Drug detail lookup (drug class, brand names, availability, CSA schedule, pregnancy risk)
• Drug interaction checker — 2-drug and 3-drug, severity classification
• Pill identifier by imprint code, shape, and color
• Drugs A-Z browsing
• Drug class navigation (Statins, Benzodiazepines, Fluoroquinolones)
• Condition browsing (diabetes, hypertension)
• News reading and category filtering
• User reviews and ratings
• Authenticated action: sign in, then read the My Med List page (task 14; genuinely gated behind real login now, not a no-op)

Templates

40 Jinja2 templates modeled on real drugs.com layout: drug detail with tabbed sections (Uses, Warnings, Before taking, Dosage, Side effects, Interactions, FAQ), drug status sidebar (Pregnancy/Lactation risk, CSA Schedule, Approval History), interaction checker with severity badges, pill identifier, A-Z index with category browse, and news with category filter tabs.

Registration

Registered in all three required locations:

File	Change
websyn_start.sh	Added drugs_com to SITES=(...) — index 16, port 40016 (after merriam_webster, index 15)
control_server.py	Added drugs_com to SITES list
Dockerfile	EXPOSE 8101 40000-40016 covers port 40016

HuggingFace assets

The seed DB and pill images are packaged as drugs_com.tar.gz (5 MB, clean — no macOS AppleDouble sidecars; DB md5 d3d228f3fc0b3b880e149ac35550e3c5) and submitted to the ChilleD/WebHarbor HuggingFace dataset as discussion #39 (https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/39). .assets-revision is pinned to that PR's ref (refs/pr/39) in the interim, since it's a strict superset of main (all 16 existing site tarballs unchanged, plus this one) — safe for every other site to build against right now. After #39 merges, .assets-revision should be re-pinned to main or the merge commit. Fetch locally with:

./scripts/fetch_assets.sh drugs_com

Review history

This PR went through a human review round followed by several rounds of independent automated re-review (deliberately varying which tool/model reviewed each round, to avoid one reviewer's blind spots repeating across rounds). Every finding was verified against the actual code before being acted on — see the comment thread below for the full reasoning on what was fixed and, in a few cases, why a real-but-out-of-scope finding was deliberately left alone (e.g. patterns shared by all 17 sites in the repo, or issues not reachable by a browser-driving agent). Net result: the auth/CSRF/answer-leak/seed-content-integrity design described above is the product of that process, not the original submission.

Non-goals

• Real FDA API calls at HTTP-request time (the label-search call happens only at build/seed time, never from a runtime route handler)
• Image assets in git (heavy assets are in HuggingFace as drugs_com.tar.gz)

Verification

Check	Result
py_compile sites/drugs_com/app.py	✅
./scripts/build.sh webharbor:dev	✅
GET /health → drugs_com alive	✅
GET http://localhost:40016/ → 200 (all 17 sites → 200)	✅
POST /reset/drugs_com + md5sum match	✅ d3d228f3fc0b3b880e149ac35550e3c5
All 21 task routes manually verified against the current seed data	✅
Real login + CSRF flow (wrong password rejected, correct password + token succeeds, unauthenticated access to gated routes redirects to login) verified against a running container with real cookies	✅
Full-table sweep for cross-drug duplicate content (the seed-content-integrity fix)	✅ none found

[YuanDaoze]

Review

Tested on a fresh worktree of this branch. HF asset (drugs_com.tar.gz, 5 MB) pulled directly from the dataset PR ref (refs/pr/13), the other 15 sites' assets hard-linked from upstream/main. Built as webharbor:rev9, run on alt ports :8221 / :42200-42215 to avoid colliding with the dev container.

What works ✓

Mechanical

• Three-place registration in sync (websyn_start.sh SITES, control_server.py SITES, Dockerfile EXPOSE 8101 40000-40015).
• All 16 sites alive after boot. drugs_com on :42215 returns 200.
• Byte-identical reset holds: md5(instance/drugs_com.db) == md5(instance_seed/drugs_com.db) == bfc94f8ff61fbbd553ae496217588bad both before and after POST /reset/drugs_com — matches the value claimed in the PR description.
• Seed gating is textbook-correct: a single top-level gate in seed_database() (if Drug.query.count() > 0: return) protects all 11 sub-seeds, with an explicit comment explaining why per-row gates aren't enough. This is the cleanest implementation of the AGENTS.md idempotency rule I've seen.
• Build context clean — no .db files, no static/images/, no scraped_data/ committed. .gitignore and .dockerignore cover runtime data and tarballs properly.
• tasks.jsonl has 21 tasks, all five required fields on every line, port matches 40015.

Functional depth

• 169 routes, 49 templates, 7080-line app.py. Most thorough single-site mirror in the repo.
• /drug-interactions?drugs=ibuprofen,warfarin (GET-style, supports comma-separated multi-drug input) returns categorized results: drug-drug + drug-food (vitamin K, cranberry juice) + drug-alcohol, with severity badges (Major / Moderate / Minor). The 3-drug variant ?drugs=alprazolam,oxycodone,fluoxetine also works.
• /pill-identifier has imprint + shape (round/oval/capsule/oblong/+5 more) + color (12 colors) form fields, real pill JPGs for 43 drugs from HF + inline-SVG fallback for the rest via the _pill_svg.html macro.
• /drug-classes/, /conditions/, /news/ all populated. Drug detail page at /<slug>.html has full sections: black-box warnings, uses, before-taking, dosage, side effects, interactions, FAQ accordion, related drugs.
• Login (alice.j@test.com / TestPass123!) works. /my-med-list shows alice pre-seeded with ibuprofen + metformin + atorvastatin (matches PR description) and her 3 saved drugs are listed with Rx/OTC badges, edit/remove actions, and a "Check all 3 medications for interactions" CTA.
• /search?q=... aggregates across drugs / classes / conditions — 24/57/48 results for ibuprofen / diabetes / antibiotics.

Task quality

• 21 tasks span 8 categories: drug detail lookup, search, interaction checker (2-drug + 3-drug), pill identifier (imprint, shape+color), A-Z browsing, drug class, condition, news, authenticated med list, reviews/ratings.
• Difficulty spans simple lookup (--0 ibuprofen drug class) → multi-step (--14 login + add drug to med list) → multi-result aggregation (--13 first 3 white oval pills).
• Spot-checked 5 risky tasks; all answerable from canonical pages:
  • --3 pill imprint I-2 → /pill-identifier?imprint=I-2 returns Round/Oval results with Ibuprofen
  • --9 Statins class → /drug-classes/statins lists Atorvastatin, Lipitor, Simvastatin, Rosuvastatin, Crestor
  • --15 lisinopril pregnancy → black-box warning text on /lisinopril.html explicitly mentions "FDA Pregnancy Category D"
  • --8 3-drug interaction → returns major/moderate/minor verdicts
  • --14 lisinopril is correctly NOT in alice's pre-seeded list (so the task can verify the add-to-med-list action took effect)
• No count-leak or single-item-catalog leaks found in spot checks.

Should-fix (non-blocking)

1. .assets-revision pinned to main, but drugs_com.tar.gz is in HF PR ref refs/pr/13 and not yet merged.

A fresh git clone + ./scripts/fetch_assets.sh will silently skip drugs_com (the file isn't on main) → drugs_com won't have a seed DB → the site won't boot. This is the same problem as PR #32 (gov_uk).

Fix one of:

• Get the HF PR merged to main and bump .assets-revision to that commit SHA (cleanest);
• Or temporarily pin .assets-revision to a fork/PR ref that includes drugs_com (the merriam_webster PR uses this pattern via repo: YuanDaozeiii/WebHarbor until the upstream HF PR merges).

2. macOS ._* AppleDouble files in drugs_com.tar.gz.

The tarball was created on macOS with default tar, which embedded 51 AppleDouble metadata files (1:1 with real files). They're 4 KB each and tar tzf triggers a stream of LIBARCHIVE.xattr.com.apple.provenance warnings. They don't break anything functionally (extracted as ._foo, ignored by Flask), but they bloat the tarball and pollute static/images/pills/.

Fix: re-create with COPYFILE_DISABLE=1 tar czf drugs_com.tar.gz drugs_com/ and re-upload. Same problem affects gov_uk's tarball, so a one-line README note in the asset workflow doc would help future contributors.

3. PR base is 3c408d8 init, way behind upstream/main.

A rebase is needed before merge. With 187 commits in this branch, an interactive squash to ~5 logical commits would also help reviewability.

4. requirements.txt is unused and mis-pinned.

The file lists deps with no version pins and is never read at build time (the Dockerfile pip-installs explicitly with locked versions). Either delete it (it's noise) or make it the source of truth and have the Dockerfile use it. AGENTS.md says "Lock dependency versions in Dockerfile's pip install — the image must be reproducible", so the current Dockerfile approach is correct; the requirements.txt file is dead code.

5. _pill_svg.html SVG fallback occasionally renders for drugs that DO have a real JPG in the asset bundle.

Spot-checked /ibuprofen.html — page contains 12 inline SVG <svg> elements but static/images/pills/ibuprofen.jpg returns 404 (only 43 drugs have real JPGs out of ~250). For those 43 the page does load the JPG. Working as designed; just noting for completeness.

Bottom line

Genuinely impressive depth — 169 routes, full interaction checker with drug-food/drug-alcohol, working pill identifier, real-data pregnancy categories, accordion FAQ, etc. Seed idempotency design is the cleanest in the project. The only real friction is on the asset-distribution side: .assets-revision needs to point to where drugs_com.tar.gz actually lives, and the macOS tar junk should be cleaned up before re-upload. Once those land, this is solidly mergeable.

[boyugou]

Review

Tested on a fresh worktree of this branch. HF asset (drugs_com.tar.gz, 5 MB) pulled directly from the dataset PR ref (refs/pr/13), the other 15 sites' assets hard-linked from upstream/main. Built as webharbor:rev9, run on alt ports :8221 / :42200-42215 to avoid colliding with the dev container.

What works ✓

Mechanical

• Three-place registration in sync (websyn_start.sh SITES, control_server.py SITES, Dockerfile EXPOSE 8101 40000-40015).
• All 16 sites alive after boot. drugs_com on :42215 returns 200.
• Byte-identical reset holds: md5(instance/drugs_com.db) == md5(instance_seed/drugs_com.db) == bfc94f8ff61fbbd553ae496217588bad both before and after POST /reset/drugs_com — matches the value claimed in the PR description.
• Seed gating is textbook-correct: a single top-level gate in seed_database() (if Drug.query.count() > 0: return) protects all 11 sub-seeds, with an explicit comment explaining why per-row gates aren't enough. This is the cleanest implementation of the AGENTS.md idempotency rule I've seen.
• Build context clean — no .db files, no static/images/, no scraped_data/ committed. .gitignore and .dockerignore cover runtime data and tarballs properly.
• tasks.jsonl has 21 tasks, all five required fields on every line, port matches 40015.

Functional depth

• 169 routes, 49 templates, 7080-line app.py. Most thorough single-site mirror in the repo.
• /drug-interactions?drugs=ibuprofen,warfarin (GET-style, supports comma-separated multi-drug input) returns categorized results: drug-drug + drug-food (vitamin K, cranberry juice) + drug-alcohol, with severity badges (Major / Moderate / Minor). The 3-drug variant ?drugs=alprazolam,oxycodone,fluoxetine also works.
• /pill-identifier has imprint + shape (round/oval/capsule/oblong/+5 more) + color (12 colors) form fields, real pill JPGs for 43 drugs from HF + inline-SVG fallback for the rest via the _pill_svg.html macro.
• /drug-classes/, /conditions/, /news/ all populated. Drug detail page at /<slug>.html has full sections: black-box warnings, uses, before-taking, dosage, side effects, interactions, FAQ accordion, related drugs.
• Login (alice.j@test.com / TestPass123!) works. /my-med-list shows alice pre-seeded with ibuprofen + metformin + atorvastatin (matches PR description) and her 3 saved drugs are listed with Rx/OTC badges, edit/remove actions, and a "Check all 3 medications for interactions" CTA.
• /search?q=... aggregates across drugs / classes / conditions — 24/57/48 results for ibuprofen / diabetes / antibiotics.

Task quality

• 21 tasks span 8 categories: drug detail lookup, search, interaction checker (2-drug + 3-drug), pill identifier (imprint, shape+color), A-Z browsing, drug class, condition, news, authenticated med list, reviews/ratings.

• Difficulty spans simple lookup (--0 ibuprofen drug class) → multi-step (--14 login + add drug to med list) → multi-result aggregation (--13 first 3 white oval pills).

• Spot-checked 5 risky tasks; all answerable from canonical pages:

  • --3 pill imprint I-2 → /pill-identifier?imprint=I-2 returns Round/Oval results with Ibuprofen
  • --9 Statins class → /drug-classes/statins lists Atorvastatin, Lipitor, Simvastatin, Rosuvastatin, Crestor
  • --15 lisinopril pregnancy → black-box warning text on /lisinopril.html explicitly mentions "FDA Pregnancy Category D"
  • --8 3-drug interaction → returns major/moderate/minor verdicts
  • --14 lisinopril is correctly NOT in alice's pre-seeded list (so the task can verify the add-to-med-list action took effect)

• No count-leak or single-item-catalog leaks found in spot checks.

Should-fix (non-blocking)

1. .assets-revision pinned to main, but drugs_com.tar.gz is in HF PR ref refs/pr/13 and not yet merged.

A fresh git clone + ./scripts/fetch_assets.sh will silently skip drugs_com (the file isn't on main) → drugs_com won't have a seed DB → the site won't boot. This is the same problem as PR #32 (gov_uk).

Fix one of:

• Get the HF PR merged to main and bump .assets-revision to that commit SHA (cleanest);
• Or temporarily pin .assets-revision to a fork/PR ref that includes drugs_com (the merriam_webster PR uses this pattern via repo: YuanDaozeiii/WebHarbor until the upstream HF PR merges).

2. macOS ._* AppleDouble files in drugs_com.tar.gz.

The tarball was created on macOS with default tar, which embedded 51 AppleDouble metadata files (1:1 with real files). They're 4 KB each and tar tzf triggers a stream of LIBARCHIVE.xattr.com.apple.provenance warnings. They don't break anything functionally (extracted as ._foo, ignored by Flask), but they bloat the tarball and pollute static/images/pills/.

Fix: re-create with COPYFILE_DISABLE=1 tar czf drugs_com.tar.gz drugs_com/ and re-upload. Same problem affects gov_uk's tarball, so a one-line README note in the asset workflow doc would help future contributors.

3. PR base is 3c408d8 init, way behind upstream/main.

A rebase is needed before merge. With 187 commits in this branch, an interactive squash to ~5 logical commits would also help reviewability.

4. requirements.txt is unused and mis-pinned.

The file lists deps with no version pins and is never read at build time (the Dockerfile pip-installs explicitly with locked versions). Either delete it (it's noise) or make it the source of truth and have the Dockerfile use it. AGENTS.md says "Lock dependency versions in Dockerfile's pip install — the image must be reproducible", so the current Dockerfile approach is correct; the requirements.txt file is dead code.

5. _pill_svg.html SVG fallback occasionally renders for drugs that DO have a real JPG in the asset bundle.

Spot-checked /ibuprofen.html — page contains 12 inline SVG <svg> elements but static/images/pills/ibuprofen.jpg returns 404 (only 43 drugs have real JPGs out of ~250). For those 43 the page does load the JPG. Working as designed; just noting for completeness.

Bottom line

Genuinely impressive depth — 169 routes, full interaction checker with drug-food/drug-alcohol, working pill identifier, real-data pregnancy categories, accordion FAQ, etc. Seed idempotency design is the cleanest in the project. The only real friction is on the asset-distribution side: .assets-revision needs to point to where drugs_com.tar.gz actually lives, and the macOS tar junk should be cleaned up before re-upload. Once those land, this is solidly mergeable.

Thanks for the review! Will fix accordingly soon

[boyugou]

Update — merged upstream/main, addressed review feedback, drugs_com moved to port 40016

Why the port changed. upstream/main merged merriam_webster (#44) into the 16th slot (index 15 → port 40015) — the slot this PR originally used. I merged main into this branch and moved drugs_com to index 16 / port 40016 so the two coexist; merriam_webster keeps its released 40015. Registration is back in sync (websyn_start.sh, control_server.py, Dockerfile EXPOSE 8101 40000-40016), the 21 task URLs now use 40016, and GitHub reports no conflicts with base. The PR title/description have been updated to 40016.

Review feedback addressed (the 3 new commits):

• AppleDouble junk → scripts/extract_assets.sh now packs with COPYFILE_DISABLE=1 and strips ._*/.DS_Store, so no <site>.tar.gz embeds macOS sidecars (the 51 ._* files came from com.apple.provenance xattrs). Regenerated a clean drugs_com.tar.gz — 0 AppleDouble entries, DB md5 unchanged → re-submitted as a HuggingFace dataset PR (https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/39), replacing the earlier junk tarball.
• requirements.txt kept (it's the new_site.py scaffold convention and merriam_webster ships one too), but the genuinely-unused Flask-WTF was pruned.
• .assets-revision pin: will be bumped to the merged HF SHA once that HuggingFace PR lands (currently inherits upstream's pin, which predates drugs_com).

Self-review fix (not in the original review): task --10 asked what the ibuprofen FAQ says about taking it "on an empty stomach", but the generated FAQ has no such entry (that guidance is in the Dosage section) — reworded to target an FAQ entry that exists. Also added .venv to .dockerignore (repo uses uv).

Verification on this branch (full 17-site image):

• All 17 sites → 200; all drugs_com task routes → 200.
• Byte-identical reset: POST /reset/drugs_com → md5(instance) == md5(instance_seed) == bfc94f8ff61fbbd553ae496217588bad.
• Task-14 med list correct (alice: ibuprofen / metformin / atorvastatin, not lisinopril); interaction checker returns Major for ibuprofen + warfarin.

Paired asset PR: ChilleD/WebHarbor HuggingFace dataset — https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/39 (clean tarball). Merge that first, then bump .assets-revision.

Note for maintainers — port coordination: drugs_com now claims 40016. PR #24 (CarMax) still references 40015, which merriam_webster (#44) already occupies on main; whichever of the remaining site PRs merges next will need its own free index.

[MufanQiu]

Review — drugs_com (PR #9)

Verdict: REQUEST CHANGES.

We deployed the environment locally and ran the full review checklist; every finding below was independently reproduced, and fidelity was checked against the real upstream site.

Issues (summary)

• BLOCKER: .assets-revision SHA presence on HuggingFace not verifiable locally (CI fail-closed risk)
• MAJOR: Answer-leak: 5 tasks fully solvable on the /search results page (no detail-page click)
• MINOR: Difficulty too low / illusory auth: most tasks are 1-2 actions; the one auth task is a no-op
• MINOR: AppleDouble '._*' junk (48 files) ships inside the image via the pinned tarball
• MINOR: Data inconsistencies: brand-as-generic duplicate rows and a CSA value in the availability field
• NIT: CLAUDE.md quick-start has an inconsistent port (49015 vs 49016/40016)

Mechanical checks: PASS (with nits)

• Reset invariant holds. instance_seed/drugs_com.db and instance/drugs_com.db are byte-identical (md5 bfc94f8ff61fbbd553ae496217588bad), and a fresh boot leaves the instance md5 unchanged.

Visual fidelity: PASS

Faithful reproduction of the dark-blue header, "Know more. Be sure." tagline, A-Z browse bar, health-category tiles, drug-images strip, featured drugs, and medical-news sections.

Functional depth: PASS (with nits)

Deep, DB-backed behaviour. Search is scored (not strict-AND) with brand/condition/class boosts, an exact-match "Best match" card, a "Did you mean?" fuzzy fallback ("ibuprofin"→ibuprofen), and working multi-word queries.

Task quality: FAIL

21 tasks, all solvable against the DB and reachable in the UI, atop a large catalog with strong distractor density (251 drugs, 95 classes, 68 conditions, 80 news, 76 interactions, 103 pill images, 748 reviews).

Required fixes before approval

1. Eliminate the search-results answer-leak (F1, now ≥10 tasks). In templates/search.html, remove drug class, brand names, CSA-schedule chip, rating/review_count, AND the uses/description blurb from both the .drug-card (lines 169-201) and the "Best match" card (lines 16-52).
2. Fix illusory auth and raise difficulty (F2). Make the login handler actually validate alice.j@test.com / TestPass123! and gate @login_required routes (or remove the auto_login shim) so Task 14's sign-in is genuinely exercised.
3. Repack and re-pin assets (F3 + F5 + instance_seed sidecar). Rerun scripts/extract_assets.sh (strips ._*/.DS_Store, packs COPYFILE_DISABLE=1), upload a clean drugs_com.tar.gz, and update .assets-revision to a MERGED HF SHA that contains it.
4. Clean up data inconsistencies (F4). Set suvorexant.availability to Rx and move C-IV into csa_schedule only; fix the five inverted brand-as-generic duplicate rows so generic_name is the real generic (or drop the duplicates).
5. Fix the CLAUDE.md quick-start port typo (F6). Change the verify curl from 49015 to 49016 to match -p 49016:40016 (docs-only).

[boyugou]

Update — addressed all 6 points from the review above, plus a self-initiated deeper pass

Went through each finding; all confirmed as real issues on inspection and fixed. Then ran three independent Sonnet-based review passes over the resulting diff (task-by-task re-audit, auth security review, DRUGS_DATA data-quality sweep) before calling this done, since search/auth/data are exactly the areas most likely to regress silently. Summary below.

Review findings

BLOCKER — .assets-revision not verifiable. Confirmed: it was still pinned to a pre-drugs_com upstream SHA. Fixed by pinning to refs/pr/39 on the ChilleD/WebHarbor dataset (https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/39) — verified this ref is a strict superset of main (all 16 existing tarballs unchanged, plus drugs_com.tar.gz), and verified ./scripts/fetch_assets.sh drugs_com actually resolves it and extracts a working seed DB. Will re-pin to main (or the merge commit) once that HF PR merges.

MAJOR — search-results answer-leak. Confirmed by reading templates/search.html: both the "Best match" exact-match card and the .drug-card list results rendered drug_class, brand_names, the CSA-schedule chip, avg_rating/review_count, and a uses/description snippet — all without a click-through. Stripped all of it from both card types and from the popular_for_query sidebar (same issue via a brand-name + availability badge). Search results now show only the drug name and a link into the detail page; every field still exists, just gated behind a click.

MINOR — illusory auth. Confirmed: a before_request hook silently logged every request in as alice regardless of the /login form, so task 14's "sign in" step was a no-op. Removed that hook and rewrote /login and /register to do real bcrypt-checked auth, matching the pattern already used by every other site in the repo (bbc_news, allrecipes, amazon, github, coursera, merriam_webster). Verified end-to-end (wrong password rejected, correct password succeeds, anonymous access to @login_required routes now redirects to /login) both via the Flask test client and against the real container with cookies.

MINOR — AppleDouble junk still shipping. Traced back to the same root cause as the BLOCKER — the pinned revision didn't include the (already-cleaned) tarball from the first review round. Re-verified the packed tarball has 0 ._*/.DS_Store entries; reachable now via the refs/pr/39 pin.

MINOR — data inconsistencies. Confirmed both:

• suvorexant's availability field was "C-IV" (a CSA schedule value) instead of "Rx" — fixed to match the pattern every other controlled-substance row follows.
• 5 rows were brand-as-generic duplicates (januvia/jardiance/xarelto/xeljanz/xanax), each duplicating an already-correct generic row (sitagliptin/empagliflozin/rivaroxaban/tofacitinib/alprazolam). Removed all 5 — none were referenced by any task or by pill/interaction/pregnancy-risk data. (Kept janumet, yaz, yasmin — legitimate distinct real-world branded products, not the same bug.)

NIT — CLAUDE.md port typo. Fixed (49015 → 49016).

Additional issues found during the self-review pass (not in the original review)

• CSRF protection was missing. A dedicated auth-security review pass flagged that drugs_com was the only one of 17 sites without CSRFProtect(app) — every other site (including the two the auth rewrite was modeled on) has it. Added CSRFProtect, added a csrf_token() hidden field to all 19 classic <form method="post"> submissions across 14 templates, and @csrf.exempt on the 3 JSON/AJAX-only endpoints (matching the exact pattern other sites use for their own AJAX endpoints). Verified: a form POST without a token now gets rejected (400), the same form with its token succeeds, and the exempted JSON endpoints are unaffected.
• 4 more drug-classification errors, found by a targeted sweep of the full DRUGS_DATA table cross-referenced against each class's own description text: vancomycin was filed under "Cephalosporins" (it's a glycopeptide, not a beta-lactam — contradicts the class's own description), colchicine under "Xanthine oxidase inhibitors" (it doesn't inhibit xanthine oxidase — that's allopurinol), and glipizide/nitrofurantoin under "Sulfonamides" (an antidiabetic sulfonylurea and a nitrofuran, respectively — neither is a sulfonamide). Added 4 correctly-described classes (Glycopeptide antibiotics / Antigout agents / Sulfonylureas / Urinary anti-infectives) and reassigned all four. None of the 4 are referenced by any task.
• 2 dead comparison suggestions on the Compare Drugs page referenced drugs that don't exist in the catalog (orlistat/phentermine, and one leg of a diabetes pairing) — removed those two broken suggestion pairs.
• Docs vs. reality: two CLAUDE.md files described task 14 as "add lisinopril to alice's med list," but the actual task in tasks.jsonl is read-only (sign in, then list what's already saved). Fixed the docs to match; flagging for you separately in case you'd rather extend task 14 into a genuine write/add-action task — that's a benchmark-design call I didn't want to make unilaterally by editing the task file itself.
• Added an explicit order_by(DrugImage.id) to the pill-identifier queries for deterministic "first N results" ordering (was relying on implicit SQLite row order).

Consequence: reseeded DB (twice)

Both the data fixes and the class reassignments required regenerating instance_seed/drugs_com.db from scratch and re-verifying the idempotency invariant each time (seed_database() on the new frozen seed leaves the instance byte-identical). Final counts: 246 drugs (251 − 5 duplicates), 780 reviews. Final seed DB md5: 22d3eeb82d7f2804592f0a9a0832485c (was bfc94f8ff61fbbd553ae496217588bad before any of this round's fixes — expected, since the underlying data changed). Re-uploaded the corrected tarball to the same HF PR rather than opening a new one.

Full re-verification after all of the above

Via a fresh Docker build + the real control plane (not just the Flask test client):

• All 17 sites → 200.
• POST /reset/drugs_com → md5(instance) == md5(instance_seed) == 22d3eeb82d7f2804592f0a9a0832485c.
• All 21 tasks' routes re-checked against the new DB by an independent review pass (no FAILs; nothing broke from removing the 5 duplicate drugs or reclassifying the other 4).
• Real auth flow re-verified against the running container with actual cookies (wrong password → rejected; correct password → session → /my-med-list shows alice's 3 meds, not lisinopril).
• CSRF re-verified: unauthenticated form POST without a token → 400; the same form with its token → succeeds; exempt JSON endpoints unaffected.
• Re-swept templates/search.html output structurally (not substring grep) to confirm the .drug-card/.exact-match-card blocks contain nothing but name + navigation links across multiple queries (ibuprofen, semaglutide, atorvastatin, antibiotics).

[boyugou]

Update — ran two independent Codex (GPT-family) reviews as a second opinion, fixed what they found

The prior round of fixes (search-leak removal, real auth, CSRF, seed-data cleanup) was checked by Sonnet-based subagents before landing. To get a genuinely independent second opinion — different model family, not just a different prompt — I ran two Codex reviews against the same diff: one a targeted re-check of those specific changes, one fully open-ended ("explore however seems useful, don't work off a checklist"). Both surfaced real issues.

CSRF exemption gaps (targeted review, then verified by hand). Two routes were @csrf.exempt despite genuinely consuming request.form/request.values from a real browser form or FormData submission, not just JSON — meaning the csrf_token field already present in those forms was silently never validated:

• /my-med-list/toggle — removed the exempt; the JSON-only call site now sends the token via an X-CSRFToken header (Flask-WTF validates this for non-form requests).
• /<slug>/review/<id>/helpful — removed the exempt; both FormData call sites now append csrf_token alongside the vote.

Verified live against the running container: toggling a saved drug without a token now returns 400; with the header, 200.

Open redirect in login() (both reviews independently flagged this). request.args.get("next") went straight into redirect() with no validation — POST /login?next=https://evil.com/phish would redirect there after a successful login. Added a same-site-only check (single leading slash, no scheme/netloc) and use it in place of the raw query param.

7 drugs referenced drug-class names with no matching class definition (open-ended review) — memantine/trimethoprim/lecanemab/ezetimibe/ursodiol pointed at class names that didn't exist in the class table at all, so drug_class_id silently ended up NULL for each (no error, just a quietly missing "Drug class" line on those 5 pages). Added the 5 missing classes. umeclidinium and ulipristal referenced near-miss variants of existing classes that didn't match exactly — reclassified both into the correct existing classes (Anticholinergic bronchodilators; Progestogens).

Three smaller bugs (open-ended review):

• 404 handler's "Popular drugs" section ordered by a nonexistent Drug.rating_count (typo for review_count), caught by a broad except and silently emptying the section.
• Two templates (drug_side_effects.html, drug_dosage.html) referenced a DrugImage.dosage_form field that was never a real column — cleaned up the dead reference in both (one was already saved by a working fallback; the other always rendered the same static fallback text regardless of drug, now made explicit).
• /api/interaction-check — a JSON endpoint, confirmed unreferenced by any template and unreachable through the UI, so no task depends on it — silently reported "alcohol" as unrecognized instead of resolving it the way the main interaction-checker page already does correctly. Fixed for consistency; zero regression risk since nothing calls it today.

One finding from the open-ended review I did not act on: it flagged that some browse routes' canonical URL (via url_for()) resolves to the .html variant rather than the "cleaner" extensionless one. This is a style/consistency question, not a functional bug — both URLs work, nothing breaks — so I'm leaving it for a maintainer's judgment call rather than unilaterally re-ordering route decorators across the file.

Consequence: the 7-drug class fix touches seed data, so instance_seed/drugs_com.db was regenerated again and the idempotency invariant reverified. New seed DB md5: 4af8a6688ec3b1d291b493abbfbdd7c8. Re-uploaded to the same HF PR discussion (https://huggingface.co/datasets/ChilleD/WebHarbor/discussions/39).

Full re-verification: fresh Docker build, 17/17 sites → 200, POST /reset/drugs_com → md5(instance) == md5(instance_seed) == 4af8a6688ec3b1d291b493abbfbdd7c8, and both CSRF fixes + the open-redirect fix confirmed against the actual running container (not just the Flask test client).

[boyugou]

Update — ran 4 parallel Codex reviews from distinct angles, fixed what held up

After the earlier two-review pass (targeted + open-ended), I ran four more Codex reviews in parallel, each scoped to a different concern: security deep-dive, benchmark task-quality/leak audit, performance and code-quality, and data-integrity/cross-field consistency. Every finding below was independently verified against the codebase before I acted on it — most security and performance findings were confirmed real but judged hardening-only with no benchmark impact, and left alone (reasoning included). Two findings were genuine bugs worth fixing immediately.

Three more answer-leaks, beyond the search-results one fixed earlier (task-quality audit):

• Homepage had a dedicated "New Drug Approvals" box listing article titles by recency — directly answering task --11 ("find the most recent article in that category") without ever visiting the news section. It also had a second, display:none-hidden grid section repeating the same titles in raw HTML (dead code — removed). The visible box now shows only a category link, no titles. Also excluded that category from the homepage's general "Latest Medical News" feed, which was independently leaking the same title via its category badge.
• /drug-classes index listed each class's top-3 drugs by name — directly answering tasks like --16 ("list 3 benzodiazepines") without opening the class page. Removed the list; this also drops an N+1 query (104 extra per-class lookups) the performance review flagged separately, so it's a two-for-one.
• /search's "Popular for <query>" sidebar showed up to 8 drug names for a matched condition/class, leaking --6/--9/--20 the same way. Removed the block and its backing query.

(One nuance: the audit also flagged plain search results surfacing statin/diabetes/hypertension drug names when searching those exact terms — I did not touch that. Task --9's own wording says "Search Drugs.com for drugs in the Statins drug class," so search legitimately returning statin-class drugs by name is the intended mechanism, not a leak.)

OpenFDA content-mismatch root cause (data-integrity review) — this was the most serious finding. fetch_openfda_label() accepted whatever the FDA label-search endpoint returned with no check that it actually described the queried drug. Confirmed live: querying polyethylene glycol returns a label whose own openfda.generic_name is ["NAPROXEN"] — a completely unrelated drug — because FDA's search does fuzzy/tokenized matching, not exact matching. This had silently seeded at least 4 drugs (polyethylene glycol, tobramycin ophthalmic, zoledronic acid, insulin aspart) with a different drug's entire uses/warnings/dosage/side-effects text. Added a match check (the returned label's own generic name must equal, contain, or token-subset-match the query) before accepting it; on rejection, seeding falls back to the existing synthetic-content generator. Verified the 4 known-bad drugs are clean post-fix, then swept the full 246-drug table for any other cross-drug duplicate uses/warnings/dosage text — none found.

XSS escaping inconsistency (security review). compare_drugs.html and interaction_checker.html both built their autocomplete dropdown via string-concatenated innerHTML without escaping the API-provided label, unlike base.html's autocomplete, which already does. Current data source is seeded drug/brand names, not user input, so not exploitable today — but the pattern was inconsistent and a latent risk if a future contributor routes less-trusted text through the same endpoint. Added the same escapeHtml() used elsewhere to both call sites.

Confirmed real but not actioned, with reasoning:

• N+1 query patterns in search/condition/account pages and the (unreferenced, dead-code) /api/interaction-check's O(n·m) brand lookup — real, but at 246 drugs the practical latency is negligible, and this is a benchmark environment where task correctness matters, not throughput. Fixing all of these would touch a lot of surface area for no benefit to what the benchmark measures.
• Canonical URL drift (url_for() resolving to the .html alias on a few routes), no visible logout link in the header, "Advertise" and "About" resolving to the same page, hardcoded SECRET_KEY — real, but each is either purely cosmetic, a convention shared by all 17 sites in the repo (not a regression here), or not reachable by a UI-only browsing agent (a forged session via the shared secret requires source-code access and out-of-band crypto, not something a browser-driving agent can do).
• IDOR and SQL injection: reviewed specifically, no findings — every owner-scoped route correctly filters by current_user.id, and no query builds SQL from unparameterized user input.

Consequence: seed content changed again (both from the leak-fix query changes and the openFDA re-fetch), so instance_seed/drugs_com.db was regenerated and idempotency reverified. New seed DB md5: 77e17b36c03fde357c3248895aff65d7. Re-uploaded to the same HF PR discussion.

Full re-verification: fresh Docker build, 17/17 sites → 200, POST /reset/drugs_com → md5(instance) == md5(instance_seed) == 77e17b36c03fde357c3248895aff65d7, all three leak fixes confirmed absent from the rendered pages, and the login/CSRF flow reconfirmed against the running container.

[boyugou]

Update — large data-quality + settings-persistence round, independently re-verified

Commit d1f096f is a substantial change (~1732 lines in app.py, ~20 templates, a reseed) that landed on this branch outside the usual incremental-fix flow this thread has followed. Before accepting it, I ran 6 independent, parallel re-reviews covering every angle that scale of change warrants: scope-compliance (did anything outside sites/drugs_com/ get touched inappropriately), app.py correctness on the full diff, data-integrity on the reseed, all 21 tasks against the new state, template-diff correctness, and an end-to-end re-verification of every previously-established guarantee (auth, CSRF, the answer-leak fixes, the OpenFDA content-integrity guard, byte-identical reset). Summary below.

Scope check: no violation. 6 root-level files were touched outside sites/drugs_com/ (.assets-revision, AGENTS.md, CLAUDE.md, CONTRIBUTING.md, README.md, websyn_start.sh). Every one of them is a stale 15/16-site or 40000-40014/40000-40016 port-range doc reference being corrected to the true 17-site state — the kind of correction that was already overdue on this branch. websyn_start.sh's SITES array, site ordering, and every other site's port assignment are untouched (the only change there is a one-line comment). No sites/<other-site>/ directory appears in the diff. .assets-revision's new pin (a resolved commit SHA rather than the branch ref) was verified to be the actual current head of refs/pr/39 — not a stale/older commit.

Data change: review count 780 → 716, expected and correct. DrugReview gained a (drug_id, user_id) uniqueness constraint, and the seeding logic deduped a popular-drugs list that had accidentally listed 16 drugs twice. Reproduced the old and new algorithms independently and the numbers match exactly: 32 literal duplicate rows removed, 32 more trimming over-represented drugs to a uniform review count. Verified no drug lost all its reviews, alice's 3 benchmark-critical reviews (ibuprofen 9/10, lisinopril 7/10, hydrochlorothiazide 10/10) are untouched, and every drug's cached avg_rating/review_count still matches its actual review rows.

Consequence for task --12: atorvastatin's rating is now 7.0/10 with 4 reviews (this PR's own earlier description had listed a since-superseded 7.1/10, 8 reviews snapshot — now corrected). If a reviewer has already pinned a verifier/rubric to the old numbers, it needs updating to match.

Genuine improvements this round made (not requested, but confirmed as strict improvements after tracing the diff): account_settings/account_subscriptions — previously flash-only no-ops, a gap an earlier review round flagged and deliberately left unfixed since no task exercised it — now actually persist to the database; the homepage's featured-drug cards had a rating/review-count badge removed (an answer-leak for task --12 this round happened to also close); /logout is now POST + CSRF-protected with a visible header link (previously a bare unlinked GET route); a session/DB-ledger-backed dedup was added for the review "helpful" vote buttons, correctly scoped to this app's actual single-process/multi-threaded deployment model (traced the lock ordering across every handler — no deadlock risk, bounded memory via FIFO eviction, no leak).

One accepted, non-blocking gap found in the new vote-dedup logic: logging out and then re-voting "helpful" anonymously in the same browser can double-count a vote by +1 until the next login reconciles it. Self-healing, bounded to +1 per review per browser, and no benchmark task reads the helpful_count field — left as-is rather than adding more complexity for a non-task-affecting counter.

Full re-verification after accepting this round: all 21 tasks re-audited against the current app.py + current seed data (20 PASS, 1 CONCERN — task --12's updated answer, noted above, not a bug); every previously-established guarantee re-confirmed by actually running the app and driving requests through it (not just reading code): real auth (wrong password rejected, correct password + CSRF token succeeds, unauthenticated access to gated routes redirects to /login), CSRF on every checked route (/login, /my-med-list/toggle, the review-helpful endpoint, /account/settings/save, /logout — all reject a missing token with 400 and accept a valid one), the OpenFDA content-integrity guard (still wired in, and a full-table sweep of the new seed data found zero cross-drug duplicate uses/warnings/dosage text), the three answer-leak guardrails (homepage, /drug-classes index, /search sidebar — all still absent), and the byte-identical reset invariant against the new seed md5.

New seed DB md5: d3d228f3fc0b3b880e149ac35550e3c5. Uploaded to the same HF PR discussion (already reflected in .assets-revision's current pin). PR description and this comment reflect the corrected review/condition counts and md5.

[boyugou]

Update — independent fresh-eyes review round (not diff-scoped)

Rather than only re-checking the previous large round's diff, this pass ran 5 independent reviews from angles chosen to not assume anything already fixed is still correct, plus a live re-verification of every benchmark task against a freshly built image:

1. Live task re-verification — built the image fresh, ran a new container, and drove all 21 tasks with real HTTP requests (not just reading code): every task resolves to a stable, unambiguous answer, POST /reset/drugs_com stays byte-identical, and real login/CSRF flows behave correctly (wrong password rejected, missing CSRF token → 400, correct login succeeds).
2. Full-table data-integrity sweep — referential integrity, cached-aggregate correctness, uniqueness, cross-drug content duplication, and value-domain checks run across all 246 drugs (not sampled). Clean.
3. Security-focused fresh audit — SQL injection, SSRF, path traversal, XSS, auth/session, mass-assignment, rate-limiting, secret handling, each verified by tracing the actual code path rather than trusting prior claims. Clean for this offline single-tenant benchmark's threat model.
4. Concurrency/thread-safety audit — traced every module-level mutable state and lock acquisition site against the confirmed single-process/multi-threaded deployment model. No race, deadlock, or memory leak; found that a coarse _sqlite_write_lock already serializes every writer, making the finer-grained locks belt-and-suspenders (a design note, not a bug).
5. Template/frontend audit — url_for references, XSS escaping consistency, CSRF token presence on all 22 POST forms, accessibility, orphaned templates, and JS DOM references, checked across all 49 templates.
6. Fresh full-file app.py logic audit — read all 8446 lines end-to-end for business-logic bugs, error-handling gaps, and edge cases, independent of the prior round's diff.

Two minor items came out of this, both addressed in 588cbab:

• drug_detail.html had an orphaned JS click handler for a #share-btn that no longer exists on the page (harmless — guarded by if (shareBtn), no console error — but dead code). Removed.
• avg_rating is a cached column computed with Python's round() (banker's rounding). For a drug whose true mean ends in exactly .x5, this can differ by 0.1 from SQLite's ROUND(AVG(rating), 1) (which rounds half-away-from-zero) — the site itself is internally consistent (every page reads the same cached column), but a reviewer writing a deterministic verifier by recomputing the average via raw SQL could pin a value 0.1 off from what the site actually displays. Documented in sites/drugs_com/CLAUDE.md so this doesn't bite a future verifier.

No task-affecting bugs, no data-quality problems, and no reseed/repackage was needed this round — everything else came back clean, including re-confirming that the prior large round (d1f096f) introduced no scope violations and no regressions.

[boyugou]

@MufanQiu @YuanDaoze Finished a major round of revision. Please feel free to take a look when you have a chance. Thank you!