Skip to content

STATUS.md

Latest commit

 

History

History
110 lines (92 loc) · 6.25 KB

STATUS.md

File metadata and controls

110 lines (92 loc) · 6.25 KB
Raw

Winnow — Status

Last updated 2026-06-16 · v0.2 (four phases + POC demo, watchlist, extra detectors)

What Winnow is

A focused alternative to last30days-skill: same "what are real people saying in the last 30 days" concept, but it ranks by authenticity (not raw engagement) and covers the sources last30days skips (App Store + Play Store reviews, Pantip/Thailand). One Python engine, two surfaces (agent skill + web dashboard).

Build state — plan complete (Phases 1–4)

Phase Tasks Status
1 Engine depth & trust T1 Play Store hardening · T2 Reddit feed · T3 account-age trust · T4 SQLite store · T5 name→id ✅
2 Distribution T6 self-contained skill · T7 .skill build + install docs · T8 CI ✅ (T8 deferred — gated)
3 Web dashboard T9 scaffold · T10 engine API · T11 brief view · T12 trend chart · T13 HTML export ✅
4 Differentiators T14 competitor diff · T15 YouTube + X adapters ✅

v0.2 additions

  • POC demo mode — winnow --demo runs on built-in seed data (no network/keys); the seed exercises every detector. Dashboard has a Try the demo button (/api/demo) that seeds the store and shows brief + trend in one click.
  • Watchlist + alerts — engine/watchlist.py compares the last two stored runs per topic and alerts on trust drop / rating drop / astroturf spike.
  • Extra authenticity detectors — rating-text-mismatch (5★ + negative text) and a corpus rating-polarization warning (1/5-star manipulation signature).
  • Pantip search API — real keyword search, gated on PANTIP_AUTH, RSS fallback.
  • Gated paths — tested on placeholder payloads, NOT yet verified live. X, Reddit-OAuth, and Pantip-search fetch() are exercised against recorded sample payloads, which proves the parsing given a correct response — it does NOT prove the live request succeeds. Honest confidence: Reddit-OAuth likely works (documented API); X is experimental (sends only one GraphQL feature flag and a fixed query id — real X needs the full set, so it will likely return nothing until fixed against a live response); Pantip-search is experimental (token format / response shape unconfirmed; RSS covers Pantip regardless). These need a real-key run to confirm and fix.

69 engine tests passing. Web typecheck + lint clean.

Architecture (see ARCHITECTURE.md)

adapters (6) ─► authenticity.score_corpus ─► fusion.rank (trust-gated)
   │                                              │
   │                          ┌───────────────────┼──────────────┐
appstore playstore            ▼                   ▼              ▼
pantip reddit            --store SQLite     render md/html/json  compare
youtube x                      │                   │              │
                               └──► winnow-web (Next.js dashboard)
  • Engine: engine/winnow.py (CLI) + engine/lib/* (zero runtime deps, Python 3.9+).
  • Authenticity spine (authenticity.py): burst, near-duplicate, thin-text, anon/generated-handle, new-account, low-karma (last two via Reddit OAuth).
  • Skill: skills/winnow/SKILL.md (thin), engine symlinked in; build-skill.sh bundles.
  • Web: winnow-web/ — brief view, trust bars, reputation-trend chart, HTML export.

Sources

Source Path Auth Live status
App Store reviews iTunes RSS none ✅ verified
Play Store reviews batchexecute RPC none ✅ verified
Reddit search.rss (keyless) → OAuth enrich optional .env ✅ RSS verified · OAuth likely (unverified)
Pantip (Thailand) RSS (keyless) → search API optional PANTIP_AUTH ✅ RSS verified · search API experimental
YouTube ytInitialData scrape none ✅ verified
X / Twitter GraphQL search requires AUTH_TOKEN+CT0 cookies ⚠️ experimental (not verified live)

Run it

python3 engine/winnow.py "Notion" --sources appstore,reddit --limit 25
python3 engine/winnow.py "Notion" --vs "Obsidian" --sources appstore   # competitor diff
python3 engine/winnow.py "Slack" --sources appstore --store            # persist for trends
python3 engine/winnow.py --demo                                        # POC seed data, no keys
python3 engine/winnow.py --seed-store && python3 engine/watchlist.py   # seed + reputation alerts
cd winnow-web && pnpm dev                                              # dashboard :3000 (Try the demo)
python3 -m pytest -q                                                    # 69 tests
bash build-skill.sh                                                     # dist/winnow.skill

Open threads

  • CI (T8): pytest-on-push workflow. Blocked by hook + global rule; needs an explicit "add CI" instruction.
  • Brand disambiguation: DONE - --context-include / --context-exclude filter keyword sources (reddit/youtube/x/pantip); the host model supplies the terms (SKILL.md). Review sources are unambiguous and never filtered.
  • Cross-source engagement normalization: DONE - fusion.rank now scores engagement per-source (log-normalized vs each source's top item), so YouTube's millions no longer crush a 50-helpful review, PLUS a per-source diversity floor (min_per_source=2) so a zero-engagement source (Reddit RSS) is not shut out of a mixed run. Mixed appstore+reddit run went from 7:1 to 6:2.
  • Dashboard feature parity: DONE - the web UI now exposes competitor diff (/api/compare + CompareView, "Compare vs…" input) and disambiguation ("Exclude terms…" input), matching the engine/CLI.
  • X query-id rotation: the GraphQL query id is hard-coded with an env override (X_SEARCH_QUERY_ID); will need updating when X changes it.
  • git repo: DONE - shipped to https://github.com/SUDARSHANCHAUDHARI/WinNow (private), main is the source of truth.

Bugs caught by live-run discipline

burst false-positive on recency slices · stale-date leak (2018 items in a 30-day tool) · Reddit spam results (sort=new firehose) · manufactured-tie verdict in compare · brace typo in a test fixture. None would have surfaced from unit tests alone.