Changelog

All notable changes to soapbox.media are tracked here. Format follows Keep a Changelog; versioning follows SemVer.

Pre-1.0 minor versions correspond roughly to development phases of the pre-launch build leading into the November 2026 US midterms.

v0.40.2 · 2026-10-15

Changed

• Public Figures roster refresh for the midterms. The roster had not changed since it was seeded on 2026-06-23, so /figures read like a June cast list. A scan of the 2,646 transcripts from the last 14 days ranked people by episodes mentioning them; the adopted default adds the Senate-map and administration names above the 60-episode line: Ken Paxton, James Talarico, Mike Rogers, Susan Collins, Troy Jackson, Sherrod Brown, Jon Ossoff, Marco Rubio, Pete Hegseth, Mike Johnson. Two ASR alias fixes on existing rows: "mom donnie" and "zohran" for Mamdani, "kla harris" for Kamala Harris. Trump stays held (he appears in roughly a third of all episodes). The roster is data in the figures table; npm run seed:figures-midterms applies this change idempotently, followed by npm run detect:figures -- 30, npm run score:figures -- 12000 15, and npm run refresh:boards.
• Figures scoring cap 1,200 -> 3,000 per daily run (FIGURES_SCORE_LIMIT). The old cap was really 1,000 until v0.39.1 (PostgREST's row cap on the anti-join) and took about 70s; 3,000 fits the 240s budget with the pool's deadline as the backstop, and lets the roster backfill clear in a day or two instead of a week.

v0.40.1 · 2026-10-14

Fixed

• Hourly snapshot cron killed at 300s after the backlog drain. Refilling the Index window (v0.39.1's drain scored 22K mentions) more than doubled the rows behind fetchScoreRows to 51K over 37 days, and the deep sentiment_scores -> classifications -> episodes -> channels join, keyset on a random uuid, went from 70-160s to 80-145s per scan with intermittent statement timeouts. writeHomeSnapshot ran that scan up to four times (the 60s memo expired between consumers once a scan outlived it), so every run from 16:35 UTC on was killed at Vercel's limit with no usage_log row, and the homepage stayed on the 14:36 snapshot (Index 0.0) for the rest of the day. The scan is now the same index-driven traversal as the score stage: episodes in the window via the published_at index, their classifications via the episode_id index in chunks, their scores via the unique FK in chunks, assembled in JS against small channel and issue lookup maps (src/lib/db-scan.ts). Validated against the old join on a 7-day window: identical rows (9,015) and Index (-0.948). Measured: the whole 37-day dashboard in 22.7s, versus 57.6s for the old join on 7 days alone. The memo now lasts 5 minutes so one scan serves the blended, independent, and legacy dashboards plus the breakdown.
• Snapshot runs can no longer vanish. The route inserts its usage_log row up front (beginScriptRun, a red "started and never completed" message) and overwrites it with the outcome at the end (completeScriptRun). A run killed at the function limit now leaves that row behind instead of nothing - the gap that hid the score cron's death for three weeks and this one for three hours.

v0.40.0 · 2026-10-14

Added

• Precomputed /figures and /advertisers boards. Both pages are force-dynamic and recomputed their overview on every request: a month of figure_scores (up to 60 pages) for /figures and every confirmed ad_segment in the window for /advertisers. Measured 2026-10-12: 28s time to first byte cold for /figures (3s warm) and 10 to 11s for /advertisers on every load. Their inputs only change when the daily figures and ads crons run, so a new daily cron, /api/cron/boards (12:20 UTC, after both), now writes the figures board and all three advertiser windows into dashboard_snapshot rows - the home-snapshot pattern. The pages read one row and fall back to the live computation only when the row is missing or older than 48 hours, so a missed cron day degrades to slow, never to stale. The cron records a usage_log row either way, with error_message on failure. npm run refresh:boards forces a refresh. Per-figure and per-advertiser drill-downs stay live (narrow queries).

v0.39.1 · 2026-10-14

Fixed

• Score cron silently dead since 2026-09-25; homepage Index read 0.0 from 10-08. runScore opened every run by paging the ENTIRE classifications table (deep join, OFFSET .range()) plus the entire sentiment_scores table just to diff the unscored set: 249K + 229K rows by October, a 7-10 minute preamble that Vercel killed at the 300s function limit before a single score or usage_log row was written. Run durations crept from 190s (Aug 20) to 270s+ (Sept 19); logged runs then thinned out and vanished, scores trickled to zero on 10-02, and the hourly snapshot cron kept publishing an Index of 0 with empty issue and mover lists. Same all-time-scan + OFFSET anti-pattern as the June-August snapshot freeze (v0.36.2), in the one scan that fix did not touch. The pending set now comes from a bounded, index-driven lookup (fetchPendingClassifications): episodes published or ingested in the last SCORE_LOOKBACK_DAYS (30) via the published_at index, their classifications via the episode_id index in chunks, and the already-scored ids via the unique FK index in chunks - every call bounded, full pages split (the PostgREST 1000-row cap), page overflow throws. Measured: 13s for the lookup plus 720 fully hydrated rows against a 22K backlog, versus a function kill. Newest episodes first, so a backlog refills the Index's 7-day window before older history. Anything unscored older than the window is abandoned on purpose; npm run score -- <limit> [lookbackDays] sweeps wider.
• Score stage now has a wall-clock deadline and fails loud. runScore passes the stage budget to the lookup and to the scoring pool, so it always returns in time to record its run. A thrown stage now lands in usage_log.error_message for every per-stage cron (recordPipelineRun derives it from the stage detail), and the score + figures routes answer 500 so Vercel's cron log shows the failure too. Before this a stage error only lived inside raw_summary, where nothing surfaced it: the figures cron failed one day in three for weeks with no visible failure.
• Figures cron timed out about one day in three (Sept 29, Oct 2, 4, 10, 11): its unbounded !left + .is(null) anti-join over all of figure_mentions hit the statement timeout, and its .limit(1200) was silently capped at PostgREST's 1000 rows. It now uses the same bounded lookup over the last FIGURES_PENDING_LOOKBACK_DAYS (14); 3.9s measured.
• Snapshot cron fails loud on an empty Index window. When the 7-day window holds zero scored mentions while classified episodes exist in it, the run is recorded with an error_message (red on /admin/pipeline) and answers 500, instead of quietly publishing 0.0 every hour for a week. The snapshot is still written, so the page shows the truth.

v0.39.0 · 2026-08-26

Added

• Advertisers: full-history views, honestly bounded. A second backfill pass covered every remaining transcript ($1.28; zero failures under the v2 confirm prompt), so detections now exist back to 2022 in a few channels' backfilled tails - but dense panel coverage only begins May 2026, and showing the sparse remnant would misread our coverage ramp as advertiser behavior. All ad views therefore clamp to AD_COVERAGE_START_ISO (2026-05-01; pre-coverage detections stay in the DB, unshown). On that foundation: the /advertisers board gains a window selector (30 days / 90 days / All since May 2026, ?w=), and each advertiser page now shows its full covered record - totals with a "since <month year>" line and a monthly placement-history chart (reads stacked by carrying-channel lean, house diverging palette: blue left / neutral gray middle / red right, Recharts v3 via the shadcn ChartContainer with tooltip + legend). Drill-down show lists and placement bars span the covered window; receipts stay the most recent 12. Middle segment of the placement split bar darkened (slate-300 → slate-400) after a contrast check flagged it as near-invisible against the track.

Added

• Advertisers: the advertiser × lean map. New public /advertisers board showing who underwrites online political media - each advertiser's detected host-read sponsor messages over the last 30 days, the shows carrying them, rough read airtime, and a reach-weighted placement-share bar across left/middle/right channels (same reach factor as the Index). Per-advertiser drill-down (/advertisers/[slug]) lists carrying shows and links every detection to its receipt: the transcript excerpt, source link, and timestamp where available. Honest scope, stated on the page and in a new /methodology section: host-read sponsorships only (produced/dynamically-inserted spots carry no sponsor-read language and are out of scope by design), estimates rather than verified contracts, never full transcripts.
• Detection pipeline. Two-phase, mirroring figures: a free regex anchor pass (validated by scripts/probe-ad-detection.ts at 94% precision / ~12% missed-ad rate on 300 transcripts) inserts candidate segments as verdict='pending' into the new ad_segments table, idempotent on (episode_id, char_start); a Haiku confirm pass settles each candidate to ad / self_promo / not_ad and resolves the brand + domain, fixing transcription mangles from context ("hymns.com" → Hims & Hers). Settled rows are tombstones - never re-judged, so detect re-runs cost nothing. Brand identities funnel through canonicalAdvertiser() (spelled-out URLs, "dot com" speech, known-brand seed map) onto one slug per advertiser in the new advertisers table. Daily /api/cron/ads (12:10 UTC), bounded by ADS_CONFIRM_LIMIT + the stage time budget (~$0.15/day at current volume), logged to usage_log like every stage. npm run backfill:ads -- <days> seeds history (hard-capped at 30K LLM calls per run).

Fixed

• /log "Couldn't load episodes" (statement timeout). The episode table paged episode_pipeline_summary with count: "exact"; that view GROUP-BYs every classification + sentiment score (183K rows each) before it can sort, page, or count, and the count query crossed the statement timeout as the tables grew (the page query itself was at ~6s and climbing). getEpisodeTablePage now pages the base episodes table (index-backed, flat cost - ~0.8-1.3s for every sort/search/page shape), then fetches classify/score counts for just the ~50 visible rows via classifications with a one-to-one score embed. (Query-shape note: starting that count from sentiment_scores with a nested episode filter let the planner walk the whole table and time out; starting from classifications rides the episode_id index.) Search now resolves matching channel ids first and folds them into a single-table or(); channel-column sorts order by the embedded to-one column; the scored_count sort maps to classify_status (the view aggregate doesn't exist on the base table).
• Same view retired from the aggregate paths. getDashboardData's episode total and getSystemStats' episode count + duration sum now read episodes+channels directly - the duration sum was OFFSET-paging the view, re-running the full GROUP BY on every page, and was most of the snapshot cron's runtime. The view remains only for the per-channel episode list, where the channel_id filter pushes down before aggregation.

v0.37.0 · 2026-08-25

Added

• Public Figures pipeline revived (daily cron). The v0.33.0 feature shipped as a one-time backfill and froze on 2026-06-23, leaving /figures with empty 30-day windows. New daily /api/cron/figures (11:40 UTC) runs the regex detect pass over the last 3 days of transcripts (free, idempotent on (episode_id, figure_slug, char_offset)) and then Haiku favorability over unscored mentions, found via a server-side anti-join (figure_scores!left + .is(null)) instead of paging every scored id into memory. Bounded by FIGURES_SCORE_LIMIT (1200) + the 240s stage budget; ~$0.5/day at current volume, logged to usage_log (cost folded into /admin/costs like score-emerging). scripts/detect-figures.ts is now a thin CLI over the shared stage logic (detectFigureMentions), keyset-paginated. The June-Aug gap was backfilled the same day (detect + attended score run).

Removed

• Trending Names (dead since v0.14.1). The home-page tease was removed in v0.14.1 and readTrending has had zero callers since; the writer kept running until it silently died in June. Deleted src/lib/trending.ts, the unscheduled /api/cron/trending route, scripts/refresh-trending.ts, the refresh:trending npm script, the Trending-Names accuracy probe (scripts/probe-entity-context.ts), and the orphaned trending_v1 row in dashboard_snapshot. (The entity/event burst probes stay - they are the basis of the living /emerging discovery funnel.)

v0.36.2 · 2026-08-25

Fixed

• Two-month home-page data freeze (2026-06-30 → 2026-08-25). The homepage served the June 30 dashboard_snapshot row for two months. Root causes, all in fetchScoreRows: it OFFSET-paginated the ALL-TIME sentiment_scores deep join, which (a) crossed Supabase's statement timeout once the table passed ~99K rows, failing every snapshot refresh into a swallowed console.error, and (b) carried a silent maxPages = 50 cap that - because the PK is a random uuid - had already been computing the public Index from a ~27% pseudo-random sample of scores. Every consumer windows to at most windowDays + 30 days, so the fetch now pushes that date bound into the query (!inner embeds) and keyset-paginates on the PK (.gt(lastId), flat cost per page). Hitting the runaway-guard page cap now THROWS instead of silently truncating.
• Same all-time-scan pattern fixed across every drill-down read path: fetchScoreRowsFiltered (issue/topic/channel pages), fetchIssueExcerptRows (mover breakdowns), fetchFigureOverviewRows + getFigureDetail (figures). All were fetching all-time rows and windowing in JS; all now push the window into the query with keyset pagination. Issue-page TTFB ~18s → ~2-6s. The channel drill-down also drops the "fetch up to 50K episode ids, resend as an IN-list on every page" pattern for a .eq("episode.channel_id", ...) on the !inner embed.
• Snapshot refresh moved to its own cron (/api/cron/snapshot, 30 min after each score run). Score saturates its full 240s budget under backlog, so the old tail-call refresh in the score cron ran past the 300s function limit and was killed mid-write. Refresh failures now land in usage_log.error_message (via recordScriptRun, which gained errorMessage
  • a cron source) so /admin/pipeline shows them instead of a console nobody reads.
• writeHomeSnapshot now runs ONE scan, not four. React cache() only dedupes inside a Server Component render - it never worked in Route Handlers or CLI scripts, so the blend + independent + legacy getDashboardData calls each re-scanned. A short-TTL module memo now shares one unfiltered scan across all cohort variants (cohort filtering is JS-side and PUBLIC_COHORTS covers every cohort). Local refresh: 176s → 81s.
• Ingest wall-clock budget + daily rotation. Ingest was the only stage with no time budget; full-list runs crept to 263-296s against the 300s function limit, so Vercel killed it mid-run on 6 of the last 16 days - silently (no usage_log row) and always cutting the same low-priority tail. It now stops cleanly at 260s (INGEST_TIME_BUDGET_MS), logs channelsSkippedAtBudget/stoppedAtTimeBudget, and rotates the channel start point daily within each platform group (preserving the YouTube-before-podcast dedup ordering) so any cut is spread across the list.
• MCP search_mentions offset capped at 5000. The offset was client-controlled and unbounded over the 183K-row deep join - a statement-timeout on demand. Deep history should narrow by published_after/published_before instead.

Added

• scripts/health-check.ts (pipeline freshness + cron-run history at a glance) and scripts/probe-drilldowns.ts (times every drill-down read path against the live DB - the regression harness for this release's fixes).
• Prepared-but-not-applied migration 20260825120000_episodes_published_at_index.sql (standalone episodes(published_at desc) index backing the new date-bounded queries). Apply manually via the Supabase SQL editor.

v0.36.1 · 2026-06-26

Changed

• Signup CTA redesign. The popup now leads with the L/R spectrum (a blue-to-red gradient accent + gradient "Send me the report" button, the brand signature) and pulls the current report out of the body copy into a standalone "this month's issue" card (title + receipt count). The report email-gate button matches the same gradient. Introduces the palette and makes the report feel like a discrete item you receive, not a clause in a paragraph.

v0.36.0 · 2026-06-26

Added

• Monthly-report signup + report gate (lead capture). A site-wide popup (EngagementPopup) invites engaged visitors (3+ distinct pages and 30s+ on the site this session; excludes admin/auth/report routes; shown once, respects dismissal/subscription) to join the free monthly Soapbox Report list. The CBS / Bari Weiss report is now email-gated: a public preview (headline, finding, the cross-network gap chart) stays open and crawlable, and the full depth (deep-dives, 136 receipts, timeline, methodology) unlocks after signup. Both flows post to /api/report-signup, which de-dupes by email into the new report_leads table (RLS-on, service-role) and sets an sb_unlocked cookie. Signup + popup-shown events also fire to PostHog. Adds shadcn dialog.

v0.35.0 · 2026-06-24

Added

• Receipts at volume on the CBS / Bari Weiss report. The report now carries a "Browse the receipts" section: 136 scored mentions of the administration from CBS, ABC, NBC, and PBS across five topics, grouped by story (collapsible), each with the model favorability score, an intensity meter, and a ▶ deep-link that opens the source video at the exact moment (deterministic timestamp via quoteStartSeconds). The Venezuela and immigration deep-dive quotes also link to their timestamps. This makes verifiability, not a few hand-picked quotes, the report's backbone, the hallmark we want across reports and the site. Built by scripts/bw-build-receipts.ts from the cached scores; receipts live in src/app/reports/cbs-bari-weiss/receipts.json.

v0.34.0 · 2026-06-24

Added

• Soapbox Report: "Did CBS News tilt toward Trump under Bari Weiss?" A one-time, transcript-based fact-check at /reports/cbs-bari-weiss. Measures CBS's on-air framing of the Trump administration against ABC, NBC, and PBS on the same stories, and finds no broad pro-administration tilt: CBS was, if anything, more critical than its peers on most major topics, and the one apparent exception (the Venezuela operation) dissolves on close inspection of the receipts. Story-matched quotes with source links, plus honest disclosure of the before/after window-sensitivity and the per-quote scoring limits.

v0.33.1 · 2026-06-22

Fixed

• Figure detail pages now link each show to its channel page. On /figures/[slug], the most-effusive / most-critical show rows and the quote receipts link the channel name to /channels/[id]. getFigureDetail now carries channelId through the per-channel stances and receipts.

v0.33.0 · 2026-06-22

Added

• Public Figures: a favorability board for individual people, a new top-level section. Tracks how favorably vs critically the panel talks about a curated roster of ~15 figures (politicians, officials, media + tech), on a −5 (hostile) to +5 (effusive) axis that is DELIBERATELY SEPARATE from the left/right Soapbox Index - favorability toward a person crosses party lines, so it never feeds the needle (new figures / figure_mentions / figure_scores tables; the Index computation does not reference them).
  • Detection is transcript-level (src/lib/figures.ts, npm run detect:figures): whole-word alias matching over transcripts, nearby-mention dedup, an ad-read pre-filter, a per-episode-per-figure cap, and free deep-link timestamps from YouTube caption markers.
  • Scoring is a separate Haiku pass (scoreFigureMention, prompt version v1) with an explicit attribution rule (a host quoting a critic is scored on the HOST's stance, not the quoted view) and an intensity gate that excludes passing name-drops / benchmark references from the favorability aggregate.
  • New /figures and /figures/[slug] pages (slate→emerald FavorabilityGauge, never the red/blue needle) showing net favorability, the most-effusive vs most-critical shows (volume-floored), and the sharpest quotes both ways with source deep-links.
  • New MCP tools list_figures and get_figure_detail (public guide updated).
  • Trump is held for a later release (he saturates ~20% of the corpus). v1 ships on a manual spot-check; a favorability gold set gates any published absolute number, and supporting_quote-recall caveats are documented.

v0.32.6 · 2026-06-22

Added

• Admin add-channel warns when a YouTube channel is a likely cross-platform mirror of a show already tracked on another feed. "Resolve & draft" now checks the candidate's recent uploads against existing active channels by publish date + normalized-title containment (either direction) and flags the best match in the preview. This catches the gap that let a duplicate get added: the ingest dedup (dedup.ts) only links feeds with the same channel name and exact normalized title, so a differently-named mirror (Dan Bongino vs The Dan Bongino Show) or a title-format difference (YT "Host | Title | Show 1408" wrapper, or a podcast "Ep. N -" prefix) silently double-counts in the Index. Detection is best-effort and never blocks adding. An audit found 6 such pairs already double-counting (Piers Morgan, Ben Shapiro, Lincoln Project, NYT/The Daily, Bongino, Benny Johnson) - reconciliation tracked separately.

v0.32.5 · 2026-06-22

Fixed

• Admin "Add a YouTube channel" (and cron ingest) no longer crash on duration-less videos. parseIsoDuration in src/lib/youtube.ts called .match() on contentDetails.duration, but YouTube omits that field for live streams, premieres, and upcoming/scheduled videos (the type annotation wrongly declared it non-optional, so tsc never caught it). Deep-ingesting a channel whose recent uploads included one such video threw "Cannot read properties of undefined (reading 'match')" after the channel row was already inserted, leaving a channel with 0 episodes. parseIsoDuration now accepts a missing value and returns 0 (those videos fall below MIN_DURATION_SEC anyway), and both videos.list call sites use optional chaining with honest types. The Add path was exposed because, unlike Resolve, its getRecentUploads call is not wrapped in a try/catch; the same parser also guards the daily cron ingest.

v0.32.4 · 2026-06-21

Fixed

• Close a double-count hole opened by v0.32.3's directional dedup. Because a YouTube ingest now defers to no one, a shared episode on an equal-reach dual-platform show could double-count if the podcast row happened to be processed before the YouTube row within a run (the podcast ingested the episode, then YouTube re-ingested it with nothing to dedup against). The channel ingest order was only reach desc, which is undefined for equal-reach sibling rows (the podcast row often inherits the YouTube subscriber count). Both ingest paths now process higher-priority platforms first (YouTube before podcast, via a stable sort on PLATFORM_PRIORITY), so the YouTube copy always lands first and the podcast defers. Only the rare cross-run case (a podcast copy ingested a full day before YouTube's appears) remains, which a future retroactive-supersede pass would close.

v0.32.3 · 2026-06-21

Changed

• Cross-platform dedup now deterministically prefers YouTube. When a show runs on both YouTube and a podcast feed, a shared episode (same title + date) used to be kept on whichever platform was ingested into the DB first (timing/order dependent, nondeterministic for equal-reach rows). Dedup is now directional: a podcast episode defers to its YouTube sibling, but a YouTube episode is never skipped because of a podcast sibling (loadSiblingEpisodeKeys in src/lib/dedup.ts, gated by a PLATFORM_PRIORITY map). This only affects overlapping episodes; YouTube-only clips and podcast-only full episodes are untouched, so coverage for podcast-heavy shows (e.g. Piers Morgan, The Lincoln Project) is unchanged. The kept YouTube copy carries the real subscriber reach and supports deep-link mention timestamps.

v0.32.2 · 2026-06-21

Fixed

• Podcast ingest collapsed network-distributed shows onto a single episode row. Both ingest paths (scripts/ingest.ts, src/lib/pipeline.ts) built an episode's source_url from episode_url/episode_permalink first. For shows distributed via a network (SiriusXM, Ark Media, MSW Media, etc.) PodScan returns a generic homepage there (e.g. https://www.siriusxm.com/), identical for every episode, so the (channel_id, source_url) upsert overwrote one row on each run instead of appending. Affected shows showed exactly one episode, perpetually replaced with the latest. Now both paths use the existing episodeSourceUrl() helper, which prefers the unique per-episode audio URL (then guid). Three shows were affected (The Megyn Kelly Show, Call Me Back with Dan Senor, The Daily Beans); their stuck rows were deleted and the shows re-ingested via the new npm run reingest:podcast recovery script. Pre-fix history beyond what PodScan still returns is unrecoverable.

v0.32.1 · 2026-06-20

Docs

• Synced the public MCP guide (/mcp) with v0.32.0: added the new get_issue_breakdown tool, noted the YouTube view-count fields on list_channels and get_channel_detail (typical per-video views, views-per-subscriber, runaway over/under-performers), and corrected the tool count (nine → ten).

v0.32.0 · 2026-06-20

Added

• Per-video view-count collection (Phase 0). New episode_metrics table banks each active YouTube episode's view count as a daily snapshot for its first 45 days, capturing the view-growth curve. This is the dataset we need to later decide whether (and how) to fold realized per-video views into reach weighting, instead of relying solely on the blunt channel-level subscriber proxy (a dormant large channel is over-weighted, a viral small one under-weighted). Three collectors:
  • getRecentUploads now rides statistics along on the videos.list call it already makes, so the ingest stage snapshots each new YT episode's view count for free (no extra API calls).
  • A new metrics cron stage (10:20 UTC daily, own 300s budget) snapshots every active YT episode inside the 45-day horizon, once per UTC day. Quota is trivial (~horizon/50 batched videos.list calls).
  • npm run backfill:views records a one-time current-cumulative baseline for all existing YT episodes (~121 batched calls for ~6k episodes).
• Decoupled and Index-safe by design. Nothing in src/lib/aggregate.ts reads episode_metrics; the reach algorithm and the published Soapbox Index are untouched. The table is a pure write-only producer (RLS-on/no-policies, service-role only). YouTube-only - podcasts expose no reliable per-episode metric. Forward-only: age_hours is stored on every snapshot so heterogeneous-age backfill readings stay alignable on a curve.
• View-count transparency in the MCP. Two read-only DB views (episode_view_latest, channel_view_stats) back new MCP fields/tools: list_channels and get_channel_detail now expose, for YouTube channels, typical_views (median per-video views over mature uploads aged 14-90 days) and views_per_sub (the divergence between actual per-video reach and subscriber count), and get_channel_detail adds each channel's runaway over/under-performing videos (performance relative to the channel's own norm, mature videos only to avoid penalizing fresh uploads). All null for podcasts and thin channels. get_methodology discloses these are transparency-only and NOT used in the Index weighting (which still weights by subscriber reach).
• get_issue_breakdown MCP tool. Exposes the home-page "who is moving this issue" breakdown: per-show signed contribution to an issue's lean over the window, each with a representative supporting quote (receipt) and source link, siblings collapsed. Pairs with get_movers (which issues moved) to explain a swing with receipts.

v0.31.1 · 2026-06-20

Fixed

• Intermittent "Couldn't load …" errors on the lazy-load receipt panels (the channel-page per-issue mentions, the episode receipts, the /log episode table) that cleared on a hard refresh. Root cause: a transient server-to-Supabase blip made the read route 500, and since nothing retried, a single blip stranded the panel until the user reloaded. Confirmed not a query bug - the failing query is cheap and fully indexed (EXPLAIN: ~1ms). Fix is at the one choke point every Supabase request flows through (db.ts noStoreFetch): idempotent reads (GET/HEAD) now retry on a network throw or transient status (408/429/500/502/503/504) with a short backoff (150ms, 400ms). Writes never retry (POST/PATCH/DELETE) - the classifications insert is non-idempotent, so retrying a write whose response was merely lost could double-insert. Gating is unit-tested (db.test.ts).
• Added error logging to the channel-issue and episode mention routes so any residual post-retry failure surfaces in runtime logs with its real message (previously the 500 body was returned but never logged).

v0.31.0 · 2026-06-20

Added

• Mention timestamps: jump straight to the moment in the video. Each classified mention now carries a start_ts (whole seconds into the episode), and the receipts UI renders a "▶ mm:ss" deep link that opens the YouTube video at that moment (&t=<seconds>s) - in both the activity-log episode receipts (EpisodeMentions) and the channel-page per-issue receipts (ChannelIssueBreakdown). Requested by a beta user: makes it far easier to go watch the source.
• Timestamps come from the transcript, not the model. Supadata native captions are now kept as timestamped segments (transcripts.segments jsonb; the YouTube fetch drops text=true so the same 1-credit call returns { text, offset } chunks). At classify time we match each mention's supporting_quote back to a start time deterministically (src/lib/transcript-timing.ts, unit-tested). The model never emits a timestamp, and there is no extra classify cost.
• MCP search_mentions exposes start_ts + timestamp_url. The data product now returns the quote's start second and a ready-made deep link (falls back to the plain episode link for podcasts and unlocated quotes). Documented on the public /mcp page and in the tool description.

Notes

• Forward-only: new YouTube episodes get timestamps as they classify; no backfill of historical mentions. Empirically ~85% of YouTube mentions locate (the model's quote start is reliable; its tail drifts, so we anchor on a short normalized prefix). A miss renders the mention as before, just without a timestamp. Podcasts deferred: PodScan word-level timestamps are inconsistently populated and podcasts lack a universal timestamp deep-link.

v0.30.2 · 2026-06-19

Added

• Channel names in the "who's driving it" breakdown link to the channel page. Each show's name (bar label and quote receipt) now links to /channels/[id] (the lead row for a canonical show). channelId is carried through ShowContribution; guarded so it no-ops on snapshots written before the field.

Fixed

• Corrected stale "independent-only" copy that conflicted with the live three-needle design (the headline needle blends both cohorts; independent and legacy each get a sub-needle). Updated the MCP tool descriptions (get_index, search_mentions, methodology field), the cohort.ts and fetchScoreRows comments, and the CohortBadge note. The methodology page and home copy were already correct.

v0.30.1 · 2026-06-19

Changed

• Every "Biggest movers" row is now expandable to its "who's driving it" breakdown. v0.30.0 only showed the breakdown for the single top mover, so the motivating example (public-health) vanished from the home page once it fell to #3. Each mover row is now a shadcn Collapsible that expands in place to the per-show contribution chart, with a link out to the full issue page. All movers' breakdowns are precomputed into the home snapshot (HomeSnapshot.moverBreakdowns) so rows expand with no live query; the page falls back to a live compute when the snapshot predates the field.

v0.30.0 · 2026-06-19

Changed

• Reach weighting moves from log10 to sqrt, system-wide. The Index is an audience-impact measure, but log10(reach) turned a 95x audience gap into a ~1.4x weight gap (one notch of intensity outweighed it). reachFactor is now 7 x sqrt(reach / 1e7) (anchored so 10M reach still maps to 7, keeping displayed weights and the published scale familiar): a 100x larger audience now carries ~10x the weight, so a mega-channel genuinely moves the needle without a single show running away with it. Single source of truth in src/lib/scoring.ts; weightedLean, the emerging-board weight, and /admin/homelab all inherit it. Measured impact on the live Index is moderate (overall ~0.2, most issues < 0.7).
• Dual-platform shows are weighted by their YouTube audience. For any show tracked on both YouTube and as a podcast, the podcast row's reach is synced to the measured, daily-refreshed YouTube subscriber count rather than an editorial podcast estimate (e.g. Joe Rogan: 14.5M estimate -> 21M subscribers). New scripts/sync-reach.ts (audit by default, --apply to write). This matters more under sqrt, which leans harder on reach accuracy.
• Methodology page updated to describe sqrt weighting honestly, including that reach-estimate error now matters more (a 2x error is ~1.4x weight, not ~5%), which is why measured YouTube counts are now preferred.

Added

• "Who's driving this issue" breakdown. A new diverging, reach-weighted contribution chart (IssueMovementBreakdown) decomposes an issue's weekly lean into each show's signed pull - the per-show fractal of the Index contribution chart. Shows are canonicalized (a show on both YouTube and podcast collapses to one bar), each carries both its editorial L/M/R badge and its stance-colored bar (they can diverge - a left show can post a right-coded bar), with the audience reach and a representative quote + source link. Renders for the top mover on the home page and on every issue page (7-day window). Shared src/lib/canonical-show.ts helper backs both the chart and reach-favoring.

v0.29.2 · 2026-06-18

Fixed

• OG social card crate now renders on production. The v0.29.1 attempt (read the crate from public/ via fs) still failed on Vercel, where the serverless function does not reliably bundle public/ files for fs reads. The card now sources both the crate and the wordmark by fetching the deployed CDN assets and inlining them as data URIs (which Satori renders reliably), removing the fs-based loaders entirely.

v0.29.1 · 2026-06-18

Changed

• Higher-fidelity crate everywhere. Swapped the wooden-crate mark to the polished source render across the header, app icon (src/app/icon.png), MCP icon, the OG card, and the brand download assets (logo PNGs + 256px icon). public/favicon.ico is left as-is (indistinguishable from the new crate at 16px without an ICO toolchain locally).

Fixed

• OG social card was missing the crate on production. It read the crate from src/app/icon.png, which the Vercel serverless function cannot fs-read (the wordmark, read from public/, rendered fine, which is what surfaced it). The card now reads the crate from public/, so it renders on the live card again.

v0.29.0 · 2026-06-18

Changed

• The wordmark now renders in its real typeface, Protest Strike, everywhere. The "soapbox" logotype was previously set in Geist Black on the live site (a mismatch with the actual brand). It is now centralized in a single <Wordmark> component (self-hosted Protest Strike, the two-color split soap #C8202F / box #114A8A) used by the header, the account and admin login screens, the welcome screen, and the brand page, so the font can no longer drift. The OG social card embeds the pixel-identical wordmark PNG (avoids forcing a display font through Satori). The brand page typography section now documents the two-typeface system: Protest Strike for the wordmark, Geist for everything else.

Added

• Needle + wordmark assets on /brand. The signature Index needle has its own section (live gauge) and a downloadable self-contained soapbox-needle.svg (with a note on rotating it to any value). The wordmark ships as real assets rather than something to retype: an outlined vector soapbox-wordmark.svg (glyphs as paths, no font dependency) and a transparent soapbox-wordmark.png (2000px).

v0.28.0 · 2026-06-18

Added

• Brand guidelines page (/brand): a public, self-serve brand kit so the logo, color, type, and voice stay consistent everywhere Soapbox appears (starting with the social channels). Includes downloadable transparent-PNG logo assets (1024 / 512 / 256px plus the favicon, served from public/brand/), the two color systems kept deliberately separate (the wordmark colors Soapbox Red #C8202F / Soapbox Blue #114A8A versus the semantic data palette where red is always right and blue always left), click-to-copy color swatches, a Geist type specimen, and the voice rules (name usage, tagline, the no-em-dash rule, color-meaning lock). Linked from the footer.

v0.27.0 · 2026-06-17

Added

• Consolidated channel-expansion CLI (npm run channels -- <subcommand>, scripts/channels.ts) toward a 250-show panel. One repeatable loop replaces the fragmented one-off discovery/seed scripts: discover (YouTube featured-channels + iTunes search + an --editorial file.json curated-list path, persisted to the new channel_candidates table, deduped centrally vs the live panel), vet (quality predicates), relevance (a cheap batched Haiku screen that demotes off-topic / foreign-language candidates, biased to keep when uncertain), review (ranked CSV + table for manual sign-off), approve, promote (onboards only approved candidates; dry-run by default), plus audit and prune for panel cleanup. Approval is manual: nothing reaches the panel without a human flipping a candidate to approved.
• channel_candidates table (migration 20260617120000): durable, re-runnable candidate store with a status lifecycle and central dedup key.
• Podcast onboarding parity: addPodcastChannel() in src/lib/channels.ts (the podcast counterpart to addYouTubeChannel), plus addYouTubeChannelById() for id-based onboarding from featured-channel discovery. PodScan field normalizers consolidated into src/lib/podscan.ts.
• Pure, unit-tested libs: src/lib/channel-dedup.ts (one name matcher, replacing four scattered normalizers) and src/lib/channel-vet.ts (reach/recency/lean predicates), each with a Vitest suite.
• Curation helpers: siblings (excludes candidates already in the panel on the other platform, plus same-platform duplicates the matcher missed), recover (re-resolves stale/misresolved podcast feeds to their live feed via the freshness-anchored resolver, pruning any with no live feed), review --describe (batched site-voice one-line descriptions to speed manual review), and a --platform review filter.

Changed

• Reach floor 300K -> 200K, YouTube only. Podcast reach is editorial and is never floor-gated. SUB_FLOOR now lives in channel-vet.ts as the single source of truth.
• Duration floor 180s -> 126s (MIN_DURATION_SEC) to admit curated short-form (e.g. NowThis Impact). Methodology + admin copy updated.
• INGEST_PER_CHANNEL 3 -> 2 (cost lever): keeps a ~250-show panel within the ~$1k/mo budget before the expansion lands.
• Retired the folded discover-channels / discover-podcasts scripts (and their npm aliases) in favor of channels discover.

Fixed

• Channel-dedup precision: nameMatches matched on a shared longest token, over-matching channels that share only a generic first name or word ("David Frum" vs "David Pakman", "Glenn Beck" vs "Glenn Greenwald", "Lincoln Project" vs "Chris Cuomo Project"). Replaced with full-substring containment (shorter name fully inside the longer, >= 5 chars and multi-word or >= 7 chars). The siblings sweep also now catches same-platform duplicates the old matcher let slip. Locked with tests.
• Podcast feed resolution: addPodcastChannel now freshness-anchors - searches query variants (normalizing curly apostrophes that broke search), keeps only title-matching feeds, picks the newest-episode one, and refuses any feed whose freshest episode is > 120 days old. The prior "first search hit" onboarded dead/abandoned feeds (2018-era Anderson Cooper, empty feeds).
• Podcast episode filter: addPodcastChannel no longer drops episodes whose PodScan duration is missing/0 (podcasts are long-form; the duration floor now applies only when a duration is known), so feeds without duration metadata still backfill.

v0.26.0 · 2026-06-16

Added

• homelab2: a staging redesign of the home page at /admin/homelab2 (admin Basic Auth; the public home is untouched). A scrolling, panel-based "enterprise analytics dashboard with a realtime feed", built as the experiment ground before promoting pieces to the live home. Progressive disclosure: the needle anchor and public hooks lead, analyst/B2B depth follows.
  • Hero: retains the master Index needle + the two cohort sub-needles, but gives EACH needle its own labeled trend (fixing the old ambiguity where one sparkline sat under the sub-needles yet plotted the master), plus an interactive index chart with a 7/14/30-day horizon toggle and the cohort lines overlaid.
  • Reworks the three inversions: "why the index is here" now leads with mention volume (bold number per diverging bar, chart-bar-negative style); "top issues" is volume-led (ranked horizontal bars, lean as color).
  • Elevates the signature slice: "Two Americas" (independent-vs-legacy trend
    • per-issue divergence), and ties in the v0.25.0 emerging board ("what's breaking" with favorability + cohort coverage).
  • Plus reworked depth panels (ownership quadrant, issue heat grid, channel momentum, cross-talk) and a clean scale strip + channel landscape.
  • Charts standardize on the shadcn ChartContainer system; a Reveal wrapper (IntersectionObserver) mounts panels on scroll so charts animate into view.
• Cut from the old 14-card lab: polarization strip (exposed the bimodal calibration weakness), lit fuses, the standalone receipts tile, audio-vs-video.

Notes

• Staging queries live, but leans on the precomputed home snapshot (readHomeSnapshot) for the master/issues/movers/breakdown data and reuses getHomelabData's twoConv for cohort trends - so it runs ~2 heavy pulls, not ~6. Promoting any panel to the public home is a separate step (its aggregates go into writeHomeSnapshot).

v0.25.0 · 2026-06-16

Added

• Favorability + cohort-coverage signal on the /emerging board. Emerging topics now carry a read on how the conversation is landing, not just mention counts. Two deliberately separate axes:
  • Favorability (the credibility anchor): a per-mention sentiment score toward the topic itself (-5 critical .. +5 favorable + intensity), Haiku, aggregated reach x intensity weighted via the existing weightedLean(). A distinct axis from the L/R needle (events have no left/right poles), so it gets its own neutral-to-emerald gauge, never the red/blue needle. Each row shows a compact "Reaction" gauge; the expanded panel adds a favorability summary and per-receipt favorability chips.
  • Cohort coverage (the honest "who's amplifying" overlay): L/M/R share of voice, normalized to mentions-per-active-channel so it can't be skewed by panel composition, plus a cohort-over-time trend. Uses the blue/red lean palette on purpose (this IS the lean axis).
• New discovery_topic_scores table (keyed on the stable discovery_topics.id, not the ephemeral candidate id), a score-emerging cron stage (Haiku, gated to the top-N board candidates, bounded + idempotent like score), and the new prompt surfaced read-only at /admin/prompts.

Notes

• Favorability is a new axis the L/R gold set does not cover. v1 ships validated by a manual spot-check (scripts/check-emerging-favorability.ts); a dedicated favorability gold set gates it before it becomes load-bearing (i.e. before the MCP server exposes emerging, deliberately deferred to mature first).

v0.24.2 · 2026-06-15

Changed

• Serve pipeline-scale stats and the social/OG card from the home snapshot instead of recomputing them per request (perf). <SystemStats> (on /log) now reads the cron-precomputed snapshot rather than running 5 count queries plus a full per-episode duration scan on every render, falling back to a live read only if the snapshot predates the field. The dynamic Open Graph image (opengraph-image.tsx) reads the snapshot's dashboard data instead of the live ~10-15s deep-join aggregate, so the shared-link preview card generates fast on a cache miss. HomeSnapshot gains a systemStats field, written by the score cron / refresh:snapshot.

v0.24.1 · 2026-06-15

Changed

• Generalized scripts/backfill-issues.ts so any newly-promoted issue can be backfilled without editing the script. New env overrides: BACKFILL_SLUGS (which slugs to insert), BACKFILL_KEYWORDS (a cheap transcript keyword pre-filter so only plausibly-relevant episodes get the expensive classify call), and BACKFILL_TARGET_TABLE (a precomputed episode_id allowlist, used when the inline keyword scan exceeds the data-API statement timeout on the full transcript table). Used to backfill the new "Gerrymandering & redistricting" issue (Elections & Democracy) over a 30-day, keyword-targeted window.

v0.24.0 · 2026-06-15

Changed

• Extracted the pure scoring math into testable modules (src/lib/scoring.ts and src/lib/emerging-score.ts) - no DB, no env, so the product's headline numbers can be unit-tested in isolation. aggregate.ts and discovery.ts now import from them. In the process, consolidated three pieces of duplication: the Soapbox Index scale (clamp(lean * 2, -10, 10)) was inlined 14 times and is now one toIndexScale(); the Biggest-Movers eligibility/ranking rules became moverHasEnoughMentions / moverIsInteresting / moverScore; and reachFactor is now single-source (shared by the Index and the emerging score). Verified byte-for-byte behavior on the home Index, movers, and drilldowns.

Added

• Tests for the math that matters (scoring.test.ts, emerging-score.test.ts): the reach weighting -> weighted lean -> -10..+10 Index scale, the movers OR-rule and max(|leanΔ|/2, |log2(ratio)|) ranking, the recency half-life decay, the breaking-velocity rule, and the momentum-blended emerging score. 30 tests total across the three suites, all green.
• CI gate (.github/workflows/ci.yml): typecheck + tests + build on every push to main and on PRs - making the checks non-optional (Vercel builds on push but doesn't run the tests). ESLint isn't configured, so no lint step yet.

v0.23.1 · 2026-06-15

Added

• Vitest test framework (lite) - the first automated tests in the repo. Node environment, tests colocated as *.test.ts, @/ alias mirroring tsconfig, no jsdom/testing-library yet (add when we want component tests). Scripts: npm test (one-shot) and npm run test:watch. Chosen over Jest for native ESM/TS + speed; the API is Jest-compatible so tests stay portable. First suite locks the L/R lean convention (src/lib/lean.test.ts, 12 tests): the red=Right/blue=Left palette, L+/R+ labels, the 0.05 aggregate-lean neutral band, and the sentiment-chip flips-at-exactly-0 rule. Foundation to expand to the aggregation
  • scoring math next.

v0.23.0 · 2026-06-15

Changed

• Foundation: one source of truth for the L/R lean language (P1). The site's signature visual - the L+/R+ sentiment chip, the L/R/M lean chip, the lean text-color, and the intensity meter - was copy-pasted across ~12 files (formatLean x6, leanColor x6, sentimentChip x2, leanBadge/inline lean maps x5, IntensityMeter x2). Consolidated into src/lib/lean.ts (formatLean, leanColor, sentimentChipStyle, leanChipStyle, LEAN_NEUTRAL_BAND) + shared src/components/Lean.tsx (SentimentChip, LeanChip, IntensityMeter). Every call site now imports the single definition, so the red(=Right)/blue(=Left) palette and the lean threshold can never drift. No rendered output changed (verified across /channels, /issues, /log, /emerging, channel + issue pages).
• One app-wide TooltipProvider in the root layout; removed the four redundant local providers (channels page, EpisodeDataTable, EmergingIssuesTable, FreshnessBadge). Closes a latent gap where CohortBadge's tooltip needed a local provider ancestor - the global now covers every surface.
• Remaining raw <button>s → shadcn Button (P2): the /welcome secondary link, and EmergingIssuesTable's sort-header + row expander (matching the pattern EpisodeDataTable already used). The one intentional exception is ChannelIssueBreakdown's full-width expandable issue row - a stacked-content layout trigger, not a Button-shaped control - left as a semantic <button>.

v0.22.1 · 2026-06-14

Fixed

• /emerging: the "flame" in the intro copy now renders the actual flame icon (it was the literal word, and ran into the next word with no space).
• /emerging tooltips are now the site shadcn Tooltip, not the native browser title (breaking badge, up/down movement, "Last mentioned", and the receipt lean chip), matching the project's no-native-tooltips rule.
• Footer nav no longer wraps mid-phrase ("For AI / Agents") after the longer tagline shipped earlier today: links are whitespace-nowrap and the group wraps as whole items, right-aligned.
• Full sweep of native browser title= tooltips -> shadcn. Every remaining native title hint across the site is now the styled shadcn tooltip via a new shared InfoTip helper and one app-wide TooltipProvider in the root layout: intensity meters (EpisodeMentions, ChannelIssueBreakdown), the cohort stacked bar (PanelBalance), channel platform links + the issue-activity bars, truncated cells in the /log table (EpisodeDataTable), the footer links, and the admin PipelineHealth / homelab cards. The site is now free of native browser tooltips.
• Copy/metadata that still said the board refreshes "daily" updated to "through the day" (it moved to a 3-hourly rebuild in v0.22.0).

v0.22.0 · 2026-06-14

Changed

• /emerging now ranks by momentum, not just decayed volume - the board was surfacing big-but-over stories above genuinely emerging ones. Ranking by decayed reach-volume alone kept concluded races and fading stories on top (a decided Senate primary at #1 on flat velocity; a fading topic at #2 on 0.25x week-over-week), while a story breaking today sat at #12. The sort key is now decayed weight x smoothed week-over-week momentum ((recent7 + k)/(prior7 + k), k=3 Laplace smoothing so a tiny topic with a near-empty prior week can't get an explosive ratio; magnitude stays the anchor). Validated on the live set: a breaking trial goes #4 -> #1, a "happening today" event #12 -> #3, and fading high-volume topics sink. Tunable via EMERGING_MOMENTUM_SMOOTHING.
• Dropped the opaque "Weight" column from the board. It was an internal score that, once rank stopped tracking it 1:1, only invited "why is #1 not the highest number?" Rank (the momentum-blended order) plus mentions / episodes / channels and the breaking badge carry the story; the table now defaults to rank order and columns stay click-sortable.
• The /emerging board now refreshes every 3h instead of once a day (discover cron 0 11 * * * -> 50 */3 * * *, just after each 2-hourly classify harvest). The board's candidate set was rebuilt only at 11:00 UTC, so off-taxonomy topics harvested the rest of the day (often ~180) sat unclustered and off-board until the next morning - the freshest YouTube signal was a day late. Re-clustering is one cheap Haiku call (~$0.03/run, ~$0.25/day). Note: composition can jitter slightly between intraday rebuilds; rank movement is unaffected (day-over-day).

v0.21.0 · 2026-06-14

Added

• "Breaking" signal on /emerging. A flame badge marks issues whose attention roughly doubled week-over-week (mentions aired in the last 7 days vs the prior 7), or that appeared fresh, off a non-trivial base. This surfaces what's accelerating now, distinct from the rank-movement arrows (position change) and the weight/rank (accumulated size) - a genuinely breaking story can be small in total volume but spiking. Thresholds are tunable constants (BREAKING_MIN_RECENT = 8, BREAKING_RATIO = 2) with a recent-count floor so a 3-on-1-mention blip doesn't flag; validated on the live set it marks the accelerating handful (e.g. a fresh trial 35-vs-9, a White House event 12-vs-4) and skips the noise. Computed per cohort in computeBoardRanks, mirroring the home Biggest Movers week-over-week volume idiom. Badge is amber/flame, deliberately not red (red = "Right" lean site-wide).

v0.20.2 · 2026-06-14

Changed

• /emerging now suppresses dead news: an issue drops off the public board once its most recent mention (in that cohort) is more than 10 days old. The recency decay (v0.20.0) down-weighted stale topics but a high-volume one could still linger mid-board; an "emerging" board carrying a month-old outbreak or a three-week-old shooting reads as stale. The gate applies only to the public board - the /admin/discovery review queue stays comprehensive, so a human can still promote a once-big theme that has quieted. Cutoff is a tunable constant (BOARD_MAX_STALE_DAYS); validated on the live set, 10 days drops the clearly dead candidates (6 of 39) while keeping the board full. Copy updated to say so.

v0.20.1 · 2026-06-14

Fixed

• /emerging receipts now lead with the most recent quotes, not the highest-reach ones. The expand panel sorted by channel reach, so a topic that was active yesterday could show only 1-2 week-old receipts (its fresh mentions came from smaller channels and fell below the 12-row cap), which read as a stale, non-credible board. Receipts now sort most-recent-first, tie-broken by reach so the most influential quote still leads within a given day.

Added

• "Last mentioned" freshness line on each /emerging row. Because the board ranks on decayed volume, each row now shows how recently the issue was actually discussed (e.g. "Last mentioned yesterday"), making a row's currency explicit instead of implied - and a genuinely stale topic obvious at a glance.

v0.20.0 · 2026-06-14

Changed

• /emerging recency is now a continuous decay, not a binary boost - the board was ranking by stale accumulated volume. Each member topic's reach contribution now halves roughly every 7 days of episode age (reachFactor x 2^(-ageDays/7)), replacing the old "x1.5 if the episode aired in the last 7 days, else x1" step. The step barely moved anything: classify harvests off-taxonomy topics long after an episode airs, so the boost fired on only ~17% of members while member episodes ran to ~180 days old, and the board ranked by whatever had the most accumulated mentions - it looked frozen. With decay, genuine bursts surface and last month's chatter fades (validated on the live pending set: a fresh trial story climbs #10 -> #4, a stale-but-huge geopolitics topic settles #1 -> #3). The half-life was tuned against the live data, not guessed. The build-time and board-time weights now share one topicWeight() helper so they can't drift.

Added

• Rank movement on /emerging. Each issue shows how it moved since the last daily refresh: a green up arrow, a muted down arrow, "new", or a steady dash. Because the daily rebuild regenerates candidate UUIDs, movement is keyed on a normalized theme (the same token-set key used to de-fragment labels), so a continuing issue keeps its history across rebuilds and minor LLM relabeling. Ranks are snapshotted per cohort each rebuild into the new emerging_rank_history table; the live board diffs against the previous refresh (not the latest snapshot, which would read as no movement).
• Board freshness line on /emerging showing when the candidate set was last rebuilt (relative + absolute US Eastern), plus copy explaining the recency weighting and the movement column.

v0.19.0 · 2026-06-14

Changed

• New home hero + site-wide tagline: "Voice, reach, and influence in political media, measured." Replaces "Online political media, quantified" as the master positioning line. The hero headline, page <title>, OG and Twitter cards, the dynamic OG image, and the footer now all read from one source of truth (src/lib/brand.ts: SITE_TITLE, TAGLINE, TAGLINE_FOOTER, META_DESCRIPTION) so they can never drift apart again.
• Hero subtext now carries live scale counters. Reads "We've analyzed N hours of political audio from M independent and legacy shows and scored K mentions, and counting." Numbers come from the home snapshot (hours of audio and total scored mentions added to HomeSnapshot.scale, written by the score cron via getSystemStats), with a live fallback before the first cron populates the field and a number-free fallback if both paths fail.

Added

• src/lib/brand.ts - canonical brand copy constants, mirroring version.ts as a single source of truth.

v0.18.3 · 2026-06-14

Fixed

• Serve /favicon.ico (was 404). Added a real multi-size public/favicon.ico (16-256px, built from the crate mark). Beyond fixing the missing site favicon, this is the one fetch-verified difference between our MCP connector and a custom connector that does render a logo in claude.ai (Higgsfield: its registrable domain serves /favicon.ico 200; its MCP host 404s; ours 404'd everywhere) - evidence that claude.ai resolves a custom connector's icon from the registrable domain's favicon, not serverInfo.icons. Pairs with the v0.18.2 serverInfo icons.

v0.18.2 · 2026-06-14

Added

• MCP connector now advertises the Soapbox crate logo. The server's initialize response (serverInfo) now carries the MCP icons + websiteUrl metadata (SDK 1.26.0 Implementation schema), pointing at a public, unauthenticated asset (public/mcp-icon.png, the 256x256 crate mark). MCP clients that read server-advertised icons (and claude.ai's connector list, if/when it renders them for custom connectors) can show the brand mark instead of a generic globe. Note: the icon is read at connection time, so an already-added connector must be removed and re-added to pick up the new metadata.

v0.18.1 · 2026-06-12

Added

• /admin/costs now captures manual CLI spend, and reconciles against billing. Every cost-incurring CLI run (classify, score, backfills) writes a usage_log row via a new recordScriptRun helper, tagged source: cli|manual. Previously only the daily cron was logged, so terminal spend (e.g. a ~$200 backfill) was completely invisible - the page showed $134 all-time when real spend was far higher. The dashboard now splits the recurring (cron) run-rate (what the budget gauge tracks) from one-off / manual spend (shown separately so a backfill can't trip the budget alarm), and tags each run row with its source. Optional reconciliation: with an Anthropic Admin key (ANTHROPIC_ADMIN_KEY, sk-ant-admin..., org-only) the page shows actual billed cost vs our estimate via the Admin cost_report API (src/lib/anthropic-billing.ts); it degrades to a configure hint when the key is unset.

Changed

• Centralized Anthropic token pricing. New src/lib/pricing.ts (MODEL_PRICING + cache-aware estimateCostUsd) replaces the per-million token rates that were hardcoded in four places (pipeline classify + score, classify.ts, score.ts, backfill-issues.ts). Rates verified against current Anthropic pricing - no number changed, this is drift-proofing. The score CLI also gained a bounded-concurrency pool (matching prod SCORE_CONCURRENCY) so a large manual score drain finishes in minutes.

v0.18.0 · 2026-06-12

Added

• Taxonomy expansion: 5 new issues + crypto. Found by bucketing the off-taxonomy harvest (discovery_topics) by distinct channel breadth over 90 days - durable categories the event-grained /emerging board structurally can't surface (it clusters one-off events like a primary race, not standing themes). All nest under existing locked Topics (no new Topic), building out the two thinnest: Economy & Work (2 to 5 issues) and Health Care (1 to 2). New issues: Trade & tariffs, Housing & homelessness, Government spending & debt (Economy), Public health & medical establishment (the MAHA / vaccines / FDA axis, distinct from health-care coverage; Health), and Veterans & military affairs (Foreign Policy & National Security). The existing AI & tech issue was broadened to "AI, crypto & tech" to cover crypto and digital assets. Migration 20260612120000_taxonomy_expansion. 30-day history was backfilled for the five new issues (crypto is go-forward only - a widened existing issue can't be backfilled without duplicating its mentions). The Trade and Veterans definitions were tightened after a sample-validation pass to filter conflict-shipping and general-military noise, and inflation's definition dropped "housing affordability" so that signal routes to the new Housing issue.

• MCP setup guide (/connect): non-technical-first onboarding. A dedicated, step-by-step page (get access, set password, add connector, sign in, ask your first question) pulled out of the marketing-heavy /mcp page. Leads with the point-and-click Claude and ChatGPT custom-connector flow; developer configs (Claude Code, Cursor, VS Code, Claude Desktop) live in their own tab. New reusable CopyField (copy-to-clipboard) and ConnectGuide components. Step 1 surfaces the "Add promotion code" path so beta testers with a free code can self-serve to $0. The post-payment /welcome screen and /account now link straight into the guide.

Changed

• /mcp trimmed to the pitch. The long per-client config dump and the duplicate pricing block were replaced with a concise "Getting connected" section that links to the new /connect guide. Marketing (the pitch, example queries, tool list, data boundaries) and the top pricing CTA stay.

Fixed

• backfill-issues.ts no longer times out at panel scale. It loaded every transcript's full text and filtered the date window in JS, which hit a Postgres statement timeout once the panel grew (~6.7k transcripts). It now filters the window at the DB and fetches text only for the episodes it will process, and runs the classify pass through a bounded-concurrency pool (matching production's CLASSIFY_CONCURRENCY) rather than serially (~14h to ~1.5h at the current size).

v0.17.0 · 2026-06-11

Changed

• Support page redesigned for conversion (campaign-ready). Studied top political-donation landing pages and applied the patterns. Cut the four paragraphs of appeal to a punchy headline + one line + three value bullets; added an on-brand hero visual (the live, animated Soapbox Index needle - "what your support funds"); and moved the ask high (first on mobile, sticky on the right on desktop). The donation widget is tightened to ActBlue norms: a six-tier amount grid ($10-$500) with a "Popular" anchor and a higher default, a monthly nudge, a more prominent CTA, and a "Secure checkout via Stripe" trust line. L/M/R per-audience variants kept for campaign refcodes. The default/anchor amount ($50) is the single biggest A/B lever (in DonationWidget).

v0.16.1 · 2026-06-11

Changed

• Freshness badge: custom tooltip + moved next to the logo. The header "Updated {relative}" badge now uses the site's shadcn tooltip on hover (showing the absolute last-pipeline-run time in US Eastern) instead of the native browser title, and sits beside the logo rather than crowded in with the nav links.

v0.16.0 · 2026-06-11

Added

• Channels page: cohort tabs + search. /channels now has All / Independent / Legacy tabs (matching /emerging) plus a name search box, over the existing Left / Middle / Right lean buckets. Client-side filtering of the server- provided list (new ChannelsBrowser); tab labels carry live counts.
• One site-wide freshness badge in the header. A single consistent "Updated {relative}" badge (with the absolute last-pipeline-run time in US Eastern on hover, and a subtle live dot) on every page, driven by the most recent cron run (getDataFreshness).

Changed

• /channels stat: "Largest single show" -> "Episodes ingested · last 24h" (getPanelStats now counts episodes created in the last 24 hours).
• Removed the now-redundant per-page freshness treatments (home TrustStrip "Last updated", /log SystemStats "Latest data", /emerging board "Updated {date}") in favor of the single header badge.

v0.15.3 · 2026-06-11

Changed

• Needle animation: a touch more pronounced, and now on the cohort needles. Bumped the spring overshoot (~7% -> ~16%) so the settle reads more clearly, and extended the animation to the two Independent / Legacy sub-needles. The three needles now settle in a slight stagger (main, then +140ms, then +260ms) for a subtle cascade instead of firing in lockstep. Still dependency-free and reduced-motion-aware; the stagger applies only to the entrance - value-change re-settles are immediate.

v0.15.2 · 2026-06-11

Added

• Subtle settle animation on the home Index needle. The hero needle now eases into its reading like an analog meter: a dependency-free, under-damped spring (~7% overshoot, ~0.8s) sweeps it in from center on load and re-settles whenever the Index refreshes. Implemented as a tiny client island (AnimatedNeedle) that writes the SVG transform via a ref (no per-frame React re-render), so the gauge chrome and all per-issue sub-needles stay static, server-rendered SVG. Honors prefers-reduced-motion (snaps instantly); SSR / no-JS render the correct final angle. No three.js (a 3D engine for a 2D needle) or GSAP dependency - just a ~40-line requestAnimationFrame spring.

v0.15.1 · 2026-06-10

Changed

• Speaker-attribution limitation investigated and parked. Followed up the beta tester's attribution concern with two offline experiments: a 90-transcript classify battery (v1 vs a stance-in-classify v1.1, with a v1-vs-v1 noise floor and an Opus refute-judge) and a separate stance-stage prototype (v1.2) run over real production mentions to measure aggregate impact. Result: genuine opposing-position quotes are only ~8% of mentions, and correcting them barely moves the per-channel-per-issue scores the site reports - mean |delta| < 1 pt on the -10..+10 scale, no strongly-stanced channel changed its read, and the only sign-flips were near-zero legacy-news pairs (BBC / Bloomberg on Iran). A naive fix would also mislabel guests the show agrees with, risking net harm. Decision: parked. The /admin/prompts maturation backlog now carries a status tag plus the full finding.

Added (offline only - not wired into the deployed pipeline)

• Reproducible attribution-eval harnesses scripts/eval-attribution.ts and scripts/eval-stance-impact.ts, and experimental prompt modules src/modules/classify/experimental.ts (v1.1) and src/modules/classify/stance.ts (v1.2). Production still runs the v1 prompts; eval outputs are gitignored.

v0.15.0 · 2026-06-10

Added

• /admin/prompts: pipeline prompts + models, now versioned. A read-only admin page revealing the exact system prompts and models the classify and score stages run (claude-sonnet-4-6 / claude-haiku-4-5), with per-stage prompt version labels (both start at v1), max_tokens, descriptions, and the dynamic inputs. Templates are rendered from the LIVE prompt builders with placeholder inputs (classifyPromptPreview / scorePromptPreview), so the page can never drift from production. Versions live next to the builders (CLASSIFY_PROMPT_VERSION / SCORE_PROMPT_VERSION in src/modules/*) - bump on any change. Catalog in src/lib/prompts.ts.
• Maturation backlog on the same page. Surfaces known validity gaps in the current prompts as the candidates for the next versions (gold-set-gated). First entry, from a beta tester: speaker attribution - neither stage tracks who is speaking or whether the host endorses vs. rebuts a quote, so a host playing an opposing position Y to attack it currently scores as aligned with Y, misattributing the show's stance X.

v0.14.1 · 2026-06-10

Changed

• Episode/publish dates now render in a fixed editorial timezone (US Eastern), not the viewer's local zone. A viewer in Hawaii (UTC-10) saw June 9 episodes dated 06/08 because toLocaleDateString used the browser's timezone, and the 10:00 UTC daily ingest lands at their local midnight - so the log looked like it had almost nothing from "today." Dates are a fact of the record, so they're now pinned to one zone via a single source of truth (DISPLAY_TZ + formatDateET in src/lib/utils.ts), applied across the log/episode tables, the emerging + channel-issue receipts, the home "as of" label, the trust strip, the index/volume chart axes, and the OG image. (Admin/operator timestamps stay in their own frame for UTC/cron reasoning.) This also unifies a prior inconsistency where some surfaces pinned UTC and others used local.

Removed

• Trending Names (BETA) removed from the home page - superseded by the new /emerging board. Dropped the home-page section + its data read, deleted the TrendingNames component, and removed the now-pointless daily /api/cron/trending schedule. The trending lib + cron route remain in the repo (unscheduled) so the feature can be revived or wired to MCP later.

v0.14.0 · 2026-06-09

Added

• /emerging cohort tabs: All / Independent / Legacy. The emerging-issues board now splits by channel cohort, surfacing where independent and legacy outlets diverge (the platform's core thesis). The per-cohort cuts are recomputed from each candidate's member topics, not the stored all-cohort stats - so weight, rank, mentions, episodes, and channels are correct per tab, and an issue that only independents are discussing doesn't show up under Legacy. Cohorts partition cleanly (every channel is exactly one), so All = Independent + Legacy (verified: top candidate 76 + 92 = 168 topics). Computed in one pass over discovery_topics (new getEmergingBoard()), and the active tab flows into the receipts fetch (/api/emerging/[id]/receipts?cohort=) so an expanded row shows that cohort's quotes. Tab labels carry per-cohort counts; new EmergingBoard component on shadcn Tabs.

v0.13.1 · 2026-06-09

Changed

• /emerging polish from live review. (1) Renamed the page from "Emerging topics" to "Emerging issues" throughout - "Topics" is the reserved category level in the taxonomy; these candidates are issue-level. (2) Added a rank column (1, 2, 3 ... by weight, the canonical trending order). (3) Added an "Updated {date}" badge above the table (newest candidate created_at = last daily refresh). (4) Redesigned the expanded receipts to match the site's other receipts panels (EpisodeMentions / channel-issue): a grid with a colored chip + quote + episode link, instead of the flat gray list. Since off-taxonomy mentions aren't sentiment-scored, the chip colors by the source channel's lean (blue L / red R / gray M) - honest, and it restores the visual variety. New lean field on /api/emerging/[id]/receipts.

v0.13.0 · 2026-06-09

Added

• Public "Emerging topics" board (/emerging, linked from top nav). Surfaces the discovery pipeline's findings publicly: auto-detected, machine-clustered topics the shows are discussing that aren't in the issue taxonomy yet, in a sortable table (weight / mentions / episodes / channels). Each row expands to reveal real episode receipts - the exact supporting quotes with channel, episode title, date, and outbound link - lazy-loaded from a new route /api/emerging/[id]/receipts (highest-reach episode first, one per episode). Shows only pending candidates; promotion into a tracked Soapbox issue stays human-gated in /admin/discovery. Copy frames them honestly as raw, daily, not-hand-curated signals. New getEmergingIssues() + EmergingIssuesTable, mirroring the existing expand-for-receipts design (EpisodeMentions).

Changed

• Discovery cron now runs daily (0 11 * * *, was weekly Mon) so the public /emerging board refreshes every day. Cost is negligible (~$0.03/run Haiku).

Added

• Channel page: per-issue mentions expand inline instead of linking away. Each issue row on a channel page (/channels/[id]) was a link to the system-wide issue page. It now expands in place to reveal the exact scored mentions for that channel on that issue - the same supporting-quote + sentiment + intensity receipts already shown per episode, plus the source episode (title, date, outbound link), ordered strongest-first (|sentiment| x intensity). Lazy-loaded on expand from a new route /api/channels/[id]/issues/[slug]/mentions, scoped to the same last-30-days window as the mention count on the row (counts match: the route starts from sentiment_scores with the same channel/issue/window filters and no cohort filter, mirroring getChannelDrillDown). The system-wide issue page is still reachable via a link inside the expanded panel. Mirrors the existing EpisodeMentions expansion design (ChannelIssueBreakdown component).

v0.11.1 · 2026-06-09

Fixed

• Issue discovery silently regressed to zero candidates as topic volume grew. It worked at launch (42 candidates from ~106 topics, v0.6.39), but the Haiku clustering pass ran with max_tokens: 4096. Once the off-taxonomy pile grew to a full 250-label batch, the JSON output (dominated by member_indices arrays) truncated mid-array; extractJson returned null and buildDiscoveryCandidates returned {candidatesCreated: 0} with no error. Because the rebuild deletes pending candidates first, the first post- regression run wiped the originals and every run since produced 0 - so the queue sat empty despite 14k+ harvested topics. Raised max_tokens to 16000 (~4x the truncation point, within Haiku's safe non-streaming range) and added hard guards: clustering now throws on stop_reason === "max_tokens", a null parse, or a non-array result, so this can never fail silently again. A single npm run discover now yields 82 candidates from 14,342 topics.

Changed

• Discovery clustering de-fragments near-duplicate labels before ranking. Topics were grouped by exact normalized string, so phrasing variants of one theme ("la mayoral race spencer pratt" vs "spencer pratt la mayoral race") counted as separate candidates and split a hot theme's frequency across the top-250 cut. Grouping now uses a normalized token-set key (lowercase, strip punctuation, drop stopwords, de-dupe + sort tokens) so word-order and punctuation variants collapse; the most common surface form is shown to the model. The LLM still does the semantic merge - this just stops obvious duplicates from diluting the ranking.
• Review queue is capped to the most significant candidates. A full rebuild was producing ~82 candidates - too many to review. Clustering still runs over the full top-250 label set, but only multi-channel themes (MIN_CANDIDATE_CHANNELS = 3, drops single-show obsessions) ranked by weight are persisted, capped at MAX_PENDING_CANDIDATES = 40. Topics in dropped themes stay unclustered and are reconsidered next run, so nothing is buried.
• /admin/discovery surfaces the refresh result. The "Refresh candidates" button now shows a persistent status line ("Clustered N topics into M candidates", a no-op note, or the error message) instead of silently refreshing the (previously always-empty) list.

v0.11.0 · 2026-06-08

Changed

• Visual consistency pass + design-token migration. The whole site (public and admin) now reads color from one semantic token palette instead of ~620 scattered literal Tailwind grays.
  • Token foundation: globals.css stock shadcn tokens were retuned from pure-neutral to the exact Tailwind v3 gray ramp (HSL), and an extended ink-* text ramp (strong/body/muted/faint/faintest) + a subtle surface were added (wired in tailwind.config.ts) to preserve the page's 6-step text hierarchy that stock shadcn (2 text tokens) would flatten. Every token value is pinned to the literal gray it replaces, so light mode is pixel-identical.
  • Gray to token migration: ~620 text/bg/border/divide-gray-* utilities mapped to text-foreground / text-ink-* / text-muted-foreground / bg-card / bg-muted / bg-subtle / border-border / border-input etc. via one centralized mapping table. Dark CTAs (bg-gray-900) now use bg-primary.
  • Dark-mode scaffold: an inverted .dark token block was added (dormant - not wired to a toggle yet), so theming is now a one-knob override.
  • Card adoption: 29 hand-rolled card containers across 18 files now use the shadcn <Card>, giving cards a uniform subtle shadow-sm. Clickable <a> cards and the login <form> card keep their elements (Card cannot be an anchor/form) but match the styling.
  • Radius + focus unified: stray rounded-sm/rounded-xl collapsed into the rounded-md/rounded-lg scale; per-input focus:ring-gray-300 overrides removed so every control uses the shadcn focus-visible:ring-ring treatment.

Note

• Light mode is intended to be pixel-identical except for the one deliberate polish: previously-flat cards now carry a uniform subtle shadow. The chart --chart-* lean palette and the custom SVG gauges are unchanged. The OG image (opengraph-image.tsx, inline-hex via Satori) was deliberately excluded from tokenization since Satori does not resolve CSS variables.

v0.10.0 · 2026-06-08

Changed

• Site is now 100% shadcn/ui, on Recharts v3. Upgraded Recharts 2.15.4 to 3.8.1 and patched src/components/ui/chart.tsx for v3's tighter Tooltip/Legend prop types. Converted every hand-built UI surface to shadcn primitives:
  • Charts standardized on the shadcn chart pattern: IndexAreaChart and VolumeAreaChart now use ChartTooltipContent (no more hand-rolled tooltip components) and reference colors via --chart-* tokens. New shared src/components/ui/sparkline.tsx (Recharts LineChart + ChartContainer) replaces the bespoke inline <svg> sparklines in TrendingNames and IssuePreview.
  • Tables moved to the shadcn Table family: BiggestMovers, EpisodeDataTable (TanStack logic unchanged, only the markup shell), PipelineHealth, the /mcp tools table, and the inline-bar data lists IssueContributionsChart and IssueActivityByTopic (bars kept inside cells, full-row links preserved via a stretched-link anchor).
  • Buttons / inputs / selects / textareas converted to shadcn Button, Input, Select, Textarea, and Label across SubscribeButton, DonationWidget, login-form, AdminNav, the admin login / add-channel / discovery forms, and the gold-set label form.
• New chart color tokens in globals.css (--chart-right/-left/ -neutral/-index/-muted plus numbered aliases): one home for the house lean palette (red = right, blue = left, gray = neutral), referenced via ChartConfig.color = "var(--chart-*)" instead of scattered hex literals.
• CLAUDE.md: "shadcn first" rule. Before building any UI feature, check the shadcn library (via the shadcn MCP) first; never hand-roll a primitive that shadcn provides. The only sanctioned hand-built UI is the custom SVG gauges (SoapboxNeedle, SubNeedle) and 1-D position markers, which have no Recharts/shadcn equivalent.

Note

• All conversions are behavior-preserving: existing Tailwind classes, the house palette, copy, links, and data flow are unchanged; primitives carry the original styling via className. /admin/homelab (a decision-mock scratch page) was intentionally left on raw Recharts.

v0.9.2 · 2026-06-08

Changed

• Banned em dashes project-wide. Added a hard style rule to CLAUDE.md (never use the long U+2014 dash anywhere; use commas, colons, parentheses, or a spaced hyphen) and scrubbed every existing em dash from the site copy, components, lib, comments, CLAUDE.md, and this changelog. Mechanical replacement: spaced em dash to " - ", attached to "-". (Minus signs in scores like -5 are a different character and were left alone.)

v0.9.1 · 2026-06-08

Changed

• Support page copy rewritten in a warmer, Wikipedia-style appeal - reflective and second-person ("think back over this past year of politics… what is political media actually saying?"), with the "$5, $20, $50, or whatever feels right" ask, an independence/threat framing, and a closing thank-you. L/M/R headlines kept as the targeted hook over a shared appeal body; widget made sticky. Footer link now reads "Support our work."

v0.9.0 · 2026-06-08

Added

• Support donations (workstream 3). Wikipedia-style "fund the referee" contributions to Soapbox - pay-what-you-want one-time (Stripe payment mode
  • submit_type: donate) or monthly (ad-hoc recurring price), no account required (/api/stripe/donate, DonationWidget). New /support page plus L/M/R persuasion variants (/support/left|middle|right) reframing the ask for donors who already give to candidates across the map, sharing one SupportLanding + copy map. /support/thanks confirmation. Clearly labeled not a political contribution and not tax-deductible (for-profit LLC) - goes to Soapbox/Breakfastball operations, never to candidates. Footer now links Support.

v0.8.6 · 2026-06-08

Changed

• /mcp page: subscribe in, beta-access out. Replaced the "request a beta access key" CTA (and the bottom "Get a key" section) with the $300/mo pay-first Subscribe card (new reusable SubscribeButton). Connect instructions updated to the OAuth sign-in path - no YOUR_ACCESS_KEY headers; just the server URL + browser login. claude.ai/ChatGPT note updated from "OAuth on roadmap" to "supported."

v0.8.5 · 2026-06-08

Changed

• Subscription flow switched to pay-first (Model B). No account needed to subscribe - the friction of "sign up before you can pay" is gone. Flow: /pricing → Subscribe (anonymous) → Stripe Checkout collects email + card → webhook checkout.session.completed provisions a Supabase user from the email (provisionUserByEmail - finds existing or creates + sends a Supabase invite/set-password email), links the subscription, and grants entitlement → new /welcome page handles the set-password landing and shows the connect-your-agent instructions. Stripe's hosted checkout can't create a login in our system, so provisioning happens server-side in the webhook keyed by email. Removed the account-first checkout (no more login gate on /pricing); getOrCreateCustomer/syncSubscription replaced by provisionUserByEmail + linkSubscription + syncSubscriptionByCustomer.
  • Supabase config needed: add https://www.soapbox.media/welcome to the Auth redirect-URL allowlist (invite link lands there). Invite emails use Supabase Auth's email - fine for testing; custom SMTP for scale.

v0.8.4 · 2026-06-08

Fixed

• Checkout now fails gracefully instead of hanging. An unhandled Stripe error (e.g. account/mode mismatch) returned a generic 500 with no JSON, so the Subscribe button hung. Wrapped customer + session creation in a try/catch that returns the error message (502) and logs the key mode (sk_live_/sk_test_ prefix only, never the secret) for diagnosis. Also a fresh deploy to snapshot the current production Stripe key.

v0.8.3 · 2026-06-08

Fixed

• Checkout hung with "No such customer" after going live. A test-mode Stripe customer id (from local test-mode checkout) was stored in the shared subscriptions table and reused in live mode, where it doesn't exist. getOrCreateCustomer now verifies the stored customer exists in the current Stripe mode and recreates it if stale (handles test↔live and deleted customers). The stale test row was also cleared.

v0.8.2 · 2026-06-08

Fixed

• Checkout STRIPE_PRICE_ID not configured in prod - the price id was read at module load (export const PRICE_ID = process.env…), which can be inlined at build / survive build-cache and miss a runtime env var. Now read at request time via priceId(). (If it still reports unset after this, the var genuinely isn't in the Production environment scope.)

v0.8.1 · 2026-06-08

Changed

• Deploy bump to pick up the production Stripe env vars (STRIPE_SECRET_KEY, STRIPE_PRICE_ID, STRIPE_WEBHOOK_SECRET) - Vercel only injects env on a fresh deployment, and the v0.8.0 build predated them. No code change.

v0.8.0 · 2026-06-08

Added

• MCP access monetization - $300/mo Stripe subscription (workstream 2). Full subscribe → entitlement → gating chain on top of the OAuth identity:
  • /pricing (subscribe) → /api/stripe/checkout (Checkout, subscription mode, client_reference_id = Supabase user) → /api/stripe/webhook (signature-verified; checkout.session.completed + customer.subscription.* → upserts the subscriptions entitlement row) → /account (status + post-checkout landing) via /api/stripe/status.
  • Gating: MCP tools now require the mcp scope, which verifyMcpToken grants only to active subscribers - OR everyone while MCP_OPEN_BETA is on (default), so current testers aren't locked out. Flip MCP_OPEN_BETA=false to enforce paid-only; static MCP_ACCESS_KEYS stay exempt (comped/demo). Entitlement reads are DB-only (lib/entitlements.ts), no Stripe SDK in the MCP hot path.
  • New subscriptions table (RLS-on, service-role only). setup-stripe-product.ts creates the idempotent $300/mo product+price. New dep: stripe.
  • Config still needed before live transacting: STRIPE_SECRET_KEY (test in dev), STRIPE_PRICE_ID (set), STRIPE_WEBHOOK_SECRET (from stripe listen / dashboard webhook).

v0.7.9 · 2026-06-08

Changed

• Charting standardized on Recharts v3 + the shadcn chart component. Upgraded recharts 2.15.4 → 3.8.1 and re-pulled the v3 shadcn chart.tsx (now ships ChartTooltipContent/ChartLegendContent + initialDimension). Our code used none of v3's removed APIs, so the bump is clean - typecheck + build green, home/issue/homelab chart pages render without errors. CLAUDE.md now mandates the shadcn chart pattern (ChartContainer + ChartConfig + ChartTooltipContent) for every new/touched chart, for consistent tooltips, animations, theming, and a11y.
• shadcn MCP server configured (.mcp.json) so components/charts are added via the registry MCP (npx shadcn@latest add …) instead of hand-copying. (Loads on next Claude session.)
• Note: existing chart components (IndexAreaChart, VolumeAreaChart, homelab cards) now run on v3 but keep their current custom tooltips; they adopt the full shadcn ChartTooltipContent pattern as they're rebuilt in the home-page redesign. Visual QA of the multi-axis Composed chart on /admin/homelab recommended.

v0.7.8 · 2026-06-08

Fixed

• MCP get_index / get_movers were very slow - they called getDashboardData() live on every request (full paginated deep-join + rolling-sparkline recompute, ~10s). Now they read the precomputed dashboard_snapshot for the default 7-day window (indexed single-row read, ~sub-100ms), matching what the website home page already does. Non-default windows still compute live. Confirmed working end-to-end via a real claude.ai connector (OAuth + audience hook validated).

v0.7.7 · 2026-06-08

Added

• MCP OAuth - consent screen + end-user accounts (part 2). The app now has end-user authentication (it was service-role-only before). New /oauth/consent screen handles Supabase's OAuth authorization handoff: reads authorization_id, requires a logged-in session, shows the requesting client + scopes, and Approve/Deny via supabase.auth.oauth.{getAuthorizationDetails,approveAuthorization,denyAuthorization}. Login uses the shadcn login-03 block wired to Supabase Auth (email/password + Google/Apple social OAuth, password reset), reused on a standalone /login. New browser Supabase client (supabase-browser.ts, session-persisting) kept separate from the server data client. Added shadcn card + label primitives.
  • Remaining dashboard step: set Authentication → OAuth Server → Authorization Path = /oauth/consent and confirm Site URL, then the claude.ai/ChatGPT connect flow is end-to-end. Enable Google/Apple providers if you want the social buttons live.

v0.7.6 · 2026-06-08

Added

• MCP OAuth 2.1 - resource-server half (part 1 of real MCP auth). The MCP endpoint now validates Supabase-issued JWTs as an OAuth resource server (RFC 9728): new /.well-known/oauth-protected-resource metadata route advertising Supabase as the authorization server, JWKS token validation (jose) with RFC 8707 audience binding, and the spec-compliant 401 + WWW-Authenticate challenge that lets claude.ai / ChatGPT web connectors discover the auth flow. Dual-auth during migration: existing MCP_ACCESS_KEYS static keys (Bearer or x-api-key) keep working unchanged, so no demo customer breaks. JWT path fails closed until the Supabase OAuth Server + access-token-hook (audience) are configured. Part 2 (consent page + end-user login) and subscription gating land next. New dep: jose.

v0.7.5 · 2026-06-08

Changed

• Trending Names ranking, after an accuracy investigation. The v0.7.4 breadth-only ranking crowned "New York Times" (#1 on 119 shows) - but a context probe showed NYT runs at only ~1.5–1.9× its baseline: omnipresent, not trending. Inversely, the genuine breaking event (the Meet-the-Press / Kristen Welker walkout) was absent - real (0→10 shows, accelerating) but too fresh to clear the breadth bar under 1×/day ingest. Findings: (1) burst floor raised 1.4→1.7 to strip flat perennials; (2) a higher bar (2.5×) for major media outlets, which are structurally always-cited, so they only headline if genuinely the story; (3) volume floor (≥40 mentions, ≥8 shows) + extended stoplist to kill niche/common-word leaks. Ranking stays breadth-led (it implicitly suppresses the extraction noise floor) among entities that clear the rising bar. Result: an all-real, coherent leaderboard (Maine Senate race, California politics, Hamas/Lebanon) instead of NYT on ubiquity. Known beta limits unchanged: ASR-misspelled canonicals (e.g. "Plattner") and a 1–2 day lag on breaking events - both await entity-linking/NER on the roadmap.

v0.7.4 · 2026-06-08

Added

• Trending Names (BETA) on the home page. A named-entity burst tracker: people, orgs, and places surging across the tracked panel this week, ranked by how many shows picked them up, each linking back to the channels discussing them. Validated via two probes (n-gram → entity) before build; entities self-canonicalize the story (no clustering/LLM needed) and the baseline auto-suppresses ad reads. Extraction is cheap & deterministic (Title-Case runs over transcripts, ~$0) with edit-distance variant merging (folds ASR misspellings like Platner/Plattner), a sentence-initial-ratio filter + stopword/first-name blocklists to strip capitalized non-entities, and breadth-gated rising rank. Persisted to dashboard_snapshot (trending_v1), refreshed daily by /api/cron/trending (10:30 UTC, after ingest). Labelled experimental - canonicalization is still maturing. Roadmap path: PERSON/ORG/PLACE typing (ORG opens a corporate-comms cohort) and an MCP tool.

v0.7.3 · 2026-06-06

Added

• /admin/homelab - home-page redesign lab. All 14 candidate cards for the home-page overhaul rendered against live data (90-day pull), grouped by proposed zone, so the v1 cut and ordering can be chosen by looking: Pulse, Battlefield, Heat Grid, Ownership Map, The Gap, Two Conversations, Risers & Faders, Megaphone treemap, Lit Fuses, Strips, Receipts, Cross-Talk (string-match v1 over scored quotes), Polarization Strip, and Audio-vs-Video. Deliberately heavy/unoptimized - it's a decision tool; chosen cards get snapshot-backed production implementations.

v0.7.2 · 2026-06-06

Changed

• Editorial reach calibration for 77 placeholder podcasts. 72% of the podcast panel carried the seeder's 300k default (the PodScan fields the seeders read are never populated, so the "fallback" was really the default). Replaced with researched editorial estimates anchored to the ~28 publicly-corroborated shows, using chart positions, publisher announcements, public rankers, YouTube presence, and a Podchaser powerScore prior (calibrated: ρ=0.57 - useful as prior, not as truth). Values bracketed to coarse tiers (50k–5M); biggest corrections: Theo Von 300k → 5M, Up First 300k → 4M, Benny Show/Piers Morgan 300k → 2M. Total podcast panel reach 23.1M → 39.7M. Apply script (scripts/apply-reach-estimates.ts) only touches exact-300k rows, so anchors are untouchable and re-runs are no-ops.
• Methodology page: "Audience reach: how we measure it." New section disclosing the split measurement model - YouTube subscriber counts from the YouTube Data API (auto-refreshed daily at ingest) vs editorial weekly-listener estimates for podcasts (no public per-show measurement exists at panel scale; commercial aggregator audience fields verified unusable). Documents the anchoring approach, tier bracketing, review cadence, and why log10 weighting compresses estimation error.

v0.7.1 · 2026-06-06

Added

• /mcp page - public walkthrough of the MCP server. Who it's for (campaign teams, media buyers, consultants, pollsters, comms shops, journalists), eight persona-mapped example questions the dataset answers today, the nine-tool reference table, the data boundaries (excerpts + source links, never full transcripts), and copy-paste connect configs for Claude Code, Claude Desktop (mcp-remote bridge), Cursor, and VS Code - with an honest note that claude.ai/ChatGPT web connectors need OAuth (roadmap). Beta keys are free, requested via [email protected]. Footer now links the page as "For AI Agents".

v0.7.0 · 2026-06-06

Added

• Public MCP server (/api/mcp/mcp, Streamable HTTP) - external AI agents can now query Soapbox data directly: campaign managers / media buyers / consultants connect their own agents and ask arbitrary questions instead of being limited to our charts. Nine read-only tools: get_index, get_movers, list_issues, list_channels, get_issue_detail, get_channel_detail, search_mentions (the workhorse - filtered quote-level search with sentiment, source links, pagination), get_issue_trend (weekly volume/sentiment series), and get_methodology (scoring scale + live panel stats, for citation). Auth via MCP_ACCESS_KEYS (comma-separated bearer keys, fails closed) so demo keys can be issued per-customer; OAuth/Stripe deferred until demo interest proves out. Transcript policy enforced at the data layer: mention-level verbatim quotes + episode source links only - full transcripts are never exposed (PodScan/Supadata license them to us, not through us). New deps: mcp-handler, @modelcontextprotocol/sdk, zod.

v0.6.82 · 2026-06-05

Changed

• Score throughput resized for the expanded panel: SCORE_LIMIT 240 → 720. The channel expansion tripled episode intake (~470 eps/day, ~10 mentions/ep ≈ 4,700 mentions/day) while score capacity stayed at 8×240 = 1,920/day - every cron run saturated its cap and ~2,000 unscored classifications piled up, showing as partially-scored episodes in /log. 240 mentions took ~35s per run, so 720 (~105s) fits comfortably inside the 240s stage time budget; new capacity is 8×720 = 5,760/day. Cost impact is negligible (score is Haiku, ~$0.65 per 1,000 mentions).

v0.6.81 · 2026-06-03

Changed

• Tighter pipeline cadence for a fresher site. Now that transcribe/classify/ score run through concurrency pools, processing latency - not cost - is the thing to cut (cost tracks episode volume, which is unchanged). Transcribe + classify go from every 4h → every 2h; score from every 6h → every 3h (score also refreshes the home snapshot, so the needle now updates 8×/day instead of 4×). Capacity check: score 8×240=1,920 mentions/day vs ~1,400 steady-state; classify 12×60=720 episodes/day vs ~230. Ingest stays 1×/day - its 3-episode cap is per-run, so more frequent ingest would over-sample high-volume channels past the 3/day "stance per audience" cap.

v0.6.80 · 2026-06-03

Added

• Cohort dropdown on admin add-channel. The /admin/channels form now has an Independent / Legacy selector (defaults to Independent), so a channel's cohort is set at add-time instead of defaulting to independent and being fixed up afterward (as 60 Minutes + Real Time had to be). Threaded through AddChannelInput.cohort → the insert in addYouTubeChannel.

v0.6.79 · 2026-06-03

Added

• Admin add-channel now auto-drafts the description. A new "Resolve & draft" step on /admin/channels resolves the handle, reports floor/dup status, and generates a one-sentence rationale in the site's house voice (Haiku, grounded on the channel's own YouTube description + recent video titles + assigned lean) - pre-filling the editable field so the admin edits rather than writes from scratch. previewYouTubeChannel() / generateChannelRationale() in src/lib/channels.ts; generation never blocks the add (falls back to a template on error).

v0.6.78 · 2026-06-03

Fixed

• Transcribe no longer strands episodes on a transient blip. runTranscribe previously marked every failure transcript_status='failed' and only ever re-queried 'pending', so a one-off Supadata outage (2026-06-02, ~95 episodes failed in a single run) permanently abandoned episodes whose captions were perfectly fetchable. getVideoTranscript now returns a discriminated result distinguishing a terminal "no captions" (206 / empty / bad video) from a transient error (5xx / 429 / network). Transient failures leave the episode pending and bump a new episodes.transcript_attempts counter, retrying up to MAX_TRANSCRIPT_ATTEMPTS (3) before giving up - so blips self-heal while a genuinely-broken video still terminates. Backfilled the 89 stranded episodes.

v0.6.77 · 2026-06-01

Changed

• Cohort badge (mic/tv) now uses the same styled shadcn tooltip as the L/M/R lean badge instead of the native browser title, so hover tooltips look consistent on /log and /channels. Added a TooltipProvider (150ms delay, matching /log) around the /channels list to host it. Also swapped the em dash in the cohort tooltip labels for a middle dot.

v0.6.76 · 2026-06-01

Changed

• /log legend row: lean + cohort legends left-justified, status-dot legend right-justified, to visually separate the category legends from the pipeline-status legend.

v0.6.75 · 2026-06-01

Changed

• Consolidated the /log table legends: the L/M/R lean legend, the cohort (mic=independent / tv=legacy) legend, and the status-dot legend now sit together in one right-aligned row directly above the episode table. The cohort legend moved out of the "Episode receipts" heading.

v0.6.74 · 2026-06-01

Added

• Cohort legend (<CohortLegend>) on /channels and /log - defines the mic = independent / tv = legacy icons next to the channel list + episode table.
• Cohort breakdown on the panel cards:
  • Panel balance gains two stacked bars - shows-by-cohort and reach-by- cohort (independent vs legacy) - alongside the existing L/M/R bars. StackedBar generalized to per-segment colors.
  • Panel scale gains a "By cohort" line with the icons + each cohort's show count and combined reach (getPanelStats now returns channelsByCohort / audienceReachByCohort).
  • Both gated on legacy actually being present.

Changed

• Home subheadline: "legacy institutions" → "legacy media"; em dash → comma.

v0.6.73 · 2026-06-01

Added - LEGACY COHORT LAUNCH 🚀

• Legacy media is now live alongside independent. Flipped PUBLIC_COHORTS to ['independent', 'legacy'], which simultaneously:
  • Blends the master Soapbox Index across both cohorts (≈L+0.1, reach- weighted, volume-capped).
  • Reveals the two sub-needles under the master - Independent (≈L+0.5) vs Legacy (≈R+1.5) - with the caption "same issues, same scoring."
  • Shows the cohort icon (mic = independent, tv = legacy) on /channels and /log, and surfaces the 9 legacy channels + their episodes.
• Copy reframe. Home headline → "Where is online political media leaning right now?"; subheadline introduces the independent-creators vs legacy- institutions split. Site title, social meta, footer, and OG image updated from "alternative media discourse" → "online political media, quantified."
• Methodology gains a "Cohorts: independent vs legacy" section.

v0.6.72 · 2026-06-01

Added (gated - invisible)

• Independent vs Legacy sub-needles under the master Soapbox Index on the home page. Two compact needles (<SubNeedle>, reusing SoapboxNeedle at a smaller size) showing each cohort's Index - so the blended master headline arrives with the split that explains it. Gated on PUBLIC_COHORTS.length > 1, invisible until the flip.
• The home snapshot (writeHomeSnapshot) now also computes and stores per-cohort indices (HomeSnapshot.cohorts), so the sub-needles read from the precomputed row - no extra per-request work. Field is optional for backward compatibility with older snapshots.

v0.6.71 · 2026-06-01

Added (gated - invisible)

• Cohort badge (<CohortBadge>): a small icon + hover label marking a channel/episode as independent (mic) or legacy (tv), placed next to the L/M/R lean badge on /channels and /log. Gated on PUBLIC_COHORTS.length > 1, so it renders nothing while the site is independent-only and appears automatically when legacy is exposed at launch. Threaded cohort through the channels query and the EpisodeTableRow (the view already exposes it).

v0.6.70 · 2026-06-01

Changed

• Taxonomy + methodology copy refresh. Dropped the "alt-media" abbreviation from the /issues taxonomy page (intro + activity-card header) and the /methodology page; reframed around the platform (YouTube + podcasts). De-dried the taxonomy intro and removed the "(not yet bucketed)" TODO leaking into the "Political figures & parties" group.
• Methodology now documents volume normalization. The /methodology Index section explains the two deliberate choices - audience-reach weighting and the 3-episodes/day per-channel cap - framing the Index as "stance per unit of audience" rather than who posts most. (Previously only the reach-weighting formula was disclosed; the cap was undocumented.)

Fixed

• scripts/drain.ts rides out transient blips - a stage round now retries with backoff (up to 5 consecutive errors) instead of crashing the whole drain on a one-off Supadata/Supabase fetch failed.

v0.6.69 · 2026-05-31

Performance

• Parallelized transcribe too. The transcribe stage was still serial (the slow part of a full drain). Now runs through the same mapPool at concurrency 8 - each Supadata call is multi-second, so the request rate stays ~2/s, well under the 10/s Supadata limit. TRANSCRIBE_LIMIT raised 40→100 (the wall-clock budget remains the real cap; the pool stops pulling at the deadline). Transcribe rounds drop from ~3–6 min to well under a minute.

v0.6.68 · 2026-05-31

Performance

• Parallelized classify + score (cron throughput). Both stages processed episodes/mentions one-at-a-time; the classify cron's ~90/day capacity was the pipeline bottleneck. They now run through a bounded-concurrency worker pool (src/lib/concurrency.ts → mapPool): classify at concurrency 10, score at 15, sized for an Anthropic Max-tier account. Per-run limits raised accordingly (CLASSIFY_LIMIT 15→60, SCORE_LIMIT 80→240); the per-stage wall-clock budget is still the real cap, so runs finish under the 300s function limit (the pool stops pulling new work at the deadline).
  • Net: classify throughput ~5–6× per run (~90/day → ~500+/day at the same cron cadence), bounded now by the Anthropic tier rather than the serial loop. Counters are mutated inside the pool, which is safe (single-threaded).
  • New npm run drain (scripts/drain.ts): loops the parallelized stages until the backlog clears - used to drain the legacy seed immediately rather than waiting ~1–3 days for the crons.

v0.6.67 · 2026-05-31

Added (foundation - invisible)

• Channel cohorts: independent vs legacy. Groundwork for an independent-vs-legacy comparison and a blended master Index. New channels.cohort column (default independent, indexed); all 86 existing channels backfilled to independent. episode_pipeline_summary view gains cohort.
• All public reads are now cohort-aware, gated by a single control point (src/lib/cohort.ts → PUBLIC_COHORTS = ['independent']). The Index (fetchScoreRows), issue/topic drill-downs, channel list, panel/system stats (shows, episodes, hours), and the /log feed all filter to the public cohort. This lets legacy channels be seeded and ingested invisibly - legacy data accumulates but never surfaces until we flip PUBLIC_COHORTS and ship the comparison UX. Zero behavior change now (every channel is independent). Non-political legacy content stays a non-issue: it classifies to no-signal and never enters scoring/weighting.
  • Known follow-up: the secondary scale totals (transcripts/classifications counts in SystemStats) are still whole-pipeline; tighten at launch.

v0.6.66 · 2026-05-31

Performance

• Drill-down pages (/channels/[id], /issues/[slug], /topics/[slug]) were ~7s - now DB-filtered. Each called fetchScoreRows() - the full ~17K-row sentiment_scores deep join - then filtered in JS for the one channel/issue/topic. They pulled the entire table to show a single slice (the same problem the home page had, never fixed for the drill-downs).
  • New fetchScoreRowsFiltered() anchors on classifications and filters at the DB via the indexed issue_slug / episode_id columns, returning only the rows in scope (e.g. iran-conflict 2,137 rows, a channel ~700, vs 17,673 every time). Same ScoreRow shape and scored-only semantics as fetchScoreRows; paginated for hot issues that exceed 1,000 rows.
  • getIssueDrillDown filters by issue_slug; getTopicDrillDown resolves the topic's child issues then filters by their slugs; getChannelDrillDown resolves the channel's episode ids then filters by them. The downstream JS filters become no-ops on the already-scoped set, so the numbers are unchanged - only faster. Stays live (no snapshot/staleness).

v0.6.65 · 2026-05-31

Performance

• /log now server-paginates its episode table. The page was loading the entire ~2,000-row archive every request to power client-side search/sort/paginate - ~1.3s TTFB that grows with the archive. (Measured: the underlying episode_pipeline_summary view runs in ~64ms - the DB was never the bottleneck; the cost was fetching + serializing the full row set.)
  • New GET /api/episodes endpoint: sort, search, and pagination run in Postgres (getEpisodeTablePage - .range() + count: 'exact'), returning only the ~25 rows a page shows plus the total count. Search is sanitized before the PostgREST or() filter; stage columns sort by their underlying status field.
  • EpisodeDataTable gained a serverSide mode (TanStack manual sorting/filtering/pagination + debounced search + abortable fetch). /log uses it; the per-channel table keeps client mode (small, preloaded sets). Expandable per-episode receipts (v0.6.64) work unchanged.
  • /log TTFB no longer scales with the episode count - it fetches one page regardless of archive size. Trade-off: the table now hydrates client-side (a brief "Loading episodes…") rather than being in the initial HTML.

v0.6.64 · 2026-05-31

Added

• Expandable per-episode receipts on /log. Each scored episode row now expands to show exactly what the pipeline classified and scored: every issue mention with its sentiment chip (L+/R+, the home Index convention), a 1–5 intensity meter, and the supporting quote the model flagged - plus an episode net-lean summary. Delivers on the page's "receipts, in the open" promise and gives the operator a fast lens to spot mis-scores or bad issue-mappings before scaling the channel set.
  • Lazy-loaded on expand via a new GET /api/episodes/[id]/mentions route (one episode at a time), so the table never eager-loads the full classifications join. Mentions are sorted strongest-first (|sentiment| × intensity).
  • Built on the existing shadcn data table (shadcn <Table> primitives + TanStack getExpandedRowModel); a caret appears only on episodes that produced classifications. New <EpisodeMentions> sub-row component.
  • Quotes are excerpts only, never full transcripts.

v0.6.63 · 2026-05-31

Fixed

• Cron classify silently stalled - transcripts.id doesn't exist. The scheduled classify stage reported pendingFound=0 on every run for >24h while 68 transcribed episodes sat ready. Root cause: runClassify did .select("id, …") / .order("id") on the transcripts table, whose PK is episode_id - there is no id column. The query 400'd every run, the error was swallowed (const { data } = with no error check), the loop broke, and the empty result read as "queue empty." Broken since v0.6.47 (the "add ORDER BY" fix used the wrong column name); masked because the CLI catchup (scripts/classify.ts, episode-first since v0.6.48) did the real draining.
• Fix: cron classify is now episode-first, mirroring the CLI. Query the episodes table for classify_status='pending' AND transcript_status= 'fetched' (cheap - no text), then load each transcript's text on demand inside the loop. This eliminates the ≈80MB "pull every transcript" payload that also caused the response-size/timeout fragility, and the pending-episode query now checks its error and throws instead of silently reporting an empty queue - so this class of stall fails loud, not silent.
• Drained the 68-episode backlog (classify + score) so the Index reflects current data.

v0.6.62 · 2026-05-30

Added

• "What alt-media is talking about" card on /issues. The issues page was the only main page with no data card above its list - a static taxonomy reference with no live signal. Added a topic-level attention rollup above the taxonomy: the 23 issues' mention volume aggregated into the same 11 topics the list is grouped by, ranked by mention count, each with a volume bar, the topic's volume-weighted lean tint, and a deep link to its /topics/[slug] page.
  • Reads per-issue volume/lean from the existing dashboard_snapshot (one row, no heavy join) via readHomeSnapshot(), with a live getDashboardData fallback when the snapshot is absent. So the page stays fast and adds no new DB aggregation.
  • New <IssueActivityByTopic> component (pure presentational, prop-driven - same pattern as PanelBalance / PanelScale). Bars + headline use raw mention count ("how much is this discussed"); lean tint uses volume-weighted lean so the direction matches the Index basis.
  • Deliberately distinct from the home page's "Biggest movers" (a lean-swing leaderboard) - this is an attention-volume distribution, answering the /issues reader's question "which areas are hot, which should I open?"

v0.6.61 · 2026-05-30

Performance

• Home page TTFB: precompute the dashboard instead of recomputing per request. v0.6.60's cache() fix only deduped the double fetchScoreRows call within one render - it can't cache across requests, and cache() is per-render scope only. Direct prod timing after v0.6.60 still showed ~9.5s TTFB on / (every visitor recomputed the full ~17K-row deep join from scratch), while /channels and /issues stayed ~0.4s. Root cause is structural: the home page re-aggregates all history on every hit, but the underlying data only changes when the daily pipeline runs.
  • New dashboard_snapshot table (migration 20260530120000_dashboard_snapshot.sql): one JSONB row per window key (home:7) holding the precomputed { dashboard, breakdown }. Service-role only.
  • writeHomeSnapshot() computes getDashboardData() + getIndexBreakdown() once (sharing the per-request fetchScoreRows cache → one DB pass) and upserts the row. Called at the end of the score cron (the last data-producing stage) and the manual /api/cron/pipeline run; best-effort so a snapshot failure never fails the cron. Also runnable ad hoc via npm run refresh:snapshot.
  • readHomeSnapshot() + home page now reads that single indexed row (~sub-100ms). Falls back to the live computation when the snapshot is missing or unavailable (first deploy / before first cron / pre-migration), so the page never breaks. <IssueContributionsChart> takes the breakdown as a prop (live fetch retained as fallback).
  • Net: home / TTFB drops to ~0.4s for every visitor, with no cold-miss cliff on deploys (unlike a request-level cache). Removes the 17K-row join from the request path entirely; scales as the panel grows. Delivers the "cached SQL views / materialized aggregates" TODO that was noted in src/app/page.tsx.

v0.6.60 · 2026-05-30

Performance

• Home page TTFB ~15s → expected ~4s. Direct prod timing showed 14.6–15.8s TTFB on /. Root cause: fetchScoreRows() was called TWICE per render (once by getDashboardData() for the dashboard, once by getIndexBreakdown() via the sibling <IssueContributionsChart> server component) - each paginating the full 17K-row deep join independently, ~35 round trips to Supabase apiece.
  • Wrapped fetchScoreRows with React cache() so all server-component callers within one render share the same Promise. Halves the work on the home page; no behavior change.
  • Bumped fetchScoreRows pageSize 500 → 1000. The 500 cap was added in v0.6.3 because Vercel's edge→Supabase route returned short pages on big response payloads and the old length < pageSize terminator interpreted that as end-of-data (silent truncation). v0.6.51 fixed the terminator to only stop on truly-empty pages, so short pages no longer truncate; can safely go back to 1000-row pages. Halves round-trip count again (17 pages instead of 34). Per-row payload is small (~300 bytes, no text), so a 1000-row page is ~300KB - comfortably under response cap.
• Combined effect: home page does ~17 round trips instead of ~70. Other pages that call fetchScoreRows once each (/issues/[slug], /topics/[slug], /channels/[id]) get the page-size win (about 50% faster) but not the dedup win (they only call it once).
• Re-time after deploy with curl -o /dev/null -s -w "TTFB %{time_starttransfer}s\n" https://www.soapbox.media/ to verify.

Notes

• The longer-term play is materializing the rolling-window aggregation in Postgres (a view or materialized view computed by cron) so app reads a small result set instead of all 17K scored rows. The original aggregate.ts comment from v0 flagged this as the "v1 will move this" path. v0.6.60 buys time but doesn't replace it - at 100K+ scored rows the per-render scan will still be slow even with dedup + bigger pages.

v0.6.59 · 2026-05-30

Fixed

• /log scored status falsely rendered "done" for un-classified episodes. Regression from v0.6.54's status-cascade rewrite. When an episode hadn't been classified yet, classification_count = 0 and scored_count = 0, so sc >= cc evaluated 0 >= 0 → true → "done". The pre-v0.6.54 code caught this with a leading cc === 0 ? "na" branch; my rewrite only caught the cc=0 case when classify_status='processed' (mapping to "no-signal") and let cc=0 + classify_status='pending' fall through. Result: 132 episodes were rendering as scored=green on /log when they were nowhere near scored. Visible on the activity log as rows with transcribed=gray, classified=gray, scored=green - a logical impossibility the cascade should have prevented.

  Fix: explicit classified === "pending" ? "pending" guard before the sc >= cc check in getEpisodeTableRows. The score column now faithfully cascades: can't be scored before classified, can't be classified before transcribed.

v0.6.58 · 2026-05-30

Fixed

• Podcast reach auto-refresh removed - PodScan's audience_size is unreliable for the panel's purposes. v0.6.57's reach-refresh pass attempted to hit PodScan's /podcasts/{id} endpoint and pull pickPodscanReach from the response, but the immediate post-deploy refresh exposed the gap: zero of 44 podcasts updated. Probing the endpoint directly showed audience_size IS exposed - just nested at reach.audience_size (not top-level where the helper looked) - but the values are wildly off from publicly-reported listener estimates:

  • Joe Rogan Experience: DB 14.5M vs PodScan reach.audience_size: 4.7M
  • Mark Levin Show: DB 7.0M vs PodScan reach.audience_size: 100

  Our stored numbers align with Edison-style weekly-listener estimates; PodScan's appears to be its own internal-tracking metric (a lower bound, often missing entirely). Auto-refreshing from PodScan would crash podcast reach 50–70% to less-real numbers, so podcasts are now intentionally NOT in the refresh path. Removed getPodcastById + pickPodscanReach calls from runIngest and scripts/ingest.ts; the 41-channel YouTube refresh (which works perfectly, daily) is the entire auto-refresh story now.

Changed

• Honest copy on /channels. Intro paragraph: "YouTube subscriber counts refresh daily during the ingest pass; podcast audience estimates are editorial and reviewed at panel-add time." (Previously: "Reach figures refresh daily from the YouTube Data API and PodScan" - half wrong.)
• <PanelScale> freshness label: now reads "YouTube subs refreshed Xh ago · podcast reach editorial" - was "Reach refreshed Xh ago," which implied podcasts were also auto-refreshed.

Notes

• getPodcastById helper stays in src/lib/podscan.ts - it's a clean by-id lookup that may be useful for other contexts (e.g., verifying a candidate matches what's in the panel during admin add-flow); just not for reach refresh.
• The 44 podcast rows still have their reach_updated_at backfilled to created_at (17 days old). That's accurate - we genuinely haven't refreshed them. The PanelScale label correctly reads off MAX(reach_updated_at) which is now-today (the YT refresh time), so the visible signal is right.
• Memory [[podcast-reach-editorial]] written so this gotcha isn't re-discovered next time someone tries to wire PodScan to channels.reach.

v0.6.57 · 2026-05-30

Fixed

• channels.reach was set-once-at-seed and never refreshed. Three call sites wrote reach (seed-channels.ts, channels.ts add-flow, enrich-legacy-wishlist.ts); none refreshed it. 84% of the panel was carrying 17-day-old subscriber/listener counts. The /channels intro paragraph claimed reach was "pulled live" - technically true, at seed time only. Index math weights by log10(reach), so stale reach = mildly wrong weights.

Added

• Reach refresh piggybacked on the daily ingest cron. The ingest pass already iterates every active channel; now it also refreshes each channel's reach in the same loop. YT is batched via getChannelDetailsBatch (one API call for up to 50 channels, ~1 quota unit each - free tier handles 10,000/day); podcasts are per-row via the new getPodcastById helper in src/lib/podscan.ts (PodScan has no batch endpoint). Failures are logged-and-skipped - a transient API blip on one channel must not abort the whole ingest pass. Only positive reach values overwrite the stored stat; a 0 / null response keeps the existing number so a lookup miss doesn't zero out a known channel.
• channels.reach_updated_at column. New TIMESTAMPTZ with now() default. Backfilled to created_at for existing rows (conservative "at least this stale" floor; the seed scripts didn't track it). Bumped on every refresh attempt - even when the number didn't change - so staleness-detection isn't misleading.
• Freshness signal on <PanelScale> - top-right of the card now reads "Reach refreshed Xh ago" (MAX(reach_updated_at) across active channels). Same relativeTime shape as the existing "Latest data" timestamp on <SystemStats>.
• /channels intro paragraph tightened - now reads "Reach figures refresh daily from the YouTube Data API and PodScan during the ingest pass" instead of the previous "pulled live" wording, which is honest about cadence.

Changed

• CLI npm run ingest now also refreshes reach (mirrors the cron path via the same helpers). Per-channel log line includes the before→after delta when reach changes (reach: 5,990,000 → 6,012,000 ↑ 22,000) so manual catchup runs print visible movement.

Notes

• New migration add_channels_reach_updated_at - non-destructive ALTER TABLE … ADD COLUMN, backfill from created_at, set NOT NULL + default now().
• pickPodscanReach (same field-fallback as seed-podcasts.ts's pickReach) is now duplicated in three files (seed-podcasts.ts, pipeline.ts, scripts/ingest.ts). Worth extracting to src/lib/podscan.ts in a follow-up cleanup once the dust settles.

v0.6.56 · 2026-05-30

Changed

• Stat cards re-homed by reader question, not by convenience. /log's System Scale was carrying the panel-composition stat ("Combined audience reach", added in v0.6.54) which actually answers "is this panel representative?" - a /channels question, not a /log question. The /log reader is asking "is the pipeline running?". Moved the reach number off /log and onto a new <PanelScale> card on /channels where it belongs.

Added

• <PanelScale> card on /channels - composition stats (shows tracked, combined audience, platform rows, largest single show). Same visual shape as <SystemStats> on /log so the cards rhyme, but the question they answer is different. Sits ABOVE <PanelBalance> so the page reads magnitude (raw numbers) → distribution (stacked bars) → list (per-lean show grid).
• New getPanelStats() aggregate helper. Channels-table only - no episode/classification/score queries. Returns shows tracked + L/M/R count, audience reach + L/M/R split, platform row count + YT/Pod split, and the largest single show by max reach. Mirrors the unique-show methodology of <PanelBalance> and the old getSystemStats.audienceReach field so all three surfaces agree on the same number.

Changed (cont.)

• /log System Scale trimmed to 4 pipeline-only stats (was 5): shows tracked, episodes analyzed, hours of audio, issue mentions. Grid shifted from lg:grid-cols-5 to md:grid-cols-4 - same breathing room per stat. getSystemStats still computes audienceReach + audienceReachByLean for any downstream caller; it's just not displayed on /log anymore.

v0.6.55 · 2026-05-30

Added

• Panel balance badge on /channels. Two stacked horizontal bars (count + reach) show the L/M/R distribution side by side so the asymmetry between editorial-intent-balanced counts and what-the- landscape-looks-like reach is visible at a glance. Current state: shows are 36% L / 14% M / 50% R but reach is 28% L / 15% M / 57% R - right-leaning shows carry larger average audiences (2.77M vs 1.93M L), so reach skews right. Badge says this plainly rather than letting the intro paragraph imply uniform balance. Asymmetry sentence renders dynamically - only shown when avg-reach ratio across cohorts ≥ 1.25×, so it'll quiet down if the panel rebalances.
• The honest copy explicitly notes that log10(reach) weighting in the Index dampens the asymmetry but doesn't erase it - a methodology cue for readers comparing the published Index to their intuition.

v0.6.54 · 2026-05-30

Added

• "No signal" status on the public activity log. ~8% of processed episodes (161/1941 today) are off-taxonomy - classified successfully but produced no political-issue mentions (sports, true crime, celebrity, etc.). These previously rendered as the same gray dots as "pending" episodes, with the scored column tooltip saying "Not applicable" - confusing because gray reads as in-progress, and a "complete but empty" episode isn't in-progress. New no-signal status with a hollow outlined dot (border-only, transparent fill - reads as "registered but empty") on both the classified and scored columns when classify_status='processed' and classification_count = 0. Tooltip: "No political signal · issue taxonomy didn't match." Added to the visible legend.
• Combined-audience reach stat on /log. Headline number for "how big is this panel?" - sum of unique-show reach (max per show across platform rows, so dual-platform shows aren't double-counted; matches the methodology for the by-show comparison from yesterday's enrichment script). Sublabel breaks reach out by editorial lean (L · M · R), same shape as the existing show-count sublabel - surfaces cohort balance on the same surface.

Changed

• episode_pipeline_summary view: added classify_status. Migration add_classify_status_to_pipeline_summary_view - non-destructive CREATE OR REPLACE VIEW. Column had to be appended at the end of the SELECT (Postgres can't reorder existing view columns; only append). The view's only consumer (getEpisodeTableRows) updated to select it.
• Hours-of-audio stat reformatted. Was 1.4K (compact) which read like a placeholder; now 1,433 (full number) with sublabel ≈ 60 days continuous instead of the static Long-form, Shorts filtered. Confirmed 100% of episodes have duration_sec - the data was always plumbed; just the formatter obscured it.
• Issues-mentions sublabel: dynamic count + folded sentiment-scores stat. Was hardcoded Across 15 issues (stale - taxonomy is at 23). Now reads the active-issue count from issues table and renders Across N issues, all sentiment-scored. The standalone "Sentiment scores" stat was dropped to make room for combined-audience - post-v0.6.53 score == mentions for the autonomous-cron steady state, so the standalone number wasn't pulling its weight.

v0.6.53 · 2026-05-30

Fixed

• CLI scripts had the same .range() family bug the cron path got fixed for in v0.6.51 - the previous audit pass (v0.6.52) only covered src/, not scripts/. Caught by the catchup drain itself: the classify stage drained cleanly (393 → 0, added 4,758 new classifications), but scripts/score.ts told the catchup loop "queue drained" while 5,809 classifications were actually unscored. Root cause: both pagination loops in score.ts had no .order() AND the data.length < pageSize early-out
  • so the script only ever read page 0 of classifications and sentiment_scores, scored the 200-ish overlap in page 0 across 3 catchup iterations (600 scored), then page 0 showed "all scored" → "drained" sentinel fired. Same dual-bug as the original v0.6.47.
• scripts/score.ts - added stable .order("id", asc) on the classifications loop and .order("classification_id", asc) on the sentiment_scores loop (UNIQUE constraint makes it a valid pagination key); removed both data.length < pageSize early-outs. Same canonical pattern as aggregate.ts:155-209.
• scripts/classify.ts - happened to work in the catchup drain (the filtered pending set fits in a single page below the response cap), but carried both the non-unique-sort-key bug and the short-page early-out. Added id as a stable tiebreaker after published_at; removed the short-page early-out. Future-proofs against the panel doubling.

Notes

• Full audit now extended to scripts/ directory; both CLI surfaces (classify.ts + score.ts) and one already-correct file (transcribe.ts uses single .limit(), not a paginated loop) conform.
• The 5,809-classification score backlog this leak created will be drained separately via npm run score -- 8000 on the fixed v0.6.53 code (~$4 in Haiku, ~30 min wall-clock).

v0.6.52 · 2026-05-29

Fixed

• Audit-pass: remaining .range() antipatterns surfaced by the post- v0.6.51 grep -n "range(" src/ sweep. Three callers had subspecies of the same family of pagination bugs. None were currently breaking the cron (that was v0.6.51), but each would have bitten silently as the panel keeps scaling - so fixing all of them is part of "runs autonomously."
  • src/lib/audit.ts paginatedSelect - the generic helper used by /admin/channels-audit had both halves of the v0.6.47/v0.6.51 bug: no .order() and a data.length < pageSize early-out. Hardcoded .order("id", ascending: true) inside the helper (all three callers use tables with an id PK; the helper's contract is now unambiguous
    • "I paginate by id") and dropped the short-page break.
  • src/lib/episodes.ts getEpisodeTableRows - had empty-page-only termination ✓ but ordered by published_at DESC alone, which isn't unique. Two episodes posted in the same second could re-cross page boundaries and appear duplicated in the /log table. Added .order("id", descending) as the stable tiebreaker after the business order; UI behavior unchanged when published_at values are distinct (the common case), now deterministic when they collide.
  • src/app/channels/page.tsx - single-call .range(0, 999) silently truncates at 1000 active channel rows. We're at 85 today but the scale-out target is ~200 unique shows (2–3 platform rows each, easily 400–600), well within the lifetime of this code. Converted to the canonical paginated loop (.order("id") + empty-page-only break); JS-side groupByShow → maxReach already re-sorts so the user-visible order is unchanged.

Notes

• Repo now has 8 .range() callers, all conforming to the audit pattern in [[pagination-stable-order]]: stable .order(<unique_key>) AND data.length === 0 as the only loop terminator. The pattern is duplicated across 5 files (aggregate.ts ×2, discovery.ts, pipeline.ts ×3, audit.ts, episodes.ts, channels/page.tsx) - a good candidate for extraction into a shared helper if/when scope allows.

v0.6.51 · 2026-05-29

Fixed

• Cron classify + score short-page early-out → silent backlog stall (round two). Same pendingFound=0 symptom as v0.6.47, different half of the same pagination antipattern. v0.6.47 added the required ORDER BY but kept if (data.length < pageSize) break; as the loop terminator. That early-out fires on any short page - and Vercel's edge→Supabase route hits a response-size cap before the row cap on runClassify's deep-join query (each row carries full transcript text). Once transcripts grew past the response threshold (1,779 rows as of today), the first page came back short, the loop exited, the in-memory array only held the oldest already-processed rows, and the JS filter to classify_status='pending' returned []. Result: 3 of every 4 classify cron runs today found 0 pending despite 393 actually pending (08:30/12:30/16:30 UTC; only the 00:34 + 04:34 runs processed work). Fix: terminate on empty page only - matches the canonical pattern at aggregate.ts:155-209 (v0.6.3) and the getSystemStats pagination at aggregate.ts:450-461. Applied to all three paginated loops in pipeline.ts (runClassify transcripts, runScore classifications, runScore sentiment_scores).

v0.6.50 · 2026-05-29

Added

• Mention-volume signal alongside lean in "Biggest movers." The home card now ranks issues on two orthogonal axes - lean swing (L↔R movement) and mention-volume swing (attention shift) - and shows both. A row earns its spot if |leanΔ| ≥ 0.5 OR volumeRatio crosses [0.67×, 1.5×]; both numbers display so visitors can see which signal (or both) put it there. Ranking uses max(|leanΔ|/2, |log2(volumeRatio)|) so a 2-point lean swing and a 2× volume swing carry equal weight, and the existing MOVER_MIN_MENTIONS = 25 floor applies on both windows so neither axis fires on thin samples. Cap moved into getDashboardData (6 rows) - the home page just renders data.movers directly now. Mobile keeps the original 3-column layout for readability; desktop expands to 6 columns (adds Last week / Mentions / Volume).
• Per-issue mention-volume sparkline on /issues/[slug]. New <VolumeAreaChart> component (neutral gray, non-negative y-axis, no zero reference line - counterpart to <IndexAreaChart>) renders alongside the existing lean trend in a 2-up grid. Answers the question the lean chart can't: "is anyone actually talking about this issue right now?" Powered by a new rollingVolumeTrend() helper in aggregate.ts that mirrors rollingLeanTrend's windowing but keeps mid-series zero days (a stretch of zero is a real "issue went silent" signal - lean is just undefined at 0/0, volume isn't); leading-only zero days are trimmed so the chart starts at first activity.
• IssueMover extended with currentMentions, prevMentions, volumeRatio (week-over-week mention-count ratio). IssueDrillDown gains volumeTrend: { values, dates }. No new pipeline cost - both surfaces are derived from the existing fetchScoreRows() data.

v0.6.49 · 2026-05-29

Changed

• scripts/discover-socialblade.ts handles markdown + smarter triage. Added a markdown-table parser (auto-detected by extension or content) so Social Blade pages saved via a browser markdown-clipper extension work directly - previously only HTML was supported. Tightened the bucketing: beyond "in panel" / "legacy" / "candidate", the script now flags "non-US/non-English" (Cyrillic / Devanagari / Burmese / CJK scripts; known Spanish/Bengali/Hindi outlets) and "non-political" (gaming, true-crime, finance-tutorial) so the actionable candidate list isn't drowned by 100-row globals. Name normalization strips "The X Show" / "X Podcast" boilerplate to catch Social Blade ↔ panel mismatches (Ben Shapiro ↔ "The Ben Shapiro Show", etc.).

• docs/legacy-media-wishlist.md - appended a "From Social Blade Top 100 News (US, 2026-05-29)" section with cable / broadcast, digital-native, local-affiliate, and ambiguous-cohort entries surfaced by the scrape.

v0.6.48 · 2026-05-29

Changed

• CLI classify is episodes-first. The old scripts/classify.ts paginated the entire transcripts table with text embedded in the SELECT - a 1700- row × ~100KB/row payload that hit Postgres's statement_timeout once the panel hit ~80 channels. Refactored to query episodes (no text) filtered on classify_status='pending' AND transcript_status='fetched', then load each transcript on demand inside the loop. Orders by published_at DESC so the most-recent backlog drains first. The cron path in pipeline.ts may benefit from the same treatment if/when it starts timing out at larger scale - for now its 300s function budget masks the inefficiency.

Added

• scripts/discover-socialblade.ts - one-time parser for saved Social Blade "Top by category" HTML pages (politics, news, etc.). Direct fetch is blocked by Cloudflare, so you save the page from a browser, point this script at the file(s), and it: extracts every /youtube/channel/UC… link with name + sub count, filters to ≥300K, dedups against the existing panel by both channel ID and normalized name, and prints a sorted candidate list. Run via npm run discover:socialblade <file.html> [file2.html] ….

v0.6.47 · 2026-05-29

Fixed

• Cron classify + score paginated without ORDER BY → silent backlog stall. Two of today's four scheduled classify runs (08:30 and 12:30 UTC) reported pendingFound = 0 despite 564 episodes actually being pending. Once the transcripts table grew past 1000 rows, PostgREST's .range() pagination returned non-deterministic pages - some runs got pages where every row was already classify_status='processed', so the cron silently decided there was nothing to do and exited in 9 s. This is the exact pagination gotcha called out in CLAUDE.md; the CLI scripts had stable .order("episode_id") since v0.6.29 but pipeline.ts never got the same treatment. Fixed in all three paginated reads (transcripts, classifications, sentiment_scores) by adding stable PK ordering. Backfill drained manually via CLI after the fix shipped.

v0.6.46 · 2026-05-28

Fixed

• Add-channel flow now requires a lean rationale. The 21 channels seeded today (and Sam Harris) were missing classification_rationale, so they showed up on /channels without the one-sentence descriptions every other channel has. Backfilled all 22 by hand in the project's editorial voice; threaded the field through addYouTubeChannel and the /admin/channels add form (required, validated, with a placeholder example) so the gap can't recur.

v0.6.45 · 2026-05-28

Added

• /admin/channels - admin flow to add a channel + deep-ingest history. Editor enters a YouTube handle/URL + L/M/R lean; the server action resolves via the YT API, enforces the 300K subscriber floor, inserts, and deep-ingests the last 30 episodes. The cron then catches up transcribe→classify→score automatically. Shared logic in src/lib/channels.ts (addYouTubeChannel, extractYouTubeHandle) so the CLI tool and the admin UI go through the same code path. Page also shows the 20 most recently added channels. AdminNav adds a new Channels tab; the existing audit moves to the Audit label.
• Per the channel expansion strategy, the existing 8 newly-seeded channels (Valuetainment, TPUSA, Knowles, Klavan, Indisputable, Legal AF, Katie Phang, Talking Feds) were seeded via SQL + npm run backfill:channel-history and are catching up via cron. Panel is now 56 unique shows (69 channel rows).

v0.6.44 · 2026-05-28

Changed

• Cron stages run multi-times/day to fix the backlog dynamic. Ingest stayed daily (10:00 UTC); transcribe and classify now run every 4h (6×/day), with classify offset +30 min; score runs every 6h (4×/day). Same total work per day, smoother throughput - with the v0.6.43 time-budget guard each run completes cleanly, so the only knob needed is frequency. At 48 channels this keeps the pipeline caught-up (transcribe 240/day vs ~148 ingest/day; classify ~90/day vs ~40/day transcribed). Empty runs are free.

Added

• Channel expansion strategy draft (docs/channel-expansion-strategy.md) for the 48→200 scale-up: curation criteria, sourcing ladder, ~$870/mo cost model at 200 channels, throughput requirements (hourly classify), phased rollout, and open editorial decisions (reach floor, lean balance target, cost ceiling). Not implemented - review artifact.

v0.6.43 · 2026-05-27

Fixed

• Classify cron 504 after the taxonomy grew to 23 issues. This morning's scheduled classify ran the full 300s on a 15-episode batch and was killed mid-batch (12 episodes done, no usage_log row) - the larger taxonomy makes each episode slower and produce more mentions, so a fixed CLASSIFY_LIMIT can overshoot. Added a wall-clock budget (STAGE_TIME_BUDGET_MS = 240s): the classify loop stops when the budget is hit and always completes cleanly, processing as many episodes as fit. CLASSIFY_LIMIT stays as an upper bound; the run now reports stoppedAtTimeBudget. (Adapts automatically as the taxonomy keeps growing.)

v0.6.42 · 2026-05-27

Added

• Topic drill-down pages (/topics/[slug]) - the deeper Phase 2 read path. getTopicDrillDown rolls a parent Topic's child issues into a topic-level lean + 30-day trend (same reach×intensity weighting as the Index, so the numbers stay consistent across issue/topic/overall). Each topic page shows the needle, trend chart, and its child issues ranked by share of voice. The /issues topic headers now link to them. ScoreRow carries issue_topic_slug (added to fetchScoreRows).

v0.6.41 · 2026-05-26

Two-level taxonomy - Phase 2 (read path) + discovery integration + staged classify-broadening. (Parent Topics contain child Issues; see docs/taxonomy-v2-design.md.)

Added

• Issue taxonomy page grouped by Topic. /issues now lists issues under their parent Topic (Foreign Policy, Health, Rule of Law, …), making the two-level structure visible. Index/scoring unchanged.
• Discovery promote is Topic-aware. Promoting a candidate now requires picking a parent Topic; the new child issue is created under it (issues.topic_slug). discovery_candidates.assigned_topic_slug records it.
• 7 gap-filling issues staged (inactive). Health care, Social Security & Medicare, Justice/rule-of-law, Government corruption, Gun policy, Drug policy, Race & discrimination - to cover the empty/thin Topics classify is currently blind to. Staged active=false with draft L/R anchors, so they do NOT affect classify or the Index until the anchors are reviewed and activated.
• Migrations taxonomy_v2_topics_layer and taxonomy_v2_gap_issues_and_discovery_topic (DB; additive only).

v0.6.40 · 2026-05-26

Changed

• Admin login screen + menu (replaces HTTP Basic Auth). /admin/* is now gated by a cookie session instead of the browser Basic Auth dialog. New /admin/login form checks ADMIN_PASSWORD and sets an httpOnly cookie (value = SHA-256 of the password, 30-day expiry); middleware redirects there when the cookie is missing/invalid. New /admin landing menu (Pipeline · Costs · Channels audit · Discovery) and a Log out control in AdminNav. Same password as before; cron auth (CRON_SECRET) and the public /eval/label tool are unaffected. After deploy, existing sessions are logged out and must sign in via the new form.

v0.6.39 · 2026-05-26

Emerging-issue discovery with admin oversight - the fixed 16-issue taxonomy no longer silently misses new topics (e.g. it would now surface something like a "Trump anti-weaponization fund" for review).

Added

• Harvest (Phase 1): the classify pass now also returns substantive political topics that don't fit the taxonomy (OffTaxonomyTopic - label + quote), stored in the new discovery_topics table. Applies to both the cron runClassify and the CLI. Marginal token cost; no extra LLM pass. Off-taxonomy episodes (0 taxonomy mentions) are exactly where new issues hide.
• Cluster & rank (Phase 2): src/lib/discovery.ts + src/modules/discover merge recent off-taxonomy labels into candidate themes via one Haiku pass, score each by reach × recency × frequency, and rebuild the pending discovery_candidates set. Triggered by a weekly cron (/api/cron/discover, Mondays 11:00 UTC) and npm run discover.
• Review queue (Phase 3): /admin/discovery (Basic-Auth) lists ranked candidates with example quotes + counts, and offers Promote (form → new taxonomy issue, with human-written L/R positions), Merge (into an existing issue), or Ignore. Added to AdminNav.
• Migration discovery_tables (discovery_topics, discovery_candidates; RLS-on/no-policies per convention).
• One-time harvest-only backfill npm run discover:backfill (re-runs classify over recent transcripts writing ONLY off-taxonomy topics, never duplicating classifications) to populate discovery without waiting for days of cron. Initial backfill of 40 episodes harvested 106 topics → 42 candidates.

Guardrail

• Discovery proposes, a human disposes - the system never edits the taxonomy on its own; only the admin Promote action (which requires the editor to write the L/R positions) creates an issue. Decided candidates' source topics stay linked so dismissed themes don't resurface.

v0.6.38 · 2026-05-26

Fixed

• Cross-platform episode duplication. Shows tracked on both YouTube and a podcast feed (e.g. The Rubin Report) publish the same episode to both, which was ingested twice and double-counted in the Index. New src/lib/dedup.ts matches a cross-post by show + normalized title + publish date; ingestion (both the cron runIngest and the CLI scripts/ingest.ts) now skips an episode already present on a sibling channel. Because channels ingest reach-desc, the higher-reach copy is kept and the re-post is skipped. Backfill: removed the 18 redundant copies already in the DB (with their classifications/scores; transcripts cascaded) - all were lower-reach podcast copies; where only one copy was processed, that one was kept regardless of reach. No remaining cross-platform dup groups.

v0.6.37 · 2026-05-26

Fixed

• Cron split into per-stage jobs to fix a 300s timeout. After v0.6.29 made classify do real work, the combined nightly pipeline exceeded Vercel's 300s function limit - the 2026-05-26 run returned 504, classified 73 mentions, then was killed before score (left them unscored) and before writing usage_log. The four stages now each run as their own cron with a full 300s budget: /api/cron/{ingest,transcribe,classify,score}, staggered at :00/:15/ :30/:45 past 10:00 UTC. Stage logic was extracted unchanged into src/lib/pipeline.ts (stages never call each other, so they split cleanly - see ARCHITECTURE.md). The old /api/cron/pipeline endpoint is kept for manual full runs (logs as source "manual"). Each stage logs its own usage_log row.

v0.6.36 · 2026-05-25

Changed

• Methodology page de-hyped toward a lab-notebook voice. Rewrote the intro from marketing framing ("the way you'd want it measured", "source of truth") to a factual statement of what the page documents; softened "hand-curated" → "curated". The rigorous middle (formulas, channel-skew honesty, known limitations) and the bottom "Why this exists" mission section are unchanged - the goal was to keep hype away from the method. Per reader feedback that the page mixed marketing jargon with the actual methodology.

v0.6.35 · 2026-05-25

User-feedback clarity pass on two charts (methodology rewrite tracked separately).

Changed

• Biggest Movers redesign. Added column sub-headings (Issue · Last week · This week · Change). Replaced the ambiguous ↑/↓ delta with a neutral ←/→ arrow showing direction of movement on the left–right axis, decoupled from position (which keeps its L/R color). Added a one-line decoder so it's clear an issue can move right yet still sit in left territory. Per user feedback that the chart was hard to decipher.
• Index contributions chart caption rewritten in plain language. Removed the inline Σ(...) formula (now linked to Methodology) and clearly distinguishes the bar ("how much the issue moved the Index") from the number ("average lean"), since a reader found the old wording opaque.

v0.6.34 · 2026-05-25

Fixed

• Logo alignment nudge. Wordmark moved up another 1px (−2px total) to sit centered against the crate icon.

v0.6.33 · 2026-05-25

Fixed

• Header on mobile + logo alignment. The nav row didn't wrap, so on narrow screens it overflowed and overlapped the wordmark. The header now wraps (nav drops below the logo on mobile) with a tighter mobile gap. Also nudged the wordmark up 1px so it sits centered against the crate icon.

v0.6.32 · 2026-05-25

Fixed

• Issue/channel trend charts: width + vertical range. The chart was capped at max-w-md, so inside the wide drill-down cards it filled only the left half. IndexAreaChart now takes a maxWidthClass prop (default max-w-md for the home hero; "" on the drill-downs so it fills the card). It also takes includeZero (default true): the home Index keeps its 0-anchored range, but issue/channel charts now fit to their own data - an entity that sits far from neutral (e.g. a channel at L+4.8) uses the full chart height instead of squashing the line into a third with dead space above it. The zero reference line is hidden when not anchoring to zero.

v0.6.31 · 2026-05-25

Changed

• /log table polish. Status legend moved from the pagination footer to above the table (right-aligned), so it's visible before scrolling. Date column now numeric (MM/DD/YYYY) instead of spelled-out month. Both changes also apply to the "Recent episodes" table on channel pages (shared component).

v0.6.30 · 2026-05-25

Added

• Trend charts on issue and channel pages. The home-page IndexAreaChart (Recharts/shadcn) now also appears on /issues/[slug] ("How this issue has trended") and /channels/[id] ("How this channel has trended"), showing the entity's rolling lean over the last 30 days. New reusable rollingLeanTrend() helper in aggregate.ts (same daily-rolling, trailing 7-day-window logic as the home sparkline, scoped to a single issue/channel); getIssueDrillDown / getChannelDrillDown now return a trend series. The chart is hidden when there are fewer than 2 points.

v0.6.29 · 2026-05-25

Fixed

• Classify reprocessing loop (head-of-line blocking). The cron + CLI classify queue was defined as "transcripts with no classification row." An episode that yields 0 mentions never got a row, so it stayed "pending" forever and was re-sent to Sonnet every run. The first 15 pending happened to be genuinely off-taxonomy (sports, true crime, celebrity, stale/junk clips), so they permanently clogged the CLASSIFY_LIMIT=15 batch - ~$1/run for 0 new classifications, while newer classifiable episodes behind them were never reached and the backlog never drained. Diagnosed from live data: 142 pending, 350K input tokens → 60 output tokens across 15 episodes (model correctly returning []).
• Fix: new episodes.classify_status column (migration add_episode_classify_status). It's set to processed after each classify attempt regardless of mention count, and the pending queue keys off it. 0-mention episodes are recorded as done and never re-sent; the batch advances and the backlog drains. Backfill marked the 946 episodes that already had classifications as processed; the never-reached remainder stay pending so they're classified properly (not skipped). A partial index keeps the pending-queue scan cheap. Applied to both runClassify (cron) and scripts/classify.ts (CLI).

v0.6.28 · 2026-05-25

Changed

• "Biggest movers" now requires a minimum sample. An issue must have at least MOVER_MIN_MENTIONS (25) classifications in both the current and prior 7-day window to qualify as a mover. Previously a quiet week could produce a large, noisy lean swing and headline the card on a thin sample (e.g. an 18-mention week outranking a 400-mention one). The swing is only trustworthy once each side of the comparison has enough rows behind it. Verified against live data: the change correctly drops a thin-week issue and promotes a large-sample swing to #1. Mention counts on the issue drill-down (30-day window) were spot-checked against the DB and match exactly.

v0.6.27 · 2026-05-25

Housekeeping: finish the v0.6.26 dead-code removal and track the dev guide.

Removed

• IndexSparkline.tsx and EpisodeList.tsx - v0.6.26 emptied these to stubs but never git rm'd them. Nothing imports either; deleting the files completes that release's intent.

Added

• CLAUDE.md is now tracked in the repo - the working guide for Claude Code (commands, release ritual, guardrails, infra facts). Previously untracked/local-only.

v0.6.26 · 2026-05-25

Pre-beta audit: stale content, dead code, docs.

Changed

• Activity is back in the top nav (it was footer-only) now that /log is a real surface.
• Methodology page refreshed: removed the stale Bannon transcript-coverage limitation; corrected "15 issues" → 16 (incl. the Iran conflict); replaced the unvalidated "85–90% accurate" claim with an honest note that scoring is model-produced and being calibrated against an independent human gold set.
• ARCHITECTURE.md + README brought current: Supadata (not the old scraper), RLS-on/no-policies + service-role + no-store fetch, shadcn/TanStack UI, real cron batch limits, the UNIQUE(classification_id) constraint, episode_pipeline_summary view, gold-set tables, and the /admin surfaces.

Removed

• Dead code: IndexSparkline (replaced by IndexAreaChart) and EpisodeList (replaced by EpisodeDataTable), plus the now-unused getRecentEpisodes / getEpisodesForChannel / attachPipeline helpers in episodes.ts. (The two component files are emptied here; git rm them.)

v0.6.25 · 2026-05-25

Fixed

• /log header sort-arrows overflowing. A constant sort arrow on every column header spilled into the neighbouring header on tight columns. The arrow now appears only on the actively-sorted column (standard data-table pattern), with overflow-hidden on the header cells as a safety. Headers fit cleanly.

v0.6.24 · 2026-05-25

Home-page trend chart + /log header fix.

Added

• Interactive Index trend chart on the home page - a Recharts area chart (via shadcn chart primitives, src/components/ui/chart.tsx) replacing the static SVG sparkline under the needle. Shows the 30-day rolling Soapbox Index with hover tooltips (date + L/R value), a neutral zero baseline, an L/R-oriented y-axis, and a range caption. Adds the recharts dependency.

Fixed

• /log table header overlap. v0.6.22's column percentages fit the cell contents but not the header words, so narrow headers ("Category", "Transcribed") overflowed and their sort arrows bled into the next column. Rebalanced the widths (still summing to 100%) so every header fits cleanly.

v0.6.22 · 2026-05-24

Online gold-set labeling + /log table polish.

Added

• Online scoring-calibration tool (/eval/label) to replace the CSV gold set. Multiple independent labelers score the same blinded items (lean-coded source only, no channel name / model score / ids) on sentiment (−5…+5), intensity (1…5), confidence (1–3), + notes - instructions and the three calibration examples are built into the page. Shared link + name to start; forward-only and resumable. New gold_items / gold_labels tables (migration 20260524000002), seeded by npm run seed:gold-set (same stratified sample as the CSV exporter; model answer frozen per item). Submissions go through a server action on the service-role client - no client-side DB access; the page is noindex.

Fixed

• /log table no longer scrolls horizontally. Switched the fixed-layout column widths from pixels (which summed wider than the container and forced a scrollbar) to percentages that sum to 100%, so the table always fits. Long titles/channels truncate with tooltips. Also aligned the page back to the site-standard width, made channel names link to the channel page, and swapped native title tooltips for Radix tooltips.

v0.6.21 · 2026-05-24

/log cleanup: admin split, shadcn/ui, real data table.

Added

• shadcn/ui adopted as the component system (the codebase already used cn, clsx, tailwind-merge and shadcn-style markup). Added theme tokens to globals.css + tailwind.config.ts (additive - existing literal-gray pages unaffected), components.json, and src/components/ui/: button, input, table, badge, dropdown-menu.
• Episode receipts → a real data table (EpisodeDataTable, TanStack Table + shadcn). Columns: category (L/M/R), date, channel, video, type, length, and Transcribed / Classified / Scored status (colored dots with Radix tooltips) - all sortable, with search, pagination, and a column-visibility menu. Channel names link to the channel page. The channel drill-down's "Recent episodes" reuses the same table (Category + Channel columns hidden).
• episode_pipeline_summary view (migration 20260524000001) computes per-episode classify/score counts in Postgres, so /log loads one light result set instead of thousands of join rows.
• Admin nav (AdminNav) across the gated /admin/* tools.

Changed

• Pipeline health moved to /admin/pipeline - it's operational detail for internal consumption, not public. The public /log is now scale + searchable episode receipts only.

v0.6.20 · 2026-05-24

Data integrity: one score per classification, enforced.

Fixed

• Duplicate sentiment scores. Overlapping score runs (CLI + the cron's score stage + the daily cron) raced: each read a classification as unscored and inserted, with no unique constraint to stop them. Result was 257 duplicate score rows across 172 classifications, double-counting in the Index and per-issue/channel aggregations. Deduped (kept earliest per classification; scores 8,031 → 7,774, now exactly 1:1 with classifications).

Added

• UNIQUE (classification_id) on sentiment_scores (migration 20260524000000) - duplicate scores are now structurally impossible.

Changed

• Score insert → upsert with onConflict: classification_id, ignoreDuplicates in both scripts/score.ts and the cron runScore, so overlapping runs no-op cleanly instead of erroring against the new constraint.

Note

• Also closed out the failed-YouTube recovery: 369 of 370 episodes that the broken cron had marked failed were re-transcribed via Supadata and flowed through classify + score. Only 1 was genuinely caption-less.

v0.6.19 · 2026-05-24

/log reworked into the public pipeline + scale transparency page.

Added

• Pipeline health on /log. New PipelineHealth component surfaces the usage_log data that previously only lived on /admin/costs: four per-stage status cards (ingest/transcribe/classify/score) showing each stage's current health in plain English plus a small last-7-run trend strip, and a detailed recent-runs table with per-stage counts and any error message. Operators can see at a glance which stage is broken; users get real transparency. Shows no cost/token data - that stays on the operator-only /admin/costs.
• Per-episode pipeline progress in the receipts list. EpisodeList now shows each episode's progress through all four stages (Ingested → Transcribed → Classified → Scored) with done/failed/pending/partial state, instead of a single transcript-status badge. Makes the failed-YouTube backlog and where each episode stalled visible at a glance. Applies on both /log and the channel drill-down.

Changed

• System scale moved from /channels to /log and redesigned. New lineup: shows tracked (with L/M/R split), episodes analyzed (of total ingested), hours of audio, issue mentions, sentiment scores, and coverage-since date. Dropped "words transcribed" (redundant with hours). Header now shows data freshness. Counts are live, so they reflect the post-dedup classification total correctly.
• Fixed a latent early-break in the hours-of-audio pagination (same class of bug as the classify/aggregate pagination issues): now advances by rows returned and stops only on an empty page.

v0.6.18 · 2026-05-24

Cron is end-to-end. Cleanup + the real root cause.

Fixed

• Cron transcribe now succeeds. The actual blocker was operational, not code: SUPADATA_API_KEY was missing/empty in the Vercel runtime, so getVideoTranscript threw and every YouTube episode was marked failed. (Masked until now because the v0.6.16 key/cache bugs kept the cron from ever reaching the Supadata call.) Set the value in Vercel; a seeded episode transcribed cleanly (succeeded: 1, ~1.4s round-trip, Supadata credit consumed). With this, the full pipeline runs unattended: ingest → transcribe → classify → score.

Changed

• runTranscribe no longer swallows errors in a bare catch {} - failures (missing env var, Supadata outage) are now logged. This is what would have surfaced the SUPADATA_API_KEY problem on day one.

Removed

• Temporary transcribe diagnostic logging (served its purpose locating the Supadata-key failure).

Note

• The v0.6.17 platform-by-map change is retained as a robustness improvement, but it was not the root cause - the missing env var explained the failure on its own. No evidence the channel embed was actually broken.

v0.6.17 · 2026-05-24

The last link in the cron chain.

Fixed

• Cron transcribe failed every YouTube episode without calling Supadata. Once v0.6.16 fixed the key and the cache, the cron could finally see pending episodes - but it still marked them all failed in ~50ms, never reaching Supadata. Root cause: the channel:channels!fk(platform) embed didn't expose .platform reliably at runtime, so row.channel?.platform was undefined and every episode flunked the === "youtube" guard. Replaced the embed with a direct channel_id → platform map in runTranscribe, which is embed-shape-proof. (The classify/score stages still use channel embeds; auditing those is a separate follow-up.)

v0.6.16 · 2026-05-24

Pipeline reliability release. Three compounding bugs kept the production cron from doing useful work while the CLI worked fine; all are fixed here.

Fixed

• Stale cached reads (the big one). Supabase-js issues reads as fetch GETs, and the Next.js App Router caches fetch by default - so every server-side read (the cron and server components) was frozen at the first snapshot taken after each deploy. The cron reported identical results across separate runs (pendingFound: 1504 twice while the live table was at 552) and never saw its own writes or the CLI's. db.ts now forces cache: "no-store" on every Supabase request so reads always hit the live database. force-dynamic on the route did not reliably cover the client's fetches; forcing it at the client is the durable fix.
• classify dedup pagination. scripts/classify.ts built its "already classified" set with .limit(50000), which Supabase silently caps at the project Max Rows (1000). Once the classifications table grew past ~1000 rows the dedup set was incomplete, so episodes were re-classified on every pass - catastrophic under a loop (a catch-up run reclassified 234 episodes ~95× into 24k duplicate rows before being caught). Now paginates via .range() and terminates only on an empty page. The runaway duplicates were cleaned up out-of-band.

Added

• scripts/catchup.sh - full-pipeline drain that runs ingest, then loops transcribe/classify/score until each queue empties, with hard per-stage iteration caps so a logic bug can't run away unattended.

Operational (no code)

• Corrected Vercel's SUPABASE_SERVICE_ROLE_KEY: it held a legacy anon JWT, not a service-role key. With RLS enabled on all tables and zero policies, an anon key reads/writes nothing - which is why the cron saw an empty database while the CLI (real service key) worked. Swapped to the new sb_secret_… key. (RLS-with-no-policies is a latent landmine to address separately.) CRON_SECRET rotated.

Removed

• Temporary transcribe diagnostic logging from the cron route.

v0.6.14 · 2026-05-14

The actual fix for YouTube transcripts: swap the unmaintained scraping library for a managed transcript API.

Background

The youtube-transcript npm library had two compounding problems:

1. Library bug. Returned "Transcript is disabled on this video" for videos that demonstrably have captions on YouTube. Documented issue since mid-2024 - the library does HTML scraping and breaks whenever YouTube changes the embedded ytInitialPlayerResponse shape.
2. Cloud-IP blocking. Even when the library worked, YouTube throttled responses from Vercel and GitHub Actions egress pools. Strip the captionTracks field silently, library reports "disabled," scraper-aware infrastructure has a bad day.

A half-day of misdiagnosis (v0.6.4 through v0.6.13) chased these two intertwined issues separately. v0.6.4 flipped ordering to oldest-first hoping caption-timing would resolve. v0.6.13 moved transcribe to GH Actions hoping IP rotation would. Neither fixed it because they were both treating symptoms of two problems as if they were one.

Changed

• Transcripts now fetched via Supadata (https://supadata.ai), a managed YouTube transcript API. We hit GET /v1/transcript with the YouTube watch URL; they handle scraping, proxy rotation, library maintenance - everything we were doing badly. ~$17/mo on the Pro plan for our ~3000-transcript/month volume. Uses mode=native so we only fetch existing captions, never pay for AI generation.
• youtube-transcript package removed from dependencies. Was the source of the cascading failures.
• Transcribe stage re-enabled on Vercel cron. The reason we moved it to GH Actions in v0.6.13 (cloud-IP blocking) doesn't apply when we're calling Supadata's API rather than scraping YouTube directly. Vercel cron now handles the full pipeline again: ingest → transcribe → classify → score in a single 10:00 UTC run.
• GH Actions transcribe workflow demoted to manual-only. Kept around as an escape-hatch trigger for ad-hoc catch-up runs, but no longer scheduled. Now requires SUPADATA_API_KEY repo secret.

Setup steps

1. Add SUPADATA_API_KEY to Vercel environment variables (Settings → Environment Variables → New).
2. Add SUPADATA_API_KEY to GitHub repo secrets (only needed if you'll use the manual workflow).
3. Push v0.6.14. Tomorrow's 10:00 UTC Vercel cron will use the new path.

v0.6.13 · 2026-05-14

Architectural fix for the YouTube-on-Vercel transcript problem: transcribe stage moves off Vercel onto GitHub Actions.

Diagnosed

The youtube-transcript library was reporting "Transcript is disabled on this video" for videos that actually have captions available on YouTube. Manual spot-check of three failed-from-Vercel videos showed two with auto-generated captions, one with owner-uploaded captions - all present and visible on the YT site. Gregg's home network successfully transcribed the same videos.

Root cause: YouTube's anti-scraping behavior silently degrades the watch-page response for IPs it flags as suspicious. The page loads, but the captionTracks field in the embedded JSON is stripped out. The library can't distinguish "captions were never there" from "captions were hidden from this IP" and reports both as "disabled." Vercel's egress IP pool is flagged; home networks and CI runners generally aren't.

Changed

• Transcribe disabled on Vercel cron. Without this, Vercel's 10:00 UTC run would mark today's pending YT episodes as failed, poisoning the queue before GH Actions runs at 10:30 UTC. The cron now skips the transcribe stage entirely and writes a no-op stage record to keep usage_log shape stable.
• New .github/workflows/transcribe.yml runs daily at 10:30 UTC, invoking npm run transcribe -- 100 from GitHub's runner IP pool. Also exposes a workflow_dispatch trigger so it can be run manually from the Actions tab. Requires two repo secrets: NEXT_PUBLIC_SUPABASE_URL and SUPABASE_SERVICE_ROLE_KEY.

Pipeline architecture, post-v0.6.13

10:00 UTC (Vercel) ingest → classify → score 10:30 UTC (GitHub Actions) transcribe

End-to-end latency from publish → scored:

• Podcast (PodScan inline transcript): ~24h
• YouTube (GH transcribe → next-day Vercel classify): ~25h

Setup steps to activate

1. Push v0.6.13.
2. In the GitHub repo: Settings → Secrets and variables → Actions → New repository secret. Add:
   • NEXT_PUBLIC_SUPABASE_URL
   • SUPABASE_SERVICE_ROLE_KEY
3. Manually trigger once via Actions tab → "Daily transcribe" → Run workflow, to validate the runner IPs work for the YouTube scraper.
4. Confirm success rate is high (>80%) before relying on the schedule.

v0.6.12 · 2026-05-14

Transcribe throughput bump: TRANSCRIBE_LIMIT 10 → 40 per cron run.

Changed

• TRANSCRIBE_LIMIT raised from 10 to 40. Diagnostic SQL showed ~49 YouTube episodes ingested per 24h vs only 10 transcribe attempts per cron - the pending pool was growing by ~90/day with the rest of the pipeline starved for fresh content. 40 attempts at ~1s each adds ~30s to cron wall time, still well inside the 300s function budget. Won't fully close the gap with daily YT ingest, but cuts the daily growth rate substantially while the v0.7 retry mechanism is built.

Known limitation

At ~30-40% transcribe success rate (legitimate disabled-captions long-tail) and 40 attempts/day, throughput is still loss-making relative to ~100 YT ingests/day. The real fix is v0.7: retry-after- N-hours so fresh same-day caption failures get a second chance, and possibly a hybrid newest+oldest ordering strategy so today's discourse isn't permanently stuck behind a slow-burning backlog.

v0.6.11 · 2026-05-14

Root-cause fix for the 81% score-stage failure rate observed on May 13 and May 14 crons. Diagnosed via the logging added in v0.6.9.

Fixed

• Haiku's positive-number "+" prefix is now tolerated. Score output was arriving as {"sentiment": +4.2, "intensity": 3} - Haiku "helpfully" prefixing positive numbers with a plus sign. JSON spec doesn't allow a leading +, so JSON.parse rejected the entire response. Added normalizeLlmJson helper that strips leading + in JSON value positions (after : , or [, before a digit). Doesn't touch + inside string literals.
• Score prompt updated to explicitly instruct "no leading + on positive numbers" plus three worked examples (negative, positive, zero). Prevents the issue at the source; the parser fix is the defensive belt to the prompt's suspenders.

Impact

This was the silent failure path that produced ~13 of 16 failed score attempts on each of the last two cron runs (~80% loss rate). With v0.6.11 deployed, the next cron should score at the ~98% success rate we saw in yesterday's local catch-up run.

v0.6.10 · 2026-05-14

Sparkline expansion. Adds context to the home page trend line without changing its visual character.

Changed

• Reference lines at ±5 in addition to the existing dashed-zero line. Gives the eye a magnitude anchor at a glance - previously every value just floated relative to whatever the data range happened to be.
• Endpoint date labels under the chart. "Apr 20 ───── May 14" so the time range is readable without crossing back into the surrounding copy. Drives off a new sparklineDates field on DashboardData, populated in parallel with the values themselves.
• Range summary beneath the dates: Range L+0.3 to L+1.4 · rolling 7-day index. Replaces the previous "24-day history · rolling 7-day Index" label, which only said how many days of data existed without saying what was in it. Sorted most-L to most-R for natural reading across crossings of zero.

Technical

• aggregate.ts now returns sparklineDates: string[] alongside sparkline: number[] - same length, same order. Days with no data are skipped in both arrays so they stay in sync.

v0.6.9 · 2026-05-14

Score-stage error logging. Mirrors the transcribe-stage logging added in v0.6.4. Diagnostic for the 81% score failure rate observed in the May 13 and May 14 cron runs.

Fixed

• Score errors now surface in logs. runScore was using bare catch {} and ignoring failed sentiment_scores inserts silently. Now logs [score] <ErrorClass> for classification <id>: <message> on any thrown error from scoreClassification (Anthropic API errors, JSON parse failures, rate limits), and a similar line for Supabase insert errors. Next cron run will tell us why score has been failing 13/16 attempts.

v0.6.8 · 2026-05-13

Hotfix for v0.6.7: the OG image build failed on Vercel because Satori (the renderer behind next/og) does not support <text> SVG elements. The three gauge endpoint labels (L 10, 0, R 10) have been moved out of the SVG and rendered as regular HTML below the gauge using flex positioning. v0.6.7 tagged but did not successfully deploy.

v0.6.7 · 2026-05-13

OG image visual alignment with the live site.

Changed

• OG image now matches the actual home page identity. Previous v0.6.6 render used a flat horizontal gradient bar for the needle and omitted the crate logo. Replaced with the same half-circle gauge SVG the home page renders (identical geometry, gradient stops, tick marks, and needle), and added the wooden-crate logo to the top-left brand row. Visiting the live site after seeing a share now feels continuous rather than disjoint. The crate is inlined as base64 from the 256×256 favicon asset (~56KB payload, much smaller than the 1024×1024 source).

v0.6.6 · 2026-05-13

Social-share polish and brand attribution.

Changed

• Page title rewritten. Was "Soapbox: The FiveThirtyEight of Alternative Political Media" - that framing was useful internally as a north star but is the wrong thing to put on a tab, invites premature comparison, and the product should stand on its own. Now: "Soapbox · Alternative media discourse, quantified".
• Meta description aligned with home page hero. Identical wording so the social-share blurb matches what visitors see on landing.
• Footer tagline updated to match: "alternative media discourse, quantified" replacing the older "alt-media discourse, updated daily."
• Built by Breakfastball LLC · © 2026 attribution added as a second footer row, gray and small, so it doesn't crowd the nav.

Added

• Dynamic Open Graph image at src/app/opengraph-image.tsx. When the site URL is shared (iMessage, Twitter/X, Slack, LinkedIn, etc.) the preview card now shows the live Soapbox Index value, a needle bar, channel + episode counts, and the as-of date - generated at request time via next/og's ImageResponse and cached for an hour per URL. Every share becomes a data preview of the current state of alt-media discourse. Twitter card configured to use the same image via summary_large_image.

v0.6.5 · 2026-05-13

Home page UX polish and scoring-evaluation tooling.

Changed

• Home page hero copy tightened. "alt-media" → "alternative media" throughout. Subtext replaced with a sharper one-line explainer: "Soapbox is a data platform that uses language models to quantify what major alternative media is saying about US policy issues. We ingest and process new episodes daily."
• Logo trimmed. Crate dropped 36px → 32px with a tighter gap to the wordmark; the previous size felt heavy against the wordmark weight.
• "Why is the Index where it is?" moved to the home page. The per-issue contribution chart now lives directly under the hero needle/number so the explanation stays adjacent to the headline it explains. Methodology page links back to it. Window aligned to the same 7-day rolling period as the Index number above it.

Added (tooling, not user-facing)

• Independent scoring validation package. New eval/ directory with LABELING_INSTRUCTIONS.md (a 4-page methodology brief for an outside labeler) and scripts/extract-gold-set.ts (stratified sampler that emits two CSVs - a clean labeler version with channel names blinded to lean, and an internal answer key with model scores). Run with npm run eval:extract-gold-set. Designed to validate the Haiku scorer against independent human judgment; output feeds the v0.7 prompt audit.

v0.6.4 · 2026-05-13

Transcribe reliability fix. Cron's transcribe stage was burning its TRANSCRIBE_LIMIT on the freshest YouTube uploads of the day, which typically don't have auto-captions generated yet. Those failed and got marked permanently failed (no retry logic). Older pending episodes - which actually do have captions ready - were starved.

Fixed

• Transcribe order flipped to oldest-pending-first. Both src/app/api/cron/pipeline/route.ts and scripts/transcribe.ts were ordering by published_at DESC (newest first), exactly the worst order given YouTube's caption-generation latency. Flipped to ASC. Trade-off: ~24h latency between an episode being published and being transcribed, which is fine for a trailing 7-day Index aggregate.
• YT transcript errors now logged. getVideoTranscript was using bare catch { return null } which made every failure invisible. Now logs error class + message so Vercel logs tell us why a fetch fails.

Known followup (v0.7)

Failed transcripts are still permanently failed - no retry. When transcript_attempts + transcript_last_attempted_at columns are added, "failed" becomes retryable until N attempts spread over M hours. Tracked in memory under v0.7 queue.

v0.6.3 · 2026-05-12

Critical data-correctness fix: fetchScoreRows() pagination was silently dropping ~46% of sentiment_score rows in production, causing channel drill-down pages to show 0 mentions when the underlying data was present.

Fixed

• fetchScoreRows() pagination bug. The terminator condition if (data.length < pageSize) break interpreted any short Supabase page as end-of-data. Vercel's edge→Supabase route was returning short pages (response-size cap hitting before the row cap, due to the deep nested join), causing premature termination. On the live site this manifested as the prod Soapbox Index reading from 1,314 of 2,444 scores while local dev saw the full set. Channel drill-downs for recently-classified channels showed empty. Fix: terminate only on empty responses (not short ones); add explicit order("id") so pagination is deterministic; cut pageSize from 1000 to 500 to keep individual responses comfortably under any size cap; add a 50-page safety bound to prevent runaway loops.

v0.6.2 · 2026-05-12

Operational tuning: cron batch limits raised so we actually keep up with the daily ingest backlog.

Changed

• Cron batch limits raised - CLASSIFY_LIMIT 2 → 15, SCORE_LIMIT 30 → 80. Original limits would have taken ~75 days to burn down a single day's 150-episode ingest backlog. New limits target a 1-week catch-up rate while staying ~45s clear of the 300s function timeout. Stage timing observations documented inline in src/app/api/cron/pipeline/route.ts.

v0.6.1 · 2026-05-12

Same-day branding + transparency-surface polish on top of v0.6.0.

Added

• Brand identity - wooden-crate logo + red/blue soapbox wordmark (red #C8202F on "soap", blue #114A8A on "box") replacing the plain text mark. Logo source-of-truth at src/assets/logo-crate.png, served through next/image with priority + blur placeholder (~5KB delivered at retina). Favicon auto-detected from src/app/icon.png (256×256).

Changed

• Activity moved to footer - the /log link lives in the footer alongside Issues / Channels / Methodology rather than the top nav. Activity is a transparency surface, not a primary destination.
• Trust strip totals aligned with /channels SystemStats - both now report cumulative channel + episode counts rather than mixing in-window counts with all-time. "Episodes in window" → "Episodes tracked".
• .media removed from header - top-of-page brand mark is now the wordmark alone; the .media TLD was redundant next to the logo.

v0.6.0 · 2026-05-12

Post-MVP foundations release. Same-day as v0.5.0; bundled because all of this work shipped in a single extended session.

Added

• Admin tooling (Basic Auth gated via middleware against ADMIN_PASSWORD):
  • /admin/costs - Anthropic spend dashboard. Daily/weekly/monthly burn vs $1k budget cap, 30-day daily bar chart, recent-runs table. Backed by a new usage_log table written from the cron pipeline.
  • /admin/channels-audit - three views to guide channel curation: publishing cadence per show (last 14 days), L/M/R coverage gaps by issue, and "mentioned but not tracked" report scanning supporting quotes for candidate voices.
• PostHog product analytics - client-side init + manual pageview capture for the App Router. Autocapture / heatmaps / web vitals on; session recordings off.
• Public /changelog page - renders CHANGELOG.md directly via react-markdown so the file remains the single source of truth. Footer version pill links here.
• Public /log activity feed - paginated 50/page; every episode the pipeline has ingested with status badges + link to source. Receipts for transparency.
• Per-channel episode list on channel drill-down pages - last 25 episodes for that show with publish date, duration, transcript status, source link.
• External-link affordance per channel - every channel card on /channels and the drill-down page links out to YouTube or Apple Podcasts.
• Shared Header + Footer components - DRYed out the inline JSX across all pages; nav changes are now a one-line edit.
• Version surface - v0.x.y pill in every page footer linking to the on-site changelog. src/lib/version.ts is the single source of truth.

Channel curation

• Added 9 channels: Shawn Ryan Show (R), Real America's Voice / RAV (R), The Rubin Report (R), Hodgetwins (R), More Perfect Union (L), Democracy Now! (L), Heather Cox Richardson (L), Aaron Parnas (L, below 500k YT threshold but flagged for cross-platform reach), Call Me Back with Dan Senor (M).
• Pinned PodScan IDs for Joe Rogan and Ben Shapiro to fix stale-feed resolution that was returning episodes from 4-10 months ago. New podscanPodcastId field on the SeedChannel schema bypasses search-based resolution for high-importance shows.
• Total channel rows: 60 (40 unique shows after grouping; ratio reflects shows tracked on both YouTube + podcast).

Changed

• Daily-cadence framing site-wide. Replaced "this week" / "weekly" copy with trailing-7-day-window language; Soapbox Index methodology refactored to use a rolling 7-day window updated daily by the cron.
• Auto-generated headline on home page below the needle, driven by the same per-issue contribution data shown on /methodology. Headline links to the contribution chart for "see why" drill-down.
• Em-dash sweep across all user-facing text. Replaced with appropriate punctuation (colons before lists, parens for asides, commas in flow). Per Gregg's site-wide style choice.
• Channels list grouped by show - same name across YouTube + podcast collapses to one card with platform indicators, eliminating the visual "duplicate channel" problem.
• Status badge clarity on activity log - "pending" renamed to "awaiting transcript" so casual visitors understand it as expected latency, not a bug.
• Hero subtext rewritten - sharper framing of why soapbox exists (alt-media now shapes US political discourse; not measured at scale; Soapbox listens above your personal algorithms).
• Issue contribution chart added to /methodology with auto-generated narrative explaining which issues are pulling left vs right.

Technical

• Vercel Cron /api/cron/pipeline endpoint runs the full pipeline at 10:00 UTC daily (6 AM ET). Writes a usage_log row at completion.
• src/middleware.ts enforces HTTP Basic Auth on /admin/*.
• Added react-markdown, @tailwindcss/typography, posthog-js.
• ARCHITECTURE.md - comprehensive live source-of-truth document. Maintained per non-trivial commit.

Vexes documented for vNext

• Cross-platform same-content duplicates: shows that publish identical content to both YouTube and podcast feeds get ingested twice. Future fix: dedup by (show + date + duration).
• Stale-feed PodScan resolution: name-search resolution can pick wrong feed when a show has changed feeds. Workaround in v0.6.0: explicit podscanPodcastId field on SeedChannel. Future fix: smart resolver that prefers the feed with most recent episodes.
• Reach is a snapshot at ingest time - need periodic re-fetch.
• Issue taxonomy fixed editorial - emergent-topic detection deferred.
• Twitch streamer ingestion still deferred.

v0.5.0 · 2026-05-12

Initial public release after the 5-day MVP sprint.

Added

• 49 hand-curated alt-media channels balanced across Left, Middle, and Right political-publishing posture.
• End-to-end pipeline: ingest (RSS + YouTube Data API), transcribe (PodScan inline + youtube-transcript for YT), classify (Claude Sonnet 4.6), score (Claude Haiku 4.5), aggregate.
• Soapbox Index: single L/R number for the trailing 7-day window of alt-media political discourse. Updated daily via Vercel Cron at 10:00 UTC (6 AM ET).
• 16-issue taxonomy with explicit left and right positions per issue (Iran-conflict added on launch day after Israel-Gaza and Trump/GOP were absorbing related content).
• Dashboard pages: home (Soapbox Index + sparkline + biggest movers + top issues), issue drill-downs, channel drill-downs.
• Methodology page with live "Why is the Index where it is?" per-issue contribution chart linkable from the home page.
• Auto-generated narrative headline below the needle driven by the same contribution data.
• System-scale stats banner on /channels showing hours of audio analyzed, words transcribed, issue mentions classified, etc.
• External "Visit on YouTube" / "Find on Apple Podcasts" link out per channel.
• Vercel Cron: full pipeline runs daily, idempotent against re-runs.

Technical foundation

• Next.js 14 (app router) + TypeScript + Tailwind + Geist sans on Vercel.
• Supabase Postgres backend with paginated reads + service-role server client.
• Claude Sonnet 4.6 for classification, Claude Haiku 4.5 for scoring.
• PodScan.fm for podcast transcripts (inline with episode metadata).
• YouTube Data API v3 + youtube-transcript npm package for YT.

Methodology disclosures

• Filter: YouTube videos under 3 minutes excluded from ingest (filters Shorts).
• Known gap: Bannon's War Room and Charlie Kirk Show (podcast feed) aren't transcribed by PodScan; metadata only.
• Classifier directional accuracy: ~85–90% at the per-row level.
• L/R position assignments per issue are editorial and reviewed quarterly.

Known limitations / vNext queue

• Channel reach is a snapshot at ingest time; needs periodic re-fetch.
• Issue taxonomy is fixed-editorial; emergent-topic detection is a v0.6+ design challenge.
• No admin tooling yet (cost dashboard, channel management); planned next.
• Episode-level transparency surfaces (per-channel episode list, public daily ingest log); planned next.