What we’ve shipped

In CSB John 1:26 "answered" is the same word KJV uses, but tapping it opened the KJV fallback instead of going straight to the Greek. Words in a quote that a translation moves — "'I baptize with water,' John answered them" versus "John answered them, saying, 'I baptize with water'" — now match.

The note column referenced a translation that was never passed into the editor, so Vercel typecheck failed and selah.so stayed on the previous release.

(e.g. "John 1 · WEB") so two panes of the same chapter are easy to tell apart.

"Pick translation" sits on the left; the settings gear sits on the right. The old empty footer is gone.

Set two favourites in Settings (web and iOS). The reader opens the first pane in your Primary edition, and a second column uses Compare instead of guessing.

when the link does not already name one.

with quiet P / S labels, and a gear opens Study settings on the web.

Yesterday's filter did not match the real error and 31 more were recorded across two readers before the morning check caught it.

Sidebar, search, and the page layout were rendering as bare unstyled content after a midsummer CSS cleanup; the Fumadocs theme is restored.

Opening a prayer no longer repeats “Active” in the header and footer; status lives in Overview, where you can still move it. Short titles sit closer to the folder chip instead of leaving a blank second line.

Reading plan (and the other tool sheets) no longer sit under a tall empty band above the name; Catch up rows are tighter; Mark read shares the line with today’s passage; and the help control sits beside the progress line instead of claiming a whole empty row.

, and you can open the translation picker while a tool sheet is up — chapter previous/next stay next to the passage control instead of stranded across the spacer.

Trying NASB or another Plus-only translation no longer shows a blank prompt.

Note cards go to the note, prayer cards land on that prayer in the prayer board, and reading-plan cards take you to your plans.

Files you upload can be marked as your own or as someone else's, with the speaker's name and a kind. Until now everything imported was treated as your own conviction, so a commentary you uploaded could end up on your belief map as something you believe.

A belief moves out of "Others" on its own the moment you write anything about it yourself.

Knowledge has a third view, Activity: a plain, day-by-day account of what has been read and what it found. While an import is running it sits at the top and tells you how far along it is. If something needed another look, it says so there rather than nowhere.

The map's sense of how developed a belief is was counting notes you'd recorded from other people, so reading widely on a subject looked the same as having worked it out yourself.

Most people using Selah are studying rather than preaching, and the assumption pushed it to read personal notes as though they were teaching.

Repairing an orphaned note — one carrying a date with nobody attached — through an unrelated Kind edit cleared the stored date but left it sitting in the date field. Name an originator without closing the popover and the field came back holding that ghost; closing then quietly discarded what you could still see.

In the Kind chip, opening "From someone else" and setting a date without naming who it came from saved that date on its own. The note then showed a bare date with no context, and Selah had no speaker to attribute the teaching to. The date field now waits until you have named someone, and says so.

Everyone who joined before today gets one pass through what changed: that long documents are read to the end, and that a belief you hold under two names is one entry again. It says plainly that nothing you wrote was altered, and links straight to your map so you can check it yourself.

Testing against the largest affected map showed the previous safeguard was too blunt: built to stop two different ideas being fused into one entry, it was also refusing plainly identical names that happen to contain "and", like "Law and Prophets" beside "Law and the Prophets". The repair now settles simple naming differences on its own, first: plurals, small words like "the", and reversed orderings such as "Works and Faith" beside "Faith and Works". The safeguard itself now blocks only the genuinely dangerous shape, a single idea being folded into a two-idea entry. The repair also runs until it stops finding anything, picks the better-established name as the survivor every time, and a congregation at a city is no longer treated as the city itself.

When the repair runs more than once, a later round can fold away an entry that an earlier round had already folded others into, leaving those earlier pointers aimed at a dead entry. Everything a user sees was unaffected, but the trail used to undo a fold could dangle. Pointers now follow through to the surviving entry automatically.

Reviewing the repair's proposals surfaced merges that would have flattened real distinctions: the Second Death of Revelation 20 into death generally, salvation by faith into salvation, glory into the glory of God, and God's wisdom into wisdom. The underlying rule is now stated both ways round: a broad theme and a specific form of it are different entries no matter which one Selah met first.

A merge that turned out to be already done was being counted as a failure; it is now reported as skipped, so a failure count of zero means zero things actually went wrong.

Selah was creating single entries like "Light and Darkness" for what are plainly two separate things, which meant neither could be followed on its own. It now keeps them separate, while still recognising that some names genuinely contain "and" and belong together, such as "Faith and Works" or "Sodom and Gomorrah".

Tribulation, which believers endure now, was being combined with the Great Tribulation, which is one particular period. Those stay apart, along with the same shape elsewhere.

If a note mentioned both of the entries being combined, the repair stopped partway. It now folds the two records into one, keeping the stronger confidence, the earlier date, and any human judgment over an automatic one.

A dry run against a real map surfaced two faults before anything was changed: it could accept a match the model picked from further down the list, well past the point where two beliefs stop being the same thing, and it could chain merges so one belief pointed at another that had itself been folded away. Both are closed, and it now refuses to merge across kinds that can never be the same thing, such as a person and a book of the Bible.

When we re-sourced the KJV and fixed 3,451 verses that had been silently cut short mid-sentence, the corrected text didn't appear for anyone already reading those chapters. The old text was cached with no expiry, so there was nothing to make the app look again. Chapters now refresh on their own within a day.

If you started the trial-intent onboarding flow, checked out with Stripe, but didn't finish the wizard in one sitting, coming back later showed the "Start your 14-day Plus trial" screen again — every time you logged in — even though you'd already paid.

Sermon, Study note, Devotional, Book, Podcast, Conversation — rename any of them, add your own, delete the ones you don't use. Nothing is locked.

/notes gains a kind filter alongside status and folder; combine all three.

If Selah Plus can't charge your card, you'll see it right away — not just buried on the billing settings page. A quiet banner appears across the app with a one-click way to update your card, and it stays out of the way once you've seen it. Your Plus access keeps working while Stripe retries; this just makes sure you actually notice.

Setting a note's type used to require naming a person first — the type only existed to describe someone else's material, so your own work had nowhere to go. Kind now stands on its own: Sermon by itself means you wrote it, Sermon · Pastor Mike means you heard it. Saying who it came from is an optional second step, not a precondition.

Ask it about your sermons or what you've been getting from podcasts and it can answer, because the kind now reaches it for every note rather than only for notes you'd credited to someone else.

Every companion response now carries a Report action next to Copy and Retry. It goes straight to us — no form, no ticket, one click.

A line under the composer says Selah can be wrong and isn't a substitute for professional or pastoral counseling. It quotes scripture and reflects on your journals and prayers, which is close enough to counsel that the boundary is worth naming out loud.

, taken from the note's own headings, as you scroll. It appears only once you've scrolled into a section, and only in a note that has headings.

Reading and note-taking stopped being two places. Open a note next to the chapter you're in and it's the real editor — slash commands, scripture blocks, the lot — saving as you type, with the passage still on screen. Clicking a note from the reader used to throw away the chapter you were reading; it doesn't anymore.

The column opens on your recent notes with a "This passage" filter beside it, so you can pull up something you wrote weeks ago about a different book without leaving the passage. The count is honest — if nothing here touches this chapter it says zero rather than hiding the tab.

Notes take the right-hand column and scripture stays left. Tap a verse reference inside your own notes and the passage jumps there, with a back chip to return to where you were.

The record control lives in the note's header, so capturing a session doesn't mean leaving the notes you're teaching from.

Flip the note out of editing and the text steps up to reading size with no caret and nothing to accidentally type into — for when you're speaking from it, not writing it.

The previous release fixed the "Write a thought" button; clicking the thought area itself still went down the update path and reported "Couldn't update — this highlight has no thought yet." Both routes save.

Selecting text and pressing Write worked; opening an existing highlight and pressing "Write a thought" did nothing at all — no message, nothing saved. Both paths save now.

The last release surfaced a message to you but the diagnostic detail never reached us, so a report still couldn't be traced. It does now.

It used to roll the highlight back and stop — the highlight appeared, vanished, and the Write button did nothing, with no message. Whatever went wrong is now reported, both to you and to us.

A highlight you made while writing turned bright yellow the moment you flipped to read it, and dividers changed shape too — both were falling back to the browser's own defaults instead of the app's.

, scaled to the larger reading type rather than reused from the editor.

Two failures used to do nothing at all — no message, no explanation — which is indistinguishable from a dead button. They now say what went wrong.

In the read view — the one you use when you're teaching from a note rather than writing it — verse references did nothing. Only the editing view responded. Both work now, by click or keyboard.

Converting an inline reference to a full passage produced a block that looked interactive and wasn't — in either the editing or the reading view. Tapping its reference now moves the passage, and tapping a specific verse number goes to that verse.

With Word Study up, picking Notes used to tell you there wasn't room and suggest hiding the sidebar. The 360px in the way belonged to the panel Notes was replacing, so it now closes it and takes the space.

In some conditions the editor showed "Saved · just now" for text that never reached the database. The save indicator was telling you the truth about its own state and nothing about your note. Fixed, and the tests that cover it now check storage rather than the indicator.

Highlighting silently did nothing while the notes column had focus, with no error to explain it.

"Hebrews 4:12–5:2" filed itself under Hebrews 4 and was missing from Hebrews 5.

The passage label on each row was sage-on-gold at 10px, which measured under the accessibility floor, and untitled notes were indistinguishable from each other. Rows now carry a one-line excerpt.

Picking Notes goes straight to the column beside the passage instead of the narrow panel with a button in it offering to do that.

, and it remembers the width on that screen. Arrow keys work on the handle too (hold Shift for bigger jumps, Home/End for the extremes). The passage keeps its reading width no matter how far you drag.

now, the way it does on the phone, instead of appearing inside the passage and pushing the verses down while you were reading them. Tap it to go back, or dismiss it.

Starring a note did it every time: star it, keep typing, and the next sentence never saved. Typing while a save was still in flight did it whenever the connection was slow. The warning was wrong — nobody else had touched the note — and the Refresh button it offered reloaded over the paragraph you were in the middle of.

When the companion or Telegram adds to a note you have open, what you were typing used to be replaced without warning. Both now survive: Selah's addition goes on the end, your sentence stays where it was, and your cursor doesn't move.

A note edited in a way that overlaps what you wrote now pauses saving and offers Keep mine or Use the latest, holding your unsaved work until you pick. Nothing is discarded on your behalf.

Renaming a note while offline used to drop the rename silently as soon as any later save succeeded. Whatever didn't land now rides along with the next save.

They saved and opened normally, but search returned nothing for them and Selah answered as though you had never written them. Every note is now indexed the moment it's created, and the notes already affected have been repaired.

A note's title is part of how it gets matched, but a rename never reached the index — and editing a note's body used to drop its title out of the index entirely, so the longer you worked on a note the harder it was to find by name. Titles now stay current.

Notes were indexed three different ways depending on how they were created, so whether you could find a note by its title was effectively luck. Every note has been re-indexed the same way.

Labels, inputs, and the type choices sat almost on top of each other, so the three settings read as one dense block. The panel now has room to breathe and each setting reads as its own group.

Sometimes you know a verse but not its address. Select the words as you jotted them — "grace teaches us" — and convert the selection: Selah finds Titus 2:11-12 and puts the real passage in its place. Founder-only for now.

A scripture block takes either form in the same box: a reference like john 3:16-21, or the wording as you remember it. Typing a reference is instant and exact; a phrase is searched against the Bible text first, and only if those words appear nowhere does Selah work out what you meant.

A reference in a sentence and a full passage quote were separate blocks that happened to hold the same verse. They're one block. Show full passage opens the verse out; Shrink to reference folds it back. The wording carries across either way, so nothing reloads and nothing is lost.

A reference in the middle of a sentence now reads as part of the sentence — quiet text with a soft underline — instead of a chip that interrupted it. It matches how references already looked in Stream, the reader, and shared notes, so scripture looks the same wherever you meet it. Full passages keep their gold edge: that's when the verse itself is the point.

— refresh it in your current translation, pick a different verse, switch how it's shown, or take it out.

The source panel used to open with three fields where two were greyed out, which read as broken rather than "not yet". It now asks only who taught or wrote the note; the type and date appear once you've named someone, and the option to clear only shows when there's a source to clear.

The type choices (sermon, book, podcast, conversation) previously gave no indication of which one was set. The active one is now marked, tapping it again clears it, and a type you typed yourself comes back the way you left it.

Content you've attributed to someone else no longer shows up in your Belief Map, and no longer shapes your daily devotional or proactive check-ins as if it were your own conviction.

A note like "preparing a sermon on Romans for next Sunday" has a shelf life. Selah's check-ins were still reaching for those after they'd passed; now they only use what's still current, the same way the rest of the app already did.

Your answers about your role, your translation, your tradition, and what you're studying were saved but effectively invisible when Selah went looking for what it knows about you. They now surface like everything else it remembers.

When you opened a related note in the Connections panel, the text appeared as an unstyled wall. It now renders with proper spacing, headings, and paragraph rhythm — the same as a normal note.

When you save notes from a sermon, a book, or a podcast, you can mark who taught or wrote the content — a pastor, an author, a host — and when they said it. Selah's companion then treats that content as someone else's teaching ("Pastor Mike taught…") instead of folding it into your own beliefs, so a sermon you wrote down never quietly becomes your position in a later conversation. Your own reflections in the same note still read as yours. Set it from the "+ Source" chip in a note's top row.

When a recording's speaker is marked as someone other than you, turning it into a note now fills in who taught it and the date automatically. You can edit or clear it.

Your message appears right away, and a quiet confirmation lets you know it's on its way.

When we answer, you'll hear about it — a dot on the "Need help?" link, plus a Telegram message or email if you're not in the app.

The sidebar (and the mobile drawer) now has a support link right above Settings. It opens a panel where you can send a message directly — replies show up right there, no email needed.

New pages under selah.so/features show what each part of Selah actually does — the reader, reading plans, journal, notes, library, knowledge map, prayer board, and companion — each demonstrated with the real interface, not a screenshot. A "Coming soon" section on the notes page previews recording a sermon, yours or someone else's, into a searchable note.

About 1 in 9 verses (3,451 of them, across 45 of 66 books) were silently cut short — Esther 8:17 used to end mid-sentence, dropping "upon them." Every verse is now complete, re-sourced from the same provider already powering the Berean and World English translations.

When Selah's background extraction gets unusable output from the model on one of your notes, it retries once, and if it still can't, it leaves your knowledge graph untouched rather than writing a partial record. The Knowledge "Sync" button now tells you when some notes couldn't be processed and offers a Retry, instead of silently showing a stale graph.

" question now offers Instagram and Facebook alongside Reddit, YouTube, a friend, and search, and dropped the em dash from its wording.

Instead of ending on a pricing screen, new people choose a reading plan, then step straight into the real Reader for a guided first study: select a verse, highlight it, write a thought (an example is filled in if you're not sure what to say, or skip it), and finish on a quiet "that's the heart of it" screen. The plan you pick begins the next morning in your own timezone. The Plus offer shows once, after that first saved thought, and never blocks the Reader.

The first 100 people to subscribe to Selah Plus get a permanent founding-member number and an invite to a private Telegram group to share feedback and talk directly with the team. A cancellation frees the slot for the next new subscriber.

Every recording hit an error because of how keyword hints were sent to the transcription service. Recordings transcribe reliably again, with a fallback so a bad keyword can never sink a transcript.

The product-analytics funnel showed landing-page visitors but zero signups, even though people were signing up. Anonymous visitors were tracked with an in-memory id that vanished the moment they navigated through signup, so the anonymous visit and the finished signup looked like two different people. Visitors now carry a durable first-party analytics id, so a landing visit stitches to the account it becomes.

The Google sign-in prompt used to fire on the marketing homepage on top of the cookie banner, before anyone showed any intent to sign in. It now appears only on the sign-in and sign-up pages.

The welcome email now points to one clear next step instead of three. The six feature-education emails dropped their bullet lists and Bible verses for a couple of sentences and one button. A few subject lines that read as guilt-driven, over-personifying the AI, or confrontational were replaced with plainer ones. The trial-ending email now states the exact price and billing period you'll be charged instead of just "your card."

"

so we can measure how people find Selah. Accepting the cookie banner additionally turns on analytics and advertising cookies (Google, Meta, Reddit); Decline opts out entirely, and Global Privacy Control is still honored automatically. Banner copy and the privacy policy were updated to match.

It used to concede original languages wholesale to Logos. It no longer does: Selah's word study — tap any word for its Hebrew or Greek, Strong's number, parsing, and lexicon entry from BDB, Abbott-Smith, or LSJ — is now a stated strength, next to a reading experience that feels natural and notes and highlights that are first-class rather than a sidebar. Logos's depth is reframed as weight (powerful, but heavy and years to learn), and the honest concession is narrowed to the one thing Logos genuinely owns and Selah doesn't: a library of books to buy. Two new comparison rows, a dedicated original-languages section, and two new FAQs.

The Belief Map's promised follow-through: every belief card now carries a "Sharpen this belief" button (and the Companion a /sharpen command) that starts a Socratic challenge session on that belief — iron sharpens iron. Selah gathers what you actually hold — your own words, your scripture anchors, the places your own notes already wrestle — and presses with one challenging, scripture-anchored question at a time. It will never tell you the doctrine is right or wrong, never grade your reasons, and won't budge even if you ask it point-blank for a verdict: the belief is yours to own before God and the text. When the sharpening lands somewhere, it offers to save the whole arc as a study note.

Selah now reports signups, trial starts, and purchases back to Meta the same way it already does for Reddit — matched to the ad click server-side, so campaign performance stays measurable even though conversions finish days later in a Stripe webhook. The browser pixel is consent-gated (nothing loads until a visitor accepts) and only tracks page views; every conversion is sent server-side.

A layout toggle arranges every concept on an organized ring, grouped by kind — doctrine, themes, people, places — with quiet group labels and connection arcs through the middle. Tap a belief and its connections light up in sage across the circle. Freeform stays the default.

A search control on the map flies the camera to a concept and opens its card — including concepts currently hidden by an active focus filter.

Drag concepts where you want them in Freeform; your placements survive reloads, and new concepts settle naturally around them. A quiet "Reset arrangement" action brings back the organic layout.

The belief card's new button opens the on-page Companion dock with a starter prompt already typed — edit and send. A draft you were already writing is never overwritten.

A caption reads like "125 beliefs · 4 you're wrestling with · Jesus Christ is your most connected," with tap-throughs. Concepts added this week carry a gold tick; beliefs you're wrestling with carry a warm underline. One Focus menu replaces the old row of filter chips.

Asking Selah to rewrite or remove a section — from the dock or the companion screen — previously dead-ended with a request to paste the note's ID from the URL. The companion now knows which note it's reading and which notes a search returned, so edits just work. Section edits also got stricter: if Selah can't tell which note you mean, it asks instead of guessing — it never rewrites a section of a note you didn't name.

Zoomed out, lesser labels rest as dots and fade in as you zoom — 200 concepts no longer collide into an unreadable pile, and labels cannot overlap at any zoom level. The map settles with a short exhale on load, flexes around your hand when you drag, and lights a concept's connections on hover. At rest it is perfectly still, and reduced-motion preferences are honored throughout.

The legacy hand-rolled companion loop is gone. Every surface — web dock, the Companion screen, Telegram, and the admin test console — now runs exclusively on the AI SDK 7 engine that went GA in 0.39.28.0. Renaming a chat now rides the same v2 API as everything else.

Mark today's reading done in the reader and the stream reflects it immediately, no reload — the two views now read from one shared source instead of tracking the day separately. Reopening the reader is also snappier: today's plan is cached, so it no longer re-fetches every time.

tsx, /api/companion/chat (GET/POST/PATCH; rename PATCH ported to /api/companion/v2 first), CORE_TOOL_SUBSET, dead normalizeCompanionHistory`. Net −2,604 lines.

A source-pin test now asserts the legacy symbols and flag strings stay gone.

When you chatted alongside a note, the note's full text was silently re-attached to every message you sent and stored that way, so the model saw the same note once per turn (and your own bubbles showed the pasted note on reload). Attachments now live in one deduplicated context block — each note appears exactly once per reply, re-read fresh so the companion always sees your latest edits — and your messages are stored clean. Telegram conversations benefit too: the replayed history is now capped, so week-long threads stop ballooning.

Highlight a line in a note (or a verse in the Reader), then click into the companion to type — your highlight used to vanish the instant you started writing, so "what does this mean?" arrived with nothing attached. The selection now stays put until you send, clear it, or move to another page, so your question lands on exactly what you were looking at.

Web dock, the Companion screen, and Telegram all route through the AI SDK v7 engine (runCompanionTurn) with the full 27-tool surface — prayers, devotional pivots, heartbeat config, and reader grounding tools are now available on the engine path, with the same prep-mode and grounding gates as before. companion_v2 and companion_v2_engine default ON; both remain readable as per-user kill switches until the legacy loop is deleted.

metadata`.

ts) on both the legacy loop and the v2 route; never drops the current turn, never opens on an assistant row.

lock actually moved to ai@7.0.17 / @ai-sdk/anthropic@4.0.9 / @ai-sdk/react@4.0.18 (PR #567 had bumped package.json + package-lock only; bun installs were still resolving ai@6, which also explains the 23 phantom tsc errors and the provider 404). tsc is now clean. Verified: 2320 tests, companion-chat eval 100%, spike:ai-sdk GO (cache read 5769 tokens on turn 2).

Plans now sit two to a row as cards, and "Build your own" — tell Selah a goal, get a plan drafted around it — leads the page instead of hiding below the catalog. The most-picked plan sits beside it with a quiet badge.

Hover a card under "Your plans" (on a phone, the control is always visible) and a two-tap Remove clears it out. Plans you've actually read stay put — those are history, managed from settings.

Open a tool panel and it rises as a sheet; tap the passage behind it and it slides away, without that tap also selecting a verse or jumping you elsewhere. Type into a note or word-study field and the panel lifts clear of the keyboard so you can see what you're writing.

The bar at the foot of the reader is now a floating, color-coded card — each section of the Bible in its own hue — and pulling up opens a compact chapter strip that keeps your target chapter centered as you sweep. On a trackpad or mouse the whole book is one swipe away.

Marking the final day read brings a quiet gold moment — "You finished what you started." — with a link to the full story on your Stream. Before, the reading bar simply disappeared.

If earlier readings are unread, the message closes with "3 earlier readings are set aside — they will keep. Read at your own pace." Being behind is no longer invisible, and never shamed.

The collapsible companion panel now runs on the same engine as the full companion screen and knows what you're looking at: highlight a verse or open a note and your next message carries that context, so "what does this mean?" lands on the passage in front of you instead of starting cold.

Asking for Obadiah, Philemon, 2–3 John, or Jude used to error out after several silent retries — the generator kept inventing chapters those books don't have. It now corrects itself on the first try.

Today's reading is remembered for a few minutes instead of re-fetched on every visit, and marking a day read anywhere clears it immediately — including the moment you switch accounts on a shared device.

When the companion wants to replace a section of a note — the one edit that's hard to unwind — the conversation pauses with a small card: "Selah wants to rewrite the section 'Application' of your note." Allow lets it happen; Not now skips it and the conversation carries on. Saving thoughts, prayers, and additions never ask — only full rewrites do. The step trail records what you decided, and answering works even if you come back to the conversation later.

Verses Selah sends now arrive as set-apart quote blocks instead of italics, and a full chapter comes through collapsed with a "show more" — where before it was dropped and only the reference survived. Scripture references in Telegram messages also link straight to that chapter in your reader.

The reader's book bar is now a true scrubber: press and slide to flip through the 66 books, color-coded by section — Law, History, Wisdom, Prophets, Gospels, Letters — then pull up to open a book and sweep through its chapters. Let go and you're there. On a phone, resting your thumb at the screen edge keeps a long book flowing past; with a mouse or trackpad the bar becomes a ruler for the whole book, so Psalm 119 is one sweep away instead of a dozen.

A tap still jumps instantly — and a small "Back" button follows it for a few seconds, returning you to the exact chapter you were reading.

Select multiple captures in your Journal and add them together, to a new note or an existing one, instead of one at a time.

Any companion session can become a real note from the session's ⋯ menu — your words set apart as quotes, Selah's replies as prose, with the summary and captures included. The note opens beside the chat, joins your library, and Selah can find it in her memory right away.

The note panel beside the chat now has an edit toggle: type directly, it saves as you go, and "Open in editor" is still there for the full page. If Selah edits the same note while you're typing, a calm banner lets you keep your version or take hers — nothing is lost either way.

Onboarding asks for one genuine thought — if you teach, about the passage you're teaching next — saves it as your first capture, and the companion responds to it with a single question in your own words. You arrive in the app with your Stream already started.

Every capture now has "Add to note": start a new note, pick a recent one, or take one of Selah's suggestions based on what the capture says. The original stays in your Journal.

Import three or more notes (Markdown, Word, PDF, Notion, Obsidian) and Selah maps the beliefs in them right away — you land on your Knowledge page and watch it fill in.

Settings → Privacy now downloads a zip of your whole workspace as Markdown: every note, your journal by month, and your prayer board. Free tier included.

Any note can become a clean one-page handout — scripture quoted in full, ready for print or save-as-PDF, with a small "Prepared with Selah" line you can turn off.

A shared note now previews with its title and passages, reads comfortably on a phone, and carries a quiet invitation at the end.

Tell Selah which day you teach and a card keeps your current note, the days remaining, and this week's captures in one glance.

New pages compare Selah and Logos plainly — including where Logos is still the better tool — with walkthroughs for bringing your study over from Logos or Obsidian ([selah.so/selah-vs-logos](https://selah.so/selah-vs-logos)).

A day-two note if you haven't captured anything yet, and a weekly digest of your study week for Plus members who aren't on Telegram. Both respect unsubscribe, both send at most once per window.

After a month of steady use, one quiet question — how disappointed would you be if Selah disappeared? — so the product grows from what real users say.

An intermittent formatting quirk in the model's response could drop a week's reflection entirely — it ran hourly in production logs. Those responses now recover instead of being discarded.

A long reply containing a quote could fail to send at all (you'd get a generic error); messages now split cleanly at any length.

— the companion gives up after a bounded wait instead of leaving you watching a spinner until the request dies on its own.

If autosave can't reach the server, "Couldn't save — Retry" now shows on every screen size, not just desktop.

on the notes list, instead of a blank row.

It kept reappearing after you dismissed it, and greeted new members about a change that predated them.

The gold glow after an AI edit no longer nudges your text while it settles, the slash-command menu stays on screen on phones, the sharing panel closes with Escape, and typing in it no longer loses focus mid-word when a background save finishes.

The conversation column now uses the same comfortable reading width as Notes, small tablets no longer squeeze the chat behind the sessions rail, the message box sits clear of the iPhone home indicator, and phones get sensible margins.

— the same gentle squeeze as everywhere else in the app, and it stays quiet if your device asks for reduced motion.

The sitemap and robots file had been redirecting crawlers to the sign-in page since launch — Google could never index the site properly. Both are now public.

A thought the companion saved for you now files under that session in your Journal instead of appearing as a loose entry.

When you share something meaningful, Selah now prioritizes offering to keep it over chasing a new thread.

A short mirror of your own words can stand on its own now, instead of every reply needing a question tacked onto the end.

Opening a note used to make three round-trips before anything appeared. Now the page arrives with your note already on it — one quiet loading moment, then your words. The browser tab shows the note's title, too.

While you're on the notes list, Selah quietly fetches the editor in the background, so opening a note doesn't wait on the download.

On a phone it now appears below your selection — clear of the system copy/paste bubble — and wraps to fit small screens. Every button can be reached with Tab and Enter, and the AI clean-up sparkle explains itself when the selection is too short instead of doing nothing.

Sharing, handout export, and AI clean-up live in one "…" menu on small screens; deleting asks with properly sized Yes and Cancel; and one accent color marks your active filters.

Picking a highlight color no longer dims and blocks the page: a slim tray rises from the bottom, your verses stay visible and adjustable above it, and the tray quotes exactly what you've selected — updating live as you move the selection handles.

Selecting verses, tapping words for word studies, and resizing the study panels no longer make the page stutter. Panel drags glide, word taps respond instantly, and the page stays still while you work.

Panel handles, cross-reference markers, and footnote letters all have properly sized touch targets now; buttons give the same gentle press feedback as the rest of the app; and the reader respects reduced-motion settings and the iPhone home indicator.

New hero, a week-rhythm walkthrough (Tuesday capture, Thursday study, Sunday teach), a plain "who Selah is not for" answer, and the promise that your notes leave with you, stated up front.

When you say something worth keeping, it offers once — "Want me to save that to your notes?" — and a no means no. It won't offer during interviews or for small talk.

A fresh Stream, Notes, Knowledge, or Prayer page tells you the one action that fills it, instead of just being empty.

Same-day fixes that took the Reddit tracking from shipped-but-inert to working: the pixel loads on selah.so and reports a page visit once you accept cookies, and the server-side conversion reporting is wired to Reddit's current API.

If you unsubscribed and later let a subscription lapse, Selah no longer sends the "come back" email. Your opt-out is honored everywhere marketing email goes.

Success, invalid link, and error each show their own clear message, and the page no longer flashes an error heading before settling. A bare visit explains how to unsubscribe instead of showing an error.

Selah now reports ad performance to Reddit through both a browser pixel and a server-side Conversions API, the same two-sided setup it already runs for Google. The pixel counts page visits on the marketing pages (only after you accept the cookie banner); the server reports the real conversions — account signup, trial start, and first payment — straight from the webhooks, so they still land when a browser blocks the pixel. Dormant until the Reddit keys are set, so nothing changes for existing visitors until then.

The feature-education series (Reader, Reading plans, Pray, Journal, Companion, Knowledge) sends on days 3, 6, 9, 12, 15, and 18 after you finish onboarding — a calmer, less frequent run-up. The admin "Emails to users" timeline shows the new schedule.

Opt-in (set ADMIN_TELEGRAM_CHAT_ID to your private chat id): a new registration pings you with name/email/plan, and a Stripe plan transition (trial started, converted to Plus, canceled) pings you with the email and what changed. Dormant until the env var is set.

When Selah names a verse — in a nudge, a Scripture suggestion, or a chat reply — tapping it opens straight to that chapter and verse in the reader, instead of leaving you to find it.

Routine patches to the framework and sign-in library, plus a fix so a note's title can't slip markup into its public share page. Nothing changes in how Selah looks or works, and there's nothing you need to do.

It was showing a broken-tile icon (and the poster flashed as a broken image) because the site's security policy didn't allow the video host. Both now load correctly.

It was showing "selah.so refused to connect" because the app blocks all iframing for security; the preview now loads the email's HTML and draws it inline, so you can see every email's body in light and dark.

A verse or two inlines fine; anything longer sends the reference only and lets you open it in the app. Proverbs chapters (the daily reading in the proverbs-story-gospel plan) were the main casualty.

The proverbs-story-gospel plan was only showing the Proverbs reference; the Genesis and Luke readings for that day were missing.

A still frame of the app shows by default; tap the play button to watch the tour with sound. A short preview loop animates while the hero is on screen — on phones too, not just on hover — and stops when you scroll past.

After you sign up, you get one email every couple of days for the first two weeks, each on a single feature: studying in the Reader (word studies, people and places, highlights that hold a thought), reading plans, keeping a prayer log and watching answers come, your journal and its weekly arc, the Companion, and your belief map. The subject lines ask a question worth opening — "Do you forget to pray?", "What do you actually believe?", "Did God answer?".

The onboarding sequence reads as a timeline from signup; the rest group by when they fire. Tap any one to see its subject, preview text, audience, timing, a live render of the real email (light and dark), and its open and click rates.

An anonymous visit count runs either way; cookies and Google Analytics turn on only if you accept. It honors a browser's Global Privacy Control signal, and you can change your mind any time through "Privacy choices" in the footer.

Analytics cookies stay off until you accept, and your choice is remembered.

(founder-only). One screen that answers "how's it doing, what's it burning, what needs attention": this-month AI cost with an 8-week trend, plan counts, a needs-attention panel, AI cost by feature and by top user, and a manual infra/ops expense list. Revenue and margin are gated until billing goes live, shown honestly as "awaiting billing launch" rather than confident zeros.

The heartbeat's voice was tightened so it points you toward Scripture and prayer without sounding like a person or a spiritual participant.

A perfect verse for something you wrote yesterday should reach you even if a suggestion went out earlier in the week.

The heavy drop shadows are gone, replaced by a quiet border that deepens as you hover and a gentle press when you tap. The weekly reflection and an active session still lift off the page so they stay set apart.

Instead of a row of folder chips plus a separate "Add another" button, one control shows the folders a note is in and opens a searchable list to add, remove, or create folders.

When something you captured or prayed seems to connect with a verse, the heartbeat may suggest one, humbly, as an invitation to compare ("…may be worth sitting with"). It names the reference and you open it yourself; reply naturally if it lands or doesn't. Rolling out gradually behind a flag, off by default for now.

Its nightly review of your recent activity sometimes threw away a whole night's observations when the model returned its answer wrapped in extra explanation. Those notes are kept now.

A past day's unread reading kept announcing itself as "Today's reading"; now only the actual current day does. Plan entries that had lost their plan showed "Day 1 of 0" or a blank "(plan)" name, and a loose capture left a stray dot trailing its timestamp. All three are cleaned up.

The toggle now reads cleanly, knob inside the track.

After you capture your first thought, Selah follows up with a quiet, timed nudge — to pray through it, save a note, then see what it's learned about you. Three touchpoints across ten days, each one waiting at least a day before the next.

/help and the welcome message both use the same clear list of commands and capabilities, so you always know what Selah understands in a conversation.

Every note now has a "Use in Selah's memory" switch, on by default. Leave it on and Selah can draw on that note and bring it back when it's relevant to your study. Turn it off and the note stays saved and searchable, but Selah won't use it to remember, connect, or recall. Turning it off asks first, and nothing is ever deleted.

In the Companion, the memories Selah surfaces now match what you're actually talking about, instead of whatever you happened to save most recently.

Open a concept and you'll see your own positions grouped into "What you believe," "Still working through," "Where you've seen this," and "In your words," each with where it came from.

A long note used to lose its later passages when Selah pulled meaning from it. Now the whole note is covered.

Tap the handle to expand it to 70% of screen height — useful for reading long Word Study or Cross-References content without losing the passage entirely. Tap again to return to the compact 340px default. The handle shows a chevron that rotates to indicate state.

A slim bar below the navigation reads "✦ 100 founding seats at $9/mo — after that, Selah Plus is $12/mo" with a "Lock in pricing →" link. Dismiss it once and it won't reappear.

Long-pressing a tagged word previously showed the Word Study panel instead of the OS text-selection gesture. Tagged words are now implemented as <span role="button"> so the selection gesture reaches the OS first.

The sage-green pricing strip at the bottom of the pricing section now left-aligns its text on narrow screens, keeping the single-line layout.

The previous arc path was a 180° semicircle, which left no room for the F and R at the endpoints. The arc is now ~195°, giving the text comfortable clearance on both sides.

Telegram daily messages, the journal drawer, and the reader all pull from the same source of truth (daily_readings.passage_ref) rather than a stale snapshot stored at generation time. Users who updated their reading plan no longer see the old passage reference in the drawer or Telegram.

The previous 450-character hard cutoff silently dropped the text; longer passages now appear inline and are gracefully truncated with "… (cont.)" only when the full message would exceed Telegram's 4096-character limit.

Keyboard navigation now shows a clean sage outline with a 3px gap on every focusable element — buttons, links, inputs. The previous faint washed-out ring has been replaced by a single authoritative rule in globals.css.

The "N" key hint on the New Note button no longer looks like a persistent focus ring; its border and text now inherit the button's text color at 50% opacity, so it reads clearly on sage, secondary, and ghost buttons alike.

A 28px gradient at the right edge signals that additional tools are available by scrolling.

The Old Testament / New Testament / Poetry section names previously appeared only on hover. They now appear at low opacity at all times, making the navigation bar readable on first encounter. Hovering a section brightens the label.

Topbar buttons, dock tabs, pane tools, and the mode toggle all respond with a subtle scale-down on press.

The green "Founding member pricing — 100 seats only" strip above the pricing ledger no longer appears. The context paragraph below the ledger does the work instead.

The copy reads "lock in $9/mo · $90/yr for 24 months minimum" rather than a hard cutoff, which is more accurate while we're still in beta.

"The companion and knowledge graph are AI-powered" is now "Our app is AI-powered," which is shorter and doesn't require readers to know what the knowledge graph is.

A sage-green notice above the pricing table reads "Founding member pricing — 100 seats only. When they fill, Selah Plus moves to $12/mo permanently." The founding member card strip also shows "First 100 seats · then $12/mo" so the price increase is visible in both places.

The open-beta explanation now leads with the consequence: when 100 founding members fill, the price goes to $12/mo and stays there. Founding members lock in $9/mo for 24 months. The copy explains that the companion and knowledge graph cost real money to run, and that the $9/mo rate is what makes it sustainable — rather than building the cost into ads.

The open-beta paragraph sits in a white card directly below the pricing ledger. The founding member seal is stickered on top of the card — centered and tilted — instead of living inside the pricing column. The $12/mo · $108/yr post-launch price note moved from the pricing column into the card body so the context and the caveat live together.

Selecting a role moves directly to the companion step after a 160ms visual pause — no separate "Continue" tap required.

Picking a translation on the companion step scrolls the tradition section into view so the next question appears without manual scrolling.

The expand button is inline with the tradition grid at 44px touch target height, replacing the small text link that appeared below the grid.

Nine translation options now lay out in a single scrollable row with snap points and a right-edge fade instead of wrapping into three rows.

Completing onboarding sets formatChoiceSeen so the home screen shows the first-day hero and companion invitation instead of immediately redirecting to the reading plan.

The drawer opens automatically when a tour step needs to spotlight a nav item — previously the spotlight fell back to a centered popup with no visual anchor.

The three-phase capture → organize → companion cycle replaces the old four-step journey. Each card names a real moment: capture a verse mid-sermon, find themes surfacing from your notes, let the companion ask the question you haven't asked yourself yet.

Chat bubbles animate into view as you scroll the landing page, demonstrating a real exchange between user and Selah before signing up.

The reader, prayer board, notes, and plan tiles now stack in a single centered column — each section has room to breathe. The responsive grid adjusts without breakpoints.

List and grid note cards now use the elevated card shadow pattern from the design system — they lift subtly and press on tap, matching the visual weight of cards elsewhere in the app.

The note title and all H2/H3 headings inside the editor are now set in Lora (serif), giving the writing surface a distinct journal-like character separate from the navigation UI.

Active status and folder filters now use bg-sage text-white — the same vocabulary as filter strips everywhere else — replacing the previous ring/opacity variants.

The add-to-folder button reads "Add folder" when no folder is assigned and "Add another" when one or more are present, replacing the ambiguous "Folder" label.

When no notes exist, the placeholder now reads "A blank page is an invitation." The icon sits in a soft gold circle rather than appearing inline with low contrast.

"Selah is one place." is now "Your faith life, finally in one place."

All four action buttons in the note editor (favorite, publish, reformat, delete) now have min-h-[44px] min-w-[44px] padding, meeting the iOS/WCAG minimum for interactive touch targets.

"Last saved" timestamps now display as 09:45 AM instead of the system locale's verbose format with seconds.

The new-note shortcut now ignores Shift+N and modifier combos — only plain n while no input is focused fires the shortcut.

Error banners now carry role="alert" so they are announced immediately on insertion. Status and folder filter strips have role="group" + aria-label; the "All" button has aria-pressed.

The uppercase section label above the folder filter strip is gone — the aria-label carries that context for assistive technology without the visual clutter.

Tapping a screen-capture video on the landing page no longer triggers native player controls, fullscreen requests, or the Android video overlay. Videos are purely decorative: pointer-events: none on both .lp-vid and the knowledge plate video, disablePictureInPicture on all five video elements, preload="none" so they don't buffer on cellular.

"Get started" in the nav, hero, and final section now goes to /sign-up instead of #pricing.

FAQ answer about Telegram no longer uses "captures" or "your Stream" — rewritten as "notes" and "your study". Pricing and FAQ consistently say "knowledge map" instead of "knowledge graph", matching the section heading. Section eyebrows removed from Reader, Prayer board, Reading plan, and "How Selah is different" — down from five to one (Notes & journal, which uses the kicker intentionally).

The three-item bullet list is replaced by a single paragraph that reads naturally.

Switching to a new reading plan paused the old one, but only the active plan was shown — your other in-progress plans had no way to reappear or be resumed. The full library now renders.

Opening the picker from Settings → Daily Message led to a dead end with no back button. A "← Settings" link now appears at the top of the pick step.

The community group link in the app was pointing to an expired invite URL.

The reading plan section now shows every in-progress plan you've put on hold — not just the one currently active. Each paused plan displays how far you got and lets you resume or make it active with one tap.

A "Discard" action removes the plan from your library after an inline confirmation. Your active plan is never affected.

Choosing a reading plan now ends with a brief "confirmed" screen showing the plan name and when your first reading arrives, before continuing to the app.

Buttons squeeze a touch when you press them, cards lift on hover and settle when you click, and the home screen and onboarding reveal their lines in sequence instead of all at once. Cards now rest on a soft shadow rather than a hairline border.

Each section's heading and copy ease in from a gentle blur as they reach the screen.

The Companion's free allowance went from 5 to 7 messages a day, still shared across web and Telegram.

Close, dismiss, and help buttons across modals, drawers, the prayer board, and the reader now carry a full-size tap target even where the icon itself is small.

The weekly summary counts use fixed-width figures, so the row no longer shifts as they change.

The card now follows your cursor in every column, not just the others.

The first card you dragged could land misaligned while later drags were fine; the board now measures its columns continuously during a drag, so the first one behaves like the rest.

The capture drawer had a permanently-disabled button that read as broken. It's hidden until the feature ships.

From a user's admin page, "Start trial" grants Plus for 7, 14, or 30 days, and "Extend" adds time to an active one. The trial ends on its own, and you can remove it any time. Accounts that pay through Stripe point you to the Stripe dashboard instead.

Drop a table into a note and a small toolbar now appears while your cursor is in it — add or delete rows and columns, or remove the whole table. Before, an inserted table was stuck for good.

Sign-up and sign-in now carry a "By continuing, you agree to our Terms and Privacy Policy" line with links, and the landing-page footer links both pages.

It's a left-rail command center with a searchable, filterable Users directory. Each person shows their plan, billing status, email verification, last sign-in, and activity at a glance — no more bouncing between Stripe and Clerk to look someone up.

A Plan control on each user's page comps a free Plus plan in one click, and revokes it just as easily, for anyone without a paid Stripe subscription.

Signed-out visitors now see Google's One Tap prompt automatically — one click to authenticate without navigating away or filling a form.

It now carries read-only Billing and Identity panels (live subscription detail, a Stripe deep-link, Clerk sign-in and ban state) alongside the rename, admin, nudge, and delete actions, all in one place.

Both nav entry points are hidden by default while the features are still being designed. Enable PostHog flags nav_new_menu_enabled and nav_search_enabled (0% rollout) to show them per user.

Every surface that showed the brand name — nav, mobile drawer, onboarding header, sign-in, sign-up, and email footers — now uses Geist at weight 500, all lowercase. Cleaner, more minimal, and consistent with the rest of the type system.

Both the header and footer on the public landing page use the same Geist treatment.

The lock-and-link icons beneath the checkout button now show Link's actual brand mark instead of a hand-drawn chain — more recognizable at a glance.

The quotes read in Inter (not a serif), consistent with the rest of onboarding.

A sage-tinted marker sweeps the full height of the text under "triggers my mind to think deeper" and "treated as a tool". It comes in slowly so scanners have time to read it, and the highlighted phrase is set slightly bolder.

The Billing & plan row now says "Free" or "Plus" so you can see your plan at a glance without opening the page. On mobile the subtitle reads "Selah Plus" or "Free — upgrade anytime."

The One Tap prompt shipped last release but was silently blocked by the site's content-security policy, so signed-out visitors never saw it. Google's sign-in domains are now allowed and the prompt appears.

When the Companion sent a reading-plan or Selah link, Telegram rendered a large preview image that crowded the chat. The Companion's replies now suppress that card. Sharing Selah elsewhere still shows the preview.

A retried Stripe webhook could send the trial-ending reminder or win-back email twice. Each now sends at most once.

The fallback base URL was selah.app; it's now selah.so, so links never break if the environment variable is unset.

Reloading mid-signup no longer bounces you back to the first step.

The scatter section now reads "Your faith life, finally in one place." The Telegram community invite was expired; it's been replaced so clicking "Join the community" works again.

Tap a plan day in the reader (or the drawer footer, or the day band), and the card in the stream switches to "✓ read" without a refresh. Before, the stream sat on stale state until you reloaded.

Mark it in the reader, switch back to the stream — the band is gone. No more nag for a day you already finished.

Move a note to a folder somewhere else in the app, then open the folder picker — the folder list reflects the move. Same for the prayer board (which also refreshes when you switch tabs back).

If you skipped yesterday, this morning's message is for yesterday's day, not today's. The cron now walks the backlog before falling back to today. You stop silently losing days from Telegram.

Before, the devotional drawer's mark-read button worked but didn't tell any other surface, so cards stayed stale until reload.

v0.22.2 decoupled the two settings on the page so toggling daily-scripture wouldn't silently re-enable devotional. But the morning cron still required both to be on, so the daily-scripture toggle wasn't actually delivering for anyone who had turned devotional off independently. The cron now respects the daily-scripture toggle on its own.

Before, you could toggle "Send a daily message" on without ever setting a delivery hour, and the cron would skip you silently. Now the toggle sets 7 AM local + auto-detects your timezone if either is unset, so the message actually arrives the next morning.

The placeholder is gone. The pill is the action.

A leftover prayers.tags column reference was throwing undefined_column on every page load because dev Neon had dropped the column out of band weeks earlier. The schema now matches reality, and the orphan prayer_tags table is gone too.

Two rapid taps on a prayer's status used to dispatch two opposite requests, and the loser would overwrite the winner. Mutations now go through one path with a stable cache key.

Phrase highlights now follow you across translations. The matching tokens light up in the translation you switch to. When the auto-match isn't confident, you see the whole verse marked with a gold "adjusted" pill so you know the system isn't sure.

A highlight that spans two verses gets aligned per-verse against the target translation. If verse 16 matches cleanly and verse 17 falls back, you see the precise phrase in 16 and the whole verse in 17 — not the whole highlight collapsing to fallback.

Click the gold "adjusted" pill on a phrase highlight, the active verse stays bright while siblings dim, you select the words you actually meant and hit Save. One click in, native text selection, no modal.

Four defaults everyone gets — Active, Exploring, On hold, Resolved — plus your own custom statuses through a single picker. Same color palette, same behavior, same place on every surface.

A note about Romans 8 can sit in both Sermons and Romans without you picking one. Folder chips show in the note header with × to remove and + Folder to add another.

Threads, Insights, Teaching, Devotional — your old categories are now folders. A one-time banner explains the change; existing notes are auto-filed by their type.

Two filter strips above the list: pick a status, pick folders, get the intersection.

A glance at a capture card tells you its status and which folders it lives in.

The bot was sending it after an hour of silence following any Telegram chat. Telegram chat doesn't have a session you "start" or "wrap up" — the message read as nonsense. Devotional sessions still get the same idle check, where it actually maps to something the user is doing.

Some sign-ups in the last day created a Clerk account, then had it deleted seconds later, bouncing the user back to the sign-in page. Two affected users can retry their original invite codes and finish creating their accounts.

The memory system extracts atomic facts from session summaries (study style, theology, life context, preferences, observations) and stores them persistently. Facts are reinforced when they reappear, superseded when they change, and expired when they have a deadline. The companion and devotional engine pull the top active facts into context so each session builds on what came before.

When the cron resynthesizes a concept summary, it includes memory facts that mention that concept by name — prefixed [memory] in the snippet context passed to the LLM. "Grace" as a concept will now surface that you view it as empowerment rather than permission.

Every concept detail view now shows a relative timestamp ("Updated 3 hours ago") regardless of whether the concept has a summary yet. Previously the timestamp was only visible when a summary existed.

A new edit_count column on notes increments on every content save. The companion's activity digest now reflects the real edit depth rather than a hardcoded floor of 1.

The /changelog page now groups releases by date instead of stacking version sections vertically. When several versions ship the same day, they collapse into one accordion bullet per release with the version number tagged inline. Less scroll, easier to find what shipped this week.

Fact text is screened against known injection patterns before storage. Suspicious facts are stored with flagged_injection=true and excluded from all LLM context rather than dropped — so "I want to disregard my doubts about grace" doesn't disappear, it just waits for review. Reinforcement updates also propagate the flag if a later extraction catches it.

The stale-concept flagging logic (which marks concepts for re-summarization after a session produces related facts) now uses a shared token-based matcher instead of a substring check. Short concept names (≤ 3 chars) no longer produce false positives. Same matcher used in both the pipeline and the cron so they can't diverge.

factsVersion was set to facts.length on every update, so two simultaneous regenerations could both read 5 facts and write factsVersion: 5 — making the second write look like a no-op when it wasn't. Now incremented atomically with COALESCE(current, 0) + 1 in SQL on both the INSERT and UPDATE branches.

Postgres errors from the Neon serverless driver were arriving as "NeonDbError: null" with no SQLSTATE, no severity, no way to triage. The redactor strips message text for privacy (users write prayers); the categorical Postgres metadata never made it through Sentry's normalization. Two new tags now ride along on every NeonDbError event: dbErrorCode (the 5-char SQLSTATE — 42P01, 23505, etc) and dbErrorSeverity (ERROR / FATAL / PANIC / etc). When a database query fails in production, we'll know in seconds whether it's a missing table (migration drift), a duplicate key, a connection failure, or a timeout. Closes [SELAH-SENTRY-D](https://selah-sr.sentry.io/issues/7440039530/), unblocks triage on SELAH-SENTRY-C and SELAH-SENTRY-3 (same throw site, different captures via errorDigest matching).

relation "sessions" does not exist was killing every /api/telegram/webhook call since the stream v2 deploy. Root cause: the migration runner's regex splitter (/;\s*(?:\n|$)/) shredded DO $$ PL/pgSQL blocks into invalid fragments. Neon's HTTP driver accepted them silently (returned 200, executed nothing), the runner marked all three migrations (0026, 0027, 0028) as applied, and the companion_sessions → sessions rename never happened. Recovery migration 0032_fix_sessions_schema.sql applies all the skipped DDL idempotently — rename, new columns, index, stream_entries table, session_id foreign keys on stream_captures and notes.

Sentry's captureMessage calls (the orphan-intent breadcrumb, the tool-loop hop budget exhaustion log, the final-response fallback warning) hand a static string to event.message, which the redactor strips for privacy. Without a tag, every site looked identical on the dashboard. New messageOrigin tag derives from a fixed prefix allowlist of literal strings our code emits — companion.orphan_intent, companion.hop_budget_exhausted, companion.final_response_fallback, ai.stop_reason, ai.schema.json_parse_failed, ai.schema.validation_failed, ai.job. Each captureMessage site now sorts to its own dashboard bucket. Closes [SELAH-SENTRY-8](https://selah-sr.sentry.io/issues/7436346156/).