* feat: give a cold mailbox a rotation lifecycle so a tired one can rest and come back, instead of running at full volume until a hard band trips: send_lifecycle is warming, active, resting or reserve and decides whether cold sender resolution offers the mailbox at all, which is a different axis from risk_band deciding which worker and IP host it, so a resting mailbox is still a clean-band mailbox that keeps its warmup traffic and its reputation; the hourly rebalancer rests a mailbox at throttled and worse but never at watch, since watch is defined as the band that changes nothing a customer can feel and leaving cold rotation is very much something they feel, and a rested mailbox returns only after three clean days so one good hour cannot bounce it back to full volume; reserve is the owner's hold and is never overridden, the default is active so no existing mailbox changes on deploy, and the state never travels in a workspace archive because it is this instance's decision about sending it watched
* feat: stop a query error re-admitting rested mailboxes, make probation measure healthy time, and rotate the candidate window so no mailbox starves: sendLifecycles returned a nil map on failure and an unresolved state reads as active, so one bad query quietly put every resting and reserved mailbox back into cold rotation, and the gate is now applied only when the states were actually read, with the skip logged rather than silent; ReadyToResume measured total time resting, so a mailbox that sat unhealthy for three days resumed on its first healthy tick having served no clean time, and an unhealthy evaluation now restarts the streak; and ordering candidates by send_lifecycle_since put every never-moved mailbox equal-first, so on an install with more than one page of them the same page was re-examined forever, which a checked-at stamp and its index fix
* feat: refuse a launch whose list is known to be largely undeliverable
* feat: count the deliverable audience in SQL instead of subtracting two overlapping totals, surface the unverified advice, and stop a query error skipping the gate silently: a contact can be both suppressed and unsubscribed, so Total minus both removed it twice and inflated every share computed against the remainder, which for a fully overlapping list produced a negative deliverable clamped to zero; the unverified-list branch set a summary and remediation but neither Block nor Warn, and a preflight report only surfaces checks that did not pass, so that advice could never be displayed; and a failed audience query was treated as a pass, which is fail-open on a safety gate without even a log line saying the check did not run
* feat: measure an uploaded list at import time and report what it looks like
* feat: document the import assessment and show it in the wizard
* feat: count only sendable leads as invalid, and stop guessing which column held a malformed address: an invalid lead that was also suppressed sat in the numerator while Deliverable excluded it from the denominator, so a campaign whose sendable list was clean could project above 100% and be refused; every verification count now shares the deliverable predicate, and a row whose MAPPED address will not parse is recorded as malformed directly rather than scanning other cells for an at sign, which could pick up a notes field
* feat: refuse a launch whose list is known to be largely undeliverable
* feat: count the deliverable audience in SQL instead of subtracting two overlapping totals, surface the unverified advice, and stop a query error skipping the gate silently: a contact can be both suppressed and unsubscribed, so Total minus both removed it twice and inflated every share computed against the remainder, which for a fully overlapping list produced a negative deliverable clamped to zero; the unverified-list branch set a summary and remediation but neither Block nor Warn, and a preflight report only surfaces checks that did not pass, so that advice could never be displayed; and a failed audience query was treated as a pass, which is fail-open on a safety gate without even a log line saying the check did not run
* feat: give an organization one fused abuse posture, because every existing control watches a single subject and an actor slightly wrong on several axes sits under all of them: organizations gains risk_state, risk_score, risk_reason and an append-only risk_signals evidence blob, modelled on the warmup participant health machine that already works rather than a second vocabulary for the same idea; restricted cuts per-mailbox cold volume to a quarter and forces the free warmup pool so a risky tenant cannot spend the paid pool's shared reputation, suspended stops sending at the send gate, and watch deliberately changes nothing a customer can feel so evidence accumulates before anything is taken away; an operator's suspension outranks the derived band so a detector clearing cannot release a workspace a human suspended, transitions ride the audit spine to every teammate's dashboard, a banner says which limit is active and why rather than letting volume drop silently, and the posture never travels in a workspace archive since it is one platform's verdict reached from evidence the destination never saw
* feat: make the suspension actually stop sending, and emit the audit transitions the PR claimed: emailsend.SendEmail is only the manual and API path, so campaign and warmup sends went nowhere near the gate and a suspended workspace kept sending on its schedule, while the restricted multiplier floored every mailbox at one a day which quietly turned suspension into a trickle rather than a stop; the campaign scheduler now defers the whole campaign with a reason and the warmup task skips as org-suspended, since warmup is outbound mail from the same domains; separately the band change emitted no audit entry at all despite the entity type and the frontend spine entry both existing, so no banner moved for a teammate and there was no trail of who was restricted when, and only a real transition is logged so a detector re-recording the same finding cannot fill the feed; one of my own live tests also asserted how far out a slot lands, which depends on the hour the suite runs, and now asserts the property it was about
* feat: capture and score a signup's origin instead of discarding it
* feat: document what a signup records, and prove the origin write is on the account path with a seam test rather than only testing the scorer
* chore: drop the em dashes and trim the comments the review flagged
* feat: give an organization one fused abuse posture, because every existing control watches a single subject and an actor slightly wrong on several axes sits under all of them: organizations gains risk_state, risk_score, risk_reason and an append-only risk_signals evidence blob, modelled on the warmup participant health machine that already works rather than a second vocabulary for the same idea; restricted cuts per-mailbox cold volume to a quarter and forces the free warmup pool so a risky tenant cannot spend the paid pool's shared reputation, suspended stops sending at the send gate, and watch deliberately changes nothing a customer can feel so evidence accumulates before anything is taken away; an operator's suspension outranks the derived band so a detector clearing cannot release a workspace a human suspended, transitions ride the audit spine to every teammate's dashboard, a banner says which limit is active and why rather than letting volume drop silently, and the posture never travels in a workspace archive since it is one platform's verdict reached from evidence the destination never saw
* feat: make the suspension actually stop sending, and emit the audit transitions the PR claimed: emailsend.SendEmail is only the manual and API path, so campaign and warmup sends went nowhere near the gate and a suspended workspace kept sending on its schedule, while the restricted multiplier floored every mailbox at one a day which quietly turned suspension into a trickle rather than a stop; the campaign scheduler now defers the whole campaign with a reason and the warmup task skips as org-suspended, since warmup is outbound mail from the same domains; separately the band change emitted no audit entry at all despite the entity type and the frontend spine entry both existing, so no banner moved for a teammate and there was no trail of who was restricted when, and only a real transition is logged so a detector re-recording the same finding cannot fill the feed; one of my own live tests also asserted how far out a slot lands, which depends on the hour the suite runs, and now asserts the property it was about
* chore: drop a test that belongs to the signup-risk change and reached this branch by mistake, where the type it exercises does not exist
* feat: close the two halves of the delivery-signal loop that were still open, complaints and domain-auth refusals: a spam complaint arrives as mail long after the send succeeded, and nothing read those reports, so the strongest negative signal a sender gets never reached the complaint rate, the suppression list or the breaker; internal/pkg/arf parses RFC 5965 feedback reports worker-side alongside the existing DSN path, takes the LAST Message-ID because the reported mail's headers follow the report's own, and records only abuse-type reports so a not-spam report cannot be inverted into a complaint; separately a receiving server refusing mail because the SENDING DOMAIN failed its authentication (5.7.515, 5.7.26) was classified as SERVER_UNREACHABLE and retried forever, and it is now its own hard code that blames the domain rather than the address, leaves the recipient unsuppressed, and brings that domain's DNS re-check forward so the sweep confirms the verdict the send gate acts on
* feat: make the complaint path actually reachable, and stop it trusting the report: selectTextParts only ever picked text/plain and text/html, so an RFC 5965 report's message/feedback-report and message/rfc822 parts never reached the worker and both Feedback-Type and the reported Message-ID were invisible, which would have left this feature inert and has been quietly weakening DSN parsing too; report parts are now selected as plain text; the complainer is taken from the RESOLVED SEND rather than the report body, because a report is unauthenticated mail anyone able to reach the mailbox could forge and honouring the address it names would let a forger suppress a contact the send never went to; Original-Mail-From is no longer read as the complainer since that address is the sender; and a domain-auth refusal now releases the reservation instead of spending one of the lead's attempts, because the recipient received nothing and the problem is the mailbox's domain, so another mailbox in the pool picks the lead up
* feat: fix the same missing-parts gap in the Gmail adapter, and say plainly that Microsoft Graph cannot see reports at all: goog.extractBody took only text/plain and text/html exactly as the IMAP path did, so a feedback report synced from Gmail was as invisible as one synced over IMAP, and the tests that would have caught either called the parser directly rather than going through the adapter where it actually broke; the new tests exercise that seam, and Graph returns one rendered body with no parts so reports there are undetectable without a MIME fetch that is not built, which the docs now state rather than implying full coverage
* feat: ease a graduating mailbox into cold volume instead of handing it the full cap the day it joins a campaign: cold sending read warmup HEALTH but never whether the mailbox had actually warmed enough, so one at its 40/day warmup ceiling could send 50 cold the next morning, which is the post-warmup spike providers penalise; effectiveCap gains a min() term that starts the mailbox at 5, 10 or 20 a day by how long it warmed and adds 5 per clean day toward its own cap, freezing on a spam placement through the same union-of-freezes the warmup ramp uses, anchored by a new cold_ramp_started_at stamped idempotently on the first cold send and reset on org import because it raises a ceiling the destination never watched being earned; mailboxes that never warmed are not gated, since capping senders who never opted into warmup is a different decision from smoothing the transition out of it
* feat: anchor the graduation ramp on a CONFIRMED send, and make the hold the drawer reports the hold the scheduler applies: stamping cold_ramp_started_at at dispatch started the clock on a send the worker then failed, so a mailbox climbed on days it had not proven anything, and the stamp moves to the worker's EMAIL_SENT; separately the drawer computed its hold over every placement while ColdCeiling only counts placements after the first cold send, so a placement predating the ramp read as paused while the scheduler kept climbing, and both now go through one ColdHeldUntil
* fix: restore the Callout closing tag my conflict resolution dropped, which types:check and lint both pass and only pnpm build catches
* feat: let a warmup recipient answer the mailbox that just wrote to it, so a thread reads as a conversation rather than two mailboxes monologuing on their own ramps: warmup_tasks.target_account_id was written as nil and never read by anything, so a reply only happened when the recipient's own ramp fired AND the draw happened to land on that partner; a verified receipt now sometimes re-points the recipient's pending warmup task at the sender 25 minutes to 5 hours later inside its own warmup hours, which is a re-pointing rather than new work because only one warmup task may be pending per mailbox, it can never delay a send the mailbox had planned sooner, and it stops before the thread cap so replies cannot answer replies forever; the clock parser also moves into models.ClockMinutes so a second copy of the HH:MM parsing that silently disabled every sending window cannot drift back in
* feat: stop the reply-back drawing the reply rate twice, and stop its jitter escaping a short warmup window: the scheduler drew the recipient's reply rate to decide whether to answer at all, then the task handler drew it again to decide reply-versus-new, so a 30 percent reply rate produced a 9 percent answer rate and a directed task could send a fresh message to the mailbox it was meant to be answering; a directed task now IS the reply, and the opening-time jitter is capped to the window width so a mailbox warming 09:00 to 09:20 is not scheduled past its own close
* feat: score campaign copy where it actually matters, on the message the recipient receives rather than only the template in the editor: warmlint.Score already ran at save time through /templates/score, but nothing checked the copy after merge fields, spintax, A/B selection and AI blocks resolved, which is exactly where a clean template becomes Hi comma or picks the one spammy spintax branch; the send path now scores the rendered message and writes one campaign-feed warning per step per day rather than one per recipient, RunPreflight gains the same check across every step, warmlint gains image-heavy, many-images and attachment heuristics, and the whole thing is advisory behind Settings > Sending > Content checks; the launch dialog also finally calls POST /campaigns/:id/preflight, whose report had no UI at all and so never reached the person about to send
* feat: fix three defects the review found in the content lint: a preflight run that could not read the campaign's steps left the worst score at 100 and reported the check as PASSED, which is the one outcome a check that did not run must never produce, so it now reports failed with the reason; the send-time path read organization settings while preflight read campaign-effective ones, so a campaign that turned the check off or moved its floor was still warned, and EffectiveSettings is now exposed and used by both; and the check-then-insert dedupe let concurrent recipients of the same low-scoring step each find nothing and each write, so it becomes one advisory-locked conditional insert proven by twelve racers producing exactly one feed entry
* feat: feed each sender's per-provider spam placement back into warmup partner selection, so a mailbox failing only at Microsoft stops being handed Microsoft partners instead of waiting for an aggregate band to trip: pickWeightedPartner weighted on inverse domain frequency and routing rules alone, while PoolSpamPlacementsByProvider's segmentation only ever reached the admin overview; the new per-sender query keys on who RUNS the recipient's mail via models.ClassifyProvider rather than email_accounts.provider, because that enum collapses every custom host into smtp_imap which is exactly the bucket the signal matters in, and the weight multiplies by 1/(1+4*rate) with a five-send floor so a failing provider is downweighted but never excluded, since a sender that stops mailing a provider can never discover it recovered
* feat: stop failed warmup sends flattering a provider's junk rate, and stop a re-admitted partner sidestepping the penalty: a warmup_tokens row is written before the send leaves, so counting every token put failed sends in the denominator and understated the rate exactly when the sender was doing worst; the send count now joins tasks on status = completed, and GetPoolParticipantProviders is read unfiltered because that map only resolves a candidate's provider and never decides eligibility, so a just-unblocked partner missing from it scored an unpenalized 1.0
* feat: move the campaign content-lint work off this branch, where it landed by mistake: the pre-send scoring, its preflight check and the warmlint image and attachment heuristics belong to the content-lint change and are unrelated to per-provider partner routing, so this branch now carries only the routing signal it is about
* feat: close the warmup ramp loop so an early spam placement holds the mailbox instead of letting it climb into the wall: every band in evaluateMetrics needs a sample floor before it can trip (20 warmup sends in 7 days, 100 delivered in 30), which a mailbox in its first fortnight never reaches, so one landing in junk on day three kept adding an email a day until it had sent enough to be judged; a placement inside 48h now cuts the day about a quarter and holds the ramp where it stood, and the resume subtracts the frozen days rather than catching them up because climbing three steps in one morning is the spike the hold exists to prevent; the ramp arithmetic moves into internal/app/warmupramp so the mailbox drawer reports the target the scheduler will actually act on instead of its own private copy, and says which signal cut it and when it resumes
* feat: fix three defects the review found in the warmup ramp hold: the ramp is now elapsed days minus frozen days rather than a level held at the newest placement, because MAX(created_at) meant a second placement arriving mid-hold RAISED the held level and a mailbox landing in junk repeatedly ramped up; the early-signal cut and the health band are resolved in one shared warmupramp.Resolve so the scheduler's health gate and the dashboard's ungated copy can no longer disagree about the same mailbox; and the hold is reported for the whole 72h freeze instead of only the 48h cut window, so a mailbox between the two no longer shows a ramp that has stopped climbing with nothing to explain it
* feat: make send-time optimization actually schedule sends, instead of being a documented setting with no caller: advanced.OptimizeSendTime had zero call sites anywhere in the codebase, so an org that enabled send_time_optimization through PATCH /outreach/settings changed nothing about when its campaign mail left, and the API reference said so in a callout; the campaign scheduler now resolves the recipient's timezone from the contact's timezone custom field, then the country-code suffix of its email domain, then the org fallback, and holds the slot until that clock reaches a preferred hour, raising hardFloor so the task handler reschedules rather than sending immediately, the snap is forward-only and never crosses the campaign end date, the default flips to off so no existing workspace silently re-times its sends, and Settings > Sending gives the block its first UI
* feat: stop the recipient-hour gate deferring a send forever, and fix the window parser that silently disabled every campaign schedule: recipientSlot now searches for a moment BOTH calendars accept and yields when they never meet, because raising a hard floor at an hour the sender cannot serve made every tick re-derive it, defer, wake in the sender's window and defer again; separately parseTimeOfDay accepted only the 15:04 layout while start_time, end_time, warmup_start_time and warmup_end_time are Postgres time columns pgx renders as 09:00:00.000000, so every read parsed to 0 and effectiveWindows read that as unconstrained, leaving both the campaign sending window and its day-of-week gate off for every campaign on the legacy fields and pinning warmup to its 08:00-20:00 fallbacks; Settings > Sending also gains the MANAGE_SETTINGS gate a direct visit needs
* feat: add internal/pkg/mailhtml, a mail-oriented HTML sanitizer and text flattener, because rendering a received message body means rendering the sender's markup: Sanitize builds on bluemonday's UGC policy but keeps what real email is made of (table layout attributes, inline CSS through the property-allowlisted style sanitizer, legacy font/center, data: and https: images) while dropping script, iframe, object and the text content of style/head blocks so a marketing email's stylesheet cannot render as body copy, forcing target=_blank plus nofollow/noreferrer on links and allowing only http, https, mailto and tel; ToText flattens the same input for previews, turning block boundaries into newlines and decoding entities back to the characters they stand for so an already-escaped body does not surface as literal &; LooksLikeHTML reports whether a stored body is actually markup, which is how a body recorded as HTML by an older sync but containing no tag at all can be recognised as the plain text it really is
* feat: add internal/pkg/mailhdr for RFC 5322 header values, since headers are ASCII on the wire and every transport was writing raw UTF-8 into Subject and display names: Subject and AddressList RFC 2047-encode non-ASCII (a no-op on plain ASCII, and a bare address stays bare rather than being wrapped in angle brackets), DecodeWords reverses encoded-words with a charset hook wired to go-message so legacy encodings Go does not handle natively still decode, and Bare/BareList strip a display name down to the routable address for SMTP envelope commands where 'Ana <a@b.com>' in RCPT TO is a syntax error, promoting go-message from an indirect to a direct dependency
* feat: encode outbound Subject and address headers on all three transports, so a subject or sender name containing an accent, a currency sign or an emoji reaches the recipient as the characters the user typed instead of mojibake: SMTP and Graph were writing the raw string into Subject (only the Gmail transport encoded it) and Graph built its From by fmt.Sprintf rather than mail.Address, so a non-ASCII display name went out unencoded there too, and all three joined To/Cc/Bcc entries verbatim so an encoded display name never appeared even when the caller supplied one; additionally the SMTP envelope now takes bare addresses through mailhdr.BareList, because an API caller may pass 'Name <addr>' (the compose handler has a bareAddress helper precisely because that arrives) and passing that to RCPT TO gets the recipient rejected by the server
* feat: rewrite the IMAP body reader, which was the reason received mail from SMTP/IMAP mailboxes came back corrupted: it built one FetchItemBodySection with a hardcoded Part []int{1} and a comment saying it would adjust when recursing, which it never did, so on a multipart/alternative the text/plain bytes were fetched twice and the second copy was stored as the HTML body (plain text rendered as markup loses every line break, shows & as an entity and swallows anything inside angle brackets), and decodeIfNeeded never reversed Content-Transfer-Encoding at all, leaving quoted-printable bodies full of =E2=80=99 runs and = soft breaks and base64 bodies unreadable, while its charset detection parsed params off a media-type string that never carried any and its mail.ReadMessage call could silently eat leading body lines as headers; the reader now walks the body structure for real part paths, fetches every text leaf in a single FETCH with a server-side Partial size cap, decodes quoted-printable and base64 (tolerating a tail cut mid-quantum by the cap) then converts the part's charset to UTF-8 with go-message, skips attachment-disposition parts so a .txt attachment cannot stand in for the body, takes one part per type inside a multipart/alternative but treats sibling inline parts in mixed/related as additive, and is bounded at five text parts per message; the stored body cap also goes from 200 KB to 512 KB because 200 KB cuts real HTML newsletters mid-document
* feat: decode Gmail's raw headers and entity-escaped snippets, because the Gmail API hands header values back exactly as they arrived on the wire, so a message from a sender whose subject or display name was RFC 2047-encoded showed in the dashboard as =?utf-8?q?caf=C3=A9?= rather than as the text it stands for, and the API's own snippet field is HTML-escaped, so a preview containing an apostrophe surfaced as ' in the conversation list and, until the thread reader stopped rendering snippets as message bodies, inside the message itself; getSingleHeader now runs values through mailhdr.DecodeWords (a no-op unless the value actually contains an encoded-word, so Message-ID and the warmup token header are untouched), the comma-split fallback in getAddressList does the same for display names net/mail could not parse, and the snippet is unescaped once on the way in
* feat: fix the conversation-list snippet, which collapsed whitespace before splitting on newlines so the quoted-line and signature filters below it could never match a thing, stripped HTML with bluemonday's strict policy and then showed the escaped output verbatim so an ampersand in an HTML-only message read as & and a marketing email's stylesheet text rendered as body copy, and cut at 100 bytes with text[:100] so a multi-byte character or emoji at the boundary became a replacement glyph; it now flattens HTML through mailhtml.ToText (entities decoded, style and script content dropped) including when a sender puts markup in their text/plain part, filters quoted history and everything past the RFC 3676 signature delimiter while the text still has lines, collapses whitespace afterwards, and truncates on a rune boundary at 200 characters
* feat: make GET /unibox/:id serve a display-safe body and stop it failing outright, sanitizing body_html through mailhtml before it leaves the API so every consumer gets markup that cannot execute rather than each call site having to defend itself, degrading a body blob that cannot be read to the message's preview text with a new body_truncated flag instead of returning 500 (which made a message with a missing blob unopenable, and hit every seed, sandbox and dev-history fixture row since only the '<seed-' prefix was recognised while the sandbox uses '<sbx-' and dev history '<dev-'), and treating a stored HTML body that contains no tag at all as the plain text it really is, because mail synced before the IMAP reader addressed parts individually recorded the plain part under both bodies and serving that as HTML is exactly what collapsed a ten-line message onto one line
* feat: escape composer text before turning it into the HTML part of an outgoing email, replacing body_html: trimmedBody.replace(/\n/g, '<br />') in both the compose window and the reply composer with a shared plainToHtml that escapes the five markup characters first, so an email containing 'Terms & conditions' no longer ships a broken entity and one containing anything in angle brackets ('<see attached>', 'a < b', a pasted tag) no longer has the rest of the paragraph swallowed by the recipient's mail client as an unclosed tag, while runs of spaces survive as non-breaking spaces and bare URLs become links without eating the sentence punctuation after them; the same unescaped plain-to-HTML pattern in the campaign step editor's applyTemplate now goes through promptToHtml, which escapes as it paragraph-wraps
* feat: render the real message body in the unibox thread reader instead of the list preview, which is the whole of the reported bug: ThreadView mapped each thread row to a UniboxEmail whose body was '<p>' + escapeHtml(m.snippet) + '</p>' and MessageBubble rendered that as the message, but a snippet is a preview capped at 100 characters with every run of whitespace collapsed to one space, so a ten-line email displayed as roughly two lines on a single continuous line, and Gmail's already-escaped snippet was escaped a second time so an apostrophe read as '; each expanded message now loads its own body from GET /unibox/:id (the newest message and anything unread open on mount, older messages collapse to their preview line so a long thread does not fetch every body at once) and renders it in a sandboxed iframe carrying no allow-scripts, which keeps a sender's stylesheet from restyling the dashboard and means nothing in the message can run even though the API already sanitized it, sizing itself from the inner document as images load, with the preview kept as the fallback when a body cannot be fetched and a notice when only a preview is stored
* feat: document how a message body is read and returned, adding a 'Reading a message' section to the unibox guide covering the expand-on-open behaviour, that formatting and special characters are preserved as sent, that the conversation list preview is a summary and not the message, and that HTML mail renders in an isolated frame with links opening in a new tab, plus a paragraph in the API endpoint reference stating that GET /unibox and GET /unibox/thread return previews carrying snippet while GET /unibox/:id returns body_plain and a sanitized body_html, and what body_truncated means
* feat: add email_accounts.save_to_sent, the per-mailbox switch for filing a copy of outbound mail in the Sent folder, defaulting on because plain SMTP submission leaves nothing behind in the sender's account while Gmail and Outlook file their own copy through their APIs, making it a per-mailbox choice rather than a global one since a submission server that files the copy itself (Gmail's SMTP, Fastmail, Zoho) would otherwise end up with two of everything, which is exactly why every desktop mail client ships the same switch, and wiring the column through the Email model, the mailbox read paths and UpdateEmail so it is readable and writable from the dashboard and the API
* feat: teach the IMAP client to APPEND a sent message and the SMTP client to hand back the exact bytes it submitted, the two transport pieces the Sent-folder copy needs: AppendToSent resolves the folder from the RFC 6154 \\Sent special-use attribute first (requesting it only when the server advertises SPECIAL-USE) and falls back to matching the known names against both the full mailbox name and its leaf, since servers namespace as INBOX.Sent and localize the label, caches the result for the life of the connection, files the message flagged \\Seen and dated when it was sent, and returns a sentinel rather than an error when the account has no Sent folder at all; APPEND addresses its mailbox by argument and never touches the selected mailbox, so unlike the warmup MOVE/STORE actions it is safe to run while the sync loop is mid-fetch on the same connection
* feat: file a copy of every SMTP send in the mailbox's Sent folder, closing the gap where a message sent from Warmbly through an SMTP/IMAP mailbox existed only in the recipient's inbox: nothing appeared in the customer's own mail client, and nothing appeared in the unibox either, whose thread reader can only show messages the sync found in a folder, so a user who sent from the dashboard and then went looking for what they sent found no record of it at all; the worker now APPENDs the exact bytes the SMTP client submitted after a successful send, best effort so a failed append never turns a delivered message into a failed task, skipping warmup traffic because filing dozens of machine-generated messages a day would bury the customer's real sent mail, and skipping Gmail and Graph mailboxes entirely since their APIs file their own copy; the per-mailbox setting rides along on the add-email worker payload as a pointer so an older control plane that does not send the field is read as unset and takes the default rather than as an explicit no
* feat: expose the Sent folder copy as a mailbox setting in the dashboard, adding a 'Keep a copy of sent mail' toggle to the Settings tab of the mailbox drawer that only renders for SMTP/IMAP mailboxes (Gmail and Outlook file their own copy, so the control would be a lie there), tracked by the drawer's save bar alongside the other editable fields, and worded so the one case where it should be turned off is obvious: a provider that already saves its own copy, where leaving it on means seeing every sent message twice
* feat: document the Sent folder copy in the mailboxes guide and the API reference, explaining why the toggle exists at all (SMTP submission leaves nothing in the sender's own account, so without it a sent message shows in neither the customer's mail client nor the unibox thread), when to turn it off (a provider such as Gmail, Fastmail or Zoho that already files its own copy of anything submitted over SMTP, where leaving it on doubles every message), that OAuth Gmail and Outlook mailboxes never show the control because their APIs file the copy themselves, that warmup traffic is deliberately excluded, and that PATCH /emails/:id takes save_to_sent
* feat: add unibox_emails.body_text and its search index, because unibox search ran against search_tsv, a generated column built from subject and snippet, and a snippet is a truncated one-line preview, so searching for a phrase that appears in the third paragraph of an email returned nothing at all and read as broken search rather than as search that only covers the first line; message bodies stay in object storage where they belong, and what lands in Postgres is a bounded 16 KB plain-text rendering carried on the new-email worker event, indexed with a GIN expression index rather than a second stored generated column since adding one of those rewrites the whole table while this builds against a column that is empty on every existing row
* feat: index what a message actually says, adding mailhtml.SearchText (HTML flattened, entities decoded, whitespace collapsed, quoted history deliberately kept because a phrase someone quoted back at you should still find the conversation, truncated on a rune boundary) and computing it on all three sync paths so IMAP, Gmail and Graph mail all arrive with searchable text, writing it on insert, and widening the unibox search filter to match either the existing subject-and-preview vector or the body expression, written exactly as the new index declares it so the index is actually used
* feat: backfill the searchable text of messages that were synced before bodies were indexed, so search covers the archive a customer already has instead of only mail that arrives from now on, which would have made the feature useless on day one for exactly the people who need it; the sweep pages through unibox_emails by id, reads each body from object storage under the mailbox owner's key, renders it with the same helper the sync path uses and writes it back, at 100 rows per 30 seconds because nothing waits on it, and returns for good once a pass finds nothing left to visit, with rows whose stored body really is empty simply revisited after the next restart rather than needing a tried-and-failed marker in the schema
* feat: document that unibox search now covers message text and not just subjects and previews, in the search paragraph of the unibox guide where the old wording only promised that search stays inside the current scope
* feat: add generation.RenderThread, the shared way to put a conversation in front of a model, because every AI surface was grounding on preview snippets and a draft written from the first hundred characters of each email answers the greeting rather than the question; it strips quoted history and signatures (the earlier messages are already in the prompt on their own, so quoting them again spends the budget twice, though a reply written underneath the quote is kept rather than thrown away when there is nothing meaningful above the attribution line), spends a bounded character budget newest-message-first since the message being replied to matters most, degrades older messages to their preview line instead of dropping them once the budget runs low, and renders oldest-first so the transcript reads in order
* feat: add grounding reads to the unibox service and repository, returning message text (the stored body, falling back to the preview for mail synced before bodies were indexed) for a thread or for all correspondence with one address, kept deliberately separate from the preview queries and given their own result type so a 16 KB body can never leak into a list response by accident, capped at twenty messages whatever a caller asks for, and paired with a RenderGrounding helper so every AI surface formats a conversation the same way instead of each one rolling its own transcript loop
* feat: ground every AI writing surface in what the messages actually say, switching the unibox reply draft, the compose draft's correspondence history, the inbox agent's thread history and the assistant's read-thread tool from preview snippets to real message text through the new grounding reads, which is what makes a drafted reply answer the question that was asked rather than the first sentence of the email; the inbox agent's triviality gate also reads the reply's full text now, since a preview line cannot tell a one-word ack apart from a long message that happens to open with one, and the assistant tool returns a bounded body per message with quoted history stripped instead of a snippet field
* feat: say in the docs that AI drafting reads the messages and not their previews, in both the unibox reply-draft section (adding that quoted history is stripped and the newest messages get the most room, so a draft answers what was asked rather than the opening sentence) and the inbox agent's grounding section, where 'the full thread so far' was true of the message list but not of how much of each message the model actually saw
* feat: renumber the two new migrations to 000087 and 000088 after rebasing onto main, which landed its own 000085 (org data transfer) and 000086 (email sync state) in the meantime, so the sequence has no duplicate versions
* feat: add the two new API fields to the OpenAPI spec that landed on main while this branch was open, documenting save_to_sent on the Mailbox and MailboxUpdate schemas and body_truncated on UniboxEmail, and saying on body_html that what the API returns is already sanitized so a client can render it directly
* feat: cut the README quick start and self-hosting sections down to the install command, one paragraph of what it does, and links out to the local development, first-run, deployment and warmblyctl docs pages, dropping the recovery if/then table, the MAIL_TRANSPORT invitation note and the dependency matrix that all duplicate those pages
* feat: give warmblyctl an API-key half so agents and scripts can operate any Warmbly instance including the hosted one, adding a raw passthrough (warmblyctl api get/post/patch/put/delete <path> with --data taking a literal, - or @file, and --idempotency-key) plus eleven typed families (me, campaign, contact, mailbox, inbox, analytics, settings, webhook, apikey, template, crm) driven by one spec table that generates dispatch, flags, per-command help and the request, covering list/get/create/update/delete, sequence steps, sender pools, preflight/start/stop/test-email, contact notes/timeline/import/export, mailbox behavior/verify/send and the six warmup controls, unibox threads/reply/compose/seen/agent-drafts/scheduled sends, outreach settings, webhook secrets and deliveries, and API key self-service, authenticated with Bearer wmbly_ keys from WARMBLY_API_KEY against WARMBLY_API_URL falling back to API_PUBLIC_URL then the hosted service, printing the API's JSON untouched and surfacing the error envelope's code and request_id with Retry-After on 429, while the DB-direct operator commands and their trust model stay exactly as they were
* feat: ship two in-repo agent skills so AI assistants working against Warmbly discover the right warmblyctl half on their own, .claude/skills/warmbly-api teaching product operation over the API commands (key and URL setup including the seeded local dev key, the eleven command families, pagination and error-code and Idempotency-Key conventions, and a sending-safety section that names the six commands that put real mail on the wire and holds agents to preflight before start and the 50/day default cap) and .claude/skills/warmbly-ops teaching instance administration over the DB-direct commands (status --json as the contract to parse, the recovery command table, TTY versus -T piping, Redis-down behaviour, and org export/import handling including --dry-run first and the sensitivity of credential archives), each pointing at the other for what it does not cover, narrowing the .gitignore .claude/ rule to .claude/* with !.claude/skills/ so personal agent state stays local while the skills ship
* feat: document warmblyctl's new API half on the warmblyctl reference page, reframing the intro around the two halves and their two trust models and replacing the 'no HTTP surface and never will' line with the accurate claim that the CLI never serves HTTP while the API commands are a client of the already-gated public API, adding WARMBLY_API_KEY and WARMBLY_API_URL to the environment table with the API_PUBLIC_URL-then-hosted fallback, renaming The commands to The operator commands, and adding an API commands section covering key setup, the eleven command families, the raw /v1 passthrough with curl-style --data forms, the pagination, idempotency-key and Retry-After conventions, a warning callout naming the six commands that put real mail on the wire with preflight-before-start guidance, and a pointer to the shipped .claude/skills agent skills, plus API authentication and permissions links in See also
* feat: move the shipped agent skills from .claude/skills/ to a top-level skills/ directory so they follow the convention other repos use for distributable agent skills rather than living inside Claude Code's personal state directory, restoring the .gitignore .claude/ rule to its original form since nothing tracked lives under it anymore, and updating the warmblyctl reference's For AI agents section to name skills/ and show installing a skill by copying it into the agent's own skills directory or pointing the agent at SKILL.md directly
* feat: clear the Security Scan failure by lifting the two flagged indirect Go modules past their fixed versions, github.com/moby/go-archive from v0.2.0 to v0.3.0 for the CVE-2026-17106 tar path traversal and golang.org/x/mod from v0.37.0 to v0.40.0 for the CVE-2026-56864 and CVE-2026-56865 GOSUMDB and GOPROXY forgery pair, with the x/sys, x/text and x/tools bumps go mod tidy pulls along
* feat: take the Trivy dependency scan off the PR gate and restructure CI the way larger projects do, because a full-repo CVE scan on every pull request goes red the morning any dependency gets a new advisory regardless of what the PR touches, which is exactly how this branch failed on two indirect Go modules it never went near, moving the scan to its own security.yml running weekly, on demand, and on main pushes that change a dependency manifest, pinned to trivy-action 0.36.0 instead of @master, extracting the pnpm+Node+frozen-install boilerplate repeated across the web, admin and site jobs into a .github/actions/setup-pnpm composite action with the store cached per lockfile, and collapsing the CI Status rollup's ten hand-enumerated result checks that had to be edited in two places per new job into a single contains(needs.*.result, ...) expression over failure and cancelled, all validated with actionlint
Matthew Meszaros