selfpost/docs/audit-2026-10-12.md

Code audit — 2026-10-12

What this is. A full-tree audit of the code and infrastructure at 8f226bd (release 2.4.0): leftovers, unfinished work, defects, performance, doc drift and test gaps — explicit ones (there are no TODO/FIXME markers anywhere in the product code) and the implicit ones that only reading finds. The panel's visual design (mockups, panel.css, layout) is a separate audit and is not touched here. Nothing was changed in the tree; this file is the only output.

Method. Six parallel readers, one per zone (mail path; background roles; store/backup/composition root; HTTP surface; view and services; build/deploy/ CI/e2e), each reading every file of its zone and its tests in full, each told to find where the code does not do what architecture.md says, and to prove it (a throwaway test, EXPLAIN QUERY PLAN, a probe) rather than guess. Machine checks: go build, go vet, gofmt -l, go mod tidy/verify, staticcheck, govulncheck, go test ./.... Five claims the readers could not settle from source were settled on the test server (selfpost.mixfed.ru, running a locally built 2.3.0-post.c817f08) and against GHCR. Two hypotheses were refuted there and are not findings: saslpasswd2 -d of a missing login exits 0 (deletion is idempotent, as the code says), and a duplicate key in a texthash: map is a warning with the first entry winning, not a fatal error.

Machine results. Build, vet, gofmt, tidy, verify: clean. staticcheck: four items (an unused domainIDSet in handlers/authz.go:63, two unused assignments in handlers_outbound_pages_test.go:174,194, one S1003 in view/pages_test.go:170). govulncheck: nine standard-library findings (net/http HTTP/2, net/textproto, html/template), all fixed in the current Go patch release; golang.org/x/crypto 0.54.0 has three published fixes in 0.56.0 that this code does not call. modernc.org/sqlite is seven minor versions behind (1.53.0 → 1.60.1). go test ./...: green on Linux semantics; on Windows two tests fail for OS reasons (TestWriteLoadPrivateKeyRoundtrip expects a POSIX mode; TestFollowTailsAndRotates renames an open file) — the tests, not the product, are platform-bound.

Severity: critical — mail or data lost or the security boundary crossed today; high — a documented guarantee is false in a way an operator will hit; medium — wrong under a realistic condition, or a stated guard that is not there; low — wrong under an unlikely condition, or cost without failure; nit — hygiene. Confidence is stated where a finding was not proven.

---

Findings

High

H1. Every published release image is stamped VERSION=e2e, so the backup/restore version check never checks anything. release.yml:75-84 builds the image with VERSION=X.Y.Z and tags it selfpost:e2e; the e2e step then runs docker compose build through test/e2e/compose.override.yml:23-32, which rebuilds the same tag with args: VERSION: e2e; docker tag selfpost:e2e …X.Y.Z-amd64 pushes the rebuilt one. Confirmed: docker run ghcr.io/mixeme/selfpost:1.3.0 panel -version prints e2e. Consequences: the footer and -version lie; backup.CheckRestore compares a manifest's version to e2e, so a backup of any release restores into any other and the promise in development.md:270 and guide.md:191 ("the version is the single source … so they cannot drift apart") is false. Every image this workflow has ever published is affected; they stay as they are (owner, 2026-10-12: old tags and images are history and are not touched) — the fix is for the next publication.

H2. Renewed TLS certificates are never picked up while the container runs. postfix-config.sh:75-89 copies the mounted cert and key into /etc/postfix/tls-internal/ at start and points Postfix at the copy; postfix-cert-reload.sh runs postfix reload daily but nothing re-copies. The copy is the only tls-internal reference in the tree; on the test server the copy carries the container's start time. Scenario: Let's Encrypt renews at day 60, nobody restarts, from day 90 ports 465/587 serve an expired certificate. architecture.md:33,154-155 and supervisord.conf:92-95 say the opposite.

H3. The queue-reconcile sweep marks delivered mail bounced whenever the tailer is blind. Four mechanisms, one consequence (a row the log would have resolved as sent is closed as a failure, which the Outbound log shows and the notifier believes):

• Tailer stuck, sweep running. logtail.go:559-581: caughtUp closes after the first drain and never re-opens; reconcileLoop (:174-191) then runs every 5 min regardless. If openAt fails after a rotation (:573-576, e.g. postlogd recreated the file 0600 — entrypoint.sh:132 says this happens) the loop logs and retries forever on the old inode while every new delivery line goes unread; two minutes later reconcile() finds the rows absent from postqueue -p (they were delivered) and bounces them.
• Truncated queue listing read as complete. postfix/queue.go:64-68,81-87: Queue() caps postqueue -p at 4 MiB and appends a text note; QueueIDs() parses that text and returns the first 4 MiB of ids with no error. A deferred backlog of ~30k messages — the outage the sweep exists for — makes every row past the cut bounced while Postfix still holds the mail. queuejson.go:48-52 handles the same cap correctly for the notifier; the two drifted.
• Rotation drain window. :567-578 drains the renamed inode once more; postlogd keeps writing to it until master's reload reaches it (seconds), so lines written in that window are never read. architecture.md:372-373 claims the drain covers it.
• No status guard on close. UpdateStatus (store/sendlog.go:86-98) is WHERE queue_id = ? AND to_addr = ?; the reconcile path can overwrite a sent written an instant earlier, and a reused short queue id to the same recipient rewrites a months-old row.

H4. Application logins are case-sensitive in the panel and case-insensitive in Postfix; a tenant can take another tenant's envelope sender. app/validate.go:35-48 accepts A-Z and nothing lowercases a login; applications.login is UNIQUE under BINARY collation (0001_init.sql:91), so Ops and ops are two applications. The sender map is written as stored (postfix/postfix.go:95-115); Postfix's reject_sender_login_mismatch compares logins without regard to case (the code's own comment, milter/sender.go:35-36, and architecture.md:239-240 say so). Scenario: domain A has Ops bound to @a.com; a domain-role administrator of B creates ops in B, authenticates as ops, sends MAIL FROM:<x@a.com> with From: y@b.com. Postfix accepts the envelope (@a.com Ops matches), the sender check judges only the header, which is B's — the message leaves with A as its SPF and bounce identity. Confidence: the panel side is confirmed; Postfix's compare is taken from the code's own statement. Fix shape: lower-case at validation, COLLATE NOCASE on the column (a migration — check the live registry for a pair first).

Medium

Persistence, restore, backup

M1. The restore manifest is consumed before anything that can fail. backup.go:436-446 deletes manifest.json inside CheckRestore; main.go:282-300 runs store.Open and resyncAfterRestore afterwards. If either fails, supervisord restarts the panel with restored == false and the one-time drift-healing Resync never runs. Docs say "on the first successful boot"; the code says "on the first attempt". restore_test.go passes testNoReload=true and never fails Open/Resync, so nothing covers it.

M2. Three queries on the single connection run without the index they need. Measured with EXPLAIN QUERY PLAN against the real migrations:

• CountMessages by domain (ratelimits.go:186-190) and sendStats (stats.go:77-94) use idx_send_log_domain (domain) and apply created_at >= by scanning every row of the domain over the retention window — on every MAIL FROM of an application under a domain-level limit, with the milter, tailer and HTTP queued behind it. The baseline's comment promises a composite index for "the journal milter … on every MAIL FROM"; only the app_login half exists.
• ListQueuedOlderThan (sendlog.go:117-122) is a full SCAN send_log every 5 min (no index on status; the cutoff matches nearly every row). Mirror of what 0005 fixed for rejected: a partial index (created_at) WHERE status = 'queued' makes it a handful of rows.
• Scoped QuerySendLog (sendlog.go:214-219) sorts a domain's whole scope with a temp b-tree for every page a domain user opens.

M3. selfpost-backup with a wrong SELFPOST_DB_PATH creates an empty database and ships a "valid" backup of nothing. backup.go:394-405 opens the source without mode=ro or an existence check; VACUUM INTO on a non-existent path succeeds (measured), creates the file on disk and snapshots a 4 KiB empty database; Create and Check pass. A restore of it opens the /setup link on a host that still carries live DKIM keys and sasldb2.

M4. secretfile's KDF cap does not do what its comment says. secretfile.go:108-113,184: maxLogN=20, maxR=32, maxP=16 lets a 44-byte header demand 128·2²⁰·32 = 4 GiB and sixteen times the CPU, under a comment that says the cap exists "so a hostile header cannot make the reader allocate its way to an out-of-memory kill". The domain-import form calls NewReader on the upload. Operator foot-gun rather than attack surface (global role only), but the stated guarantee is false; the real writer never exceeds N=2¹⁵, r=8, p=1.

M5. The CLI enforces no password floor for an encrypted backup. cmd/selfpost-backup/main.go:170-187 takes any non-empty string; the panel enforces validate.MinSecretFilePasswordLen (12); security.md:291 says the minimum applies to these files. (The roadmap's "empty password accepted" item: an empty password is not accepted as an encryption password — empty means plain output by design; the gap is the floor.)

M6. The compose pin, README and guide point at an image that does not exist. deploy/docker-compose.yml:24, README.md:115 and guide.md:187 pin ghcr.io/mixeme/selfpost:2.4.0 (2.4.1 since this cut); GHCR holds 1.0.0 and 1.3.0 only (checked through the Packages API and by pulling); no 2.x was ever published, and tag v2.0.0 exists in git without a release. The quick start cannot work for anyone but the owner, who runs locally built -post.<sha> images. Decided (owner, 2026-10-12): the old tags and images are history and stay untouched; the pin points forward — it is made true by the next publication, after H1 is fixed, and until then the documents should not claim an image that is not there.

HTTP surface

M7. The last-global-user guard is check-then-act without a transaction. store/users.go:315-341 (DeleteUser: CountGlobalUsers then DELETE), handlers_users.go:277-287,317-322 (demotion), users.go:80-97 (CreateGlobalUser: UserExists then INSERT). MaxOpenConns(1) serialises statements, not pairs. Two globals deleting or demoting each other at once leave no global; /server/* answers 404 to everyone and recovery is database surgery. security.md:155 promises the guard.

M8. Login discloses whether a username exists by timing. auth/handlers.go:109-121: unknown user returns at :114 without bcrypt; a known one costs ~250 ms (cost 12). The test compares bodies only. Rate limiting slows enumeration; a dummy compare stops it.

M9. submitUserUpdate is five writes with no transaction. handlers_users.go:304-336: UpdateUser → endSessionsAfterPasswordSet → SetUserRole → ClearUserReach (error only logged) → SetUserReach. A failure after the first leaves a user renamed and signed out everywhere with the old role, or a domain user with empty reach — the state the "Assign at least one domain" check exists to prevent.

M10. Domain import trusts the export file for everything but the domain name, selector and password. domain/transfer.go:145-190: DMARCRua is not run through validate.Email (the form path is), client IPs not through parseIPList, rate-limit Mode/WindowSeconds/MaxMessages not through the form's parsers. WindowSeconds: 0 yields a limit that counts nothing (milter/ratelimit.go:61); garbage IPs make the imported application refuse every client with no word at import. app/validate.go:50-55 calls the file "untrusted input" — for the password only.

M11. Polling stops for good after one failed refresh. panel.js:202-257: the scheduled poll is htmx.ajax(...) without source, so htmx attributes it to document.body; htmx:responseError therefore reaches an element with no data-poll and nothing reschedules; htmx:sendError and htmx:timeout are not handled at all. A panel restart while Overview or Health is open freezes the page until the tab is hidden and shown. Related: an expired session answers an HX-Request with 303 /login (auth/middleware.go:14); the XHR follows it and the sign-in document is swapped into the box — no HX-Redirect on either side.

Log tailer and checks

M12. deliveryRe is unanchored; a stranger's HELO parses as a delivery result. logtail.go:85: a port-25 reject line ending helo=<41E862C00D9E: to=<victim@…>, status=sent> yields (id, recipient, sent) (probed), and UpdateStatus rewrites the matching row. The attacker must guess an id+recipient pair, so impact is bounded — but lines.go:21-24 and inbound.go:48-51 anchor at the head for exactly this reason, and this pattern, the one that writes status, does not.

M13. The sender-check deferred counter misses the case it is documented to see. sendercheck.go:50 matches only postfix/…smtpd[pid]; header/EOH events run in cleanup(8), so the warning for a panel that accepts the socket but hangs at end-of-headers is logged by postfix/cleanup. The fixture at sendercheck_test.go:64-68 puts those lines under smtpd and says it was "composed from Postfix's source rather than seen". architecture.md:286-287 claims the figure sees "a panel that is running and not answering". Confidence: plausible; one kill -STOP on the stand settles it.

M14. Resolver fall-back is hollow when the first resolver drops packets. dnscheck/resolver.go:72-90 tries resolvers in turn under one 5 s context; a blackhole eats the budget and the second is asked with an expired context. The DefaultResolvers comment promises otherwise. M15. dnscheck/server.go:26-33,42-46 turns a resolver timeout into "does not resolve" / "no PTR record" (every other check routes transient errors to Unknown), and empties ServerIPs so the SPF check goes Unknown in the same minute. architecture.md:339-340.

Infrastructure

M16. envdoc_test.go compares three hand-written lists to each other and reads neither main.go, the scripts nor guide.md. envdoc_test.go:70-136. Already drifted: ANTISPAM_DIR, INBOUND_JOURNAL_MILTER_SOCKET, INBOUND_SPAM_LOG_RETENTION_DAYS, RATE_LIMIT_MESSAGES_PER_IP, RATE_LIMIT_WINDOW_SECONDS are read by loadConfig and absent from loadConfigKeys; postfix-config.sh and postfix-wrapper.sh read INBOUND_JOURNAL_MILTER_SOCKET, SELFPOST_DATA_DIR, SELFPOST_DB_PATH, INBOUND_RELAY_ENABLE, absent from buildScriptKeys. The tests pass; development.md:353-357 and :503-504 describe a guard that does not exist.

M17. The supervisord control socket is 0770 root:selfpost and the postfix and opendkim users are in selfpost. supervisord.conf:26-29 ("only root and the two SelfPost services"), Dockerfile:68-71 (three users); confirmed on the server. A compromised smtpd can supervisorctl stop panel (all submission deferred), stop opendkim or shutdown. No escalation; a whole-stack denial from the least trusted process.

M18. DMARC maps are never ownership-normalised. entrypoint.sh:108-125 normalises five maps; dmarc_recipients/dmarc_transport/dmarc_relay_domains are created by postfix-config.sh:315-317 as root only when missing. A tar restore as root (guide, methods 3/4) yields root:root 0640, which smtpd cannot open → 451 table lookup problem on every report until the panel next rewrites them. Confidence: plausible, needs a run.

M19. logrotate state lives in the container layer. logrotate-loop.sh:42 uses the default /var/lib/logrotate/status, not under /data. On the server the state file is dated the container's start; with no entry logrotate records today and does not rotate, so a host that recreates the container daily (.env change, image bump) never rotates mail.log. The e2e uses -f and cannot see it.

M20. test.yml never compiles test/e2e. A type error there surfaces in release.yml after the GitHub Release is published and the tag checked out.

M21. deploy/traefik/extract-cert.sh writes empty PEM files and exits 0 on a jq miss (:17-25, sh without pipefail; jq exits 0 on no match); the files are 0644 until :27. Postfix TLS breaks at the next restart.

Low

• L1 -decrypt -o FILE truncates the output before the password is verified (main.go:132-148); a wrong password leaves an empty file where a good one may have been.
• L2 Backups skip symlinks silently and Check does not notice (backup.go:323-327); a certs/ of symlinks (the Let's Encrypt layout) yields an archive with no TLS material while the guide promises refusal.
• L3 Trailing bytes after the last secretfile chunk are ignored (:400-416): two envelopes concatenated open as the first.
• L4 parseTrustedProxies turns a bare ::ffff:10.0.0.1 into ::/32 silently (main.go:247-253).
• L5 tls_policy is keyed by next-hop; two inbound domains to one [host]:port with different TLS modes write duplicate keys and the first (by domain order) wins with a warning — confirmed on the server. The per- domain TLS mode in the form is not what Postfix does (postfix/inbound.go:69-72).
• L6 Multi-file map rebuilds are not all-or-nothing (inbound.go:42-55, dmarc.go:31-41); a write failure mid-sequence leaves relay_domains ahead of transport.
• L7 atomicfile.Write never fsyncs the directory and the package has no tests (atomicfile.go:40-44); every map writer goes through it.
• L8 The refusal reply names the other sender's domain (sender.go:334) — an oracle for the server's signing names, available to any authenticated application.
• L9 go-milter accumulates every message's headers for a connection's life (library session.go:173-187, cleared on ABORT only); the sender check is safe because it ignores h, and the test passes nil so would not catch a "simplification" to h.Values("From").
• L10 parsePage is not clamped to lastPage for the Outbound log (handlers_monitor.go:357-371); the inbound log clamps. Known roadmap row, confirmed.
• L11 parseAutoMultiplier accepts NaN (handlers_ratelimit.go:44-49); SQLite stores NULL, read-back is 0, recalc silently uses 2.5 (probed).
• L12 Outbound rows belong to a domain by name (handlers_monitor.go:97,345): a domain deleted and re-added shows its earlier rows to the new assignee. security.md:419 accepts this for the inbound log only.
• L13 HandleReload answers failures as plain-text http.Error (handlers_domains.go:173-184), and the flash "Configuration regenerated from the database" oversells: DMARC maps and filter-list files are skipped (architecture.md:856-859 says so for DMARC; Lists.Sync is documented nowhere).
• L14 http.Server has only ReadHeaderTimeout (httpserver.go:145-149).
• L15 The Account password change shares the per-IP login bucket (handlers_account.go:142): ten wrong current passwords lock sign-in for that IP 15 min.
• L16 Login issues a fresh token but does not destroy the session a presented cookie already holds (auth/handlers.go:123-130).
• L17 domain.Service.Delete returns a bare error after the row is gone when resync() fails (service.go:202-210); apps.Resync() is not reached, so the sender map keeps the purged logins.
• L18 A rate-limited refusal is a grey tag in the Outbound log (outbound_log.go:213, no such status in components.go:132-146); its two siblings are red, and the test pins the grey.
• L19 Recalculate rate limit on the application form submits the whole form and discards every unsaved edit (out_app.html:67, handlers_ratelimit.go:178-189).
• L20 DMARC parse rewrites <pct>0</pct> to 100 and stores a missing date_range as 1970 (parse.go:125-136).
• L21 Several mailStack instances per process with no shared lock on table rebuilds (run, newPanel, runNotifier); the notifier's EnsureServerSigned at start can write a stale table over a domain added in the same second. L22 IncrDMARCParseFailures is read-modify-write across concurrent dmarc-ingest processes (dmarc.go:163-173).
• L23 The restore-boot Resync reloads Postfix before Postfix exists (panel priority 200, postfix 300) — an error line per restore boot; the DMARC bootstrap works around the same thing with a no-op hook and does not say why (main.go:301-307).
• L24 health/processes.go:43, supervisor.go:17,30: exec.Command without context; /healthz and the 5 s status poll hang on a wedged supervisord. L25 health/cert.go:70-83 assumes the first PEM block is the leaf and grades a not-yet-valid certificate OK; machine.go:193-204 advances prevAt when only one of cpu/net succeeded.
• L26 Only successful sends count toward the notifier's 60/h ceiling (notify.go:196-207); with sendmail broken, CheckLimits spawns a 20 s process per recipient per minute forever.
• L27 v=dmarc1 matched case-insensitively (records.go:74,121); RFC 7489 §6.4 says the version must match exactly, so a lower-case record is graded OK and ignored by receivers. L28 The domain cache is keyed by name only (dnscheck.go:155-173) though the verdict depends on ExpectedDKIM/ServerIPs.
• L29 lineTime normalises a Feb 29 line read the following January to Mar 1 (inbound.go:254-261); isoStampRe accepts a zone-less stamp that time.Parse then rejects (dated now).
• L30 OpenDKIM: Syslog no with no LogWhy/Logfile — it logs nothing anywhere; no Canonicalization relaxed/simple; no OversignHeaders From (the header-append replay is the class 2.4.0 is about).
• L31 Postfix: smtpd_tls_protocols at the 3.7 default >=TLSv1, disable_vrfy_command = no (VRFY on 25 enumerates relay_recipient_maps), the stock (Debian/GNU) banner — all confirmed on the server. RATE_LIMIT_*, *_SIZE_LIMIT, MILTER_*_TIMEOUT reach postconf -e unvalidated; a non-integer passes postfix check and kills every smtpd at connect.
• L32 crashexit fires on FATAL only, which arises only from startretries exhausted within startsecs (1 s): a program that crashes after one second cycles forever and never exits the container. architecture.md:40-41 reads wider than that.
• L33 .dockerignore excludes docs, dev, bin but not test/; a developer build after an e2e run ships test/e2e/.stage/data (DKIM keys, sasldb2) into the build context. Base images pinned by tag, not digest.
• L34 e2e: stage.go:77 pulls alpine:3.20 from the internet; hostname_gate_test.go:97-101 runs with the host's real network; sender_check_test.go:141 has a fixed time.Sleep(2s) against development.md:409 "no fixed sleep".

Nits

Dead code: handlers/authz.go:63 domainIDSet; view.Facts (components.go:414); dmarc.Service.HostedSuggestion, ValidateHostedAddress (dmarc/service.go:98-109); app.Service.UpdateMode (app/service.go:132-145, superseded by SaveSettings); Settings.WithResult (view/account.go:137-145) re-implements resultFlash; KitPage()'s hard-coded User: "admin" (overwritten by the handler); dead method checks in handlers_dmarc.go:20-24 and the default: branches of HandleUserNew/HandleUserEdit/HandleServerSettings (mux patterns are method-bound); the postfix set-permissions branch in entrypoint.sh:170-172 (postfix check has already created the tree). Duplication: canViewDMARCDomain repeats lookupDomain's check; ParseForm twice on the user-edit POST; Lookup then Touch per request in auth/middleware.go:17,28; retry.go:58-67 repeats the six postconf keys as literals instead of retryConfKeys; three header-name canonicalisers with two behaviours (milter.go:124, inbound.go:89, sender.go:248). Behaviour: submitUserCreate does CreateUser then UpdateUser for the e-mail; self-demotion redirects to a page that is now 404; redirectHome gives a domain user with no outbound reach an empty 200 where the inbound list gives 404; sort.Slice in postfix/inbound.go:59 reorders the caller's slice; strings.Split(raw, "\n") in sender.go:65,104 keeps \r; lists.go:428 trims the filter key while the Lua side reads it verbatim; InDomain.WithMX labels a warn status "NOT CHECKED"; senderSession does not override Body (the library discards the session after every accepted message — harmless, inconsistent with milter.go:130-134). The handlers test package takes ~105 s on this machine (likely bcrypt cost 12 on the user-creation paths).

Doc drift

architecture.md: :167 (§ OpenDKIM) still says the From: header is not held against the login — contradicted by 2.4.0 and by § Sender check in the same file; :272-274 credits /run being tmpfs for a fresh .alive mark (it is a plain image directory; a restarted container keeps the mark); :328-329 "a deleted domain's row goes with it" — only dns: rows, limit: lingers up to 48 h; :493 and § Who may open what omit /inbound/log and /inbound/filter-lists; :800 names only migration 0002; :832-836 omits inbound_spam_log, inbound_spam_list, notify_state, sender_check_days; :838 "the panel is the only writer" — dmarc-ingest is a separate process that opens the store, migrates and writes; :856-859 vs backup.go:19-23 and httpserver.go:52-53 ("the same Resync") — the button skips DMARC maps and filter lists; the Internal env list omits ANTISPAM_DIR and INBOUND_JOURNAL_MILTER_SOCKET; § View engine calls the kit page "fixtures" (it is production data). schema-migrations.md: :20 "chain head user_version = 5" (six files; TestBaselineSchema asserts 6); the "Schema at head (v2)" table omits sender_check_days and send_log.reason/header_from. guide.md: :214 INBOUND_JOURNAL_MILTER_SOCKET "Set in: compose" (it is not); :1260 lines "deleted together with the row" (at the next six-hourly prune, and not while a reused id has a row); :720 "first successful boot" (M1). development.md: :270 version/tag coupling (H1); :353-357, :503 the env guard (M16); :404 "milter fail-open" as a covered negative (the suite now proves deferral); :409 "no fixed sleep". roadmap.md: template-data-typing describes renderDomainDetail with forty map keys — gone since 2.0 (one typed-before- render map remains, the health sampler's); structured-logging says logf "in five packages" — it is in 21 files of internal/web; review-2026-08-followups row "empty password" is misdescribed (M5). Code comments: milter.go:7-11 "it never rejects" (it tempfails at MAIL FROM); store/sendlog.go:19-22,61-65 StatusRejected/InsertRejected describe the level-2 case only; store.go:2-3 "domain-admins"; schema_test.go:15 TestBaselineIsTheOnlyMigration; health/machine.go:225 "see roadmap panel-docs … later" (shipped 1.8.0); inboundspam.go:349-350 "does not read the whole journal" (it scans it); logtail.go:466 mentions copytruncate; supervisord.conf:25 "two services", :107 "7-14 files"; entrypoint.sh:119-120 "no secrets" in /data/antispam (filter.key); docker-compose.yml:85,88 cite a guide section by a title it does not have.

Test gaps

No test for: the H1 stamp (a panel -version assertion in the e2e against the version the stand was told to build); cert re-copy (H2); follow() with an unreadable or recreated log, a truncated postqueue -p, lines written to the old inode after the switch, UpdateStatus from sent (H3); two logins differing in case (H4); M1's ordering (restore tests never fail Open/Resync or exercise the real reload hook); CheckRestore with malformed JSON; M3, M5, L1, L2; secretfile concatenation (L3); the sender-check counter with the real program name (M13); a LookupAddr timeout (M15); TestLimitRefusalsSince EXPLAINs a stand-in query, not the real one; TestOpenMigratesA20Database stops at 0004; Import with a bad rua / IPs / window 0 / mode (M10); panel.js has no test at all (M11); login timing (M8); concurrent last-global (M7); TestOriginCheck lacks the Host-with-port vs Origin case; TestServerSubtreeIsGlobalOnly skips any /server/ route not in routes.always; internal/atomicfile and internal/supervisor have no tests; TestRenderInboundMaps has no two-routes-one-next-hop case (L5); TestSenderCheckResetsBetweenMessages passes h = nil (L9); the handler test rows of review-2026-08-followups (apps/domains/users/dmarc handlers directly, internal/dmarc service) are still open.

Checked in stage 6 (2026-10-15): closed since, each by the test that now holds it: the H1 stamp (e2e image_stamped_with_version); cert re-copy (e2e renewed_certificate_picked_up); a recreated log and lines written to the old inode after the switch (TestFollowTailsAndRotates, TestFollowReadsTheRotatedFileAfterTheSwitch, TestALogRotatedWhileDownIsReadToItsEnd); a truncated postqueue -p (TestQueueIDsRefusesATruncatedListing); a sent row left alone by the sweep and by a reused id (TestCloseUnreportedClosesOnlyQueuedRows, TestUpdateStatusRewritesOnlyTheNewestRow); H4 (TestAddApplicationRefusesALoginDifferingOnlyInCase, TestServiceCreateFoldsTheLogin); M1's ordering (TestARestoreStartThatFailsIsARestoreAgain; the real reload by e2e full_backup_restore); M3 (TestSnapshotRefusesAMissingDatabase, TestBackupOfAMissingDatabaseIsRefused); M5 (TestEncryptionPasswordFloor); L1 (TestDecryptLeavesTheOutputAloneOnFailure); L2 (TestCertsLinksAreFollowed, TestCheckRestoreReadsOnlyRegularFiles); L3 (TestBytesAfterTheLastChunkRejected); M13 (e2e panel_hung_mid_message_is_counted, and the unit fixture now holds the cleanup lines as the stand wrote them); M15 (TestServerPTRTimeoutIsUnknown, TestServerAddressTimeoutIsUnknownAndKeepsTheAddresses); TestLimitRefusalsSince (the real limitRefusalsQuery is planned in TestSendLogQueriesUseTheirIndexes); TestOpenMigratesA20Database (migrates to the head, 11); M10 (TestDomainImportIsCheckedLikeTheForms); M11 (TestPanelJS: poll_test.js, unsaved_test.js); M8 (TestLoginTakesAsLongForAnUnknownName); M7 (TestTheLastGlobalSurvivesTwoAtOnce); internal/atomicfile and internal/supervisor (tests of their own); L5 (TestRenderInboundMapsOneTLSLinePerNextHop); the handler rows of review-2026-08-followups (struck in the roadmap with their reasons; users, DMARC and apps have handler tests). The five then still open, named by no step, were closed in stage 6 the same day: follow() with a log gone past blindMin and made again (TestFollowCountsALogItCannotReadAndReadsItAgain); CheckRestore with a manifest that is not JSON (TestCheckRestoreRefusesAManifestThatIsNotJSON); a Host with a port against an Origin (TestOriginCheckComparesThePort); every /server/ pattern in web.go registered through server() and listed in the routes table (TestEveryServerRouteIsGuardedAndListed - it found GET /server/health/fragment missing from the table; the route was guarded, not tested); L9 as the owner decided it, the comment on senderSession.Headers and TestSenderCheckResetsBetweenMessages passing an h that holds the earlier messages' From:.

Checked and sound

Worth writing down so the next review does not re-read it: the sender check's layered rule holds against every From: shape tried (quoted local parts with @, CFWS, encoded words decoding to addresses, groups, <>, address literals, obs-route) — the "one @ outside quotes" rule is the real backstop; extension stripping and case folding match Postfix; every error path of the check tempfails; level-2 accounting reserves under one lock and cannot be inflated by multi-RCPT or bypassed by ABORT/reconnect; the inbound verdict is trusted only under the queue-id-bound HMAC; the port-25 -o overrides are complete and read back by the e2e through postconf -Ph; the generated milter chain is exactly sender(tempfail) → OpenDKIM(tempfail) → journal(accept); /server/ is guarded at registration and per handler; scopes fail closed (1 = 0); sessions store the hash only, __Host- cookie, constant-time setup token; SaveSettings restores prior state on mid-way failure; secretfile's AAD binds header, counter and last-flag, and truncation, reorder and tamper are tested; shutdown order (tailer offset saved, HTTP drained, DB closed last) is right; rspamd map names, deny message, symbol prefixes and key-file format match the Go side byte for byte.

---

Plan

Seven stages (six, and stage 2b split off stage 1 on 2026-10-14), each a self-contained cut, each small enough for one sitting, with model routing per development.md (security/infra → Opus, UI/docs → Sonnet). "Done when" is the test that proves it, not the edit.

Agreed with the owner on 2026-10-12 (roadmap row audit-2026-10-12): stage 0 first, then stages 1 and 2 as one cut, the rest after in any order (stage 2b after 2.5.0, before the others). This is the active plan in the sense of development.md § Plan checklists: the - [ ] lines below are the implementation checklist, a line is ticked in the commit that finishes it, and the roadmap's Progress column counts them. After a context reset, start at the first unchecked line of stage 0. The owner's decisions on the contested items are written next to the steps they changed and under Decided with the owner at the end; they are not open.

Working rules that apply, from development.md and the project's memory: every step updates CHANGELOG [Unreleased] in its own commit; a version is cut only at the owner's word; main is pushed, tags are not; design changes (the only one in this plan would be a Delivery log card, and it is not planned) are mockup-first; a migration the test server has already run is never edited — the next file is added; the published 1.x tags and images are history and are not touched.

Stage 0 — release hygiene (no product code) → 2.4.2 (shipped 2026-10-14)

• Stamp the right version (H1). Make the e2e build take the version from the environment (VERSION: ${SELFPOST_E2E_VERSION:-e2e} in compose.override.yml, set by release.yml), or have the e2e reuse the image already built instead of rebuilding. Done when the e2e asserts panel -version equals what it was told, and release.yml runs that assertion before the push. — Opus
• Compile the e2e in test.yml (M20): cd test/e2e && go vet ./... && go test -c -o /dev/null . (not go test -run xxx: TestMain would still run and bring up the stand). — Haiku
• .dockerignore + test/ (L33). — Haiku
• The pin (M6): old tags and images are not touched (owner, 2026-10-12). The next publication through the fixed release.yml is the one the pin names; until then README/guide say that the current line is not yet published rather than name an image that is not there. — Sonnet
• Toolchain and deps: modernc.org/sqlite, x/crypto, x/sys; note in development.md that the image is rebuilt after each Go patch release (golang:1.26-bookworm floats, so a rebuild suffices). — Haiku
• The four staticcheck items. — Haiku
• A negative e2e case passes only on the refusal it is about (owner, 2026-10-14; seen, not found by the audit). negative_level1_ratelimit sends up to sixty messages and waits for one to be refused; it once passed on "limit tripped on attempt 1/60" when the connection was reset because Postfix was still starting. The case counts as the limit only the reply code of the limit, and only after at least one message was accepted; a reset, a timeout or any other reply fails it and says what came. Then every other negative case of the suite is read for the same mistake - any failure taken for the expected refusal - and held to its own reply code. - Opus Done 2026-10-14; what the stand showed: waitForSMTPSReady took a TCP connect for readiness, and Docker's port proxy accepts before Postfix listens - that was the reset; it now waits for the TLS greeting. Held to anvil's 450, the case then failed on its first message: 553 … not owned by user l1app, an smtpd process from before the panel's postfix reload still holding the old sender_login_maps - so the level-1 limit had never been the refusal the case saw. The first message now waits (at most 15 s) for the map; the limit trips at 51/60 with 450 4.7.1 Error: too much mail from. In the product this is a second or so after an application is saved in which its first message can get a permanent 553; not worth a change (an application is configured after it is created, not within the second). Measured 2026-10-14: applications created and used after 0, 10, 40 and 110 s of waiting (the first send two seconds after the save at the soonest) were all served by an smtpd started after the reload; the stale one only ever answered within the reload's own second. The e2e waits for the map after each application it creates (waitForSenderMap), after the main positive path failed on it once. The other cases are held to 451 4.7.1 (level 2, both MAIL FROM of one connection), 553 at RCPT (sender mismatch), 554 … Client host rejected at RCPT (no AUTH, and the relay case, whose comment claimed reject_unauth_destination - on 465 the client restriction comes first; port 25 holds 554 … Relay access denied), 550 5.1.1 (port 25 unknown recipients), any 4xx reply (panel down).

Stage 1 — mail that is shown wrong (H3, H2, M12, M13) → 2.5.0 (shipped 2026-10-14)

• QueueIDs on a truncated listing returns an error, as postqueueJSON does; reconcile already treats an error as no information. Test: a 5 MiB fake postqueue -p. — Opus
• Reconcile only closes queued rows (AND status = 'queued' in the reconcile path's update) and marks what it closed: reason = 'reconcile' on the row (the 0006 column; a new value, no schema change). — Opus
• Measure before deciding the sweep's fate (owner, 2026-10-12: the keep / remove / lost question needs real data, and today the only record of a sweep is a stdout line that dies with the container). A table delivery_log_days (migration 0007, DDL only; a row per UTC day, kept a year, the shape of sender_check_days) with four counters: reconcile_closed (rows the sweep closed); reconcile_contradicted (a reason = 'reconcile' row the tailer later rewrote from a status= line — the log wins, and the rewrite is counted); tailer_blind_seconds (follow() unable to read: reopen failing, file gone — the sender check's downtime mechanics); queued_over_lifetime (rows still queued past maximal_queue_lifetime at the day's end — what removing the sweep would leave visible). No screen for this (owner, 2026-10-12: nothing is drawn that ordinary operation will not need): the table is read off a copy of the database when the month is up. Whether a durable Delivery log card (tailer blindness, stuck rows — the two counters that outlive the decision) earns a place on Health is a separate question, asked only if the data shows those numbers are ever non-zero. Decision rule, written now: closed = 0 for a month → remove the sweep; closed > 0, contradicted = 0 → keep; contradicted > 0 → each case explained by tailer_blind means fix the tailer and keep the sweep, otherwise lost or remove. Decide after 30 days of data, or sooner if contradicted is ever non-zero. — Opus Changed by the reviews of stage 1 (2026-10-14): a fifth counter, reconcile_missed - a closed row whose delivery line was in mail.log or mail.log.1 when the sweep closed it; contradicted alone cannot see a lost line, which is never read late. queued_over_lifetime cannot show "what removing the sweep would leave visible" while the sweep runs (those rows are closed by then); it counts rows stuck despite the sweep, and what removing it would leave is closed - missed - contradicted. A contradicted row keeps the mark reconcile-contradicted. The rule reads accordingly in architecture.md § Log tailer.
• Deferred until the data is in (not a step yet): pausing the sweep while the tailer is not reading (publish "healthy" from follow, skip the sweep otherwise), and the rotation window (keep draining the old inode until it has not grown for a tick and the new file has a line). Both are fixes to the tailer's blindness and stay on the list whatever the sweep's fate; their priority follows tailer_blind_seconds.
• Anchor deliveryRe at the head of the line like its two siblings; test with the probed HELO and quoted-sender lines. — Opus
• TLS re-copy (H2): postfix-cert-reload.sh compares the mounted files to the copy and re-copies before postfix reload (atomically, same modes); e2e replaces the cert in the stand and reads the new serial off 465. — Opus
• Deferred counter (M13): on the stand, kill -STOP the panel mid-message and record which program logs the warning; widen the matcher accordingly and replace the composed fixture with the seen lines. — Opus
• Docs: architecture § Log tailer, § Queue reconcile (the measurement and the rule), § Sender check "what it costs"; guide § Delivery log, § Health. — Sonnet

Stage 2 — the tenant boundary and the registry (H4, M7, M9, M10) → in 2.5.0 (migration 0008; shipped 2026-10-14)

• Logins lower-cased at validation (owner, 2026-10-12: fold, do not refuse) and applications.login COLLATE NOCASE (migration 0008 rebuilds the unique index). Before the migration a check that no two existing logins differ only in case (there is one installation; look). Tests: Ops/ops refused; the map written lower- case. Then verify Postfix's compare on the stand once (MAIL FROM as the other case) so the doc's claim is seen, not assumed. — Opus
• Last global in one transaction: DeleteUser/SetUserRole take a tx, count and act inside it; CreateGlobalUser likewise. A concurrency test with two goroutines. — Opus
• submitUserUpdate in one transaction (store method that takes the whole change). The users pages get direct handler tests while they are rewired - a row of review-2026-08-followups, placed here 2026-10-14. — Opus
• Import validates like the form: route DMARCRua, IPs, limit mode, window and count through the same parsers; refuse the file with the validator's text. Tests per field. — Opus
• Send-log rows belong to a domain by id (L12; owner, 2026-10-12). send_log.domain_id INTEGER REFERENCES domains(id) ON DELETE SET NULL in migration 0008, backfilled once by name against the registry as it is at migration time (rows of a domain already gone stay NULL); written by the journal milter and by the sender check's refusal rows at insert, looked up from the login's domain; the domain name column stays for display and for the global view. A domain user's log, stats, auto rate limit and the notifier's refusal count select by id; a NULL row is nobody's but the global role's. A domain deleted and re-added under its old name therefore starts with an empty log. The inbound journal (inbound_spam_log, the same accepted risk in security.md:419) gets the same treatment in the same migration — my proposal, so that the two logs do not follow two rules. Tests: delete-and-re-add shows nothing to the new assignee; the old rows are still in the global log. — Opus Done 2026-10-14. DMARC reports stay by name on purpose: a report is about the name in DNS, and whoever holds the name now is its reader. Changed by the reviews of stage 2: the backfill follows the rule new rows follow (the application's domain, then the name) and gives a row only to a domain - and an application - that existed when it was written; an application's numbers are its login's rows of its domain, from its creation on. Deleting a domain now rewrites its rows' domain_id (14 s for 333k rows, measured by the review; milliseconds at this installation's size) - accepted, to revisit if a log grows that large.
• Docs: security.md § Roles (the guard) and the struck accepted risk, guide § Domain export, schema-migrations.md. — Sonnet
• Which key OpenDKIM signs with, seen and not reasoned (owner, 2026-10-13, from the Not verified list of from-header-binding). The check and OpenDKIM read From: with two parsers; that they agree was argued from the check's two rules, never run. An e2e case sends, as one application, the From: lines the check passes - quoted strings, comments, folded lines, groups, escapes, encoded words, an address extension - and reads the d= of the signature on what arrives: any domain but the application's fails the test. The lines live in one table shared with the check's unit test, so a line added to one is run by the other. A disagreement found is fixed in the check before the cut, and security.md § Mail path says the claim is tested. - Opus Changed by the reviews of stage 2: a second list, lines naming another sender in a quoted string, a comment, a quoted pair, capitals, an encoded word (from-lines-other.txt), each refused and never signed for it; the passed list gained capitals, a quoted local part and a name with a space after its @ (passed by design, signed for the own domain). L8 is half closed: refused-or-passed still tells an application which domains are hosted here. Owner, 2026-10-14: leave it - the probe is a trusted application's guess, one name at a time, as before L8; refusing every foreign address in a display name is not wanted (security.md, accepted risks). Done 2026-10-14: fifteen lines (internal/milter/testdata/from-lines-passed.txt), each signed for the application's domain alone, the signature valid. No disagreement. Groups are refused by the check, so none is in the file; a comment before the address - (the shop) a@domain, valid RFC 5322 - is refused too ("it must be one address"): strict, not a hole, left as it is.
• The 550 reply names no other domain (L8): sender.go:334 stops at "it names another sender of this server"; the refused message's page still shows the From: line as written (header_from), so the operator loses nothing. Test pins the reply text; e2e's sender-check negatives assert the reply does not contain the other tenant's name. — Opus

Stage 2b — the log tailer, and what the panel shows of a refusal → 2.5.1 (shipped 2026-10-14)

Split off stage 1 on 2026-10-14 at the owner's word: the six steps below were agreed on 2026-10-13/14, after the plan, and had made stages 1 and 2 a cut of twenty-one steps. Stages 1 and 2 ship as 2.5.0 in their first size; this stage follows as a cut of its own, with its own reviews. Five of the six change one thing, the tailer and the state it saves, and are done in this order - each stands on the one before.

• A log rotated while the panel was down is read to its end (owner, 2026-10-13; the Not done list of from-header-binding had it as "deferrals are lost", and it is wider). At start, a saved offset whose fingerprint is not the current file's means "start on the new file" (offset.go); what Postfix wrote between the panel's stop and the rotation is never read - delivery statuses, stored delivery lines, refusals, inbound journal rows, the deferred counter, and the recovery of unwritten refusals above. The tailer first opens the rotated file (mail.log.1), checks by fingerprint that it is the one the offset belongs to, reads it from the offset to its end, then goes on to the new file. The fingerprint gains the inode, which closes the roadmap's review-2026-08-followups row on collisions. Not found, or rotated twice: logged, with from when to when nothing was read. A test that stops the tailer, appends, rotates, appends, starts it, and finds both halves once. - Opus
• The deferred counter remembers a place in the log, not a second (owner, 2026-10-14, from the Not done list of from-header-binding; he chose it over leaving the case written down). After an unclean stop the tailer re-reads lines it had read, and the counter passes over every warning no later than the last one it counted (logtail/sendercheck.go, last); mail.log stamps whole seconds, so a warning of that very second, later in the file, is lost. The count and the place in the log it has reached are saved in one transaction, and what to pass over is decided by the place. Done after the rotation step above, on the same saved state - the place must stay true across a rotated file. A test that stops the tailer between two warnings of one second and finds both counted, once each.
  • Opus
• A refused message's page shows what Postfix wrote (owner, 2026-10-13: "the log box is required" - the box is the reason delivery-log-storage was built, and its absence from the picture he approved for 2.4.0 was an oversight, not a decision). Today every rejected row - wrong From, wrong client, rate-limited - has no queue id, and its page shows the box with one sentence, "This message never reached the queue, so Postfix wrote no delivery lines for it" (handlers_monitor.go). The sentence is untrue: Postfix wrote a line for each (milter-reject: END-OF-MESSAGE ... 5.7.1 From: ... from cleanup, NOQUEUE: reject: ... from smtpd). The tailer attaches that line to the row it belongs to, by client address, time and the addresses in the line, and stores it in send_log_lines like any other; the page shows it. First on the stand: read mail.log for the three kinds of refusal and write down what identifies each line. Then the mockup: the refused variant of out-message gets the box with the line in it (design: commit), then the code. The sentence stays only for a row whose line was truly never written or is older than this change, and says that. Done when the e2e shows the stored line on the page of each of the three refusals. - Opus
• A message the sweep closed does not say the receiving server refused it (review of stage 1, 2026-10-14). The history step of a bounced row reads "the receiving server refused the message permanently, or Postfix gave up ... The reason is in the delivery log below" (handlers_monitor.go); for a row with reason reconcile neither is known and the log has nothing. Such a row gets its own step: Postfix no longer held the message and mail.log never said what became of it, so the panel closed it as a failure. Wording in the mockup out-message first (design: commit), with the box of the step above. - Sonnet
• A refusal the panel did not get to write is written from mail.log (owner, 2026-10-13, from the Not done list of from-header-binding; built on the step above). The check answers 550 and writes the row a moment later, through a buffer; a panel that stops in between - a rollout, a crash - leaves an application that was refused and a log that does not know it. When the tailer meets a refusal line that has no row, it makes the row from the line (client, sender, recipient, reason), marked as recovered. The tailer resumes from its saved offset, so this covers a kill as well as a clean stop - which a flush of the buffer at shutdown would not. Done when the e2e refuses a message, kills the panel before the row is written, and finds the row after the restart, once. - Opus
• Downtime is counted when supervisord does not answer at start (owner, 2026-10-13, from the Not done list of from-header-binding). The panel asks supervisord once since when Postfix runs (sendercheck.go, postfixStarted); an error is logged and that start's downtime is not counted at all - the number the owner will decide by reads low. The question is repeated for about thirty seconds. Still no answer: the downtime is taken from the alive mark alone - an upper bound, Postfix may not have been running - and the day is marked as estimated (a column in sender_check_days, the next migration), which the card shows. The mark on the card is drawn in the mockup with the renamed column above (one design: commit). - Opus
• The Sender check card names what it counts (owner, 2026-10-13, from the Not done list of from-header-binding): the column Messages deferred becomes Connections turned away. Postfix refuses from the connection on while the check does not answer, before any message is offered, so a client that connects three times with one message is three, and a monitor that only connects is one; messages cannot be counted. Decided by the owner, 2026-10-14: since the M13 fix the figure also counts whole messages - those a panel that hangs mid-message makes cleanup defer, one per message, with a queue id - and the two are shown apart, not mixed: Connections turned away (smtpd's warning, the existing deferred column) and Messages deferred (cleanup's warning, a new column of sender_check_days in this stage's migration, beside the estimated mark). The counter writes each warning to its own column; the card has both columns, and the help topic says what each counts. The figures collected before do not mix: cleanup's case was counted only from the M13 fix on, not yet deployed when this was decided. The mockup health and help topic sender-check first (design: commit), then the page and view/help.go, word for word. - Sonnet
• Docs: architecture § Log tailer (the rotated file, the saved place, a refusal's line and a recovered row), guide § Outbound log and § Health, schema-migrations.md for the mark on sender_check_days. — Sonnet

Stage 2c — the sender check reads From: as OpenDKIM does → 2.5.2 (shipped 2026-10-15)

Decided by the owner on 2026-10-14, after reading OpenDKIM's source together (libopendkim dkim_mail_parse, called in mlfi_eoh on the first From: with SenderHeaders unset - build/opendkim.conf sets none): the signing address is the one in angle brackets, or a bare address, found while quoted strings and comments are skipped; encoded words are not decoded. The check's scan of the display name and comments for another sender of the server is a guard against two parsers parting ways - and it is what makes L8 half closed. This stage replaces the guard with the parser itself. Stage 3 and those after it move up one patch number each.

• Port dkim_mail_parse to Go, from the version the image runs (Debian bookworm's opendkim 2.11.0~beta2 source package, not upstream's develop branch), line for line, with its tokenizer (dkim_mail_first_special, dkim_mail_token). The check then decides by the address OpenDKIM will take: the application's - any address of its domain, or one of its list - passes, whatever the display name or a comment holds; anything else, a parse failure included, is refused. One From: line and one address stay required. The scan for another sender of the server in the name, in a comment and in a decoded name goes, and with it the reason that told an application which domains are hosted here (L8 closed; security.md's accepted risk struck). Whether the "one @ outside quotes" rule is still needed is decided against the port, and said. - Opus Done 2026-10-15: internal/milter/mailparse.go; the C itself compiled from the image's source agrees with the port on 200 019 lines (testdata/mailparse_harness.c). The "one @ outside quotes" rule goes: the owner's decision says the address decides "whatever the display name or a comment holds", and with OpenDKIM's parser there is no second one to part ways with. Found reading the source: OpenDKIM matches its SigningTable against "user@domain" cut to 256 bytes (MAXADDRESS), so a longer address is refused - cut there, a quoted local part could end in another tenant's domain. The check asks for header values with their leading space, as OpenDKIM does; the SigningTable reader goes.
• The port checked against OpenDKIM itself. On the stand, several hundred generated From: lines - quotes, escapes, comments nested and open, groups, brackets, encoded words, a second domain of the server in every place - are sent; for each, the domain the port names must be the d= the message arrives signed with, or both say "no signature". The shared files of lines (from-lines-passed.txt, from-lines-other.txt) are kept: the second one's test becomes "never signed for {other}" rather than "refused". - Opus Done 2026-10-15: not through Postfix, where the check refuses most of the lines first, but straight to OpenDKIM: TestPortAgainstOpenDKIM (internal/milter, test/upstream/opendkim-probe.sh) starts OpenDKIM from the image with keys for three names and plays it 385 lines - the two shared files and lines composed from display names, comments, groups, routes, literals and quoted local parts - through go-milter's client; for each, the d= OpenDKIM signs with is the domain the port predicts (its address, cut at 256 bytes, against the SigningTable), 385 of 385, 287 signed. The two lines padded to be cut at 256 bytes are signed by OpenDKIM with the other domain's key, as read in the source.
• The heuristics review: every assumption about another component, answered from its source (owner, 2026-10-14: "нам нужно полное ревью кода на предмет всей этой эвристики. Её не должно там, где можно посмотреть код и ответить точно и однозначно" and "в коде таких наших блоков должно быть обоснование со ссылкой на код компонента"; the rule is in development.md § Code that relies on another component's behaviour). First an inventory, over the whole tree (Go, build scripts, configs, tests' fixtures, docs' claims): each place SelfPost relies on what Postfix, OpenDKIM, Cyrus SASL, rspamd, supervisord, logrotate, Docker, SQLite/modernc, go-milter or a browser does, with what is assumed and what justifies it today (read, measured, nothing). It is written down as docs/assumptions-2026-10-14.md. Then each is answered from the source or documentation of the version the image runs, and the code gets the comment that cites it; a measurement stays a measurement and says so; where the source contradicts the code, the code is fixed in this stage. Known before the inventory: qmgr's expiry against the queue lifetime (the margin of queued_over_lifetime), postlogd after a rename (the rotation window), the smtpd that outlives a reload (measured, not understood), the NOQUEUE line's fields (two fixtures carry a sasl_username= Postfix does not write, one presented as seen on the stand). - Opus Done 2026-10-14: the register holds 621 entries and their answers; the citations are in the code (ce363f4, b113e72, 599149b, ad3d834). What the sources contradict is fixed by the lines below, each named by its register entries; the measurements the sources cannot replace are the last of them.
• Forged filter verdicts signed (C-R7, B-R1): rspamd 3.11.1's milter_headers skips its verdict routines - nothing removed, nothing added - for a client in local_addrs or an authenticated one (skip_local, skip_authenticated, both on by default), while the custom signature routine runs for every message; a sender on a private address, or every sender when Docker's proxy shows the bridge gateway, gets his own X-Spamd-Result signed. Both set to false; a dev trial in a throwaway rspamd 3.11.1 (rspamc --ip 10.0.0.5 with a forged header) shows it removed and rspamd's own added. The image tag pinned to 3.11.1 (C-R12); the from-rules take the envelope only, extract_from = "smtp" (C-R4); the dead prefilter = true lines go (C-R5). - Opus Done 2026-10-14: 31c2b50.
• The sender map read as Postfix reads it (B-PS1): Postfix takes the owners of the first key mail_addr_find() finds - the address, then without its extension, then @domain - not any key that lists the login; today an application holding @domain may put in From: an address a list-mode application of the same domain owns, which Postfix refuses it as envelope sender. First key wins; tests with a list and a wildcard application on one domain. The map parser reads continuation lines and blank-separated owners as Postfix does (B-PS4). - Opus Done 2026-10-14: 55aef6e; a key given twice makes the check defer, as it makes Postfix refuse the table.
• Queue lifetimes read as Postfix reads them (A-8, B-PQ1, B-PQ10, B-PQ11): expiry is one qmgr line per message, <id>: from=<..>, status=expired, returned to sender, without to= - the tailer maps it to every open row of that id, closed as bounced, and the dead per-recipient expired case goes; postqueue -p marks a forced-expired message #, which the listing parser keeps; a bare number in a time parameter takes the parameter's own default unit (d for the lifetimes, h for delay_warning_time) and conv_time() reads one number and one unit.
  • Opus Done 2026-10-14: 0ed03b2, with the next line.
• Refusals matched by what Postfix writes (A-21, A-23, A-15, A-24): smtpd's milter-reject line carries sasl_username= - the refusal is matched on the login too, so two applications behind one address sending as one sender are told apart; sasl_username= ends at a comma (sasl_sender= may follow); cleanup's to= is the last recipient only, which the rows built from it say; the IPv6: trim goes. - Opus Done 2026-10-14: 0ed03b2.
• Recipients matched by what Postfix writes (B-G5, D-S20, A-9, A-31): go-milter hands the address without brackets and parameters, so cleanAddress cuts no blank; Postfix keeps a recipient's case, so the match is exact; a rewritten recipient is matched by orig_to=; a recipient given twice is one row (cleanup drops the duplicate). Tests through go-milter's own client. - Opus Done 2026-10-14: c8f6697; the match stays case-insensitive, now with its reason from the source (D-S20).
• The log tailer's rotation and stamps (A-33, A-12): postlogd writes to the renamed file until postfix reload, so the old file is drained for a while after the switch, not once; no Postfix writes the ISO stamp isoStampRe was made for - it goes, and the fixtures take the syslog stamp Postfix 3.7 writes. - Opus Done 2026-10-14: 463d4ff; the renamed file is read for a minute after the switch.
• Inbound mail over 10 MB (C-P31): cleanup runs with main.cf's message_size_limit, never set, so 10240000 bytes, whatever port 25's -o says - a 12 MB message to the inbound relay is accepted through DATA and then refused 552. cleanup gets the largest listener limit; e2e sends 12 MB inbound. And local_transport = error:... (C-P8): an address literal of the container's own address is local today.
  • Opus Done 2026-10-14: e174537; cleanup takes port 25's limit when it is larger than main.cf's; Postfix refuses the address literal at RCPT.
• Supervision as supervisord does it (C-P42, C-V5, C-V6, C-V10, B-U2, C-P34): master leaves the shell's process group, so stopping the program never stops Postfix - the wrapper traps TERM and runs postfix stop; startsecs above the milter wait for Postfix and 1 for logrotate so a failure reaches FATAL (logrotate's not taking the container down); postfix-reload rests in EXITED after a run; a reload's exit status is read, not assumed; the dead set-permissions branch goes. e2e: restart of the program, stop of the container. - Opus Done 2026-10-14: f2ad699; startsecs is MILTER_WAIT_TIMEOUT + 10, set by the entrypoint; the reload's status is read over supervisord's XML-RPC.
• sasldb written in place (C-S3, C-Q3): Cyrus opens sasldb2 without a Berkeley DB environment, so without locks, and saslpasswd2 writes in place; an AUTH or a backup during a change can read it torn. The panel writes a copy and renames it, and the backup reads it under the same lock. - Opus Done 2026-10-15: no lock for the backup after all: it reads each file through one descriptor, and a rename leaves that file whole; the panel's own changes are serialised by a mutex.
• Health says what is served (A-38, A-65, D-P1, D-P3): a socket file is not a listener - unix sockets are dialled as inet ones are; the certificate shown is the one Postfix serves (postconf -h smtpd_tls_cert_file), with a renewal waiting named; the wording of "Postfix defers mail" and of Required TLS (mockup first). - Opus Done 2026-10-15: ea8b6bf (mockup) and the commit after it; a probe connection is silent in all three milters (libmilter read: sendmail 8.17.1.9 fetched into dev/upstream); the expiry notice waits for a renewal that is on time.
• The panel in browsers without Sec-Fetch-Site (D-B3, D-B13): under Referrer-Policy: no-referrer a form POST carries Origin: null, which the origin check refuses - every POST, sign-in included, fails in Safari before 16.4 and over plain HTTP at a name; same-origin instead, with a header test and one POST on the stand. A shown-once secret is blanked on pagehide; Pragma goes. - Opus Done 2026-10-15. Measured the same day in Chrome 152 (the desktop app's browser): a no-store page is not restored from the back/forward cache (pageshow persisted=false, a fresh load); the pagehide handler is for browsers that do restore one. Firefox and Safari not measured.
• Two processes opening an old database (D-S3): user_version is read outside the transaction, so the panel and dmarc-ingest can both run one migration; _txlock=immediate and the version re-read inside it; a test with two concurrent opens behind a held lock. - Opus Done 2026-10-15: the test fails without the fix exactly as the register said ("table settings already exists") and passes with it.
• Certificates from certbot (C-A1, C-Q4, D-guide:258, :291, :722): certbot's live/ files are relative links into archive/, so mounting live/<host> alone, or linking certs to it, leaves Postfix without a certificate and the backup without it. The deploy fragments mount /etc/letsencrypt read-only and point the paths into it; the backup follows a link under certs/ to a regular file. - Opus Done 2026-10-15: certs/ holds links into /etc/letsencrypt/live/, mounted read-only at the same path; certbot's own chmod 0755 and chgrp/chmod 0640 for a server that is not root (kept on renewal) let the panel read them; no copying hook.
• The rest of what the sources contradict (C-M7 <pct>0</pct>, C-E2 the e2e zone serial, C-D1 a release gate comparing the image's package versions with the ones cited, C-K8/K9 and the comments the answers name), and the documents: every claim of architecture.md, security.md and guide.md the answers find false or partly true is rewritten as the answers propose (D-arch, D-sec, D-guide). - Sonnet Done 2026-10-15: the release gate is checkCitedPackageVersions (e2e) against the table in development.md; the documents by three agents, one per document, reviewed. Two register proposals were outdated by this stage's fixes (a bounce's From: line, Origin: null) and not used; the nginx renewal needed --webroot in the loop, which the register missed.
• What only the stand can answer (A-43, C-K5, C-M8, C-X1, D-S7, D-B16 and the document claims marked so): postlog's set-gid in the image; the client address Postfix sees through Docker's port publishing over IPv4 and IPv6; the largest DMARC report received; the public keys of an ECDSA pair as the reload hook compares them; the longest write against busy_timeout at the retention limit; the favicon each browser takes. Each measured, dated and written beside the code it settles.
  • Opus Done 2026-10-15, each written beside what it settles: postlog is not set-gid in Debian's package (A-43); Docker's proxy hands every IPv6 client and the host's loopback over as the gateway, IPv4 keeps the client (C-K5; a new guide paragraph, and a Security note: an allow-list naming the gateway admits every IPv6 client); one DMARC report of one record so far, too little to judge the cap (C-M8); OpenSSL 3.0.22 prints one public key alike for RSA and ECDSA (C-X1); the longest write, the line prune, 1.9 s at 1 900 messages a day after a 30-day gap, under busy_timeout (D-S7, D-arch:516); about 6 lines of 146 bytes a message (D-arch:532); Edge takes the SVG favicon, the desktop app's browser the PNG (D-B16); a published port nothing listens on is refused over IPv4, accepted by the proxy otherwise (D-guide:61); deny beats allow in every mix on rspamd 3.11.1-3 (D-arch:121); rspamd runs as uid 11333 (D-sec:424).
• Docs and wording: help topic sender-check and guide § Who an application may send as (mockup first), security.md § Mail path, CHANGELOG. - Sonnet Done 2026-10-15 with the port (2cf118c the help mockup, 3584adc).

Stage 3 — persistence and restore (M1–M5, L1–L4, L7) → 2.5.3 (shipped 2026-10-15)

• Consume the manifest after the resync succeeds (CheckRestore reports; run deletes after resyncAfterRestore returns nil). Test: Resync fails → manifest kept → second boot resyncs. And a corrupted or truncated archive is refused on restore and on import with its message, nothing half written (a row of review-2026-08-followups, placed here 2026-10-14). — Opus Done 2026-10-15: cmd/panel openData runs CheckRestore, store.Open, the Resync and then backup.ConsumeRestore. A truncated archive had nothing to refuse it - the manifest is the first entry, so half an archive was a restore on half a database: every archive now ends with data/contents.json (size and SHA-256 of each data/ file; the deploy files are the operator's to edit), CheckRestore compares /data with it before the database is opened and renames it contents.checked.json, so a start after a failed one (which had the database open) does not compare again unless the manifest changed. Tests: the start that fails, a cut in the database / before the deploy files / in the last entry (database untouched, no -wal), a damaged key named; the domain import of a cut plain export, a cut and an altered .spde (each refused, nothing created). The CLI's -decrypt -o is step 3.
• Indexes: idx_send_log_domain_created_at (domain_id, created_at) — on the id column stage 2 adds, replacing idx_send_log_domain; partial idx_send_log_queued (created_at) WHERE status = 'queued'; consider (domain, id) for the scoped page. Migration 0009 DDL only; TestLimitRefusalsSince-style EXPLAIN tests against the real query strings. — Opus Done 2026-10-15, migration 0010 (0009 went to stage 2b). Stage 2 already counted by (domain_id, created_at); idx_send_log_domain stays, the global log's filter by name reads its page off it in id order. The plans of every send_log statement showed four more: the sweep now asks for queued and deferred rows - idx_send_log_open (created_at) WHERE status IN ('queued', 'deferred'), the statuses literal and ORDER BY +id (else SQLite walks the table in id order); NOT EXISTS (... refusal_line = ?1) of stage 2b scanned the table per refusal line - partial idx_send_log_refusal_line, and its row is found by idx_send_log_rejected (status literal, +queue_id); the application's level-2 count went by the domain's index - +domain_id; a page of one domain or one application - idx_send_log_domain_id, idx_send_log_app_login (id order, no sort; several domains still sort, top-N). The statements are constants/builders the test explains (TestSendLogQueriesUseTheirIndexes, which replaces the hand-written plan check of TestLimitRefusalsSince).
• CLI backup: open the source mode=ro, refuse a missing file; password floor of 12 shared with the panel (validate.MinSecretFilePasswordLen); -decrypt -o verifies the password (first chunk) before opening the output; OnWarn for skipped symlinks and Check fails when certs/ is named but empty. — Opus Done 2026-10-15. The source check is in backup.snapshotDB, so the panel has it too; mode=ro measured not to create a missing file (TestSnapshotRefusesAMissingDatabase). -o of both directions writes a temporary file beside the name and renames it whole (a backup that fails half way left a partial file too); -decrypt peeks the first chunk - the password, and the envelope's type, authenticated - before it opens the output. Since 2.5.2 links under certs/ are followed, so what is still left out and now warned of: a link under data/, a link under certs/ to a directory, a socket or the like; certs/ with no file refused by Check and Create alike. L4, in this stage's list without a line: a bare address takes /128 when written with a colon, /32 otherwise (::ffff:10.0.0.1/32 was ::/32).
• secretfile caps: bound 128·N·r to a stated budget (e.g. 128 MiB) and p to 1-2; refuse the rest with ErrHeader; reject trailing bytes after the last chunk. Tests. — Opus Done 2026-10-15: 128 MiB of scrypt's V (x/crypto v0.58.0 Key: make([]uint32, 32*N*r)), p ≤ 2, logN ≤ 30 for the arithmetic; the refusal is the existing ErrCorrupt (there is no ErrHeader), which the import shows as "damaged or incomplete". One byte is read after the last chunk; any is ErrCorrupt. TestKDFBudget (the old cap's 4 GiB among the refused), TestBytesAfterTheLastChunkRejected.
• atomicfile: fsync the directory after rename; a test file for the package. — Opus Done 2026-10-15: atomicfile.SyncDir, best effort as sasl.go's (the file is in place either way; Go on Windows cannot sync a directory), used by Write, app.replaceSASLDB (its inline copy replaced) and the CLI's -o. atomicfile_test.go: create, replace, mode, nothing left beside; an open reader keeps the old file; a failed write (missing directory, a directory in the way) leaves the old state; SyncDir works on Linux.
• Docs: security.md § Backup; guide § Backup. — Sonnet Done 2026-10-15: security.md § Backup and domain export (what is warned of, the missing database, the restore's record - integrity, not authenticity -, the KDF budget, trailing bytes, the CLI's floor and -o); guide § Restore (step 1), § Full backup (certs/, -o, the database) and § Encrypting (the first block, a cut file, the floor).

Changed by the reviews of stage 3 (2026-10-15; technical and security, reports in the session's scratchpad): H1 - a restore of a server with DMARC on was refused on every start: the archive records no owner, tar as root leaves the files root's, and the entrypoint handed back to the panel only the maps it named, not the three DMARC maps, which the comparison then could not read; the entrypoint now hands every file beside queue/ in /data/postfix to panel:selfpost 0640 (e2e: an extra map written before the backup, unpacked as root, panel:selfpost 640 after). The comparison reads regular files only (Lstat, size first, LimitReader: a link or a pipe in a crafted archive is refused, not followed or waited on); a record unpacked again wins over the mark of an earlier check, and ConsumeRestore removes both; the mark is synced; a selfpost.db-wal/-shm/-journal beside a freshly unpacked database is refused. secretfile: r ≤ 64 as well (N=2 with a huge r asked 384 MiB of B and XY). The panel names an empty certs/ (backup.ErrNoCerts) instead of "could not read". AttachRefusalLine refuses an empty line, as the old statement did. A skipped entry's kind in words. -decrypt of an empty payload is not a failure. Guide: the shell's > empties the file before a wrong password is found, a killed CLI leaves .NAME.tmp-*, unpack again into an empty directory, the comparison is made once, and step 2 of the in-place restore stepped out of . (mv . fails; older than this stage). Accepted, with the reason: hashing every data/ file delays the first restore start - SHA-256 runs at hundreds of MB/s and the database measured on the test server is under a megabyte (a 30-day journal at 1 900 messages a day is tens of MB), well inside Postfix's wait for the milter sockets; a CLI killed by SIGKILL cannot clean its temporary file, and a handler for the gentler signals buys little for a one-off command whose leftover is the owner's alone.

Stage 4 — panel behaviour (M8, M11, L10–L19) → 2.5.4 (shipped 2026-10-15)

• Login: dummy bcrypt on unknown user; a test that measures both paths within a tolerance. Account password change gets its own limiter bucket. Login destroys the presented session. — Opus Done 2026-10-15: noSuchUserHash (BcryptCost, made once) compared against for an unknown name or a failed lookup; TestLoginTakesAsLongForAnUnknownName holds the medians of five tries each within a factor of two at the real work factor. The bucket is per user (AllowPasswordChange(userID), 10 in 15 min): the guesser is whoever holds that user's session. Sign-in destroys every session cookie the request carries before it creates its own.
• http.Server: ReadTimeout, WriteTimeout, IdleTimeout. — Opus Done 2026-10-15: 30 s, 60 s, 2 min (newHTTPServer); the full backup would have been cut at 60 s, so HandleBackup moves its write deadline to two hours through http.ResponseController (no wrapper of the ResponseWriter in the chain; e2e checks the log has no failure to move it). A test checks the four are set and long enough for a form and a DNS page.
• Polling recovery: schedule from htmx:afterRequest on the box (or pass source), handle sendError/timeout with back-off; RequireAuth answers an HX-Request with HX-Redirect: /login instead of a 303. A minimal JS test (node + jsdom) or at least an e2e that restarts the panel with Overview open. — Sonnet Done 2026-10-15, read in htmx 2.0.4's source (the npm build, byte-equal to the shipped htmx.min.js, in dev/upstream/htmx-2.0.4): both - source: el, and the schedule from htmx:afterRequest, which issueAjaxRequest() fires on load, error, abort and timeout, and after a swap that removed the box fires again on the nearest parent still in the page (so the box is found by detail.target's id); detail.successful tells failure. Back-off doubles per failure in a row to 60 s. toLogin: 401 + HX-Redirect for htmx, which handleAjaxResponse() follows before it looks at the status. No jsdom (no node tooling in the repo): testdata/poll_test.js runs panel.js in node against a stand-in for the DOM and htmx with a moved clock, from TestPanelJS (skipped where node is absent - the test server's Go has none; it fails on the old panel.js with ten findings).
• parsePage clamp; NaN refused; rate-limited tag red (add the status to components.go, flip the test); the recalc button posts only what it needs or warns; HandleReload either covers DMARC maps and filter lists or the flash and the two comments say what it covers. The log's paging is tested at its edges: first page, last, past the last, an empty log (a row of review-2026-08-followups, placed here 2026-10-14). — Sonnet Done 2026-10-15: the Outbound log counts first and clamps to the last page; a number too large for an int is the last page too (strconv's ErrRange with the largest int), on both logs. checkAutoMultiplier is written "not within", so NaN and the infinities are out. The mockup drew rate-limited grey: design: commit first (out-log, and the kit's status row, which had also lost reject), then components.go. The recalc button warns: data-unsaved on both recalc buttons, and panel.js asks on submit (ev.submitter) when a field of the page differs from what it was drawn with (testdata/unsaved_test.js under TestPanelJS). HandleReload covers what the restore's Resync covers, the filter lists and the DMARC maps included, and a failed step is an error flash on Health naming it.
• domain.Service.Delete: say the domain is deleted when a later step fails, and run apps.Resync() anyway. — Opus Done 2026-10-15: past DeleteDomain both rebuilds run and the errors join under domain.ErrDeletedIncomplete; the key is removed only when the tables were rebuilt (they may still name it otherwise). The domain list says "Domain X deleted, but …; run Reload configuration" (500) - which since the step above rewrites everything.
• A status filter and a day filter in the Outbound log (owner, 2026-10-13, from the Not done list of from-header-binding). The mockup out-log has drawn a third list since the redesign and the built log never had it; the Inbound log has its counterpart built (domain, decision, day). One list: Any status, queued, sent, deferred, bounced, rejected - the statuses of a row, and rejected shows all three kinds of refusal together. He was shown the list with the refusals split by reason and chose it without the split. The mockup's own list is wrong (it offers rate-limited, which is a reason, and lacks queued and rejected): correct it first (design: commit). And a filter by day, as the Inbound log has it (owner, the same day): the same control, drawn in the mockup first. Both filters survive paging and the rows' polling. - Sonnet Done 2026-10-15: design: commit (the list corrected, the two dates), then view.OutLogFilter (domain, app, status, from, to) carried by the poll, the pages, the rows' links and the message's way back; SendLogFilter.Status, Since, Until (whole UTC days, as the Inbound log's). A status not in the list is no narrowing. Plans: a domain user's filtered page goes by idx_send_log_domain_id (with a day filter by idx_send_log_domain_id_created_at and a sort - not pinned by the plan test); the global role's status-only count walks the table (a page view or a 5 s poll while the log is open on a status - tens of thousands of rows at the retention limit, milliseconds), accepted.
• Milter sockets: "Listening" is asked of the socket (owner, 2026-10-13, from the Not done list of from-header-binding). health.CheckSocket (socket.go) stats the file and writes "Listening"; a process that died and left its socket reads ok while mail is deferred or goes out unsigned. It connects and closes instead, and a refused connection is an error that says nobody answers. Find out first, on the stand: whether the panel's user may connect to OpenDKIM's socket, and whether a connection closed without a byte ends the sender check (security.md § Accepted risks: the milter library panics on an empty packet) - if it does, that is fixed first, in this step. A test with a socket file nobody listens on. - Opus Done in 2.5.2 (stage 2c, 98e09fe), ticked here 2026-10-15: health.CheckSocket dials and closes; a refused connection says nobody answers, a refused permission is Unknown; a connection closed without a byte ends a milter session quietly on both sides (libmilter's mi_rd_cmd, go-milter v0.4.1 HandleMilterCommands on io.EOF - the comment on CheckSocket cites them). The socket nobody listens on is milter_test.go.
• The cards layout takes its phone order from the box's number (owner, 2026-10-13, from the Not done list of from-header-binding). The order on a phone is written in panel.css for four boxes to a stack (.sp-cards > .column:first-child > .box:nth-child(1..4)); a fifth box gets no rule and lands out of place with nothing failing. The box partial sets an order class from its own number (a class, not a style: the CSP allows no inline styles), the stylesheet orders by that class, and a test checks that the numbers of a rendered cards page run in order. A change to a component: the mockup's components page and theme.css first (design: commit). - Sonnet Done 2026-10-15: an attribute, not a class: the mockup's check (check.py) refuses a class theme.css defines and no page uses, and a class per number up to twelve would be eight such. data-order="N" is not a style either, so the CSP holds; theme.css and panel.css have .sp-cards [data-order="1..12"]. box_open writes it for a box with InCards (Health's eight, the kit's three) and a number. TestCardsBoxesCarryTheirNumberAsTheirOrder reads the rendered pages: each box's data-order is its number, and the stylesheet has its rule.
• A notification is owed to each recipient, not to the subject (owner, 2026-10-14; from the reviews of panel-notifications, not found by the audit). Notifier.send (notify.go) counts a subject as said when one recipient's copy was handed to Postfix; a recipient whose copy failed, or was "held back" by the hourly ceiling, is never told. Unseen while the owner is the only recipient; the first domain administrator makes it "the global administrator knows the domain's DNS broke, its administrator does not". The state of a subject is kept per recipient (the next migration), and a later check sends to those not yet told - so the ceiling holds a copy back instead of dropping it. A user who gains the right to hear of a subject while it is already open is not mailed about it. "Cleared" goes to those who were told of the problem. Tests: one copy fails and goes out at the next check, once; the ceiling; a user added mid-subject. - Opus Done 2026-10-15: migration 0011, notify_addressees (key, address, told_at), gone with the subject's row. open sends the first message to every recipient and records who was handed it; followUp, on every check while the subject is open and still wrong (DNS bad, the certificate at its level, the queue still stuck, a limit within its day), hands the current message to the owed and untold who are still recipients; closed sends the end to the told (all recipients for a subject opened before the table). A cleared copy that fails is not retried - the subject is gone with it, as before. Tests: a failed copy goes out at the next check, once; the ceiling holds one back; a user added mid-subject hears nothing of it, its end included.
• A DNS outage is "unknown", not "wrong" (M14, M15; owner, 2026-10-14 - the two findings had no step). Each resolver gets a time budget of its own, so a silent first resolver no longer leaves the second an expired context (dnscheck/resolver.go); a timeout is Unknown, not "does not resolve" or "no PTR record" (dnscheck/server.go); a failed lookup keeps the server's addresses it had instead of emptying them, so SPF does not go Unknown with it. Notifications: a domain's DNS is reported broken only on a result that is wrong, never on Unknown - checked in notify, and tested. Tests: a first resolver that drops packets, a PTR timeout. architecture.md:339-340 and the DefaultResolvers comment say what is true. — Opus Done 2026-10-15: queryEach gives each resolver perResolverTimeout (2 s) within the check's context and asks the one that answered last first (a dropping first resolver costs one 2 s wait, then is asked last); an address or PTR lookup not answered is Unknown (lookupFailed), an answered mismatch beside it still the error it is; Checker.Server keeps the last answer's addresses when the lookup was not answered. notify.dnsReading already reported faults only - now tested with SPF unknown alone and beside a wrong DKIM. architecture.md no longer has the sentence the audit cited; the guide's SELFPOST_DNS_RESOLVERS row says the budget and the order.
• DMARC reports tested where they come in (owner, 2026-10-14; rows of review-2026-08-followups). The panel parses reports that arrive by mail from anyone, and the internal/dmarc service (service.go, addresses.go) and the DMARC handlers have no tests; no other step touches them. Tests of the service, and of a report that is malformed, oversized, compressed past its limit, or names a domain the instance does not have: refused or skipped with a reason, nothing stored, the panel alive. A finding is fixed in this step. — Opus Done 2026-10-15. Since the audit the package had gained tests (bounds, unknown domain, tag mismatch, not a report, a store failure, the allow-list) and the handlers theirs. Added: malformed XML, a gzip bomb and a zip bomb (9 MiB of zeroes), a message over the limit - each refused with its reason, nothing stored; the service's Resync on and off, the hosted suggestion and its validation. The finding: both limits cut silently (io.LimitReader), so an oversized report failed as damaged; readAtMost refuses it for its size (MaxMessage 10 MiB, the message limit moved from cmd/panel into IngestMessage). The panel is not in the path: ingest is its own process per message, and a refusal exits 0 with the reason in mail.log.
• The last untyped page data (owner, 2026-10-13): healthChecks (handlers_status.go) returns a struct instead of map[string]any, so a mistyped key in healthPage / healthCards is a compile error and not an empty card on Server › Health. No template change. This closes what is left of the roadmap's former template-data-typing. — Sonnet Done 2026-10-15: healthReadings, read by healthPage and healthCards and the tests' fixtures; no map[string]any is left in internal/web outside tests.
• Upstream TLS mode validated at save (L5): saving an inbound domain whose [host]:port is already used by another domain with a different TLS mode is refused with the validator's text naming that domain and its mode; TestRenderInboundMaps gets the two-routes-one-next-hop case and the map is asserted to hold one key. Guide § Inbound domains says a next hop has one TLS mode. — Opus Done 2026-10-15: inbound.Service.checkOneTLSModePerUpstream (the domain itself may change its own mode; another port is another next hop); renderInboundMaps writes a next hop's policy line once (rows saved before keep the first by domain name, as Postfix did). Tests in inbound and postfix (TestRenderInboundMapsOneTLSLinePerNextHop).

Changed by the reviews of stage 4 (2026-10-15; technical and security): two domains on one upstream could never change its TLS mode - each save was refused for the other (medium); inbound.Service.SetTransportAs now changes the mode for every domain on the upstream when the one saving may change them all, and refuses a domain user who may not with the mode, not the other tenant's name (the security review's F1: the first version's refusal named it). The deletion's leftover key is removed by Resync (Reload configuration, a restore's start) for names no domain has. The backup's write deadline moves before Check. hasUnsavedEdit skips password and file fields (a password manager filling the export password made Recalculate ask) and compares a multiple select's options. htmx.config.timeout 30 s, so a refresh that is never answered ends and is tried again; the rejected promise of a failed refresh is handled. A filtered empty Outbound log says so. The dummy hash is made at start. A DMARC part over 8 MiB is refused for its size like the others. The Configuration box on Health says what Reload writes (design: first). Tests: the shared upstream (service and handler), the leftover key, all eight Health boxes ordered, the legacy subject. Accepted, with the reason: a notification is owed to an address (architecture.md § Notifications); SERVFAIL from every resolver is Unknown

• it is not an answer about the record; Reload repeats the restore's list of steps by hand (HandleReload and mailStack.Resync, both commented); a failed SetNotifyAddressees is logged and the subject stays said, the next follow-up finds nobody owed; the profile form's "username taken" is documented in security.md.

Stage 5 — image and Postfix hardening (M17–M19, M21, L30–L32) → 2.6.0 (shipped 2026-10-15)

• Supervisor socket: a group of its own (supervisor), panel only; postfix/opendkim keep selfpost for the file paths. — Opus Done 2026-10-15: groupadd --system supervisor, the panel its one member; [unix_http_server] chown=root:supervisor. e2e: every program RUNNING, the panel still reloads OpenDKIM and Postfix through it.
• OpenDKIM: Syslog yes (supervisord captures it) or Logfile; Canonicalization relaxed/simple; OversignHeaders From. e2e: DKIM still verifies; a message with two From: is refused before signing anyway. — Opus Done 2026-10-15, read in OpenDKIM 2.11.0~beta2's source first: neither way worked - syslog(3) writes to /dev/log, which nobody served in the container, and there is no Logfile option. internal/syslogrelay (panel -syslog-relay=/run/syslog/log, program syslog-relay before OpenDKIM, /dev/log a link to it from the entrypoint) passes the lines to stdout, severity error and worse marked. Syslog yes, Canonicalization relaxed/simple, OversignHeaders From (RFC 6376 § 3.4, § 8.15). e2e: OpenDKIM's start line in the container log; the DKIM-verify steps pass. The two-From: refusal is the sender check's (2.4.0, 2.5.2), unchanged.
• Postfix: smtpd_tls_protocols = >=TLSv1.2, smtpd_tls_mandatory_protocols likewise, disable_vrfy_command = yes, smtpd_banner = $myhostname ESMTP; integer validation of RATE_LIMIT_*, *_SIZE_LIMIT, MILTER_*_TIMEOUT at the top of postfix-config.sh with a clear exit. — Opus Done 2026-10-15: the four settings (proto/postconf.proto: >=version since 3.6); need_number / need_time (a whole number, a time value with s m h d w) for the six counts and sizes and the three milter timeouts. e2e: postconf -h reads them back; TestNumericSettingsGate starts the image with three bad values and one good set.
• logrotate state under /data/log/.logrotate.status (-s), group- readable as the directory is. — Opus Done 2026-10-15: logrotate -s /data/log/.logrotate.status (logrotate.8.in: "-s, --state statefile"); the entrypoint's find /data/log -type f -exec chmod 0640 covers it. e2e: the file is there after the start.
• entrypoint: normalise the three DMARC maps with the other five; remove the dead set-permissions branch. — Opus Done before this stage: the maps since the review of stage 3 (every file beside queue/ in /data/postfix, H1 there), the branch since stage 2c (assumptions C-P34).
• extract-cert.sh: set -o pipefail under bash, check both outputs are non-empty PEM before moving them into place, create 0600 first. — Opus Done 2026-10-15: bash, set -euo pipefail, umask 077; jq's first(...) // empty per field, refused when empty or not PEM; mktemp beside the target, chmod 0640, mv; the files that were there stay on any refusal. Run on the test server with a stand-in jq (the host has none): a good pair, a missing domain, a non-PEM field. Executable in git.
• crashexit: subscribe to PROCESS_STATE_EXITED with a restart-count threshold, or raise startsecs so a crash after 1 s still reaches FATAL; say in architecture.md what actually exits the container. — Opus Done 2026-10-15: PROCESS_STATE_FATAL,PROCESS_STATE_EXITED; five expected:0 exits of one program in 600 s shut down, as FATAL does; syslog-relay joins the spared. build/crashexit_test.go runs the script under python with CRASHEXIT_DRYRUN. architecture.md says what exits the container.
• envdoc_test made real: parse envDefault|envInt|os.Getenv("…") out of cmd/panel/*.go and ${VAR out of build/*.sh, parse the key tables out of guide.md and architecture.md, compare. — Opus Done 2026-10-15: read from cmd/panel and internal (tests aside) and build/*.sh (${VAR:-…}), documented in guide.md § Environment variables and architecture.md's internal list; both ways. It found ANTISPAM_DIR documented nowhere (added).
• Docs: architecture § Image and processes, § Configuration; security.md § Process. — Sonnet Done 2026-10-15 with the steps: the programs table (syslog-relay), what exits the container, ANTISPAM_DIR; security.md § Output and process (the socket's group).

Changed by the reviews of stage 5 (2026-10-15; technical and security): the e2e's numeric-settings test ran a container that passed every gate and never ended (the first run of the stage's e2e hung on it): the good values are now checked beside a bad one after them, and the runs are named, bounded to a minute and removed. The gates refuse what is under a parameter's minimum (a zero window kills anvil, a zero milter timeout smtpd and cleanup - src/anvil/anvil.c, src/smtpd/smtpd.c) and what passes an int, in seconds for a time value (src/global/conv_time.c reads one unit and refuses an overflow). OpenDKIM waits, five seconds at most, for the relay's socket. The relay marks its lines syslog: , writes control characters out (a forged second line), reads at once into a queue and drops past 200 lines a second or a full queue, saying how many - a relay that stopped reading would have held OpenDKIM in syslog(3). smtp_tls_mandatory_protocols too. extract-cert.sh keeps the old files' group (a cron run as root handed the key to root's group) and sets its trap before mktemp. cert-reload joins the spared programs. security.md: the restart after five panel exits, and the relay's socket. Measured, against the technical review's guess: smtpd runs with the selfpost group among its groups (Uid 101, Groups 104 998 on the test server), so before this stage it did reach the supervisor socket. Accepted, with the reason: the logrotate state file in the postfix user's directory - that user can stop rotation through the log it owns anyway; envdoc_test reads ${VAR:-} and the Go helpers, not bare $VAR or python (none today); the 600 s pruning in crashexit is not tested (the count and the spared are).

Stage 6 — documentation and roadmap sweep (no product code, no cut) (done 2026-10-15)

• Every item under Doc drift above, file by file. — Sonnet Done 2026-10-15: architecture.md (the sender check in § OpenDKIM, the limit: row's 48 hours, the inbound log and filter-list routes and who may open them, migrations to 0011 and every table, KitPage as production data, Reload configuration as a restore writes, the panel's roles counted again); schema-migrations.md (the tables and columns since v2); guide.md (a delivery's lines go at the next prune; INBOUND_JOURNAL_MILTER_SOCKET out of the operator's table, into architecture.md's internal list); development.md (the env guard as envdoc_test reads now, the deferral the e2e proves, the e2e's fixed sleeps said where they are); roadmap.md's closing paragraph. Already right by then: the .alive mark, dmarc-ingest as a second writer, the chain head, M1's wording, the version stamp.
• roadmap.md: done ahead on 2026-10-13/14 with the owner - template-data-typing struck, structured-logging dropped, review-2026-08-followups closed with each row placed in a stage or dropped with its reason. What is left: check that every Test gaps entry above is named by the step that touches its code, and add it to that step where it is not. — Sonnet Done 2026-10-15: see Checked in stage 6 under Test gaps: five were open and named by no step; each got its test in this stage (L9 as the owner decided it).
• Comments named above (milter.go, sendlog.go, store.go, schema_test.go, machine.go, inboundspam.go, logtail.go, supervisord.conf, entrypoint.sh, docker-compose.yml). — Haiku Done 2026-10-15: each said what its code does now (the journal-milter tempfails at MAIL FROM; the rejected status covers the sender check; the two roles; the log is read whole for a filter list's hits; logrotate uses create; the guide's headings as they are). TestBaselineIsTheOnlyMigration renamed TestChainStartsAtTheBaseline. supervisord.conf and entrypoint.sh were already right.
• Dead code removal from § Nits, with staticcheck added to test.yml so it stays gone. — Haiku Done 2026-10-15: gone: view.Facts, dmarc.Service.HostedSuggestion and ValidateHostedAddress, app.Service.UpdateMode (each with the test that alone kept it), the method checks of HandleDomainDMARC and the default: branches of the user and settings handlers; Settings.WithResult calls resultFlash. Kept: domainIDSet (used by the send log's authorization), KitPage()'s User (the fixtures render KitPage() without the handler); the set-permissions branch was gone already. staticcheck 2026.1 (the release for Go 1.26) runs in test.yml on both modules, clean; for the e2e module the package now builds without its tests (a test and two constants moved out of non-test files), and three findings there were fixed (ST1005, SA4006). The test moved, TestImageBuildPreservesLogrotateMode, had sat in a non-test file since it was written, and go test had never run it; run, it failed on the entrypoint (docker run without a hostname) and now asks past it.

Decided with the owner on 2026-10-12

Four items I had first put below the line; three of them are in the plan above after discussion, which is why they are recorded here with the reasoning rather than struck.

• L5 (tls_policy per next-hop): the full fix is a model change (TLS mode per upstream, not per domain) — a design item. Decided (owner, 2026-10-12): validate instead — stage 4, item 6.
• L12 (rows by domain name): decided (owner, 2026-10-12): the send log gets the domain's id — stage 2, item 5.
• L8 (the signing-name oracle): decided (owner, 2026-10-12): do it — stage 2, item 7. (The names are not public: rDNS gives the hostname only, so the reply told one tenant which other domains are hosted here.)
• L9 (go-milter header accumulation): decided (owner, 2026-10-12): locally only — a comment on senderSession.Headers saying why h is ignored, and TestSenderCheckResetsBetweenMessages passing an h that still holds the first message's From: (stage 2, with item 7). The library itself stays as vendored; an upstream patch is a separate, unhurried matter.
• The Windows-bound tests: not worth platform guards; CI is Linux.