Release

v0.9.7a

Notes for a release that is not in the published index.

Canectic BBS v0.9.7a

Mail this system sent was arriving mangled on other boards, and files received from an uplink were never passed on to downlinks. This release fixes both, along with the ANSI art they made unreadable.

Released 2026-09-20 (Handshake).

Highlights

Messages you send now arrive readable. Canectic was writing UTF-8 into FidoNet packets while telling receivers nothing, and FidoNet assumes CP437 unless a message says otherwise. Every board following that convention saw mojibake: accented names wrong, subject lines corrupted, ANSI art reduced to a wall of Γûê. Outbound mail is now encoded correctly and declares what it carries. The same fault existed in the QWK and NNTP tossers and is fixed there too.

Files received from an uplink now reach your downlinks. A file arriving over a file echo was filed correctly, queued for every subscribed peer, and then never sent — nothing ran the outbound pass for file echoes. A downlink only received anything when somebody on your board happened to upload a file. If you carry file echoes, read the upgrade note about the backlog before updating.

ANSI art is readable again on the web. Art drawn from block characters was being flattened into a single line before anything asked whether it was art. Note this does not repair messages already stored that way.

Also: new files are now browsable on the web rather than only counted, message groups carry defaults for thread mode and character set, and an embedded-mode board starts running the maintenance it had never been running.

Upgrading

  • WARNING: Embedded mode only (PROCESS_MODE=embedded): this release starts running maintenance that has never run on your board. An embedded worker recognised 7 of 33 maintenance jobs and silently completed the rest without doing them, so log purges, message and file retention, nodelist and echolist scanning, timed events and the update check have all been idle. They begin working on the first daily run after you update, and they clear the whole backlog at once. The part to check before updating is retention: if you ever set max_age_days or max_messages on a message or file area, it has been ignored until now and will be enforced on that first run. The Retention Preview on an area's admin page reports what a limit would remove without removing anything. Boards running the normal four-process setup are unaffected, because their worker has always run these.
  • WARNING: If you carry file echoes, expect a burst of outbound files the first time the tosser runs after updating. A file arriving from an uplink was correctly queued for your downlinks and then never sent, because nothing ran the outbound pass for file echoes. That queue has been growing for as long as you have been receiving files. The first outbound pass after this update sends all of it, so a downlink that has had nothing from you for months may receive a great deal at once. Nothing was lost and nothing needs doing; it is the volume that may surprise you and your links. If you would rather meter it, update at a quiet hour and watch the first pass on the peer page, which now shows pending, sent and failed counts per connection.
  • INFO: This update rewrites stored message bodies. Three parts of the system had been storing text three different ways in the same database columns, which is what allowed messages relayed between networks to be corrupted, and the update normalises them. It runs automatically during the update and asks nothing of you. On the development board it rewrote 4,707 messages across 9 areas, all of them received over QWK; your own figures will differ. The older column is deliberately left in place and unread for this release, so it holds the previous copy of every rewritten message and the change can be reversed if it ever needs to be. Messages that were already displaying incorrectly before this update are not repaired by it.

Changes

Upgrade Notes

Everything in this block is something that happens to your board on the first run after updating. Read it before updating a live system.

  • Embedded mode only (PROCESS_MODE=embedded): maintenance that has never run starts running. The embedded worker recognised 7 of 33 maintenance jobs and silently completed the rest without doing them, so log purges, message and file retention, nodelist and echolist scanning, timed events and the update check have all been idle. They begin on the first daily run after you update, and clear the whole backlog at once. The part to check before updating is retention: if you ever set max_age_days or max_messages on a message or file area, it has been ignored until now and will be enforced on that first run. The Retention Preview on an area's admin page reports what a limit would remove without removing anything. Boards on the normal four-process setup are unaffected — their worker has always run these.
  • This update rewrites stored message bodies. Three parts of the system had been storing text three different ways in the same database columns, which is what allowed relayed messages to be corrupted. The update normalises them, automatically, asking nothing of you. On the development board that was 4,707 messages across 9 areas, all received over QWK; your figures will differ. The previous copy of every rewritten message is retained in the old column, unread, so the change is reversible. Messages already displaying incorrectly before this update are not repaired by it.
  • If you carry file echoes, expect a burst of outbound files on the first toss after updating. A file arriving from an uplink was correctly queued for your downlinks and then never sent, because nothing ran the outbound pass for file echoes. That queue has been growing for as long as you have been receiving files, so a downlink that has had nothing from you for months may receive a great deal at once. Nothing was lost and nothing needs doing; it is the volume that may surprise you and your links. To meter it, update at a quiet hour and watch the first pass on the peer page, which now shows pending, sent and failed counts per connection.

Messages

  • Message bodies are now stored one way regardless of where they came from. Three importers had been storing text three different ways in the same two database columns, which is what allowed messages relayed between networks to be corrupted. An automatic migration rewrites the affected rows when you update: on the development board it touched 4,707 messages across 9 areas, all of them QWK-sourced. Nothing is asked of you and nothing needs checking afterwards. The older column is deliberately left in place for this release, unread, so the rewrite can be reversed if it ever needs to be.

  • Fix: messages received over NNTP displayed as mojibake on a UTF-8 terminal. The reader converted them a second time, from a character set they were never in.

  • Fix: messages relayed between networks were corrupted on the way out. The QWK packer decided how to read a stored message body from which database column was populated, and that test is only correct for messages that arrived over QWK. A message that came in over FTN and went out in a QWK packet had every extended character replaced with rubbish. NNTP had the same fault in a different form: outgoing articles declared no character set at all and were written in a way that discarded anything outside plain Latin-1, and incoming UTF-8 articles were stored mangled because they were never properly decoded. Outgoing NNTP articles now declare UTF-8, encode their headers correctly, and carry their text intact.

  • Fix: every message this system sent carried the wrong character encoding. FTN assumes CP437 unless a message says otherwise, and Canectic was writing UTF-8 bytes while saying nothing. Receivers followed the convention, decoded UTF-8 through the CP437 table, and got mojibake: a block of ANSI art becomes a wall of Γûê, an accented name comes out wrong, a subject line looks corrupted. It affected names, subjects and message bodies alike; art simply makes it obvious, being entirely extended characters. Outbound mail is now encoded properly and declares what it carries with a CHRS kludge.

  • Message areas have an Outbound Character Set setting, and message groups carry a default for it. Automatic (the default) sends CP437 whenever the message fits it, which is nearly all traffic, and switches to UTF-8 only for content CP437 cannot represent — where the alternative was replacing those characters with ?. An area can also be pinned to Always CP437 for an echo whose members run software that ignores charset declarations, or Always UTF-8 for a modern network that has agreed on it. Set it once on a group and every area created in it inherits it, with Push Group Defaults available for areas that already exist (unticked by default, since it replaces a per-area choice); an explicit choice on an area always wins. The setting is per area rather than per link because the encoding of an echo is something everyone carrying it agrees on, not a property of any one link.

  • Inbound messages are now decoded by what they declare rather than always as CP437, so mail from a system that correctly announces UTF-8 no longer arrives mangled. An undeclared message is still treated as CP437, which is the network default.

  • Fix: kludge lines were destroyed on every message this system relayed. Stored control lines were being read back in the wrong form, so each one was written out as undefined: with its name and contents gone. A single captured packet carried 772 of them across 157 messages. The visible effects were charset declarations vanishing (which is what made relayed art unreadable) and reply linkage being lost, so downstream systems could not thread the messages. Already-damaged lines are now discarded rather than passed on.

  • Fix: ANSI art drawn with block characters was flattened into a single line. Paragraph reflow joins consecutive lines that are close to the full width, which is how text from a wider editor gets reassembled — but every row of 78-to-80 column art looks exactly like that. Two faults combined: art detection only recognised art that moves the cursor, so art drawn as colour plus block characters read as ordinary text, and on the web the reflow ran before anything asked whether the body was art at all. Three rows of art became one line, and since the web copy of a message is rendered once when it arrives, nothing later could put the breaks back. Art detection now also recognises block-drawn art, and the web asks before it reflows rather than after, which is what the terminal already did.

    Fixing this does not repair messages already stored. Run node src/tools/maintenance.js --regen-html --area <SHORTCODE> on an affected area, and note that it can only recover art whose plain-text copy survived import intact.

    One deliberate trade: a message that is ordinary prose but carries two or more full-width rules now reads as art and will not be reflowed. Nothing about its shape distinguishes it from a drawing, and reflow was previously merging such a rule onto the end of the sentence above it, so leaving it alone is not obviously the worse outcome.

  • Thread mode can now be set on a message group. It was per area only, so a sysop wanting a whole network to read flat had to set every area by hand and set each new one again afterwards. The group carries a default that new areas inherit, an explicit choice on the area form still wins, and Push Group Defaults gained a Thread Mode tickbox for applying it to areas that already exist. It is unticked by default, like Retention, because pushing it replaces whatever each area was set to. Existing groups default to Standard, so nothing changes until you set one.

File Echoes

  • Fix: files received from an uplink were never passed on to your downlinks. A TIC file arriving from a peer was filed into the local area correctly and its outbound rows were written correctly, and then nothing ever staged them. The periodic outbound sweep, which is the safety net that catches whatever was not sent on demand, covered echomail, QWK, NNTP and email but had never included file echoes; and the inbound toss did not signal an outbound one the way the design had always described. The only things that actually staged a file echo were a local upload, the admin Trigger Event button and a manual toss, so a downlink received files only when somebody on your board happened to upload one. Nothing failed and nothing was logged. Inbound tossing now signals outbound immediately, and file echoes are part of the outbound sweep, so a board with no local uploads at all still feeds its downlinks.

  • Fix: a downlink's %RESCAN for a file area replied cheerfully and sent nothing. A file rescan writes pending rows rather than rewinding a pointer, so unlike a message rescan there is nothing that picks them up on the next pass. The admin Rescan button carried the same promise. Both now signal an outbound toss once the rows are written.

  • Fix: file echoes did not work in either direction in embedded mode (PROCESS_MODE=embedded). The embedded worker had no case for the TIC jobs, so BullMQ handed each one over, the switch fell through to its default, and the job was marked complete without the work being done. Outbound email and the outbound sweep were missing for the same reason. A test now holds the two workers to the same set of job names, since a job nothing handles is indistinguishable from one that succeeded.

  • Fix: nothing stopped a file echo circulating. A file that reads as new is imported and then hatched onward to every subscribed downlink, so in a mesh a file that is not recognised does not merely get stored twice, it goes round again and each system that fails to recognise it passes it on. Two brakes now exist where there was effectively one.

    The standard FTN guard was missing entirely: a TIC carries Seenby and Path, this system stamps itself into both on the way out, and nothing checked for its own address on the way back in. That check is now the first thing done, before the file is imported or forwarded, and it matches on the bare address so a chain that writes the address with a 5D domain on one hop and without it on the next is still recognised.

    The dupe check, which was the only brake, had a hole that defeated it exactly when it mattered. It compared filenames case-sensitively and required an exact match even when the checksum matched, so a file returning from a chain that had uppercased or renamed it read as a completely new file with an identical CRC. Filenames are now compared case-insensitively, and a file whose checksum and size both match one already held is recognised whatever it has been renamed to. Size is part of that comparison deliberately: a CRC-32 on its own is not enough to reject a file over, since across a large area the odds of some pair of unrelated files colliding are around one in a hundred.

    The two guards answer different questions and both are needed. The dupe check asks whether this area holds the file now, which retention can make false; the Seenby stamp asks whether the file has been through this system before, and travels with the file rather than living in a table that gets pruned.

  • Fix: a peer subscribed to a file area that is not carried on its network collected a dead delivery row for every file, for ever. The outbound pass needs an areatag to build a TIC descriptor and looks one up by the peer's network; with no mapping it marked the transfer failed, and a failed transfer was never retried by anything. Nothing filtered the subscriber list against that, so the rows piled up invisibly and one downlink simply never received anything from that area. Such a subscriber is now left out at the point the rows are created, on all three paths that create them (inbound tossing, web upload and terminal upload), and the mismatch is logged naming the peer, the area and the network so it can be fixed rather than silently skipped.

  • A peer's file echo delivery is now visible, and failures can be retried. The peer page shows pending, sent, failed and skipped counts per connection, lists the most recent failures with the file, area, reason and attempt count, and offers a Retry Failed button that re-queues them and runs an outbound toss. Previously a failed transfer was terminal and invisible: there was no page that showed one and nothing in the system that would try it again.

  • Fix: the command line tosser never moved file echoes. node src/tools/tosser.js --ftn enqueued the echomail and netmail jobs only, though the documentation listed it as one of the triggers for TIC processing. A sysop tossing by hand moved messages, moved no files, and was told nothing had gone wrong. --ftn now covers file echoes in both directions, as it always claimed to.

Security

  • adm-zip updated to 0.6.1, clearing two advisories against 0.6.0: extraction followed symlinks in the destination, allowing a crafted archive to overwrite files outside the target directory, and a declared uncompressed size could be used to exhaust memory. This matters here because both archive paths read files that arrive from outside the board: FILE_ID.DIZ extraction from uploaded and tossed files, and infopack scanning for echolists and filebones. The fix is a patch release inside the range already declared, so npm install picks it up with no configuration change.

Embedded Mode

  • Fix: the clean shutdown message never printed, and shutdown ended on an unhandled error instead. Both the gateway and the door server tested a bare logtoconsole when the last server stopped, but that name is a private constant inside the logger and was never in scope, so the check threw ReferenceError from inside a timer callback. The logger now exposes the setting properly.
  • Fix: an embedded board ran almost no maintenance at all. PROCESS_MODE=embedded starts the worker inside the gateway, and that in-process worker recognised 7 of the 33 maintenance jobs. Schedules are registered in both modes, so the other 26 fired on time, were handed over, matched nothing, and were marked complete. Nothing was logged at a normal log level and the queue diagnostics reported every scheduler healthy. In practice that meant no log or retention purges of any kind (bad echomail, inbound and outbound file logs, mailer sessions, notifications, relay log, tossed packets, update checks), no message or file retention limits, no periodic inbound scan, no timed events, no update check, no nodelist or echolist scanning, and admin buttons such as Rescan Area, Rethread and Batch Subscribe reporting success while doing nothing. Boards running the default four-process setup were unaffected. The maintenance jobs now live in one module both workers call, so the two cannot drift apart again, and a test holds them to the list of jobs anything can actually enqueue.

Files on the Web

  • Fix: viewing a file area did not clear its new-file badge. The area page advances your scan pointer to the newest file it listed, but it picked that file by comparing ids straight out of the database, and PostgreSQL returns them as text. After the first comparison the rest were being sorted as words rather than numbers, where "9" comes after "37", so the pointer was set to the lowest-numbered file on the page instead of the highest. Pointers only ever move forward, so the wrong value was discarded and the pointer simply stopped, leaving a count that nothing could clear. The same mistake made the terminal file browser stop marking files as new once you had opened one.
  • New files are now visible on the web, not just counted. /files has a New Files button showing the total across every area you are subscribed to, and /files/new lists those areas with a preview of what has arrived in each. Each file area has an All / New toggle, and new files are marked in the listing. The new-only listing shows the oldest unseen file first and walks forward on each visit, so a backlog larger than one page is worked through rather than skipped.

Release Engineering

  • A cut refuses when the changelog carries more than one ## Pending heading. Only the first is read and its body stops at the next ## , so entries written under a second one are dropped from the release they were written for without a word. This changelog grew a duplicate from a hand edit and nothing noticed; it was found by reading the file.
  • Fix: the GitLab release step could create an unsigned tag at the wrong commit, and nothing about the result looked wrong. glab release create does not require the tag to exist: given a name it does not recognise, it creates one, as a lightweight tag at the remote's current head. Run at the end of a cut, before anything has been pushed, that head is the commit before the release commit. So the signed annotated tag stayed on the maintainer's machine, the remote got an unsigned lightweight one pointing somewhere else, and the release page hung off that. The page was there and the assets were right; the signature that is the entire point of signing was not anywhere a reader could check it. The step now reads git ls-remote first and refuses on a tag that is absent or lightweight, naming the command to fix it.
  • A cut still does not push, deliberately, so the GitLab step normally refuses at the end of one. --publish <version> does that half on its own once the tag is pushed, and the cut says so rather than leaving it to be remembered.
  • Publishing a second time over the same release now says what happened: GitLab refuses an asset link whose name is already on the release, so the files are not replaced, though the name and notes are updated. It names the command to remove the release if replacing them is what was meant.
  • Releases are published with node scripts/publish.mjs, which sends the release files, this changelog and each release's notes to canectic.tranqsoft.com, then downloads every file back to confirm it is served byte for byte. The changelog appears at /docs/changelog and each release's notes at /releases/<version>/notes, and the changelog is refreshed between releases as well. Nothing is ever deleted from the release host, because a development build cannot be rebuilt without the stable release it is based on.