Skip to content

NDCHAT - the plan from here

Next: mount dist\NDCHAT-INSTALL-2026-09-02.img on a lab machine once (a line in the RetroCore console window: mount floppy0 0 <path> or attach fd0 <path>), do the SINTRAN side from the admin manual section 3.1, and remove its UNVERIFIED mark. Everything else from the 2026-09-02 manuals round is built, deployed on all three machines and committed (744b2b4c); see HANDOFF-2026-09-02-MANUALS.md. No sync daemon is running - start one with tools\start-relay.ps1 and expect to need the wake trick.

The manuals live in DOC/manuals/ - user, admin, build-and-deploy - and the rule for them is the spec's: change them in the same commit as the behaviour.

Found on 2026-09-02 while photographing the screens - small, all cosmetic, all real

  1. The client does not repaint its own name after /nick. The top-right corner keeps the login name (TESTER@FJELL after /nick OLAV), and the "sent" half of a private window is labelled with it too (*(SYSTEM) for a user who is KARI). The other end sees the right name; it is the client's own copy used for display that is stale. One variable, one repaint.
  2. The header shows LOBBY@trunk for a plain LOBBY once a topic is set - the topic's name field arrives qualified and the header prints it as the room. Where the @trunk comes from has not been read out of the source yet.
  3. A topic set on one machine did not reach the other. OLAV's /topic on FJELL showed on FJELL only; KARI's header on VIDDA never changed and no topic: notice arrived there. Whether topics are meant to cross a trunk is not written down anywhere - decide, then either carry it or say so in the manual.
  4. /leave on a screen terminal shows nothing. The left the room line goes to the line renderer; the header keeps the room name and the count. The others do see the leave.
  5. /list counts only local seats. LOBBY(1) with two people in LOBBY across two machines. Documented as-is in the user manual; a cross-machine count would need the peer tables.
  6. STATUS printed empty o45 - putNumber stopped at three digits and a count of 6345 wrote its hundreds as a letter. FIXED in the source (five digits), not yet built or deployed.
  7. The build stamps were never bumped - S31-0854 / B31-0731 across four builds. Bumped by hand 2026-09-02 (S02-1340 / B02-1340) and written into the build manual as a rule. A check that cannot be forgotten would be better: the linter could refuse a stamp older than the source's last change, or the stage step could write it.
  8. planc-lint.py treats every file on its command line as one link set, so CHAT + CHATSV together draw a false inBuf collision. Either lint per link set (the manual says so now) or teach it a --link-set boundary.
  9. tools\nd-deploy.ps1 stages into sync-out while start-relay.ps1 watches sync-relay; its no-daemon fallback stops every running RetroCore. Do not use it until both are fixed.

Small items left over from before

  1. Side bug seen once: the server stored "ello123" for a typed "hello123" (first character lost, one occurrence, not reproduced).
  2. D100's SYSTEM file table sits near its 256-entry cap; expect ATTEMPT TO CREATE TOO MANY FILES on any build that creates new names. Delete scratch listings when it happens.
  3. The server should CLOSE its history files on the way down, since @ABORT does not, so a repeated edit-load cycle stops stranding CHATH locks (held 0 in STATUS). Not required for correct operation; it keeps the dev loop from needing a boot.

Outstanding, 1 to n

CHATXMS - the second library (Ronny: one sitting, do not begin unattended)

Two libraries, not one - decided. CHATLIB stays pure so CHATTST keeps its 17-second, XMSG-free, machine-free test loop; CHATXMS holds the kernel calls only. CHAT and CHATSV link both. The blocker (xrAddrOf / ADDR of a BYTES parameter) is settled - measured ADDR param 1001 / ADDR here 1001. Size: CHAT.PLNC 14 call sites, CHATSV.PLNC 21, two addressing modes (by name, by magic). Interface already designed - CHATXMS-INTERFACE-DESIGN.md - seven routines, nothing compiled yet.

  1. Create CHATXMS.PLNC. EXPORT block FIRST, before any declaration - R119; skipping this once cost ten errors that still linked and still passed 139/139.
  2. xsSendM - shape A, send by magic.
  3. xsSendN - shape B, send by name through XROUT.
  4. xsRecv - shape C. Do this one FIRST - it carries the clamp and the release, and a mistake here corrupts memory rather than failing loudly.
  5. Rewire CHAT.PLNC - 14 sites.
  6. Rewire CHATSV.PLNC - 21 sites.
  7. Client to XMPSEND, dropping XMPFSND - approved. One site in CHAT.PLNC; the server already uses XMPSEND.
  8. Check every new export name is unique at 7 characters - the linker resolves a collision to whichever it met first rather than calling it a duplicate. planc-lint.py refuses them.
  9. Gate the LISTING of every module touched, not only the one edited.
  10. Deploy to all three machines and prove it by the START ADDRESS in LIST-RT-DESCRIPTION, not by the fact it ran.

Why one sitting. 35 sites across two sources of about 300KB each, roughly twenty-minute compiles, three machines running the product. A half-refactored CHATSV is worse than an unstarted one.

Latency - the 500ms target is still missed

  1. Send fewer letters per message. L.1f (client idle-sleep tuning) is built, deployed and measured: mean came down from 683ms to 592ms, floor from 514ms to 483ms. It does not reach 500ms and cannot - the remaining ~400ms is the XMSG round trip itself (measured: a bare local client command floors at 112ms against a say floor of 514ms). No amount of sleep tuning reaches that; the lever is a wire/design change that sends less. Not designed yet.
  2. XMSG send window PARKS under a big transfer. Pulling the 467KB CHATSV:LIST failed once: *** PARKED *** send window is full; 14 datagram(s) now waiting, peer re-sending because it had not seen our acknowledgement. Our receive/ack latency under load, not an ordering race and not a timeout - same family as the blocked-queue bug fixed 2026-08-27, different trigger. The identical retry worked; no permanent fix yet. -TransferTimeout also defaults to 240s, which is short for a listing this size and made the first failure look like something it wasn't.
  3. Verify the doorbell's restart flag actually clears. The code is deployed on all three machines right now (sleepWhy, waitFlags, the 20-consecutive-empty-wake disarm). Not yet observed live: if the flag is sticky, every later sleep returns instantly and the server spins. Read the server's own log for SV doorbell off during an idle stretch - present means it disarmed correctly, absent after enough idle time means sticky and this needs a real fix.

Backlog - XMSG plumbing, Ronny's order: LOW priority, chat is the product

  1. Ethernet send sequence is not persisted. HDLC keeps xmsg-link-seed.state; Ethernet keeps nothing, so a daemon restart begins at 0 while the peer is mid-run and every frame is silently discarded. The NdLinkLayer fix (2026-08-27) stops the consequence - one stuck frame blocking the queue forever - and does not close this.
  2. Re-install the D103 boot file into its disk image. The repo copy boot/XMSG-STARTEX-L03.D103.txt was corrected 2026-08-28; boot/install-boot-files.py has not been run since, so the copy on the machine still carries the old, wrong comment. Comment-only, changes no behaviour.
  3. Optional: persist name/trunks server-side instead of via rt-load.ps1. Reverses a deliberate design decision and needs a file format, a compile and a deploy. rt-load.ps1 (R.1) already does this from the Windows side and is proven on all three machines - only worth doing if that ever proves not enough. No evidence of that yet.

WHERE THE MACHINES ARE, 2026-08-31

All three run the SAME object: segment 201B, CHATSV:BRF, names FJELL / VIDDA / SKOGEN, both trunks up on each. Built on D100, carried machine-to-machine; D102 and D103 also hold the current CHAT-MON:PROG. Build files: CHATSV:MODE, CHATMN:MODE, CHATCC:MODE - CHATUI:MODE builds the screen TEST program, not the client.

Deliberately NOT doing

  • No comment-reformatting sweep of the C#. Known, mid-sentence tags and collapsed doc comments included - retrocore-csharp section 32 forbids standalone reformatting passes through working code. Fixed when the file is next edited for a real reason, not before.
  • No general window manager. The chat windows do not overlap - they are alternative contents for one rectangle. TESTUI already showed what overlapping costs: clipping, a stack order, and three places that must agree on it.
  • 7.2, the separator style - won't-fix. fullbar (hyphens) and frame (graphic line-drawing) are both library calls; fullbar only ever runs after frame has already proven the terminal draws graphics correctly on that same screen, so its portability is never actually exercised here. Matching them exactly needs a hand-written escape - the one hardcoded, non-VTM-translated line in an otherwise fully portable client - for a cosmetic gain. Hyphens stay.

The rules this plan is written under

  • A green test suite can sit on a failed compile. Ten *** ERROR lines once still linked, ran and passed 139/139. Gate every module's listing.
  • Every build failure becomes a linter check, in the same turn, proved to fail on the bad case and pass on every real source.
  • A duplicate PLANC routine name cascades hundreds of false diagnostics through the rest of the compile, not just one clean error at the header. planc-lint.py now catches it.
  • An RT-load orphans every joined client and wipes the machine name and the trunks. rt-load.ps1 -AndStart now prints the reminder itself. Quit and restart the clients.
  • Read the machine, not the screen - LIST-NAMES for free seats, LIST-RT-DESCRIPTION for the start address, FILE-STATISTICS for what actually arrived.