Updates

2026-11-02 09:58:48 -05:00
1 parent 6b9b1697f2
commit 631987c62d
@@ -7,7 +7,7 @@
 - [Reasoning effort per model](litellm-extra-body-reasoning.md) — GLM-5.3 defaults to reasoning_effort=max and cannot disable thinking; gpt-oss needs chat_template_kwargs via extra_body
 - [Reasoning vs completion tokens](reasoning-tokens-vs-completion-tokens.md) — reasoning comes out of max_tokens, and providers disagree about counting it
 - [Open WebUI image generation](owui-image-generation.md) — works via LiteLLM; the default 512x512 size is rejected by every image model
- [Local code autocomplete](local-code-autocomplete.md) — FIM via /v1/completions; check rbw first; never tune for decode; ncmoe 22 + 8 s timeout since Viseron took VRAM
+- [Local code autocomplete](local-code-autocomplete.md) — retired 2026-10-29; FIM via /v1/completions, never tune for decode, if it comes back
 - [NPU setting clamps memory](npu-setting-clamps-memory.md) — MSI NPU/AI BIOS options cost 17% DRAM bandwidth and the NPU is unusable by llama.cpp anyway.
 - [mainframe bad core](mainframe-bad-core.md) — APIC 20 (cpu10, an E-core) throws fatal MCEs; offlined by a unit, RMA candidate.
 - [crowdsec bans own devices](crowdsec-bans-own-devices.md) — "no external access" is usually crowdsec banning his own phone; http_path includes the query string
@@ -20,3 +20,5 @@
 - [Podman NFS binds go blind](podman-nfs-bind-rprivate-blind.md) — a mount landing after container start leaves containers permanently empty; podman ps looks fine
 - [dotbot force deletes live dirs](dotbot-force-deletes-live-dirs.md) — install replaces a real dir at a link target; the Claude memory dir is now one of them
 - [Discuss before adding content](discuss-before-adding-content.md) — in his notes, record what he said; suggest extras in chat, add only on agreement
+- [Container storage on /home](container-storage-on-home.md) — bind mounts since 2026-10-29; overlay origin xattrs and conmon KillMode traps
+- [Caddy 2.11.6 cuts HTTP/1.1 streams at 60 s](caddy-2116-http1-stream-cut.md) — OpenCode ECONNRESET; upgraded to 2.11.7; test --http1.1 vs --http2 first
@@ -0,0 +1,23 @@
+---
+name: caddy-2116-http1-stream-cut
+description: Caddy 2.11.6 cut HTTP/1.1 streamed responses to POSTs at exactly 60 s (OpenCode ECONNRESET); fixed by upgrading to 2.11.7 on 2026-10-29
+metadata:
+  type: reference
+---
+
+Caddy 2.11.6 added a default 1-minute idle read timeout that cancels an
+HTTP/1.1 response still streaming 60 s after the request body was read
+(caddyserver/caddy#8103). Through llm.rcjohnstone.com this ended every LLM
+stream longer than a minute: OpenCode (HTTP/1.1) reported `ECONNRESET` and
+retried, Caddy logged `aborting with incomplete response ... reading: context
+canceled`, and Strata logged `disconnect, cancel=True` at 60 s. HTTP/2 clients
+were unaffected. Fixed by recreating connor_caddy_1 on 2.11.7.
+
+Ruled out first, all measured: Caddy `write` and `read_header` timeouts (no
+change), LiteLLM (177 s stream fine over HTTP/1.1 direct), Strata, the network
+(the cut reproduced from mainframe itself with `curl --http1.1`).
+
+**How to apply:** a stream that dies at exactly 60 s through Caddy is this,
+not the model. Test with `curl --http1.1` and `--http2` before anything else.
+Caddy floats on `caddy:latest` and podman-stack-update refreshes it weekly, so
+a regression can arrive the same way.
@@ -0,0 +1,39 @@
+---
+name: container-storage-on-home
+description: Rootful podman and docker storage live on the SATA LVM via fstab bind mounts since 2026-10-29; how the move was done and the two traps it hit
+metadata:
+  type: project
+---
+
+Since 2026-10-29, `/var/lib/containers` and `/var/lib/docker` are bind mounts
+of `/home/var-lib/containers` and `/home/var-lib/docker` (fstab,
+`x-systemd.requires-mounts-for=/home`). The NVMe is kept for models. Paths and
+configs are unchanged, so podman, docker and restic see the same paths.
+
+Scripts and logs from the move: `~/podman-move-2026-10-29/` (cutover2.sh,
+fixcopy.sh, mounttest.py).
+
+**Why:** Connor wants the only fast SSD free for LLM weights.
+
+**How to apply — two traps, both hit on the first attempt:**
+
+1. Copied overlay layers do not mount on a different filesystem. Overlay
+   `index` is on by default in this kernel (`/sys/module/overlay/parameters/index`
+   = Y). Each container's upper `diff` root carries `trusted.overlay.origin`,
+   and each layer's `work/index` carries `trusted.overlay.upper`; both are file
+   handles from the old filesystem. Every `podman start` failed with
+   `overlayfs: failed to verify upper root origin`. Fix: remove that xattr and
+   delete `work/index` after the final rsync (fixcopy.sh).
+2. `podman start` run from a transient systemd unit with default KillMode puts
+   conmon in that unit's cgroup, and systemd kills every conmon when the unit
+   ends. This produced the stale-`running` state from CLAUDE.md item 3a for all
+   51 containers. Always use `systemd-run -p KillMode=process`.
+
+Also: `mount(8)` from util-linux 2.42 fails any overlay with more than 8
+lowerdirs, with no kernel message, even on healthy storage. Test overlay mounts
+with mount(2) directly (mounttest.py), and run a control against known-good
+layers before trusting a failure. And any podman call between the final rsync
+and an rsync verify touches `storage.lock` and `db.sql` and fails the verify.
+
+Jellyfin must start before Home Assistant; both bind udp/1900 and whichever
+starts second fails.
@@ -13,7 +13,7 @@ The LiteLLM stack serves two profiles with different permitted providers:
  formally banned by his office but should be treated as prohibited until the
  security office gives written approval. That restriction is about HOSTED
  PRC-origin models (Moonshot/Kimi etc.). A PRC-origin model running locally
-  (qwen3-coder-30b on llama-swap) is fine for work, since nothing leaves the
+  (a Qwen model on llama-swap) is fine for work, since nothing leaves the
  machine. Connor's ruling, 2026-10-22.
 - **Personal** — `anthropic/*`, `openai/*`, and local.
 - `zen/*` is blocked for both, see [[zen-tos-hold]].
@@ -1,10 +1,15 @@
 ---
 name: local-code-autocomplete
-description: "Local code autocomplete (FIM) works on mainframe; qwen3-coder-30b is the best model for it despite being written off for capacity"
+description: "RETIRED 2026-10-29. Local FIM autocomplete on mainframe (minuet + qwen3-coder-30b): what was measured, kept in case it comes back"
 metadata:
  type: reference
 ---

+Retired 2026-10-29: Connor found the completions not worth having. minuet was
+removed from nvim and qwen3-coder-30b from llama-swap, LiteLLM and disk. The
+full config is archived in `~/guides/llm/archive/local-tier-2026-10.md`. What
+follows is the state before removal.
+
 Measured 2026-09-18 (llmbench suite 6b, `fim`).

 **Local code autocomplete is viable on this box.** Route: a raw