kyle
bccbe59e94
straylight: prefer IPv4 in gai.conf to fix update-channel timeouts
2026-10-11 12:05:40 -07:00
kyle
fa35293331
sift: pin 682b9b8044 (restore force-pushed commit)
2026-10-09 18:43:36 -07:00
kyle and Claude Opus 5.5
6d1000c756
straylight: add libxi to nix-ld libraries
...
Hermes' cua-driver (computer-use backend) is a prebuilt binary that links
libXi.so.6; without it the PM install fails verification with exit 127.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com >
2026-10-09 18:43:18 -07:00
kyle
b5e189bc5b
llama-server: add Gemma4 31B and Muse-Glimmer-30B models
...
Both fetched to /var/lib/llama-server/models by download.sh. Gemma4 31B
uses MTP draft (spec-type draft-mtp); Muse-Glimmer uses DFlash drafter
(spec-type draft-dflash, parallel=1 for the 104 GiB TTM window).
2026-10-05 18:14:15 -07:00
kyle
d7c1092d52
orion: pin sift 682b9b8 (P3.1 search filters + multi-query)
2026-09-30 07:38:31 -07:00
kyle and Claude Opus 5.5
c8b3b21004
dixie: honcho-embed pooling cls -> last
...
Qwen3-Embedding is a causal model; its card pools the last token. With cls every
input embedded as its first token's state ("the cat sat on the mat" and "the stock
market crashed today" at cosine 1.0). A CPU side server with --pooling last
reproduced the card's example scores within 0.002 without a manual EOS. Honcho's
stored vectors and sift's embed cache are re-embedded / flushed after the switch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com >
2026-09-29 23:46:37 -07:00
kyle
03ad1de40e
straylight: add trinity-mini (Arcee Trinity Mini 26B MoE) to llama-server router
2026-09-27 00:56:41 -07:00
kyle
cb0bc87633
orion: pin sift 5a59f7a (P6 asynchronous batch fetch)
2026-09-26 23:32:13 -07:00
kyle
4fa779be9c
orion: pin sift bdc093e (P5b long-doc questions + P5c extract; P6 plan docs)
2026-09-26 20:22:32 -07:00
kyle
cd2bd3d0e3
orion: pin sift 1311cf3 (P5a PDF documents)
...
PDF extraction in a re-exec'd child, per-page cache, map/pages/
grep, cited digests. Module MemoryMax 512M -> 1G via pinned sift.
2026-09-26 10:57:22 -07:00
kyle
ec39a569de
dixie: Z tmpfiles rules use ~0755 so existing files stay 0644 (exec dropped), dirs become 0755
2026-09-25 07:49:51 -07:00
kyle
e5c47007cd
dixie: Z tmpfiles rules to relabel dropsonde2 trees 0755 recursively (fixes rsync's 0700 dirs)
2026-09-25 07:48:32 -07:00
kyle
9a5920da2b
dixie: drop uid/gid from rsyncd modules — daemon runs as dropsonde via SSH, setgroups() fails otherwise
2026-09-25 03:03:40 -07:00
kyle
2305f4099a
dixie: rsync --server --daemon sync key (no rrsync/python3 on dixie)
...
- authorized key forces rsync --server --daemon --config=/etc/rsyncd-dropsonde.conf
- environment.etc rsyncd-dropsonde.conf: cold + analysis modules, read only=false,
use chroot=false, uid/gid dropsonde, max connections 2, lock file global
- tmpfiles: /srv/{cold,analysis}/dropsonde2 owned dropsonde
2026-09-25 02:45:44 -07:00
kyle
2e4aa6ed57
dixie: dropsonde sync key — rrsync -wo /srv (write-only, hyperborea -> dixie)
2026-09-25 02:24:26 -07:00
kyle
0ac90515bf
dixie: use script attr for dropsonde-pg-setup (ExecStart heredoc+bash path broken)
2026-09-25 00:51:32 -07:00
kyle
c402c1c599
dixie: drop python3.withPackages (duckdb CLI only) to trim the closure
2026-09-25 00:27:38 -07:00
kyle
eed2821cea
dixie: slice submodule uses sliceConfig, not serviceConfig
2026-09-24 23:42:12 -07:00
kyle
43b8e2288f
dixie: listen_addresses under settings, not top-level
2026-09-24 23:41:40 -07:00
kyle
3980242da5
dixie: close the ExecStart indented string in dropsonde-pg-setup
2026-09-24 23:41:25 -07:00
kyle
4b751796c6
dixie: import analysis.nix (was never wired into the module list)
2026-09-24 23:36:31 -07:00
kyle
0c37b07099
dixie: label /srv/cold as ds2-cold, use sfdisk for wipe
2026-09-24 23:20:01 -07:00
kyle
fddaffc825
dixie: fix slice unit name, pg-setup ExecStart, listen_addresses
2026-09-24 22:27:12 -07:00
kyle
a40873685a
dixie: open Postgres 5432 on tailscale0 for dropsonde
2026-09-24 22:17:44 -07:00
kyle
3c2a272722
dixie: dropsonde2 analysis tier (HDD /srv/cold, pg17+timescale+postgis+h3, analysis.slice, duckdb/polars)
2026-09-24 22:03:44 -07:00
kyle and Claude Opus 5.5
c12b2adbfd
disk-health plan: deploy via committed master + doas (per CLAUDE.md)
...
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com >
2026-09-24 18:59:39 -07:00
kyle and Claude Opus 5.5
85d5631fa8
docs: disk-health monitoring plan (SMART/NVMe/SD card)
...
Current state verified on 2026-09-24: hyperborea NVMe already exported
via the Debian nvme textfile; no SMART on the NixOS hosts; no
disk-health alert rules. Plan: smartctl_exporter module for the NixOS
hosts, smartmon timer and SD-card kernel-error textfile on hyperborea,
orion scrape job, a disk-health rule group, verification and deploy
notes. For hand-off to another agent.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com >
2026-09-24 18:59:15 -07:00
kyle
78d30026ce
orion: allowlist vmui's /prometheus API prefix
...
Default Server URL is /prometheus; keep import/admin 403.
2026-09-24 01:56:05 -07:00
kyle
f5ef78baca
orion: read-only vmui via nginx allowlist + serve :8429
...
VictoriaMetrics stays loopback :8428. nginx :8429 allows GET/POST
to /vmui and query APIs only; import/admin/export/flags 403.
tailscale serve HTTPS 8429. Review, no switch.
2026-09-24 01:47:25 -07:00
kyle
1cbde84601
rift: linger kyle and enable user podman-restart
...
Keep rootless podman up after logind idle.
2026-09-24 01:33:19 -07:00
kyle
f32db3c88b
metrics: systemd unit-include as [.]service
...
systemd drops \. in unit-include and warns about unknown escapes.
2026-09-24 01:22:16 -07:00
kyle
81897b3f8a
metrics: keep llama-straylight out of FleetTargetDown
...
35B eviction is not a paging event. params.model stays.
2026-09-24 00:46:23 -07:00
kyle
385529da39
metrics: phase 3 straylight/rift scrapes and hyperborea units
...
Review, no switch. node-laptop for talos/titan is excluded from
FleetTargetDown. llama-straylight scrape uses params.model for
ornith-1.5-35b-a3b (router 400 without it; will load if cold).
2026-09-24 00:42:06 -07:00
kyle
b861c6ee9f
metrics: drop llama-dixie-vision scrape
...
Router /metrics cannot be up without loading a model.
2026-09-24 00:29:24 -07:00
kyle
ed5e1a1e0e
metrics: HonchoDeriverBacklog and include honcho in target-down
2026-09-24 00:21:36 -07:00
kyle
5d8e8c4335
metrics: scrape dixie helper /metrics with model param
...
Router-mode llama-server returns 400 without it.
2026-09-23 23:57:36 -07:00
kyle
b1d5aa7e64
metrics: drop alertmanager channel override
...
Webhook is already bound to ~alerts; Mattermost's channel
override wants the name, not ~alerts.
2026-09-23 23:44:24 -07:00
kyle
ec74e0b69c
metrics: phase 1 follow-up (review, no switch)
...
Honcho workspace_name keep, node_exporter restart metrics,
drop min_over_time on rule 1, Watchdog heartbeat, pin tether
to b267e2f so labelled series exist and rule 2 can fire.
2026-09-23 23:43:11 -07:00
kyle
4958493485
metrics: phase 1 orion+dixie (review, no switch)
...
VictoriaMetrics on 127.0.0.1:8428. LoadCredential for the Mattermost
webhook and the sift scrape token. Honcho scrapes keep a closed metric
name set. Meta-alerts for target-down, VM size, orion disk. Dixie gets
node_exporter, nvidia_gpu_exporter, and llama-server --metrics.
Tether labelled series still wait on titan's fix. Rule 2 is written
against reason=client_disconnect so it stays silent until then.
2026-09-23 23:22:38 -07:00
kyle
888e8e24fb
orion: pin sift 4372673 (P2.7 metrics)
...
Prometheus /metrics behind a dedicated metrics caller; JSON stays
at ?format=json. No search-source change.
2026-09-23 23:11:30 -07:00
kyle
c07c4bfe4f
docs: fix tether cancel-storm PromQL (rule 2)
...
Label-set mismatch made the ratio empty. Select reason=client_disconnect
and kind=converted; never sum() a tether series without a selector.
2026-09-23 22:50:26 -07:00
kyle
8b14e1f99f
docs: fleet metrics and alerting plan
...
Reviewed plan only: VictoriaMetrics on orion, vmalert/Alertmanager,
tailnet exporters, first five rules. No config, no deploy.
2026-09-23 22:48:51 -07:00
kyle
504d4eb3f1
orion: pin sift e62cfc8 (restricted tavily cap skip)
2026-09-23 16:38:41 -07:00
kyle
bc30315344
orion: sift tavily+searxng, cap 130, tavily no_rerank
...
Kyle chose Tavily-first with SearXNG fallback. Brave stays off.
2026-09-23 16:25:01 -07:00
kyle
75b813b10b
orion: SearXNG-only until Kyle picks sift search default
2026-09-23 16:20:38 -07:00
kyle
415589cc13
orion: pin sift 1c57845, brave_llm only for live fixture
...
RRF original-URL fix. SearXNG and Tavily stay off until Kyle picks a default.
2026-09-23 16:10:59 -07:00
kyle
573b9dcfd8
orion: disable sift brave_llm for bake-off config 1
2026-09-23 15:42:54 -07:00
kyle
c84ca93024
orion: disable sift tavily for bake-off config 2
2026-09-23 15:36:33 -07:00
kyle
91c021cff9
orion: enable sift searxng, brave_llm, tavily
...
Bake-off sources. SearXNG journal is already LogLevelMax=err.
Keys stay in /var/lib/sift/secret.env, not the store.
2026-09-23 14:14:25 -07:00
kyle
c839f7c8ae
dixie: fail over USB Wi-Fi to internal CNVi
...
NM dispatcher plus a boot oneshot keep the AX9L (mt7921u) primary and
bring up the on-box iwlwifi profile if the dongle is gone. Connection
PSKs stay on-box; this only selects the radio.
2026-09-23 14:03:30 -07:00