ark
Star
Fork
Watch

Commits on stolas-builder-fix
wisp: periphery needs `which` to find git; stolas emulates aarch64 for the multiarch builder
committed
valefar: drop the bonsai CI filter and its tunnel forward (perf live is gone); spindle guests report free pages from 128 KiB
committed
sjo1: add a 768M swapfile
committed
wisp: copy the CA bundle into /etc instead of linking it vine-nixos
committed
vine as a NixOS host, the wisp status box
committed
sharkgirl as a NixOS host, on a docker-host module split from the edge sharkgirl-nixos
committed
baal: the US-east wisp edge as a NixOS host (aarch64-linux) baal-nixos
committed
wisp: sjo1 as a NixOS host, sharing an UpCloud module with sin1 sjo1-nixos
committed
sin1: use the scripted stage 1, which handles NIXOS_LUSTRATE sin1-nixos
committed
sin1: the Singapore wisp edge as a NixOS host
committed
wisp edge: run the gatus heartbeat probes from the edge module
committed
wisp edge: trust the Docker bridges so containers reach the mesh services
committed
focalor: strata takes upstream's crash fixes (prompt-window NaN retry, checkpoint-copy OOM skip), reply-only penalties, CPU prefill share to 3072
committed
focalor: strata back to Swift 1.5 IQ3_S with niri's think1 (ana's voice LoRA was a fun failed experiment)
committed
focalor: strata bans emoji tokens by default (config logit_bias)
committed
focalor: strata serves Swift-1.5-IQ3_S-GSQ-RCO-qfn-ana (Swift 1.5 IQ3_S with ana's voice LoRA merged in)
committed
valefar: bump tangled core to db870eb4 (its master was rewritten; 6b6a60cd no longer fetches)
committed
valefar: spindle microVMs get a free-page-reporting balloon, so guests hand freed memory back to the host while they run
committed
valefar: nixos-large spindle guests get 6 GiB, a ceiling that keeps albedo's test gate laptop-sized
committed
focalor: strata splices edited prompts again beside --kv-grow (SCR was off since 0333f3c), faster prompt bookkeeping
committed
focalor: strata parks at most 2 GiB of conversations in RAM, leaving the page cache to decode
committed
astaroth: report-html renders mermaid blocks to inline svg with pele
committed
trimounts: knot-static serves git.klbr.net, built from its own repo there
committed
focalor: strata grows its K/V on demand, prefill auto, no parking on rewound prompts (30K decode 41 -> 51 t/s, 64K prefill 95 -> 60 s)
committed
valefar: spindle on tangled core 6b6a60cd (workflow timeout starts when the workflow runs, per-workflow limits)
committed
focalor: strata serves Swift 1.5 GSQ-RCO IQ3_S with niri's think1 (ISTA IQ3_XXS think1 kept as rollback), helper 2800
committed
focalor: strata's RTX 4060 holds and computes 3300 experts beside vision (120K decode +17%, 30K +35%, KL 0.006)
committed
trimounts, valefar: ci.klbr.net/bonsai/ reaches Bonsai's CI filter through the spindle tunnel, behind a bearer token
committed
valefar: a text-only filter in front of Bonsai for albedo's perf CI, on localhost until ci.klbr.net routes to it
committed
valefar: spindle guests get a perf-agent user that spindle-workflow can sudo into, so CI can run untrusted tool calls apart from its secrets
committed
wisp edge: periphery needs openssl to generate its TLS certificate
committed
stolas: a NixOS wisp edge, and the wisp mesh as a shared module
committed
focalor: strata serves Qwen under its own name, qwen3.8-flash-next-think1
committed
valefar: spindle on tangled core ad0747f (workflow caches restore across commits and save on timeout)
committed
agents: sync skills with the local copies
committed
ghostty: follow the system window theme
committed
ghostty: switch theme with the system appearance and add miku-latte
committed
buer: bring the 1 GB headscale VPS into ark buer
committed
focalor: strata's vision encoder runs on the RTX 4060 (CUDA build for sm_89)
committed
focalor: strata no longer needs /strata-fast mounted
committed
focalor: strata back on Qwen3.8-Flash-Next GSQ-RCO IQ3_XXS with niri's think1, 200K context, served as deepseek-v4.1-flash
committed
tangled: master 9e6f01a
committed
trimounts: knot blob store
committed
trimounts: knot on tangled core master (1.0), admission closed
committed
valefar: pia-exit tries every DC ovpnudp server, not one pinned address
committed
valefar: spindle allows cross-origin reads
committed
valefar: spindle on tangled 1.0, served as did:web:ci.klbr.net
committed
valefar: swap to zram first, the NVMe partition only as overflow
committed
valefar: spindle metrics on the tailnet address for victoriametrics
committed
focalor: strata keeps and splices conversations with pictures (niri back after another request: 94 s -> 0.4 s), slot folder 16 GB
committed
focalor: nvidia-smi patched with patchelf 0.15 (the driver's copy segfaults, NixOS/patchelf#152)
committed
focalor: strata streams an empty delta every 15 s while reading a long prompt (niri's harness idle timeout ignores SSE comments)
committed
focalor: strata context 320K (niri's history is 224K tokens)
committed
focalor: strata DeepSeek-V4.1 sees images (vision tower on the 4060)
committed
focalor: strata keeps the 4060 expert GPU and its cache (3b5d464 dropped them by mistake)
committed
focalor: strata Cache-Prior over the 5070 and 4060 only, swaps at window ticks (63K decode 10.9 -> 12.0 tok/s, 16K 12.7 -> 13.5, deterministic)
committed
focalor: strata Cache-Prior favours RAM-store and 4060 experts too (63K decode 10.0 -> 10.9 tok/s, misses -40%, KL no worse)
committed
focalor: strata 4060 experts follow the conversation (63K-context decode 9.0 -> 10.0 tok/s, 16K 10.9 -> 12.3)
committed
focalor: strata SCR keeps conversation slots (a compaction after another conversation 79 -> 16 s, restart restores from disk)
committed
focalor: strata reuses edited prompts (SCR: 63K compaction 77 -> 16.5 s, continuation KL ~0.015), exits on QUIT at once
committed