my nixos setup https://git.hwebs.info/nixos-config/atom?h=main 2026-08-05T00:09:56Z 2026-08-05T00:09:56Z Henry J. Webster hwebs@hwebs.info 2026-08-04T23:49:58Z urn:sha1:1f63d22ea3689cc567402dd6bf18838c9adea773 Also remove claude-code from human user packages to fully remove agent harnesses from the human account's path. Assisted-by: pi:laguna-xs-2.1 2026-08-04T23:36:13Z Henry J Webster hwebs@hwebs.info 2026-08-04T22:54:11Z urn:sha1:790f353742ecf73eb71944d5f6c17ecfaad6a75c laguna-xs-2.1 is large enough to nearly fill the workstation's display GPU at the previous 64k context and 1h keep-alive. That let Ollama keep a very large resident set around after requests and could destabilize the no-swap desktop session. Unload models quicker, limit parallelism/queueing, reserve VRAM for the compositor, and cap the service cgroup so runaway requests fail before the desktop does. Assisted-by: pi:gpt-5.5 2026-08-04T22:41:40Z Henry J. Webster hwebs@hwebs.info 2026-08-04T22:41:40Z urn:sha1:a8273dea992f0b31c59f87c5441271f97d2b729c 2026-08-04T22:14:09Z Henry J. Webster hwebs@hwebs.info 2026-08-04T22:10:10Z urn:sha1:a4c50351f5d55438a102d48c63dcbb62f9ef5e8d Serve Ollama on kusanagi's Wi-Fi LAN while removing the blanket tailscale0 firewall trust so tailnet traffic follows explicit host firewall rules instead of inheriting every listening service. Add shared LAN networking for both hosts: allow ping, publish and resolve .local names with Avahi, and scope mDNS to each machine's declared Wi-Fi interface. Assisted-by: OpenAI:gpt-5 2026-08-04T21:05:20Z Henry J Webster hwebs@hwebs.info 2026-08-04T05:02:54Z urn:sha1:df960b2ed92f9e45fc4984d19ab65c55074bb1d7 The 12B model at 32k context uses only 7.7 GB of the 7900 XT's 20 GB, so there is ample VRAM headroom for 64k with the q8_0 KV cache — agentic coding tools are context-hungry and 32k is the practical bottleneck. The keep-alive stops the default 5-minute idle unload from adding a reload stall to every resumed session on a single-user workstation. Update flake to get newer packages. Assisted-by: Claude Code:claude-fable-5 2026-08-04T21:04:51Z Henry J Webster hwebs@hwebs.info 2026-08-04T04:38:21Z urn:sha1:8f9fbf2f4a269eae143c9e279afc6608ce64b7d0 Committing directly to main let it diverge from origin/main, so the next pull became a rebase with a conflict on this very file. Recording the rule where every harness will read it keeps main fast-forwardable. Assisted-by: Claude:claude-opus-4-8 2026-08-04T04:26:57Z Henry J. Webster hwebs@hwebs.info 2026-08-04T04:26:57Z urn:sha1:6f135470d274daa3e7a7fec1cb8f36204abd90f0 There is no cross-vendor standard for agent instruction files: claude-code reads CLAUDE.md, opencode and pi read AGENTS.md, each with its own global path. Keep one canonical file (common/AGENTS.md) in this repo, ship it to /etc/AGENTS.md, and symlink every harness's global context path to it via tmpfiles for all normal users: ~/.claude/CLAUDE.md (claude-code 2.1.187 — no AGENTS.md support) ~/.config/opencode/AGENTS.md (opencode 1.15.10) ~/.pi/agent/AGENTS.md (pi 0.75.4) Paths verified against the packaged binaries; see the profile header for details and for the per-repo convention (AGENTS.md + CLAUDE.md symlink). Assisted-by: claude-code:Fable 5 2026-08-04T03:51:26Z Henry J. Webster hwebs@hwebs.info 2026-08-04T03:51:26Z urn:sha1:9acf89a3af3fd4c354fcd887397b0c035bb21494 The recurring "ollama runs 100% on CPU" regressions were a boot-order race, not a gfx-target mismatch: ollama.service only ordered after network.target, and at boot it started between /dev/kfd and the 7900 XT's render node appearing. ollama probes GPUs once at startup, found none, and silently fell back to CPU until the next manual restart — which is why the previous HSA_OVERRIDE_GFX_VERSION commit "worked": the deploy restarted the service on a long-running system, and the restart was the actual cure. The card is a 7900 XT (Navi 31), natively gfx1100 — verified that discovery with no overrides finds it fine and drops only the unsupported Raphael iGPU (gfx1036). Remove HSA_OVERRIDE_GFX_VERSION and HIP_VISIBLE_DEVICES (ollama itself warns overriding visible devices can break discovery). Real fix: systemd-tag the kfd/renderD* char devices via udev (they get no device units otherwise) and order ollama after them, so discovery only runs once the GPU exists. Assisted-by: claude-code:claude-fable-5 2026-08-04T03:12:32Z Henry J. Webster hwebs@hwebs.info 2026-08-04T03:00:49Z urn:sha1:3df910d75fa98a7d0291e6b8c6889905c6088f64 Move the yokai agent-sandbox account (user, ikai group, /srv/git drop dir, scoped sudo de-escalation, and the local-clones workflow docs) out of machines/enzo into a shared profiles/sandbox.nix that both machines import. The only per-machine value is the human owner allowed to drop into the sandbox, exposed as a `sandbox.owner` string option so the profile stays username-agnostic (enzo: hwebs, kusanagi: henz). kusanagi gains the sandbox it didn't have before; enzo's inline block collapses to one setting. Assisted-by: claude-code:claude-opus-4-8 2026-08-04T02:39:51Z Henry J. Webster hwebs@hwebs.info 2026-08-04T02:38:08Z urn:sha1:71e36370770acc7498748a890cd0365b3d351027 The RDNA3 card reports as gfx1101/gfx1102/gfx11-generic, which the ROCm build bundled with ollama-rocm has no kernels for, so ollama found no usable device and ran 100% on CPU after the nixpkgs bump. Force it to present as gfx1100 via HSA_OVERRIDE_GFX_VERSION. Also set OLLAMA_FLASH_ATTENTION=1 so the existing q8_0 KV cache actually engages instead of silently falling back to f16 and inflating VRAM use. Assisted-by: Claude-code:Opus-4.8