Add commands

Document akmods (Fedora kernel module rebuild) and lumabri (P2P MoE swarm).

Author: SimonSchubert

assets/commands/akmods.md

@@ -0,0 +1,83 @@
+# TAGLINE
+
+Rebuild out-of-tree kernel modules from akmod packages
+
+# TLDR
+
+**Check and rebuild** missing or outdated modules for the running kernel
+
+```sudo akmods```
+
+**Force rebuild** of all modules even if they failed or look up to date
+
+```sudo akmods --rebuild --force```
+
+Build and install modules only for a **specific kernel** (same format as `uname -r`)
+
+```sudo akmods --kernels [kernel_version]```
+
+Rebuild only one **akmod** package
+
+```sudo akmods --akmod [akmod_name]```
+
+Show status of the background **akmods service**
+
+```systemctl status akmods```
+
+# SYNOPSIS
+
+**akmods** [_options_]
+
+# DESCRIPTION
+
+**akmods** checks installed **akmod** packages and rebuilds out-of-tree kernel modules when they are missing, outdated, or broken for a given kernel. It is the Fedora/RHEL-family counterpart to frameworks like DKMS: an **akmod** package ships module sources and build metadata; **akmods** compiles them on the local system (often into a matching **kmod-*** RPM) whenever a new kernel appears.
+
+Typical use is after installing or updating drivers from RPM Fusion (for example NVIDIA via **akmod-nvidia**). A systemd unit may run **akmods** at boot so modules for the current kernel are ready before they are needed. Manual runs are common when a build failed, **kernel-devel** was missing during an earlier attempt, or you need modules for a kernel other than the one currently running.
+
+Builds need a compiler toolchain and matching kernel headers/devel packages for each target kernel. Secure Boot setups may also need module signing keys under **/etc/pki/akmods**.
+
+# OPTIONS
+
+**--force**
+
+> Try all akmods, including ones that failed on earlier runs.
+
+**--kernels** _kernel_
+
+> Build and install only for the given kernel version string (same format as **uname -r**).
+
+**--rebuild**
+
+> Rebuild all modules even when they already appear up to date.
+
+**--akmod** _akmod_
+
+> Build and install only the named akmod package.
+
+**--quiet**
+
+> Reduce output.
+
+**--verbose**
+
+> Increase output.
+
+# CAVEATS
+
+Failed builds often leave no usable **kmod** until headers, compiler tools, and enough disk space are available, then **akmods --force** or **--rebuild** is re-run. On ostree-based systems (Silverblue/Kinoite) module signing and rebuild timing can differ from classic package installs. The man page is sparse; log output under the **akmods** service and journal is usually the best diagnostic trail.
+
+# INSTALL
+
+```dnf: sudo dnf install akmods```
+
+<!-- packages: 2026-08-10 -->
+
+# SEE ALSO
+
+[dkms](/man/dkms)(1), [modprobe](/man/modprobe)(8), [uname](/man/uname)(1)
+
+# RESOURCES
+
+```[Documentation](https://rpmfusion.org/Packaging/KernelModules/Akmods)```
+
+<!-- verified: 2026-08-10 -->

assets/commands/index.txt

@@ -157,6 +157,7 @@ airtun-ng.md
 aislop.md
 ajson.md
 ak.md
+akmods.md
 alacritty.md
 alembic.md
 alertmanager.md
@@ -4331,6 +4332,7 @@ ludusavi.md
 lue.md
 luigi.md
 luit.md
+lumabri.md
 lumen.md
 lumon.md
 lure.md

assets/commands/lumabri.md

@@ -0,0 +1,91 @@
+# TAGLINE
+
+Peer-to-peer MoE model swarm for distributed LLM inference
+
+# TLDR
+
+**Serve** a local model directory on the swarm (tracker + bytes + optional expert compute)
+
+```lumabri serve --model [/path/to/model]```
+
+**Chat** against a tracker, using colibri engine binaries from a local tree
+
+```lumabri chat --tracker [host:7300] --engines-dir [/path/to/colibri/c]```
+
+Open the interactive front end with **saved swarm settings** (no flags)
+
+```lumabri```
+
+Generate an **operator keypair** for a signed swarm
+
+```lumabri key --out [swarm]```
+
+Serve and **join** an existing tracker while donating disk space
+
+```lumabri serve --model [./slice] --join [host:7300] --model-name [model] --donate [GB]```
+
+Chat against a **local** model directory (no mirror, no second copy)
+
+```lumabri chat --local [/path/to/model] --engines-dir [/path/to/colibri/c]```
+
+# SYNOPSIS
+
+**lumabri** [_subcommand_] [_options_]
+
+# DESCRIPTION
+
+**lumabri** runs huge mixture-of-experts (MoE) language models across a swarm of peers using the [colibri](https://github.com/JustVugg/colibri) engine. Pure C, no runtime dependencies beyond a colibri build for the engine binaries. One machine can share a model; others chat without downloading the full weights up front. Bytes that inference actually touches are fetched from peers, verified, and kept in a local mirror under **~/.lumabri** so later prompts can run offline from cache.
+
+**serve** starts a tracker (index of who holds which files) and a maintainer that answers byte-range reads. With phase 2 enabled it can also run expert execution for peers. **chat** boots a colibri engine against a model path that may exist only on the swarm: an **LD_PRELOAD** shim fetches missing blocks, and optional expert offload sends activations to peers that hold those experts. With no arguments, **lumabri** prompts for swarm address and key once and remembers them in **~/.lumabri/config**.
+
+Integrity modes include open swarms (per-MiB hashes, optional operator signatures, spot-check of expert results) and private swarms gated by **LUMABRI_TOKEN**. Design goals emphasize byte-identical output whether experts run locally or remotely, and a hardware floor that includes CPU-only machines.
+
+# SUBCOMMANDS
+
+**serve**
+
+> Share a model: run tracker/maintainer (and usually expert execution). Flags include **--model**, **--port**, **--join**, **--donate**, **--model-name**, **--key**, **--advertise**, **--cache** / **--exec-cache**, **--no-exec**.
+
+**chat**
+
+> Connect to a swarm or a local model. Flags include **--tracker**, **--engines-dir**, **--local**, **--role**, **--donate**, **--model-dir**. In-chat commands include **/swarm** and **/model**.
+
+**key**
+
+> Create an Ed25519 operator keypair (**--out** path) for signed swarms.
+
+# ENVIRONMENT
+
+**LUMABRI_TOKEN**
+
+> Shared invite secret for private swarms; required on every node when set.
+
+**LUMABRI_PUBKEY**
+
+> Operator public key (hex) for verifying signed model hashes out of band.
+
+**LUMABRI_REQUIRE_HASH**
+
+> When set to 1, refuse fetches without integrity data.
+
+**LUMABRI_VERIFY**
+
+> Spot-check fraction for phase-2 expert results on a second replica.
+
+**LUMABRI_PREFETCH**
+
+> Readahead block count for cold mirrors (default 2).
+
+# CAVEATS
+
+Requires **Linux**, **gcc**, **GNU make**, and a matching **colibri** checkout for engine binaries (**make phase2-all ENGINE=...** for peer expert execution). First cold answers are network-bound while the working set is mirrored. Misconfigured **--engines-dir** or missing keys can cause huge downloads or unverified models; the interactive default path is intentionally safer than raw flags. Tracker/maintainer/executor ports (commonly **7300–7302**, plus offsets per extra model) must be reachable for multi-host swarms. Status is a working prototype: features such as speculative drafting and key rotation remain incomplete.
+
+# SEE ALSO
+
+[ollama](/man/ollama)(1), [llama.cpp](/man/llama-cli)(1)
+
+# RESOURCES
+
+```[Source code](https://github.com/JustVugg/lumabri)```
+
+<!-- verified: 2026-08-10 -->