This is huge release with @JaredHatfield contributing a new capacity based router!

Rather than defining which models fit with other models, the capacity router
allows declaring pools of resources and how much of each resource a model needs.
The new router will automatically evict models to free up enough capacity to
load the new model. For use cases that have a big combinations of models
that can run concurrently, the new router can eliminate a lot of manual
configuration work.

The other big feature is the Help Agent in the UI can now search llama-swap's
source code as well as its knowledge base. Every release of llama-swap includes
the source code it was built from (tests excluded). This is handy for answering
questions about how llama-swap actually works and a knowledge base article
does not exist. One step closer to the singularity!

If you want to try out the new router ask the Help Agent to generate a configuration
example using your existing configuration. Did you know, it can do that!? :)

• PR #1242 router: add capacity-based model eviction: adds a capacity router where user-defined pools bound which models can be resident together and evict_cost picks what to unload when a model no longer fits, replacing the need to hand-maintain every valid combination in the matrix on systems with lots of unified memory (#1004) by @JaredHatfield
• PR #1258 docs/kb: add guides/tools for the cmd/ programs: adds knowledge base articles for wol-proxy, fake-model and vllm-wrapper (and moves the kubeswap guide under guides/tools) so the Help chat can answer questions about the tools in the llama-swap ecosystem, plus a convention for keeping each tool's README and article in sync by @mostlygeek
• PR #1259 internal/mcptools: add code search to Help agent: gives the Docs agent code__grep, code__read, code__find and code__ls so it can verify knowledge base articles against the code that implements them, adds helpagent.disable to hide Help and stop serving /api/mcp, and raises the agent loop's max iterations from 16 to 96 so the extra tool calls do not end the answer early by @mostlygeek
• PR #1265 internal/mcptools: embed source for code tools instead of git: embeds the Go and TypeScript source (without tests) in the binary via the new embed_src build tag and reads it through fs.FS, because the runtime git checkout made the code tools fail on installs without the git command line tool; code__status and code__update are removed by @mostlygeek

This release includes security dependency updates and requested quality of life changes from users:

• the UI no longer expects to be run under a root path making it easier to serve llama-swap behind another proxy
• the max request size is configurable now. Users should be free to OOM as much as they want. llama-swap ain't your mom!
• allow manual override of values in /v1/audio/voices. This is a non-standard endpoint but makes working with upstream audio engines a bit nicer

Details below:

• PR #1224 cmd/kubeswap: attach DRA ResourceClaims and mount NFS volumes: add repeatable --resource-claim and NFS --volume support so Kubernetes backends can share a DRA GPU claim and mount models without a static PV and PVC for each export (#1222) by @gabrielcosi
• PR #1225 Add voices override for speech models: let local models configure capabilities.voices so /v1/audio/voices returns a reliable list without starting the model or querying the upstream server (#1207) by @mostlygeek
• PR #1228 ui,server: serve the UI under a reverse proxy base path: use relative asset, API and redirect paths so the UI works behind a prefix-stripping reverse proxy without routing its endpoints at the proxy root (#1227) by @mostlygeek
• PR #1249 internal/server: make the request body size limit configurable: add maxRequestBodyMB for model endpoints and Tailcat, replacing the fixed 250 MB cap while retaining 250 MB as the default (#1245) by @mostlygeek
• PR #1257 AGENTS.md, go.mod: use go 1.27.3, trim AGENT rules: raise the Go version to 1.27.3, update dependencies and simplify contributor instructions by @mostlygeek

A couple of small improvements:

• support serving llama-swap behind a path on a reverse proxy. Helpful if you want to front llama-swap with nginx and serve it under a specific path.

• add support for manually defining voices for text to speech models

• a small kubeswap update

• PR #1228 ui,server: serve the UI under a reverse proxy base path: makes the browser side base-path agnostic (relative asset, redirect, favicon and PWA URLs plus an apiUrl() helper that re-attaches a stripped prefix) so llama-swap works behind a plain prefix-stripping proxy at https://host/subpath/llama-swap/ without routing /ui, /api and /v1 at the proxy root; includes an nginx and Caddy guide (#1227) by @mostlygeek

• PR #1225 Add voices override for speech models: adds a models.capabilities.voices capability listing the voice ids an upstream server supports, because the non-standard GET /v1/audio/voices endpoint returns unreliable responses; falls back to probing that endpoint on a best-effort basis when the list is not set (#1207) by @mostlygeek

• PR #1224 cmd/kubeswap: attach DRA ResourceClaims and mount NFS volumes: adds a repeatable --resource-claim name=claim flag so backends can get GPUs via the NVIDIA DRA driver (kubeswap only references the claim, pods sharing a claim share its device) and an nfs:server/export:path[:ro] kind for --volume so models can be mounted from NFS without a static PV/PVC per export (#1222) by @gabrielcosi

This release includes a new auto:orphans set for swap matrix rules. This new set automatically includes all models that were not explictly added to a set. It is defined as (orphan1 | orphan2 | orphan3 | ...). The orphans set is handy for definitions like: +always & +auto:orphans, where +always is a set of models that are always loaded and should never be swapped out.

The /audiocpp endpoint was added to make working with audio.cpp server eaiser. The audiocpp_auto model is automatically used when the endpoint is requested. audio.cpp's OpenAI compatible endpoints can be accessed directly via /audiocpp/v1/audio/*. Very handy when used with llama-swap's unified docker container which builds audio.cpp with the UI and native model downloader.

Also one tiny commit to improve compatibility with llama-server's decision model functionality.

Thanks to the contributors in this release, especially @knguyen298 for being patient and keeping the PR updated.

• PR #1026 internal/matrix: add synthetic auto:orphans set: adds a +auto:orphans selector that expands to every model not named in another set, so a catch-all set like "+always & +auto:orphans" picks up new models automatically (#1025) by @knguyen298
• PR #1223 Add /audiocpp endpoint for audio.cpp UI passthrough: serves audio.cpp's web UI and model management through a dedicated /audiocpp/ endpoint bound to the reserved audiocpp_auto model, like /comfyui/, so the unified container works without naming a model per request by @mostlygeek
• PR #1220 internal/capcompat: detect decision models from llama-server's listing: fix decision models being reported as out: [text] (and capabilities.out: [decisions] failing validation) because discovery only read /props; output_modalities from /v1/models is now used, and tools are only advertised when the model outputs text (#1219) by @nazeshinjite

I've been busy, busy, busy and have not had as much time for llama-swap as I
would like. This release includes reliability improvements, improved
compatibility and some small quality of life improvements. Thank you for the
amazing contributors for this release! Everyone who files issues or submits a
PR is greatly appreciated.

• PR #1213 Add confirmation UI and progress tracking for WoL wake: wol-proxy gets a -require-confirm flag that shows a "Start server" page instead of waking on every request, plus a retro loading page with a progress bar estimated from the last wake duration by @mostlygeek
• PR #1215 internal/router: strip browser headers before proxying to peers: fix UI chats failing with a 401 against an api.anthropic.com peer, which treated the forwarded browser Origin header as a direct browser call; Origin, Referer, Sec-Fetch-* and Access-Control-Request-* are now removed (#1214) by @mostlygeek
• PR #1212 internal/process: report loading progress from the health check: parse progress/message from not-ready health check JSON bodies, show it on the Models and model detail pages, and restart the healthCheckTimeout countdown when progress changes so long loads don't time out (#1079, #1208) by @mostlygeek
• PR #1196 internal/perf: restart loop to trigger GPU collector when the monitor still running: fix GPU stats stopping for good after nvidia-smi exited; the exit error is now logged and the collector restarts with a 5 to 30 second backoff (#1155) by @hpdkhoa
• PR #1211 internal/process: add a per-model disableKeepAlives option (#1205): fix intermittent 502 proxy error: EOF when a POST reused a pooled connection that llama-server closed after a streamed response; disableKeepAlives: true opens a new connection per request by @johnnieCR
• PR #1199 fix(metrics): show oMLX token rates in Activity: read oMLX's usage.prompt_tokens_per_second and usage.generation_tokens_per_second as a fallback so rates no longer show as unknown by @junmo-kim
• PR #1198 internal/capcompat: add Gufo inference engine prober: add a default capability prober that uses only /v1/models data as the fallback for servers without a custom prober (vLLM, Gufo, SGLang, etc.), replacing the vLLM-specific one by @berney

systemone models are mega hyped right now. This release adds support for
/v1/systemone for that fast thinking goodness(?). Also requests missing a
Content-Type will get one automatically added for the inference endpoints.

• PR #1197 Add /v1/systemone endpoint to POST JSON routes: route llama.cpp's new /v1/systemone decision model endpoint to the model named in the JSON body (#1195) by @mostlygeek
• PR #1194 internal/server: accept JSON bodies with a missing or wrong Content-Type: fix a misleading "no model id could be identified" 404 when curl -d sent JSON as form-urlencoded; a middleware now relabels JSON object bodies as application/json and returns a 400 naming the expected format for unusable bodies (#1192) by @mostlygeek

This release has a bunch of small compatibility improvements requested by the community.
Also the changelog summaries have been changed to only AI for the list below. These notes at
the top will be written by humans (or my dog) for humans. The AI's writing more correct but
fuck it we have too much AI slop in our lives already.

• PR #1189 Increase ComfyUI concurrency limit from 50 to 999: fix ComfyUI stuck on the splash screen via /comfyui because one tab makes more parallel requests than the old minimum of 50 allowed (#1188) by @mostlygeek
• PR #1187 docker/unified: compile sm_70 in the CUDA 12 image for V100: fix ik-llama-server aborting on V100 GPUs, where only sm_61 PTX without a WMMA flash-attention kernel was available (#1185) by @mostlygeek
• PR #1158 Hardware Detection: detect Intel GPUs on Linux via xpu-smi, including dedicated Arc memory, and read the Metal version on macOS across system_profiler key names by @anantshri
• PR #1183 docs/kb: document that tailcat.models accepts selectors: note that aliases, selectors, profile pins and peer model names are accepted, with a selector example by @mostlygeek
• PR #1182 Build audio.cpp with static eSpeak-ng and install phoneme data: link eSpeak-ng statically and ship its phoneme data so Kokoro, SanoTTS and Inflect phonemization works without a writable cache (#1181) by @mostlygeek

This release fixes a crash at startup on macOS 27 with Apple M6 hardware
(#1178). Hardware detection reads the CPU frequency through gopsutil, which
expects a voltage-states5-sram property that is missing on M6. It passed the
empty value straight to CFDataGetLength, which segfaulted before the server
could start listening.

As a temporary fix, gopsutil now points at a fork with the patch from
shirou/gopsutil#2163, which checks for the missing property. This will be
removed once the upstream project picks up the patch. Some Go dependencies
were also updated as required by the newer gopsutil.

• PR #1180 go.mod: use gopsutil PR #2163 to fix darwin/arm64 crash: replace gopsutil with a patched fork that guards the CPU frequency probe against a missing property on Apple M6

Maintainer note:

This release includes a bunch of small UI improvements and
requests from users. It also includes a contribution to detect DGX Spark
hardware more accurately.

I'm working on the AI sloppiness of the change log generation. The details
are correct but I need to tune it. Perhaps I'll just keep writing notes at the
top like this :)

Logging is split into three separate streams: proxy, upstream and HTTP access
logs. HTTP request lines used to be mixed in with the proxy log, which made it
hard to search. They now have their own stream and their own tab in the log
viewer. The logToStdout setting accepts a comma separated list of streams, such
as "proxy,http", and the old values still work.

Upgrading: log data has moved off GET /api/events to a new GET /api/events/logs
endpoint. Pick streams with ?stream=proxy&stream=http. Tools that read logs
from /api/events need to switch to the new endpoint.

The UI sidebar can be resized by dragging its edge, and the width is
remembered. The model page has a new header that shows the model name, a
copyable ID, its state and how long it has been ready. The API also returns a
new readySince field for models that are ready.

The Hardware page now recognizes DGX Spark (GB10) systems. Before this, it
could not count sockets on arm64, labeled the unified memory as dedicated
memory it could not detect, showed only the efficiency cores and had no power
figure. It now shows the system model, the hybrid CPU layout, shared memory
and the nominal 140 W SoC power.

The "install as app" prompt now works when apiKeys is set. The browser fetches
the web app manifest without an API key, so it got a 401. The manifest is now
served without auth, like the favicon.

The release process now checks for Claude or Codex before generating the
changelog and uses a newer model for it.

• PR #1177 internal/server: serve web app manifest without requiring auth: let browsers fetch site.webmanifest when apiKeys is set so the install prompt appears
• AGENTS.md: revise contribution guidelines for PRs: agents check pull requests and issues against CONTRIBUTING.md and ask when there is a problem
• PR #1172 Refactor logging to support multiple streams and separate log endpoints: split proxy, upstream and HTTP logs, add /api/events/logs and allow a stream list in logToStdout
• PR #1163 hw: detect DGX Spark (GB10) systems: report the system model, hybrid CPU, sockets, shared memory and nominal power on GB10 boxes by @lemassykoi
• PR #1173 ui: sidebar resize, model uptime tracking, small tweaks: drag to resize the sidebar and show how long each model has been ready
• scripts/add-changelog.sh: use gpt-6-sol: update the model used to generate changelog entries with Codex
• Makefile: detect harness before release: use Claude or Codex to generate the changelog and stop with an error when neither is installed

Tailcat is updated to v0.7.0 and now checks allowed clients in the HTTP
handler. A client that is not allowed gets a 403 instead of a TCP timeout,
which makes peer connections much easier to troubleshoot, and changes to
tailcat.allow now apply on hot reload. A bug that dropped the connection when
a node key was not recognized, which could leave clients hanging, now also
returns a 403. Peer connections recover better when the server disconnects,
and they wait longer for the first response header by default.

Upgrading: an empty tailcat.allow list now denies every client. To keep
allowing anyone to connect, set tailcat.allow to "*".

The output of cmdStop goes to the process log again, the same as cmd. This was
lost in an earlier change (#790). A cmdStop that leaves a background child
holding its output open can no longer hang the stop, and hitting that wait
limit after a successful stop is logged as a warning instead of a failure.

The Models page shows a status icon for each target of the selected profile,
and those models can be loaded and unloaded from the profile pane.

The unified Docker image adds mp3 support and native downloads to audio.cpp.
The image build also stops failing on untagged main-branch commits. Those
builds fell back to looking up the "latest" release through the GitHub API,
which rate limited the shared CI runners. The installers now build the commit
directly and fall back to git tags when the API is not available.

Release notes on GitHub now use the CHANGELOG.md entry for the tag, and
changelog entries are written as plain-language summaries that credit outside
contributors.

• release: improve generated changelog and release notes: generate readable changelog entries, credit outside contributors and use each entry as the GitHub release notes
• PR #1169 internal/tailcat: various reliability and bug fixes: return a 403 for unrecognized node keys, improve peer reconnects and raise the default time to first header
• PR #1170 ui: link profile targets and add load controls: show status and load/unload controls for profile target models by @sousekd
• PR #1165 internal/process: log cmdStop stdout and stderr: send cmdStop output to the process log and stop a lingering child from hanging it by @dividehex
• PR #1166 internal/tailcat: upgrade to tailcat v0.7.0 and use Server.Listen: check tailcat.allow per request, deny by default when empty and add a "*" wildcard
• PR #1157 docker/unified: audio.cpp add download and mp3 support: add mp3 and native download support
• PR #1156 docker/unified: fix vllm-wrapper build on untagged commits: build untagged commits directly and fall back to git tags when the GitHub API is rate limited