[github-actions]

🚀 Release ${version}

0.10.18 (2026-11-14)

Features

• agent: read host memory from /proc and cgroup v2 on Linux (#2531) (b2f1fe1)
• chart: add relabelings and metricRelabelings to the PodMonitors (#2467) (d0ae4f3)
• inferenceservice: per-service size for the model cache (#2469) (ec18285)
• metal-agent: default the model store and state dir under XDG_DATA_HOME off macOS (#2528) (883e1f8)
• metal-agent: start without llama-server when --runtime selects another engine (#2529) (4f7bb4b)
• metal-relay: preserve client source IP through the relay (#2483) (3892c27)
• multinode: drop the page cache and set vm sysctls on each member before boot (#2527) (e012e2e)
• router: gate router readiness on required local listeners (#2468) (033ca80)

Bug Fixes

• avoid stale namespace reads in quota admission (#2471) (12af4c4)
• catalog: correct llama-3.3-70b default context (fixes #2477) (#2484) (2c1dd79)
• cli: default qwen-2.5-14b to a 16K context within its trained 32K (#2478) (0dc14c2)
• config: install the FederatedCluster CRD with kustomize (#2515) (071c1ab)
• controller: provision no model cache PVC for a Metal InferenceService (#2513) (896ee03)
• default qwen-2.5-coder-7b to its trained 32,768-token context (#2497) (c2bdde2)
• keep the requested pool member from losing the slot to another member (#2536) (c9f2dfb)
• metal-agent: stop pointing EndpointSlice endpoints at a Pod that does not exist (#2537) (375340c)
• metal: charge the KV cache for local GGUFs, counting only the layers that hold one (#2496) (4fc227e)
• metal: report failed starts and act on engine exits at once (#2436) (fa154fc)
• multinode: cap memory sysctls in the controller and document the prep's limits (#2538) (79abe4d)
• preserve unhealthy running engines during adoption (#2462) (1c0b05f)
• show the first deploy's download and stop the Model claiming a cached copy (#2532) (c7ec64a)
• warn about unmatched model file patterns (#2472) (fb75656)

Documentation

• document the defilan-agent workflow for agent pull requests (#2506) (0fdda4c)
• say which checks the merge queue requires (#2510) (40259fd)
• site: fix the broken pages the link check found, and run the check in CI (#2503) (2b03f2f)
• site: link the install page to the site pages that replaced its GitHub copies (#2505) (4cb0d68)
• site: move "Install in 5 minutes" here and add a Mac path (#2501) (95893c6)
• site: publish the one-Mac quickstart on llmkube.com (#2502) (8643f2b)
• site: replace the docs placeholders with ported and generated pages (#2504) (b942257)
• site: run the one-Mac quickstart on Apple's container cluster (#2512) (401bda4)

---

This PR was generated with Release Please. See documentation.