v0.10.16

0.10.16 (2026-11-02)

Features

• modelpool: report a member scaled up by a field manager outside LLMKube (#2356) (16446a2)
• observability: alert when an InferenceService or Model spec change goes unreconciled (#2346) (87bdb35)

Bug Fixes

• agent: proxy in full duplex so a streamed completion over HTTP/1.1 is not cut (#2358) (6de423b)
• chart: pass an explicit zero to disable the watch canary and reconcile deadline (#2352) (d1b230c)
• metal: MetalNode verifier follow-ups from the #2235 review (#2345) (4d2a83a)
• metal: withdraw a MetalNode Reachable result when a spec edit changes the check (#2359) (5ecbe3e)
• reject autoscaled ModelPool members (#2323) (7a15843)
• router: evict idle budget keys and cap team keys (fixes #2337) (265ed80)

Performance Improvements

• router: rate-limit the token budget gauge republish (#2354) (998623b)

Documentation

• clarify transcript retention when cleanup is disabled (#2347) (51b59b8)
• guides: add a guide for serving vision models with spec.mmproj (#2344) (8eab510)
• note Foreman CRD upgrade order (#2348) (1b239bf)
• runbooks: read observedGeneration from conditions in the stalled-watch runbook (#2353) (e79f900)
• security: document the InferenceService trust model, host-path roots and router budgets (#2341) (0ac7589)
• upgrade: router-proxy team budgets now keep only team values active in the budget window, at most 10,000 at once; new values past the cap share a per-budget overflow budget, counted by llmkube_router_budget_team_overflow_total. A team budget's status.budgetUtilization now covers only active values, so it no longer drifts toward zero as idle values accumulate. (265ed80)