v0.9.9

0.9.9 (2026-07-22)

Features

• api: vendor-neutral gpuSharing tiers on InferenceService (#1196 stories 1+2) (#1199) (6cbd1fd)
• chart: gpuSharing pool + VRAM values wired to operator flags (#1196 story 3) (#1205) (95eed30)
• controller: bound InferenceService pod lifetime via maxPodLifetimeSeconds (#1182) (3b39ee9)
• foreman: make the verify gate optional in the issue-batch decomposition (#1166) (e4e49aa)
• metrics: emit GPUQuota usage + admission-denial metrics (#416) (#1193) (dad2687)
• quota: VRAM-based GPUQuota accounting from the gpuSharing tier (#1196 story 4) (#1200) (2c9000c)
• runtime: Blackwell-ready image pins, NVIDIA llama.cpp divert, runtimeImages overrides, platform floors (#1197) (#1204) (9d40b36)
• webhook: admission-time gpuSharing validation (#1196 story 5) (#1206) (a9de8f7)

Bug Fixes

• agent: inject current date into issue-fix prompt and anchor research queries (#1202) (#1207) (5067095)
• chart: nil-safe gpuSharing navigation in deployment args (#1209) (bbbcdcf)
• codegen: regenerate stale committed codegen and widen the sync check (#1190) (aedcb80)
• deps: bump golang.org/x/text to v0.39.0 (GO-2026-5970) (#1192) (af5c78e)
• foreman: bound reasoning-model coder turns with a per-turn token budget (#1175) (5100c11)
• foreman: count identical read_file re-reads as no-progress under the edit-free nudge (#1179) (2fad7ec)
• foreman: inject Workload.spec.intent into coder prompts (#1201) (#1203) (9a8371c)
• foreman: warn when an MCP allowlist entry matches no server tool (#1184) (0664814)
• foreman: widen coder scope-overlap gate to a bottom-quartile floor (#1181) (f07cba4)
• webhook: serve GPUQuota webhook at its configured path, with e2e coverage (#416) (#1195) (b377f52)

Documentation

• grafana: AMD (Strix/gfx1151) GPU observability dashboard (#1188) (3da1f2b)
• observability: multi-tenancy guide + GPUQuota Grafana dashboard (#1097) (#1194) (aedbd52)
• operations: GPU sharing guide (#1196 story 6) (#1208) (c9007a3)