v0.9.12

0.9.12 (2026-07-26)

Features

• add --prompt-depth-sweep to benchmark command (#1280) (eb511f9)
• foreman: support chat_template_kwargs on agent chat requests (#1285) (1028335)
• grafana: add SGLang and vLLM runtime dashboards (#1266) (07f340a)
• grafana: union llama.cpp, sglang and vLLM metrics on the inference dashboard (#1248) (7cb226b)
• hack: spine probes for models given repo access (#1274) (b7b3d9b)

Bug Fixes

• Available condition reflects Stopped and Suspended phases (#1273) (208bd79)
• disable llama.cpp prompt cache in benchmark requests (#1272) (eaa61ad)
• foreman: apply the empty-assistant guard to the preserved truncated turn (#1282) (b1f51cc)
• foreman: hold the session compaction drop point steady across turns (#1287) (febcb2a)
• foreman: three first-run failures on a clean cluster (#1295) (5ca3e64)
• gate: check non-Go changes against the shared merge-base, not HEAD~1 (#1284) (6940033)
• spine-probes: restore --max-tokens and the 'none' ablation arm (#1278) (9493201)

Documentation

• foreman: fix AgenticTask names in minimal example (#1290) (7659f29)