v0.3.8

Patch release focused on the Spiritbuun agent stall regression left behind by v0.3.7.

Fixed

• Scoped Spiritbuun qwen3-chat to --n-predict -1 so long agentic turns no longer inherit the shared 8192 token cap that could leave multi-step chats hanging mid-run.
• Added --cache-ram 0 to the shipped Spiritbuun DFlash chat launcher so the draft path stays on the intended cache configuration during extended agent sessions.

Validation

• Ran .github/skills/easyllama-provider/scripts/validate-code.sh successfully.
• Rebuilt, restarted, and warmed the spiritbuun runtime successfully.
• Ran .github/skills/easyllama-provider/scripts/test-public-endpoints.sh spiritbuun successfully against the rebuilt runtime.

Compare: v0.3.7...v0.3.8