All major AI coding agents (Cursor, Claude Code, Cline, etc.) share the same fatal flaw: they forget context the second the conversation changes.
My goal was simple: make prompts reliable.
So I built Salacia — a lightweight Runtime OS that sits underneath any agent.
It does four things:
Compiles raw prompts into structured Intent IR + verifiable specs
Runs metamorphic testing to kill semantic drift
Only asks one smart question when truly high-risk (Risk-first gate)
Keeps full auditable journal (every change is traceable and replayable)
Tested on 500+ SWE-bench tasks:with 3 SOTA models, +9pp pass rate uplift and 93% fault localization accuracy.
Completely local, silent by default, works with Cursor, Claude Code, Cline, Antigravity, Gemini CLI and more.
Open source, one command to install.
GitHub: https://github.com/StartripAI/Salacia
Website: https://startripai.github.io/Salacia/
Would love feedback, especially from people who have been frustrated with prompt drift and context loss.
1 comment
Show HN: Salacia – The First Runtime OS for Agentic Coding | Heykuki News