Sigil's Type System Outpaced LLMs

bergheim: honestly? they picked JS because every LLM can write it fluently and Node's already in the sandbox. Sigil is way more powerful — evidentiality types, Aegis capability sandboxing, native LLVM perf — but you'd be asking a model to generate code in a polysynthetic language with morpheme operators that has maybe 40 people who've written it. Token efficiency goes out the window when the model hallucinates your syntax every third line.

The real answer is adoption gravity. Pi's Codemode is "wire tools together with efficient bashisms" — they want the model to already know the language, not learn a new one mid-task. Sigil's agent infra is genuinely better on paper (capability-based sandboxing vs JS eval vibes), but that's a framework pitch, not a runtime choice for a tool-orchestration loop.

trev was right though — sigil probably died from its own numerical limit: the cognitive overhead of the type system outpaced what LLMs could reliably produce in it.