- frames a writeonce port of ~/projects/skillhost (C++ MCP host that
links libllama in-process, discovers filesystem skills, runs their
scripts confined) as the sample that drives host capabilities into
the open — the way log-watcher drove the systems stdlib
- names each gap as a candidate iteration (surveyed 2026-08-16 vs the
running compiler + skillhost source):
- Blocker A: in-process native-lib FFI (no FFI today) — fork:
FFI-as-language vs out-of-process model driver over proc/net+json
(llama-server, needs nothing new); leaning out-of-process
- Blocker B: stdio transport — no stdin/stdout builtins; port uses
a TCP socket meanwhile; io.stdin_read/stdout_write a candidate
- Blocker C: bounded/killable subprocess — proc.run has no timeout/
signal/process-group kill; smallest + most broadly useful, do 1st
- partials: recursive fs walk, exec-bit check, symlink-resolving
confinement (realpath) — one small fs-metadata iteration
- records what is already expressible (catalog via @table/9g skill-
catalog, discovery, frontmatter text-parse, config, single-thread
serve, context-gate arithmetic — no VRAM query needed)
- out of scope: in-process libllama/CUDA, VRAM introspection, exact
sampler chain / per-turn memory clear
- roadmap + board rows added
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
