chaincompiler

ChainCompiler

A compiler-compiler for cognition.

Define a notation. Gate it so the model never malforms it. Compile it into a universally-loadable skill. Compose those skills. Organize them into a navigable tree. Code validates every step.

★ GitHub · Live site · 🌼 The Hive Log

Why

LLMs learn in-context — give the model a notation and it thinks in that notation. The hard part was never running one notation; it was minting, checking, composing, and organizing them reliably. ChainCompiler is the machine for that. The point isn't any single notation — it's the factory that produces them, with a linter at every seam so the model can't drift out of its own syntax.

It is all syntax, by design. Nothing here judges whether a chain's content is good. The compilers guarantee well-formed, composable, organized artifacts. Correctness of thought is not their job — and that boundary is load-bearing.

The one idea: a closed algebra over one type

Everything reduces to a single type — the skill dir (<name>/SKILL.md), the unit any agent auto-loads. Every operation produces or consumes only that type, so composition closes and recurses:

*CC            constructs   →  skill dir       (lift a notation into the type)
skillchain     composes     →  skill dir       (a skillchain OF skillchains)
SkillTree      organizes    →  skill dirs      (a navigable tree of them)

The *CCs are constructors, skillchain is the composition operator, SkillTree is the arrangement, and validators do the job the substrate refuses to.

The loop — every *CC runs this

Roadmap

P0 Foundation ✓ · P1 Project Surface ✓ · P2 Exchange ✓ · P3 Self-Hosting & SI ✓ · P4 Marketplace ▸ · P5 Federation ✓ (plan) · P6 Skill OS ▸ · P7 Plugin ▸

Full detail in ROADMAP.md (live on the site); concrete tasks in BACKLOG.md. The image and the site are generated from roadmap.json by scripts/update_site.py — edit the data, run the script, everything updates.

The stack, layer by layer

rulecatcher catch examples.txt --scope mydsl
rulecatcher lint  newtext.txt  --scope mydsl   # non-zero on a violation → CI-gateable

honeyc norm   examples/mbr.rune
honeyc render examples/mbr.rune --as prose

chaincompiler layers     # print the stack
chaincompiler demo       # mint a whole 'triage' cognition-language in one call

[Symptom] ⇒ [Repro] ⇒ [Hypothesis] ⇒ |Localize|