🚧 Under construction — pre-release. APIs, specs, and layout change daily; nothing here is published to the VS Code Marketplace or Open VSX, and there are no stability guarantees. Star/watch to follow along — don't depend on it yet.
ScroogeLLM is not in this repo. Its source, spec, design, and research live in a separate private repository (DR-027). This monorepo hosts MinSpec (open), the shared classifier, and the extension-pack manifest.
Monorepo for the open MinSpec VS Code extension, the shared classifier engine, and the MinSpec Pro extension-pack manifest.
| Package | ID | Domain | Status |
|---|---|---|---|
packages/minspec |
aiclarity.minspec |
minspec.dev | SDD Implement (pre-release) |
packages/shared |
@aiclarity/shared |
— | Shared classifier |
packages/extension-pack |
aiclarity.minspec-pro |
— | Manifest only — refs ScroogeLLM by marketplace ID |
ScroogeLLM (
aiclarity.scroogellm, scroogellm.com) — developed in a private repo (see banner above). The pack references it by marketplace ID once published.
MinSpec — scope-adaptive spec-driven development. Classifies each change by its mechanical scope (blast radius — files touched, lines, cross-boundary spread) into a tier, then applies proportional ceremony. A tier measures how far a change reaches, not how hard it is to think through. One-file fix = one sentence of spec. Architecture rewrite = full treatment. The predicted tier is a floor (ceremony only ratchets up); you can always raise it. Works with zero AI tools installed.
ScroogeLLM — LLM proxy that minimises token spend. Anonymises PII, caches aggressively, downgrades models when the task allows. Every token counts.
MinSpec Pro — extension pack referencing both. Unlocks spec-conformance checks that use ScroogeLLM as the inference layer.
Rules every change must preserve. See CLAUDE.md for full list.
- No AI dependency — core path makes zero AI calls.
- Tiered network consent (DR-004) — Tier 0 fully offline. No
http/fetchinpackages/minspecorpackages/shared. - No lock-in — Spec Kit-compatible markdown, no proprietary format.
- Ceremony proportional to complexity.
- User override always wins — classifier suggests, human decides.
- Harness regeneration preserves user edits (merge, not overwrite).
- All LLM calls go through proxy.
- Savings auditable — raw vs actual cost logged per request.
- PII anonymisation deterministic.
- User API keys in OS keychain only.
- Proxy binds localhost by default.
- Free-tier optimisations always active.
npm test # all packages
npm run lint
npm run build
npm run validate # frontmatter check on specs/**/*.md
# Package one extension
cd packages/minspec && npm run package # → .vsix
# Publish (requires vsce token)
cd packages/minspec && npx vsce publishspecs/<product>/ SDD artifacts (requirements, design, tasks)
docs/decisions/DR-NNN.md Architectural decisions (see INDEX.md)
docs/research/ Background research
sites/minspec.dev/ Marketing site
sites/scroogellm.com/ Marketing site
packages/ Workspaces
scripts/hooks/ Pre-commit + dispatch
Spec-driven development. Three phases per product: Specify → Plan → Implement. MinSpec is at Implement. ScroogeLLM has not started Specify.
Bug fixes follow RCDD — Root-Cause-Driven Debugging — reproduce, diagnose, fix, harden. No code changes in phases 1–2.
All architectural decisions land in docs/decisions/ as DR-NNN.md.
Multi-licensed — see LICENSE and DR-018. Each package's LICENSE file is authoritative.
| Path | License |
|---|---|
packages/shared (classifier engine + contracts) |
MPL-2.0 |
packages/minspec, packages/extension-pack |
MIT |
docs / site copy / whitepaper (LICENSE-CONTENT) |
CC-BY-4.0 |
Publisher: aiclarity.