Skip to content

CFORGE_DEV.md

Latest commit

 

History

History
75 lines (65 loc) · 3.27 KB

CFORGE_DEV.md

File metadata and controls

75 lines (65 loc) · 3.27 KB

CFORGE_DEV.md — Project Context

Identity

  • Repo: Cognitive-Forge-Systems/cognitive-forge
  • Binary: cforge
  • Purpose: Portable CLI reasoning engine — enforces structured thinking on top of LLMs

Stack

  • TypeScript, Node.js
  • Clean/Onion Architecture — Domain → Application → Infrastructure → CLI
  • LLM routing via OpenRouter (OpenRouterClient)
  • Default model: openai/gpt-4o

Architecture Rules

  • Domain never imports from Infrastructure
  • All use cases injected with LLMClient interface
  • TDD-first on all domain and application logic
  • No direct pushes to main — feature branches + PRs only
  • One GitHub issue per Claude Code session
  • Conventional commits
  • Auto-compress inputs >2000 words before LLM calls

Existing Modules

  • LLMClient — domain interface for all LLM calls
  • OpenRouterClient — infrastructure implementation (fetch, no extra libs)
  • StructuredAnalyzer — 7-field structured analysis (intent, constraints, assumptions, risks, options, decisionLogic, output)
  • IterationEngine — 3-round RCCF loop (Analyze → Critique → Refine)
  • ContextCompressor — compresses large inputs, auto-triggers >2000 words, outputs summary/architecture/coreModules/risks/openQuestions
  • HarshFeedbackEngine — Anti-Yes-Man Protocol (counterArguments, logicalGaps, riskExposure, egoBiases, verdict)
  • ChatSession — interactive REPL with session state, file injection, mode toggles, architecture context permanently embedded

Planned Modules (not yet built)

  • TasteCurationEngine — LLM output filter: clarity metric, density metric, non-generic filter, insight detection. Goal: high signal-to-noise ratio. NOT a content recommendation system.
  • ScoringEngine — multi-dimension prompt/output scoring
  • KnowledgeGarden — project-scoped decision and assumption storage
  • API/SDK layer — expose cforge use cases for programmatic consumption by Donna AI and other tools

CLI Commands

  • cforge analyze — structured 7-field analysis
  • cforge iterate — 3-round RCCF loop
  • cforge compress — context compression with ratio
  • cforge critique — harsh feedback (Anti-Yes-Man)
  • cforge chat — interactive session with context + flags

Test Coverage

  • 46 tests passing across 5 test suites
  • Domain and Application layers: full TDD coverage
  • Infrastructure: contract tests with mocked HTTP

Decision Log

  • v0.1.0: Clean/Onion Architecture — domain never imports infrastructure
  • v0.1.0: OpenRouter for LLM routing — single API, multi-provider
  • v0.1.0: maxTokens default 1500 — avoids 402 on free tier
  • v0.2.0: RCCF as dual framework — prompt structure + reasoning cycle
  • v0.3.0: Auto-compression at >2000 words — prevents context overflow
  • v0.4.0: Anti-flattery enforced in prompt — no vague agreement
  • v0.5.0: Architecture context permanently in chat system prompt — prevents Python problem (LLM recommending wrong stack)

Open Questions

  • TasteCurationEngine: should scoring be numeric (1-10) or categorical?
  • ScoringEngine: which dimensions matter most for cforge use cases?
  • API layer: REST or SDK-first? What does Donna AI need to consume?

Relationship to Other Projects

  • cforge-dev uses cforge as its reasoning engine
  • Donna AI will consume cforge via API/SDK layer (planned v1.0)