cellar CLAUDE.md

Claude Code Working Memory

This repository should be built around four layers:

• Sources / adapters
• CEL OSS contracts
• Cellar/Dilipod runtime
• Agents

The durable OSS value is the context/memory/brief/receipt data plane, not one planner and not the live runtime itself.

The product direction is CEL as the context and trust layer for AI-operated software. Agents can plan in many ways; Cellar should make their context, memory, briefs, actions, verifications, and receipts reliable.

Repo Direction

• Sources and adapters should be easy to build and extend.
• OSS CEL should own context snapshots, merge contracts, memory contracts, brief assembly, transport schemas, and receipt schemas.
• The commercial runtime should own live cortex operation, policy, monitoring, compliance, and hosted execution.
• Agents should be pluggable: LangGraph, Mastra, Codex, Claude Code, GPT, Gemini, Cursor, n8n, or future in-house runtimes.

What CEL Owns

• cel-context: fused context snapshot and merge contracts
• cel-memory / cel-memory-sqlite: durable memory contract and local backend
• cel-brief: per-turn LLM brief assembly / budgeting / brief receipts
• receipt, event, MCP, CLI, SDK, and N-API schemas
• the split between dispatch proof, model-input proof, and task completion proof

What The Commercial Runtime Owns

• live cortex operation, freshness, diffs, anomalies, source prioritization
• runtime capability reporting
• adapter lifecycle, dispatch, and policy enforcement in production sessions
• audit timelines, retention, alerting, compliance exports, and governance workflows

What CEL Does Not Need To Own Right Now

• one mandatory planner
• one mandatory orchestration runtime
• retry / branching / checkpoint policy as a repo-defining concern

Built-in planners and runners can exist, but they should be treated as clients, examples, or transitional implementations unless proven otherwise.

Boundary Rules

• Keep the agent boundary generic.
• Preserve stable context, action, result, receipt, and adapter contracts.
• Keep improving AX and the shared crates even when an app later gets an adapter.
• Prefer app-specific structured truth in adapters over forcing everything through generic UI perception.
• Do not make LangGraph, Mastra, or any single runtime the identity of the platform.
• Do not design evals so they only make sense for one agent backend.
• Treat intent -> dispatch -> observed effect -> evidence as the core trust loop.