deepagent-code
About
deepagent-code targets work that spans more than one prompt: long-running tasks where the agent must remember decisions, constraints, and past failures across sessions. It builds on opencode and adds a control plane where project memory is stored as typed, versioned documents with provenance rather than prompt text, and where four graph stores (code symbols, knowledge, memory, documents) feed a shared context assembly with admission gates. During execution, users can steer live goals without aborting in-flight work, and subagents run in isolated worktrees under a generation-fenced lifecycle with review sessions; an Expert Panel mode runs bounded adversarial debate between specialist lenses. It is AGPL-3.0 with a separate enterprise distribution, BYO-model-keys across 75+ providers, and targets teams who need auditable, steerable agent behavior over codebases.