Reference

Cross-cutting reference material that both CLI users and library consumers need. Organized by topic rather than by audience — these pages describe core-agent’s configurable surfaces in the depth needed for non-trivial use.

Configuration + identity

• Configuration — every field of .agents/config.json, every CLI flag that doesn’t have a config-file equivalent.
• Providers — Gemini, Vertex, Anthropic, Anthropic-via-Vertex, mock; env vars, model IDs, gotchas per backend.
• Permissions — gate modes (ask/accept-edits/plan/yolo), allow/deny patterns, path scope, persistence.

Capabilities

• Built-in tools — the model-facing tool catalog: file, search, shell, network, planning, plus the optional lifecycle tools (mark_task_done, ask_user, schedule_next_turn).
• MCP servers — declarative third-party tool integration via MCP.
• Skills — Claude-compatible SKILL.md bundles.
• Context management — compaction, task-boundary checkpoints, agentic tool wrappers.

Runtime

• Sessions and event log — durable session storage, audit log, replay, live tail, crash-resume.
• Attach mode TUI — core-agent-tui remote operator client.
• Multi-session daemon — one daemon serving multiple concurrent sessions with per-caller auth, ACL, audit threading, and instruction overlays.
• Scion adapter — embedding core-agent under the Scion distributed-runtime layer.