vibecollab-agent

Minimal project-first agent runtime package for VibeCollab.

This package is intentionally small. It provides the official VibeCollab agent kernel, while project state and domain capabilities remain owned by other Unix-style packages.

MVP boundaries

vibecollab-agent owns:

• message and tool-call primitives

• a lightweight AgentRuntime loop (LLM -> tool -> LLM)

• package-level CLI entry points (vibecollab-agent, python -m vibecollab_agent) for isolated testing

• project context assembly from repository files

• a small ToolRegistry

• basic project tools (read_file, write_file, edit_file, list_files, grep, run_command)

• policy guard checks for write/edit operations through vibecollab-policy

• run checkpoints through vibecollab-harness

• the existing JSON file-change AgentExecutor

It should not own the product CLI, MCP server, IDE integration, dashboard, domain models, insight lifecycle, or vector search. Those remain separate packages and can be registered as tools later. The package-level CLI is intentionally thin and exists for standalone runtime smoke tests.

CLI smoke usage

vibecollab-agent tools
vibecollab-agent context --project-root .
vibecollab-agent run "Summarize this project" --project-root . --mock-response "No tools needed."
python -m vibecollab_agent tools
vibecollab package list
vibecollab package run agent tools

Use repeated --mock-response values to test tool loops without network access.

Dependency direction

Preferred direction:

vibecollab-agent CLI / vibecollab-cli / vibecollab-mcp / external agents
  -> vibecollab-agent
      -> vibecollab-policy / vibecollab-harness
      -> vibecollab-data / vibecollab-foundation (next project-context hardening step)
      -> optional domain / insight / search tool packs

The current MVP is deliberately dependency-light and keeps LLM HTTP dependencies lazy through llm_client.py.