Any issues using several claude-flows in parallel in git worktrees?

[hakin19]

Assuming other resource issues within the project are addressed (such as ports etc) via a make file, are there any issues in using claude-flow simultaneously in different worktree directories of the same project?

[roman-rr]

Parallel worktree usage has a specific problem: ruflo's state files are per-project-root, not per-worktree.

The MCP server writes to:

• .claude-flow/hive-mind/state.json — shared state
• .claude-flow/data/auto-memory-store.json — memory entries
• .swarm/hnsw.index — vector index
• .swarm/memory.db — SQLite memory

If multiple worktrees share the same project root, they'll fight over these files. If they're separate directories, each gets its own state (but no cross-worktree coordination).

However, the bigger issue: the agent/swarm/task tools don't actually execute anything. agent_spawn creates a Map entry, swarm_init writes config — no real coordination happens between parallel instances because there's no coordination to begin with.

The memory/HNSW tools do work and could theoretically be shared across worktrees via symlinks, but SQLite doesn't handle concurrent writes well.

Practical approach: Use Claude Code's native Agent tool with isolation: "worktree" — it creates real git worktrees with isolated copies and actual subprocess execution.

Full audit of what works vs stubs: Independent Audit