← ClaudeAtlas

caflisted

Bring native agent fork across agent boundaries — use when the user wants to switch agents and continue work, fork this session into another agent (Claude Code <-> Codex <-> DeepSeek Harness), escape rate limits, try a different approach, compare models, list sessions, or check cross-agent tool health. Only fork when the user explicitly asks; never fork proactively.
russeell/cross-agent-fork · ★ 1 · AI & Automation · score 67
Install: claude install-skill russeell/cross-agent-fork
# caf — native agent fork across boundaries `caf` brings an agent's native fork workflow across agent boundaries. It creates a new native session in another agent; the original session is never modified. ## When to use - "fork this session to Codex / Claude Code", "switch to X and continue", "rate-limited, switch tools" -> fork - "what sessions are there" -> list - fork failed or "check the environment" -> doctor ## Determine the source agent This skill is agent-neutral: it runs inside any of the supported agents. Always use the explicit source matching the agent that is running this skill — never an implicit source: | Agent running this skill | Source ref | |---|---| | Claude Code | `cc:last` | | Codex CLI | `codex:last` | | DeepSeek Harness | `dsh:last` | Every supported agent is a first-class fork source and target (`cc <-> codex <-> dsh`); pick the target with `--into` and let the user choose when they do not. ## Fork ```bash caf fork <source>:last --into <target> # e.g. caf fork cc:last --into codex caf fork <source>:last --at 12 --into <target> # fork through turn 12 caf list [claude|codex|dsh] [-s keyword] [--all] caf doctor ``` Bare `caf fork --into <target>` without a source is a terminal convenience (it picks the most recent session in the current directory, excluding the target agent). Inside an agent, always name the source explicitly. ## After a fork Show the two output lines verbatim in a fenced code block; do not summarize, translate, or