atman-33 Hand off the current conversation and start the next agent in one go: write the handoff document, then dispatch.mjs launches a fresh claude that reads it and continues. The baton leaves your hand and the next runner is already moving.
You write the document (judgment). dispatch.mjs owns the deterministic rest — environment detection, the initial prompt, the A→B→C fallback, and shell-safe launching.
Step 1 — Write the handoff document
Write a handoff document summarising the current conversation so a fresh agent can continue the work. Save it to the OS temporary directory — not the workspace. Capture its absolute path.
- Include a "suggested skills" section listing skills the next agent should invoke.
- Do not duplicate content already captured in other artifacts (PRDs, plans, ADRs, issues, commits, diffs); reference them by path or URL.
- Redact sensitive information (API keys, passwords, PII).
- Treat the skill arguments as a description of what the next session will focus on, and tailor the document to it.
Done when: the document is saved, its absolute path is captured, secrets are redacted, and a "suggested skills" section is present.
Step 2 — Dispatch
Run the dispatcher using its absolute path — do NOT cd into the skill directory first. Always pass --cwd set to the current session working directory (the project root, not the skill directory):
node "<absolute path to dispatch.mjs>" --doc "<absolute doc path>" --instructions "<the user's supplementary instructions from the arguments>" --cwd "<current session working directory>"
The dispatcher script is located inside the installed skill directory. Construct its absolute path from the known plugin cache location, e.g.:
C:\Users\<user>\.claude\plugins\cache\atman-marketplace\productivity\<version>\skills\handoff-go\dispatch.mjs
Options: --cwd <dir> (required — set to the current working directory of this session), --layout pane|tab (default pane), --dry-run (print what would run without launching).
The script detects the mode, composes the initial prompt, and launches with fallback:
| Mode | Condition | Action |
|---|---|---|
| A — in-session | already inside a multiplexer (herdr HERDR_ENV, or zellij $ZELLIJ / $ZELLIJ_SESSION_NAME) |
Split a pane (or new tab) running claude |
| B — new window | not in a multiplexer, but its CLI + a terminal launcher exist | Open a new terminal window hosting the multiplexer + claude |
| C — manual | otherwise | Print a paste-ready claude command + the document path |
Done when: the script exits, reports the mode it used (A/B/C), and --cwd pointed to the correct project directory (not the skill's directory).
Step 3 — Report the result
Tell the user which mode fired. For mode C (or any fallback), surface the script's paste-ready claude command and the document path verbatim so they can launch the next agent themselves.
Done when: the launch result — or the manual command for mode C — has been relayed to the user.