A Tessl meta-plugin for orchestrating Claude sub-agents correctly: what inherits, what does not, how to pass skills explicitly, and how to validate the handoff.

What this plugin provides

Kind	Name	Purpose
Skill	`delegate-to-sub-agent`	Two delegation channels (Agent SDK and Task tool), inline examples, echo-validation protocol.
Rule	`sub-agent-delegation-rules`	Concise in-context reminder card.
Script	`scripts/echo_skills.py`	Skills-echo preamble builder and validator.
Doc	`docs/SUB_AGENT_ISSUES.md`	Extended list of documented sub-agent gotchas with GitHub issue links.

Why it exists

Claude sub-agents start with a fresh context. They inherit:

CLAUDE.md
Tool definitions (all, unless restricted)
MCP servers (all — no per-sub-agent scoping yet)

They do not inherit:

Parent conversation history
Parent system prompt
Accumulated state
Skills (unless listed explicitly)

That last one bites orchestrators who assume a sub-agent will see the same skills the parent can see. It won't. This plugin encodes the explicit-passing pattern and a validation handshake that fails loudly if the handoff is wrong.

Install

tessl install jbaruch/sub-agent-delegation

Or from this repo:

tessl install github:jbaruch/sub-agent-delegation

Usage (quick)

from scripts.echo_skills import build_echo_preamble, assert_echo_matches

expected = ["govee-h6056", "face-recognition-calibration"]
preamble = build_echo_preamble(expected)
prompt   = preamble + "\n\n" + actual_task_prompt

# launch sub-agent with `prompt` ...
assert_echo_matches(first_message_from_sub_agent, expected)

See skills/delegate-to-sub-agent/SKILL.md for the full playbook, rules/sub-agent-delegation-rules.md for the short version, and docs/SUB_AGENT_ISSUES.md for the extended gotcha list.

License

MIT — see LICENSE.

docs

evals

rules

scripts

skills

README.md

tile.json