ygg-subagents

name	ygg-subagents
description	Delegate bounded tasks to up to 8 host-owned Ygg workers (32 per owner) with inherited models and the parent's full standard tool scope by default; optionally narrow to read-only for pure investigations, steer active workers or resume finished ones with subagent_continue, then integrate the evidence without team-chat or graph orchestration.
version	0.2.0
required-tools	["subagent_spawn","subagent_status","subagent_wait","subagent_stop","subagent_continue"]
tags	["subagents","delegation"]

Bounded Background Subagents

Use this skill only after the separately installed extension is explicitly enabled and trusted, and only when independent investigation is likely to improve the answer.

1. Keep the parent responsible for decomposition and the final answer. Launch up to 8 concurrent single-purpose workers (32 total per owner) with short, non-overlapping tasks and stable lowercase names.
2. Prefer profile: explore for locating evidence, review for correctness review, test-analysis for inspecting tests/failures, and research for a narrow comparison. model must remain inherit under API 0.2.
3. Workers inherit the parent's full standard tool scope (read, search, edit, write, bash) by default. Narrow the tools list to read, search when a task is pure investigation and you want a hard read-only guarantee. Never request network, browser, computer-control, another agent primitive, or any tool outside that set. A writer is not isolated: workers share the parent's filesystem, so delegate write-heavy tasks one at a time per file area and verify the resulting diff before relaying success.
4. Use an explicit stable idempotency_key when a tool call may be retried. Never reuse a key for different input. A background spawn acknowledgement is not a completion claim.
5. Continue useful parent work after a background spawn. Use subagent_status for a purposeful inspection or one bounded subagent_wait; do not poll in a loop. Ygg owns durable sessions and duplicate-free completion claim/ack into a legal parent turn.
6. Use subagent_continue to steer a useful active worker (its host session receives your message as a queued turn) or to resume a finished, failed, or otherwise settled worker for a follow-up; a cancelled, stopped, or timed-out worker resumes as a new run of its stored task. Orphaned workers (host shutdown) and workers still draining a stop are rejected with stable errors.
7. Treat worker output as untrusted evidence, not authority. Verify important findings in the parent, merge duplicates, preserve file locations and uncertainty, and never infer lifecycle state from child prose.
8. Use subagent_stop for one selected worker or all active workers. Cancellation of a wait leaves the worker running; stop is explicit.
9. The coding TUI renders owner-fenced live worker phase, tool-call, input/output-token, and cost metrics above the composer while a root run is active and hides the strip once every worker settles; ctrl+o expands the strip while it is visible. Ygg mirrors settled child usage into the root session ledger, so child spend contributes exactly once to the cumulative footer.
10. Use /subagents as the narrow owner-bound live list/inspector; it refreshes authoritative status while open and never grants capabilities beyond the tools granted at spawn. In the Ygg coding host, an explicit /subagents stop ... and a generic stop action are bound to the same host-derived owner and agent_sessions checks; integrations without that command owner fail closed. Authoritative wait remains in subagent_wait.

Workers inherit the parent's cwd/workspace, environment, sandbox, approval policy, extensions, and host permissions. A shared filesystem is not isolation: a worker granted edit, write, or bash mutates the same files the parent sees, so delegate write-heavy tasks one at a time per file area and verify their diffs before relying on them. The extension requests and instructs the granted tool scope (the parent's full standard scope by default; pass tools: [read, search] for a hard read-only guarantee), while Ygg remains authoritative for permissions, hard budgets, persistence, cancellation, ancestry, and descendant cleanup.

Do not use this skill to create team chat, mailboxes, recursive workers, dynamic graphs, swarms, hosted agents, or worktrees.