xagent-subagents

name	xagent-subagents
description	Use xagent from Codex to launch cross-provider review and worker subagents, especially Claude review passes and Cursor/GPT worker comparisons.
metadata	{"managedBy":"sheaf-agents-installer","source":"projects/agents/global/skills/xagent-subagents"}

xagent Subagents

Use this skill when Codex needs an external review opinion, a cross-provider second pass, or a delegated worker through packaged xagent.

Prefer the xagent Codex plugin launcher instead of assuming a bare xagent binary exists on PATH. In this repository, the source launcher is:

plugins/xagent/scripts/xagent

When xagent is installed as a Codex plugin in another repository, resolve the launcher from the installed plugin package and run it from the active worktree root. The xagent executable code comes from the plugin assets; the child harness working directory belongs to the active repository. Packaged xagent logs default to /Users/joyo/Sheaf/data/xagent; set XAGENT_LOG_ROOT only when intentionally validating or isolating logs somewhere else.

Sandboxed agents can invoke xagent even when they cannot complete a real run. If the central log root cannot be written, xagent reports a structured log_root_unavailable error. If the selected child harness lacks auth, network, process, or cache permissions, xagent reports the harness-specific failure. Surface those failures instead of silently switching tools.

When xagent launches the Codex harness, it passes Codex's explicit --dangerously-bypass-approvals-and-sandbox flag so the xagent-spawned Codex child does not stop for command approvals or inherit a restrictive sandbox. Use this only through xagent; do not copy that flag into unrelated workflows.

If the packaged launcher or tool cannot be invoked, surface the broken agentic infrastructure. Do not rebuild xagent ad hoc from a guessed Sheaf checkout or silently fall back to a different agent path.

Review Routing

For review tasks, prefer a Claude-backed reviewer through xagent:

plugins/xagent/scripts/xagent run --harness claude_code --model opus --subagent "<review prompt>"

Pick the model by review depth:

• opus: strongest reviewer for subtle architecture, security, correctness, or release-risk reviews.
• sonnet: balanced default for ordinary code review.
• haiku: fast, small, inexpensive reviewer for narrow diffs, copy checks, or quick sanity passes.

Do not use stale dotted model names such as claude-opus-4.8; local Claude Code accepts aliases such as opus or full names such as claude-opus-4-8. If a different model alias is needed, verify it with local Claude Code help before retrying. Do not silently downgrade after a model rejection.

Write the review prompt with the scope and output shape:

• Name the files, diff, PR, or task being reviewed.
• Ask for findings first, ordered by severity.
• Ask for concrete file/line references when available.
• Ask the reviewer to call out uncertainty instead of filling gaps.

Worker Routing

Use Cursor through xagent when a competent worker pass is useful:

plugins/xagent/scripts/xagent run --harness cursor --model composer-2.5 --subagent "<worker prompt>"

Treat Composer 2.5 as a solid worker for straightforward implementation, cleanup, alternate drafts, or exploratory passes.

For the trickiest implementation tasks, prefer a GPT or Codex-backed worker agent instead of Composer. Use the strongest available GPT/Codex worker when the task needs deep reasoning, careful repository integration, or high-stakes correctness.

Failure Handling

If xagent, Claude Code, Cursor Agent, or a requested model cannot be launched as instructed, surface the failure. Do not silently switch tools or work around broken agentic infrastructure.