refine: enforce fresh subagent invocation in orchestration guide

[shinpr]

Summary

Closes a gap in subagents-orchestration-guide that let the orchestrator continue a completed subagent via the harness's SendMessage resume path, which silently re-entered stale context and broke the workflow's per-phase isolation.

The guide's Execution Method section now states that each subagent invocation is a fresh Agent tool call, with the rationale made explicit: every phase runs in an isolated context, and a SendMessage resume reuses the prior agent's context and breaks that isolation.

Why this wording

• Positive form first. The instruction leads with what to do (a fresh Agent call) rather than a prohibition, keeping it aligned with the project's prompt principles.
• Names the failure mode with a reason. SendMessage is named because the harness actively advertises it as the continuation mechanism; closing that door requires an explicit, reasoned counter rather than relying on inference.
• Context isolation is the rationale. Subagents exist for context separation (bias removal) and context-window conservation — resume negates both, so isolation is the precise justification.
• Minimal footprint. A single sentence, edited in place. The orchestration guide is long and read by every recipe-* workflow, so the addition is kept necessary-and-sufficient.

Changes

• skills/subagents-orchestration-guide/SKILL.md: revised Execution Method line (single source).
• Synced to dev-workflows, dev-workflows-frontend, dev-workflows-fullstack (dev-skills does not bundle this guide).
• Version bump 0.22.1 → 0.22.2 (package.json + marketplace.json + generated plugin.json).

Validation

• claude plugin validate . → passed
• sync:check → no drift
• pre-commit hooks (sync, validate, skills-index) → all passed

🤖 Generated with Claude Code