What happened
PR #2098 removed the explicit Agent tool invocation parameters from SKILL.md step 4. The old text specified model, subagent_type, run_in_background, and prompt as Agent tool arguments. The replacement text says only 'Spawn the subagents with their prompt argument composed from parts 1-5 above'. While the PR author explained that Claude Code already knows how to use the Agent tool, ralphbean questioned whether the new parameter names were verified, and waynesun09 flagged the replacement as underspecified (HIGH severity). The companion PR fullsend-ai/agents#114 is open but not yet merged — the agents repo still has the old format without tools, permissionMode, or background frontmatter fields.
What could go better
The SKILL.md now relies on implicit knowledge that Claude Code's Agent tool reads sub-agent .md files and maps frontmatter fields to tool parameters. This is correct but opaque — future maintainers cannot determine how dispatch works without consulting external Claude Code documentation. The ben-alkov comment clarified the mapping but this information is not in the SKILL.md itself. Additionally, the agents repo companion PR was created on merge day (31 days after the scaffold PR opened), creating a drift window where the two locations had inconsistent content.
Proposed change
In the agents repo copy of SKILL.md (via PR #114 or a follow-up): (1) add a brief note to step 4 explaining that sub-agent .md frontmatter fields (name, model, tools, permissionMode, background) are consumed by Claude Code's Agent tool when spawning sub-agents, and the orchestrator's only responsibility is composing the prompt argument; (2) consider adding a cross-reference to the Claude Code sub-agent docs URL for maintainer context.
Validation criteria
A human reviewer or the review agent reading the updated SKILL.md should be able to understand the sub-agent dispatch mechanism without needing to consult external Claude Code documentation or PR comment history.
Generated by retro agent from fullsend-ai/fullsend#2098
What happened
PR #2098 removed the explicit Agent tool invocation parameters from SKILL.md step 4. The old text specified
model,subagent_type,run_in_background, andpromptas Agent tool arguments. The replacement text says only 'Spawn the subagents with their prompt argument composed from parts 1-5 above'. While the PR author explained that Claude Code already knows how to use the Agent tool, ralphbean questioned whether the new parameter names were verified, and waynesun09 flagged the replacement as underspecified (HIGH severity). The companion PR fullsend-ai/agents#114 is open but not yet merged — the agents repo still has the old format withouttools,permissionMode, orbackgroundfrontmatter fields.What could go better
The SKILL.md now relies on implicit knowledge that Claude Code's Agent tool reads sub-agent .md files and maps frontmatter fields to tool parameters. This is correct but opaque — future maintainers cannot determine how dispatch works without consulting external Claude Code documentation. The ben-alkov comment clarified the mapping but this information is not in the SKILL.md itself. Additionally, the agents repo companion PR was created on merge day (31 days after the scaffold PR opened), creating a drift window where the two locations had inconsistent content.
Proposed change
In the agents repo copy of SKILL.md (via PR #114 or a follow-up): (1) add a brief note to step 4 explaining that sub-agent .md frontmatter fields (
name,model,tools,permissionMode,background) are consumed by Claude Code's Agent tool when spawning sub-agents, and the orchestrator's only responsibility is composing thepromptargument; (2) consider adding a cross-reference to the Claude Code sub-agent docs URL for maintainer context.Validation criteria
A human reviewer or the review agent reading the updated SKILL.md should be able to understand the sub-agent dispatch mechanism without needing to consult external Claude Code documentation or PR comment history.
Generated by retro agent from fullsend-ai/fullsend#2098