fix: use instructions for system prompt and prompt for user message in subagents

[tlongwell-block]

Fixes incorrect handling of recipe.instructions and recipe.prompt in the subagent system. Previously, both fields were merged together and used for BOTH the system prompt AND the first user message, which was redundant and semantically incorrect.

The Bug

In subagent_tool.rs, the build_subrecipe() function merged prompt into instructions:

// Before: prompt was appended to instructions
combined.push_str(prompt);
recipe.instructions = Some(combined);

Then in subagent_handler.rs, this merged blob was used for both:

• The system prompt template (task_instructions)
• The user message (Message::user().with_text(...))

The Fix

• instructions → system prompt (agent personality/role/guidelines)
• prompt → user message (the actual task to execute)

This matches how the CLI handles recipes in extract_from_cli.rs.

Changes

File	Change
subagent_tool.rs	Remove merge logic; only append parent's extra instructions
subagent_handler.rs	Use instructions for system prompt, prompt for user message
subagent_system.md	Conditional rendering for empty instructions

Backward Compatibility

• Recipes with only instructions still work (used for both, as before)
• Recipes with only prompt still work
• Ad-hoc subagents (instructions parameter only) still work
• Augmented mode (subrecipe + extra instructions) still works

[tlongwell-block]

/goose give me a minimal review on this change

[github-actions]

Summary

This PR correctly separates instructions (system prompt) from prompt (user message) in the subagent system, aligning with how the CLI handles recipes. The fix eliminates redundant duplication and maintains backward compatibility.

🔴 Blocking Issues

None found.

🟡 Warnings

None found.

🟢 Suggestions

None.

✅ Highlights

• Clean separation of concerns: instructions → system prompt, prompt → user message matches CLI behavior (crates/goose-cli/src/recipes/extract_from_cli.rs)
• Backward compatibility preserved: Line 124 in subagent_handler.rs gracefully falls back to instructions when prompt is absent
• Defensive template rendering: The {% if task_instructions %} guard prevents empty sections in the system prompt
• Simplified logic: Removed unnecessary string concatenation in subagent_tool.rs, making the code more maintainable

---

Review generated by goose

[DOsinga]

I think goose missed this?