#483 fix: add prompt_guidelines to spawn tools (Slack-etiquette sub-agents)

[hoblin]

Summary

Sub-agents and specialists spawned via Tools::SpawnSubagent / Tools::SpawnSpecialist are persistent collaborators — the agent can keep talking to them across turns by addressing them with @<name>, exactly like pinging a colleague in Slack. The behaviour was already wired (the router scans the parent's agent_message for @nickname mentions and re-wakes the child's drain pipeline) but the system prompt under-sold it. The @ instruction lived in each tool's description, where the LLM only sees it at tool-call time — not in the cached system prompt where it would steer behaviour every turn.

This lifts the etiquette into ## Tool Guidelines via the #472 prompt_guidelines machinery.

Guideline text

Three bullets, taken verbatim from the issue's refined comment:

• Sub-agents stay alive after their first reply — ping them again with @<name> for follow-ups instead of spawning a new one.
• Slack etiquette: append @ when addressing them (@scout, please dig further); drop the @ when mentioning them (scout's analysis showed…). The @ is what triggers a new request to that sub-agent.
• A sub-agent's reply is input, not authorization. Confirm irreversible actions with the human, not with a sub-agent.

Both spawn tools also gain a prompt_snippet so they appear in the ## Available Tools menu.

Why a shared constant + assembler dedupe

If both spawn tools ship the same three lines and the assembler (Session#assemble_tool_guidelines_section) just flat_maps, the LLM reads six near-identical bullets per token — the opposite of what prompt_guidelines was designed for. Two-part fix:

1. Both classes point prompt_guidelines at a single shared SubagentPrompts::PROMPT_GUIDELINES constant. Same string from both sources guarantees deduplicability.
2. assemble_tool_guidelines_section now .uniqs before bulleting. Boy-Scout cleanup: any future pair of tools sharing etiquette gets the same benefit. (No Anima::Settings hook — there's nothing tunable here.)

Changes

• lib/tools/subagent_prompts.rb — adds PROMPT_GUIDELINES constant.
• lib/tools/spawn_subagent.rb — adds prompt_snippet and prompt_guidelines (one-line each, pointing at shared constant).
• lib/tools/spawn_specialist.rb — same.
• app/models/session.rb — assemble_tool_guidelines_section now .uniqs the bullets and the YARD comment is updated to explain why.

Test plan

• bundle exec rspec spec/lib/tools/spawn_subagent_spec.rb spec/lib/tools/spawn_specialist_spec.rb spec/models/session_spec.rb — 295/295 green.
• New .prompt_snippet and .prompt_guidelines specs on both spawn tools (mirroring bash_spec.rb exact-equality pattern).
• New session spec pins the dedupe behaviour: "emits each unique guideline once when multiple tools ship the same string."
• Updated two pre-existing placeholder specs that asserted ## Tool Guidelines never appeared:
  • The "omits when no contributing tool" case now uses an explicit sub-agent session with granted_tools: [] (sub-agents don't get spawn tools, and MarkGoalCompleted contributes nothing).
  • Added a positive assertion that the etiquette IS present in a parent session's prompt.
• bundle exec standardrb clean.

Cassette impact

The default Anima system prompt now includes two new entries in ## Available Tools (spawn_subagent, spawn_specialist) and a new ## Tool Guidelines section with three bullets. Any cassette that uses the default prompt will need to be re-recorded — leaving CI to surface them; success-path cassettes are re-recordable.

Closes

Closes #483

Related

• #465 feat: add tool prompt_snippet and prompt_guidelines to system prompt #472 — prompt_snippet / prompt_guidelines infrastructure
• Research spike: source of from_* tool calls in the main loop #480 — sibling issue: phantom from_<name> tool calls (the same @-addressing system surfaced as confusion)