Subagents

A Claude Code subagent starts with the prompt its parent wrote and nothing else. It has none of the recall /cs did, so it would edit files without knowing the fixes and risks the team has stored for them. The hooks treat it as its own agent.

What happens

1. When it starts, the subagent-start.sh hook (SubagentStart) adds a short note to its context: it is in a Cortex session, the parent's recall is not in its context, and it must run cortex_knowledge_search and cortex_memory_search on its own task before editing. It also gets the rules for finding code.
2. When it tries to edit, enforce-session.sh looks for recall markers filed under the subagent's own id (.cortex/.session-state/agents/<agent_id>/). The parent's markers do not count. Until the subagent has run both searches, its edits are refused.
3. When it reads or edits a file with linked notes, recall-notes.sh shows them to it, once per subagent. See file notes.

The session itself is shared: a subagent does not call cortex_session_start. If the checkout has no Cortex session, or the gates are declared off, the SubagentStart hook says nothing.

Writing a subagent prompt

The subagent searches on its own task, so say what the task is in terms a search can use: the error message, the feature, the files. A prompt like "fix the bug we discussed" gives it nothing to recall with.