claude-code parser: Task tool renamed to Agent — subagent calls fall through as "unknown"

[JesseVent]

Summary

Two things found while testing claude-code subagent behavior on real evals (funded Anthropic key, claude-code-haiku and claude-code-sonnet-5):

1. Parser bug. packages/core/src/agents/claude-code/parser.ts:31-55 maps the subagent tool as Task, but the pinned CLI (2.1.191, see runner.ts) actually calls it Agent. Every subagent call falls through and gets tagged "unknown" instead of agent_task. Confirmed across 3 separate runs. Doesn't affect eval pass/fail scoring (checked all 40 EVAL.ts scorers — none key off the normalized name field), but corrupts any tool-usage stats/dashboards built from these transcripts, including the AgentPrism trace viewer (pending in a separate, not-yet-merged PR).

Fix: add Agent: 'agent_task' to CLAUDE_CODE_TOOLS.tools.

2. Caveat, not a bug. Subagents don't inherit the skill catalog — documented Anthropic behavior, not an oversight. Confirmed it's inconsistent in practice: a Haiku subagent that never checked for skills wrote a deprecated RLS policy (auth.role(), matched against an invalid value — blocks everyone) and a view missing security_invoker (bypasses RLS). A Sonnet subagent, given the identical task, checked for skills on its own and got both right. Worth knowing when interpreting no-skills vs benchmark comparisons — not something to fix, since it's intended Claude Code behavior.

Repro: forced subagent delegation on a fixture eval targeting the exact traps the supabase skill's checklist names (BOLA/IDOR, views-bypassing-RLS), compared the subagent's raw SQL output with/without skill access.