Skip ctx_commands injection in subagent sessions

[movax01h]

Problem

The SessionStart hook injects the full <context_window_protection> block into every Claude Code session, including subagents spawned via the Agent tool.

The <ctx_commands> section (~400 tokens) contains instructions for user-facing commands (ctx doctor, ctx purge, ctx upgrade, ctx stats) that are useless in subagent contexts — users cannot type these commands in a subagent session.

This wastes ~400 tokens per subagent spawn. In workflows that use multiple agents (e.g., plan-mode with Explore + Plan agents), this adds up quickly.

Where

• hooks/routing-block.mjs, lines 64-78 — the <ctx_commands> section
• hooks/sessionstart.mjs, line 35 — additionalContext = ROUTING_BLOCK (always the full block, no conditional logic)

Suggested fix

1. Detect subagent context — either via an environment variable (e.g., CLAUDE_PARENT_SESSION_ID) or a field in the stdin JSON that Claude Code passes to hooks
2. When in a subagent session, omit the <ctx_commands> section from the routing block (or omit the entire routing block if MCP tools aren't available to subagents anyway)

A simple approach would be to split createRoutingBlock() into core guidance + user commands, and only include the user commands section for root sessions.

Environment

• context-mode version: 1.0.75
• Claude Code CLI

[murataslan1]

Validated the bug — ctx_commands (~407 tokens) is injected even when CLAUDE_PARENT_SESSION_ID is set:

CLAUDE_PARENT_SESSION_ID=parent-abc echo '...' | node hooks/sessionstart.mjs
→ ctx_commands present: True, token count approx: 407

The fix in routing-block.mjs is straightforward (add { includeCommands } option to createRoutingBlock()), but I want to verify the detection method first:

Question for @movax01h: Does Claude Code actually set CLAUDE_PARENT_SESSION_ID for subagent sessions? Or is there a different env var / stdin field we should check? I don't want to ship a detection that's a no-op.

If you can share the env vars available in a subagent SessionStart hook (e.g. env | grep CLAUDE from inside a subagent), that would nail the detection.

The routing-block.mjs refactor to support includeCommands is ready to go once we confirm the detection signal.