Route subagents on the spawn message when plaintext at PreToolUse - #266

What & why

Codex subagent smart routing was effectively inert: the routing hook read task_name (a path label like task_3) or fell back to the generic "Codex subagent task" string, giving the task_v1 router no task signal — so it always returned the default arm (gpt-5-6-sol).

The actual task content lives in tool_input.message, which is plaintext at PreToolUse time (Codex encrypts it at send-time, but the hook fires before that). This was confirmed empirically by capturing a real spawn payload:

"tool_input": {
    "fork_context": true,
    "message": "Explore this codebase specifically for parser-related code. Identify the parser implementation(s)..."
}

The fix

route_spawn_tool now reads message first (when it's a plaintext string), falling back to task_name / agent_name (weaker labels) then the generic default. The isinstance(str) filter means encrypted dicts (if they ever appear) skip cleanly — no regression.

task = next(
    (
        value
        for field in ("message", "task_name", "agent_name")
        if isinstance(value := tool_input.get(field), str) and value
    ),
    default_task_label,
)

How do you know it works

• New tests: test_spawn_routes_on_plaintext_message_when_present (verifies the router receives the real message text), test_spawn_falls_through_encrypted_message_to_task_name (verifies encrypted dicts fall through, no regression).
• Full suite: 1107 passed, lint clean.
• Empirically confirmed: a real Codex subagent spawn produced a 422-char plaintext message in the short length bucket (400–1200) — exactly where task_v1 can differentiate between arms, instead of the contentless fallback that always returned the default.

This is a follow-up to #251 (codex smart routing, merged) and #255 (claude smart routing, merged).

This pull request and its description were written by Isaac.