[BUG] No project-level way to lift the Task/Agent ask-first default for task categories (only per-agent description "PROACTIVELY")

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Summary

The Task/Agent tool carries a hard ask-first default ("do not spawn agents unless the user asks"). The one supported way to make delegation proactive — use PROACTIVELY / MUST BE USED in a custom agent's description field — only raises the default for that one named agent. There's no way to express "delegation is pre-authorized for these categories of work (read-heavy audits, mechanical rename/merge sweeps)" at the project level (CLAUDE.md / a SessionStart hook) such that the model treats them as pre-approved for the generic Task tool.

Result: a project that has deliberately designed a delegation-first workflow, but routes to the general-purpose agent rather than a bespoke named one, can't get proactive delegation. Its prose guidance is overridden by the tool-description default, and the human has to re-authorize interactively every session.

Distinct from existing issues

• Not #19077 / #60763 / #61993 (nested sub-agent spawning).
• Not #29423 / #8395 (config propagation into spawned sub-agents). This is about the parent's spawn-gating default, and whether project config can raise it for a category rather than for one named agent.

What Should Happen?

• Honor a standing pre-authorization in CLAUDE.md / hook output as sufficient to clear the ask-first default for the named categories.
• A settings knob: a delegation policy, or an allowlist of pre-authorized task categories / agent types.
• Extend the PROACTIVELY mechanism beyond a single agent's description to a project-level policy.

Then the tool description could read "ask first unless the project has granted standing authorization."

Error Messages/Logs

Steps to Reproduce

1. Project ships a SessionStart hook / CLAUDE.md line: "Read-heavy audits and mechanical rename/merge sweeps go to a sub-agent — pre-authorized."
2. Give the model a task squarely in that category (e.g. a 40+ file audit) without naming an agent.
3. Observed: inline execution; delegation not considered until the user explicitly says "delegate." The category guidance doesn't clear the tool's ask-first default.
4. Expected: the project's standing authorization for that category is honored.

Claude Model

Opus

Is this a regression?

No, this never worked

Last Working Version

_No response_

Claude Code Version

2.1.201

Platform

Anthropic API

Operating System

Other Linux

Terminal/Shell

Other

Additional Information

Environment

Claude Code CLI 2.1.201, Opus 4.8; project uses a SessionStart hook emitting stage-conditioned delegation guidance.