Subagent Prompt Guardrails

Motivation

The Agent tool currently encourages broad parallel delegation and says that subagent output should generally be trusted. The built-in prompts also omit a few execution and verification expectations, while the Explore and fork prompts contain unsafe or contradictory guidance.

Design

• Tell the parent agent to delegate only bounded, independent work, keep immediate critical-path work local, avoid duplicate work, and give parallel code-writing agents disjoint write scopes.
• Require the parent to review claims and code changes before integrating or relaying a subagent result.
• Simplify the general-purpose prompt and add scope, preservation, verification, uncertainty, and structured reporting expectations.
• Narrow Explore's stateful tool surface by removing task, memory, and user question tools from its allowlist. Permit shell pipelines while continuing to prohibit writes in its prompt.
• Stop requiring fork agents to commit changes unless the directive explicitly asks for a commit.

Context inheritance and the default background-execution behavior are outside this change.

Verification

Focused unit tests assert the parent guidance, built-in prompt contents, Explore's tool allowlist, and fork reporting rule. The core package build and typecheck provide the broader compile-time check.