Subagents can be given a relative effort of lower or higher

What

The Agent tool, which Claude uses to start a subagent (a helper agent that works on part of a task), can now take an effort of lower or higher instead of a fixed effort level. Effort is how much reasoning the model puts in. A relative value moves the subagent one step down or up from the current effort, and the result is passed into the agent's definition.

• The relative value is ignored for forks, remote agents and teammates.
• It is honoured only when a helper check passes. That check needs the subagent model setting to be inherit, needs CLAUDE_CODE_SUBAGENT_MODEL_FORCE to be unset, and needs no other model override to apply. It then follows the environment variable CLAUDE_CODE_SNAPPY_PELICAN if set, and otherwise the tengu_snappy_pelican flag.
• When the check fails, the relative effort is quietly removed from the request rather than rejected.
• When the check passes, the Agent tool description adds a hint: give each independent lookup effort: "lower", and use higher for pieces that need more reasoning.
• Separately, permission-prompt handling for teammate and agent spawns (canShowPermissionPrompts) moved to the resume path.

For tengu_snappy_pelican, the flag server returned off for this site's account and for an anonymous baseline, but no reading has been taken under this release. The stock Agent tool description captured under this release is 7,081 characters, unchanged from the previous release.

Why

When it applies, this lets Claude spend less reasoning on simple parallel lookups and more on hard parts. Without the check passing, a lower or higher effort is dropped without any warning.