[Feature Request] Add per-subagent effort/thinking control to align with Agent SDK

Bug Description
Feature request: per-subagent effort/thinking control — align Claude Code with the Agent SDK

We build a multi-agent plugin (dev-squad) on the Dynamic Workflow tool. It routes ~16 agent roles
per run to different models (haiku/sonnet/opus/fable via agent({model})), but there is NO way to
control reasoning effort per agent or per stage:

1. Workflow agent() silently ignores everything except model/schema/phase/label/isolation/agentType.

We probed this empirically: passing effort:'banana' (an invalid enum) completes with no error,
proving effort, thinking, budget_tokens, and output_config are dropped before reaching
the API. Silent acceptance of unknown opts also hides the limitation — an error would be kinder.

1. Subagent frontmatter only supports model: inherit|sonnet|opus|haiku — no effort field either

(same gap as github.com/anthropics/claude-code/issues/25669).

1. The session-level /effort setting propagates to all subagents but lands non-uniformly because it

is model-gated (haiku: none; sonnet: xhigh→high silent fallback; opus/fable: full range). So one
session knob cannot express "deep critic, shallow mechanical steps" — the model tier becomes the
only effort lever, which conflates capability, context window, and cost.

The Agent SDK already exposes thinking control when launching agents (options.thinking / adaptive
thinking + output_config.effort on the Messages API). Please align Claude Code's subagent surfaces
with that:

• accept effort (low/medium/high/xhigh/max) in Workflow agent() opts and in the Agent tool /

subagent frontmatter, with the same documented per-model silent-fallback semantics as /effort;

• or at minimum, error on unknown agent() opts instead of swallowing them.

Concrete impact: in a single /squad run our fable-tier critic (commit gate) and our haiku-tier
mechanical steps are forced to share the session's effort. Running the session at low to save cost
degrades exactly the high-leverage roles; running at high overpays on roles that don't need it.
Per-agent effort would let orchestrators spend reasoning where the leverage is — the same argument
that justified per-agent model selection, which already exists.

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.173
• Feedback ID: 5021e3a5-c07b-4800-b4bf-ba3eee8f6185

Errors

[]