feat(agent): read-only subagent delegation for long tasks #3383

A GAIA agent cannot ask another agent to do anything. Everything a task requires — every file read, every search, every page fetched — has to happen in the one context window, and on a local model that window is 32K or 64K. Long investigative tasks therefore hit a hard ceiling: not because the agent cannot reason about them, but because it cannot hold them.

The shape that solves this is one writer with read-only helpers. The helper does the expensive reading in its own context and returns a conclusion; only the conclusion comes back to the parent. Delegation here is a context strategy first and a parallelism strategy second, which is why it matters more on a small local model than on a frontier one.

Write capability stays with the parent. One writer avoids the concurrent-edit problem entirely, and helpers do not need it to be useful.

Acceptance criteria

• The base agent can spawn a bounded sub-agent and receive a text result.
• Sub-agents are read-only — write and edit tools are not available to them, enforced at registration rather than by prompt instruction.
• The sub-agent runs in its own context; only its returned conclusion enters the parent's history.
• Depth is capped at one level and fan-out is capped at a stated number of concurrent children, so a runaway cannot spawn indefinitely.
• Sub-agent step budget is separate from the parent's, and exhausting it returns a partial conclusion rather than failing silently.
• Cancellation propagates from parent to children, leaving no orphans.