Per-agent overrides for compaction settings #759

Today:

• execute_threshold_percentage, execute_threshold_tokens and cache_ttl can be set per model (packages/plugin/src/config/schema/magic-context.ts:1218-1275).
• profiles may only hold historian/dreamer model, fallback and variant fields (:1192-1193).
• Subagent sessions run in a reduced mode (packages/plugin/src/hooks/magic-context/transform.ts:1167).

No setting is keyed by agent name, so every session on the same model gets the same compaction policy.

Use case: children dispatched for one short task, such as a lookup or a test run, could run with a lower execute threshold or without reduction nudges. A long-lived child, such as an implementer or reviewer that gets resumed for several rounds, could keep the root's settings. Today these differ only by model.

Request: an optional per-agent override block, resolved from the session's agent name and taking precedence over the per-model value, for example:

"agents": {
  "explore": { "execute_threshold_percentage": 50 },
  "reviewer": { "execute_threshold_percentage": 70, "cache_ttl": "1h" }
}

It covers the same fields that already take per-model values. Whether it lives in a new block or as a profile field is your call.

Prior art: Claude Code 2.1.296 autoCompactWindow for each subagent.