Problem

Cron/hook turns already run at the lower cron_effort (interactive sources keep the full agent.effort), and a cron job can already override its model in jobs.yaml. But cron_effort is a single global knob for all crons — trivial/mechanical jobs (health probes, pollers) and reasoning-heavy jobs (report drafting, triage) are forced to the same effort. You either keep it low (and the heavy jobs get starved) or high (and the cheap pollers burn tokens).

Fix

Add a per-job effort: field on CronJob, mirroring the existing model: override:

• CronJob.effort (+ from_dict / save_jobs serialization).
• CronService passes job.effort or None into run_cron / run_persistent_cron.
• effort_override is threaded through run → _run_inner → _get_or_create_client, where it becomes effort=effort_override or self._base_effort_for_source(...) in the SessionSpec (also applied on the two client-retry paths so it survives a mid-turn reconnect).

Properties:

• Opt-in / no behavior change — default "" falls back to the source-based cron_effort.
• Still clamped — the override lands in spec.effort, which the backend runs through _effective_effort(spec.effort, model), so a per-job value is model-cap clamped like the default. Backend-agnostic (single SessionSpec construction point).

Testing

Added TestCronJobEffort (default, from_dict, and a save_jobs→load_jobs round-trip). Full suite: 1551 passed, 2 skipped.