hermes - 💡(How to fix) Fix Feature: support per-cron reasoning effort overrides

Summary

Cron jobs should support a per-job reasoning / thinking-level override, rather than inheriting only the global agent.reasoning_effort from config.yaml.

Motivation

Different scheduled jobs have very different reasoning requirements:

• A lightweight hourly email scan may only need low or medium reasoning to classify recent inbox items.
• A morning briefing or travel-planning job may benefit from high reasoning because it synthesizes calendar, email, weather, and trip context.
• A simple reminder/check-in job may not need deep reasoning at all.

Today, LLM-driven cron jobs can pin provider and model, but reasoning effort is effectively global. That creates an awkward tradeoff: either run all cron jobs with high thinking and pay the latency/cost tax, or lower the global setting and degrade the jobs that actually need deeper synthesis.

Current behavior

The cron scheduler reads reasoning from global config:

# cron/scheduler.py
effort = str(_cfg.get("agent", {}).get("reasoning_effort", "")).strip()
reasoning_config = parse_reasoning_effort(effort)

Job definitions support fields like provider, model, base_url, enabled_toolsets, and workdir, but there does not appear to be a first-class per-job reasoning_effort override.

Proposed solution

Add a per-cron-job field such as:

reasoning_effort: low | medium | high | xhigh | minimal | none

Resolution order could be:

1. job.reasoning_effort, if set
2. global agent.reasoning_effort
3. provider/model default behavior

This should be supported by both CLI and API/tool creation/update flows, e.g.:

hermes cron create "0 * * * *" \
  --model gpt-5.5 \
  --provider openai-codex \
  --reasoning-effort medium

hermes cron edit <job_id> --reasoning-effort high

And in the cron tool schema/API:

{
  "model": {"provider": "openai-codex", "model": "gpt-5.5"},
  "reasoning_effort": "medium"
}

Acceptance criteria

• LLM-driven cron jobs can store a per-job reasoning effort value.
• Scheduler uses job.reasoning_effort before falling back to agent.reasoning_effort.
• hermes cron list / cron inspection displays the effective or configured reasoning effort.
• hermes cron create and hermes cron edit support setting/clearing the value.
• Tool/API-based cron creation and update support the same field.
• Existing jobs remain backward compatible and continue to inherit the global setting when unset.
• Script-only no_agent jobs ignore the field or clearly mark it as not applicable.

Notes

This mirrors the existing per-job model/provider override: scheduled tasks need both model selection and thinking-budget selection to be tunable at the job level.