feat(cron): per-job reasoning_effort override in job definitions

[victor-kyriazakos]

Summary

Cron jobs can now pin their own reasoning (thinking) effort. A new optional job field reasoning_effort overrides both the global agent.reasoning_effort and per-model agent.reasoning_overrides for that job's runs. Heavy scheduled analyses can run at high while cheap recurring jobs run at minimal, without changing the install-wide default. Requested by an enterprise self-host deployment.

Jobs without the field behave exactly as before.

Design

The field validates against the canonical effort grammar (hermes_constants.parse_reasoning_effort: none|minimal|low|medium|high|xhigh|max|ultra) at the storage choke point, cron/jobs.py. Validation is spelling-only by intent: the job's model is not knowable at create time (unpinned jobs resolve at fire time, auth fallback can swap the model mid-run), and the provider transports already clamp or omit unsupported levels at request time for config-set effort. The job pins intent; the transport owns capability. Making the store model-capability-aware would duplicate the transports' live-verified capability tables and drift.

Resolution happens in cron/scheduler.py::_resolve_job_reasoning_config, called at the existing single reasoning_config production site in run_job (after the auth-fallback model swap, so a pinned effort also governs the fallback model). Precedence: job pin > agent.reasoning_overrides > agent.reasoning_effort. A stored value that no longer parses (hand-edited jobs.json) logs a warning and falls back to config resolution instead of failing the tick.

The cronjob model tool accepts the parameter on both create and update (empty string clears). It is agent-settable, unlike the model/provider pins: effort cannot redirect spend to a different model, it only tunes thinking within the model the job already resolves to. hermes cron create/edit --reasoning-effort covers the CLI. The field is not a drift-guard snapshot axis (#44585 guard unchanged): a pin is explicit, and absent-field jobs keep config-following semantics.

Inert with no_agent=True; the no-agent path short-circuits before model resolution.

Commits

• f6fa0f6 feat(cron): per-job reasoning_effort override in job definitions
• 2f13631 test(cron): scrub tracker references from reasoning-effort test docstring

Surfaces changed

• cron/jobs.py — _normalize_reasoning_effort, create/update wiring, docstring
• cron/scheduler.py — _resolve_job_reasoning_config, call-site swap in run_job
• tools/cronjob_tools.py — schema param, create + update branches, _format_job conditional key, registry handler
• hermes_cli/subcommands/cron.py, hermes_cli/cron.py — --reasoning-effort on create/edit
• website/docs/user-guide/features/cron.md, website/docs/reference/cli-commands.md
• tests/cron/test_cron_reasoning_effort.py — 32 tests

Verification

Unit (canonical runner, includes the full cron suite plus CLI/tool siblings):

=== Summary: 71 files, 948 tests passed, 0 failed, 1 skipped (100% complete) in 22.0s (8 workers) ===

Mutation checks (fixes committed first, one side mutated, restored via git checkout):

• Scheduler helper stubbed to plain resolve_reasoning_config passthrough → 4 precedence tests fail (test_job_effort_beats_global_and_per_model_override, test_job_none_disables_thinking_never_reenabled_by_config, test_garbage_in_store_warns_and_falls_back, test_job_effort_is_model_independent); restore → 32/32 green.
• Store validator stubbed to str(value) passthrough → 10 tests fail across store and tool lanes; restore → 32/32 green.

E2E against the real dev-checkout CLI with an isolated HERMES_HOME (7/7):

• hermes cron create --reasoning-effort high persists the pin in jobs.json
• edit XHIGH normalizes to xhigh; edit "" clears (stores null, same convention as workdir)
• --reasoning-effort turbo returns an actionable error naming the valid levels; nothing persisted
• a real cron.scheduler.tick() fired the pinned job and logged using per-job reasoning_effort 'minimal' at resolution time
• a direct cronjob() tool call stores the pin

Sibling-lane sweep: one reasoning_config production site in the scheduler feeds AIAgent for both the chat-completions and ACP/codex lanes; the dashboard (web_server.py) and blueprint catalog both funnel job writes through cron.jobs.create_job, so no writer bypasses validation.

Considered and not done

• A cron.reasoning_effort global config key: agent.reasoning_effort plus reasoning_overrides already govern the cron default; a third global would add precedence ambiguity with no consumer story. Possible follow-up if a fleet-level cron default is requested.
• Model-capability validation at the store (see Design above).
• Extending the effort level grammar; per-fallback-entry effort interplay changes; desktop/TUI cron editor surfaces.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #7382. This implements the same per-job reasoning_effort persistence, scheduler-precedence, tool, and CLI mechanism; the extra validation/fallback details are refinements of that same implementation.

[teknium1]

Salvaged and merged via #91244 — both your commits cherry-picked with authorship preserved, thanks @victor-kyriazakos! One change from the original: the reasoning_effort arg was removed from the model-facing cronjob tool schema (standing policy: models don't make model-configuration decisions — same pattern as the existing model/provider/base_url omission in the tool dispatch). The CLI (hermes cron create/edit --reasoning-effort), store validation, and scheduler precedence all landed as you built them.