Spend reasoning where it counts. High effort on EXECUTION turns is what caused the spiral.

The global default reasoning effort is MEDIUM (lowered from high by apply-config-deadlines.sh). Medium is the right baseline: high effort on mechanical execution turns burned tokens and looped; low effort on planning produced shallow plans. Route effort to the work:

HIGH effort (think hard, deliberately)

• Planning / architecture / decomposition (workflow-plan-first, workflow-plan).
• Hard debugging: a non-obvious failure, a subtask flagged HARD by the difficulty signal, or the second attempt after a stuck-reset.
• Designing an interface or data model with downstream consequences.

LOW / minimal effort (≤3-4 sentences, then act — see workflow-deadline-discipline)

• Reads, searches, retrieval (search_code, retrieve_related, file reads).
• Mechanical edits: applying a planned diff, renaming, formatting, obvious fixes.
• Tool routing / running the verifier / making a checkpoint.

How to apply

• If Hermes exposes a per-request reasoning effort (per-call override), set HIGH on the turns above and LOW on mechanical turns. If it does not, apply it behaviorally: on a HIGH turn, reason carefully and lay out the plan; on a LOW turn, do NOT write a long reasoning block — state the one next action in a sentence or two and execute.
• Raise to HIGH only for planning and flagged-hard subtasks; never for routine execution. This both caps spirals (less unbounded thinking on execution) and concentrates the model's reasoning budget where it changes outcomes.
• Pair with the spiral check: if a HIGH-effort turn starts to loop, check_spiral and break out (workflow-stuck-detect-reset) — more thinking is not the fix once it's circular.