Hermes Efficiency Optimization Expert

What It Addresses

Agents often edit files without checkpoints, lose track of progress after session resume, and lack a clear budget for token spend. Repeated failures also need a path from logs back into rules. hermes-efficiency turns these controls into executable commands rather than relying on prompt-level self-discipline.

How It Works

The core entry point is dispatch.py, working around a few concrete artifacts:
- Task checkpoints: pending, pending-update, and pending-done maintain PENDING.md so mid-edit context is not lost.
- Token budgeting: optimizer.py handles message classification, token estimation, and dynamic budget checks.
- Snapshots and logs: snapshot_updater.py refreshes snapshots, while memory/ records process data for review.
- Repair and evolution: repair runs built-in rules, and evolve derives new rules from error logs.

A typical flow starts with check-session in a new session, writes PENDING and pre-reads target files before editing, runs pending-done after execution, and records misses with reflect.

Boundaries

It assumes a local layout with scripts/, PENDING.md, and memory/, and depends on a Python environment. The optimization decisions are driven by optimization_engine.py, but it does not replace code review, testing, or deployment validation.