hermes-loop-master

Hermes Loop Master

Overview

Hermes Loop Master is a disciplined operating loop for coding with Hermes Agent. It is built for work where "looks done" is not good enough: production fixes, multi-step features, refactors, security-sensitive changes, and any task that may continue across sessions.

The skill changes the agent's behavior in five ways:

1. Spec first — write a compact execution contract before editing code.
2. One slice at a time — implement the smallest meaningful increment, not a bundle of nearby improvements.
3. Verify with evidence — run real commands and record what returned.
4. Review adversarially — inspect the diff for fake-done shortcuts before declaring success.
5. Leave a handoff — make the next session able to resume without guessing.

The loop is intentionally file-backed. Durable task files survive context compression, new sessions, model switches, and interrupted runs.

When to Use

Use this skill when the user asks Hermes to:

• build or modify a feature with more than two steps,
• fix a bug that needs reproduction and verification,
• work in a production or public repository,
• touch authentication, authorization, payment, data migrations, secrets, user data, or deployment paths,
• refactor code where scope creep is likely,
• continue work from a previous session,
• coordinate with subagents or external coding tools,
• prepare a PR or public release.

Use the Tiny Change Path instead of the full loop when the task is a one-file typo, formatting-only change, comment update, or trivial documentation fix. Even then, verify the exact file changed and report the diff.

Do not use this skill to bypass user approval, hide uncertainty, or perform destructive operations without confirmation.

Adaptive Modes

Classify once during the safety gate, then use the lightest mode that can still prove the work:

Mode	Trigger	Required loop	Efficiency target
Tiny	One-file typo, copy, comment, or mechanical low-risk edit	Read → patch → cheapest relevant check → diff summary. Full artifacts are optional unless the repository already uses them.	Target no more than 8 tool calls.
Standard	Normal bug fix, bounded feature, or refactor	Compact LOOP.md, RED→GREEN when behavior changes, targeted test, broader regression check, diff review, short handoff.	Target no more than 20 tool calls.
Critical	Security, auth, payment, migration, user data, production, or irreversible risk	Full artifacts, RED→GREEN evidence, positive/negative/preservation/failure coverage, Independent Oracle where available, broader verification, adversarial review, handoff.	No fixed cap; evidence completeness wins.