New flagship workflow

New

Loop!

OMH Loop lets Hermes Agent treat a high-level goal as a repeating system only after it decides whether the request is a direct task, a loopable project, a north-star ambition, external waiting, or too unclear to start. Loopable goals then clarify the arena, research the gap, plan the next move, queue a runtime tick, prepare handoff when allowed, evaluate feedback, and continue until evidence, authority, context, or external waiting says to stop.

Back to docs GitHub

Loop Engineering

In OMH, Loop Engineering is the Hermes-facing control plane for work whose right implementation cannot be known upfront. The point is no longer just prompting. The point is giving a clear goal and designing the system that decides what to find, how to process it, when to hand off work, when to wait, and when to stop.

In OMH, Loop Engineering is the operating surface for goals that need bounded discovery, action, verification, and revision.

The goal is not to write a longer prompt. The goal is to define a durable objective, then let the system keep finding gaps, choosing the next step, preparing handoffs, evaluating feedback, and deciding when to continue, wait, or stop.

In OMH, ./loop gives Hermes Agent a visible loop contract for skills, plugins, connectors, subagents, worktrees, and external memory without pretending unobserved work already happened.

Loop Engineering is not for goals that are merely large. It is for goals whose correct implementation cannot be known upfront but can be discovered through bounded cycles of definition, action, verification, and revision.

goal interview research plan tick handoff feedback again

The Loop Goal Dilemma

If the goal is too large, the loop becomes fantasy planning. If the goal is too small, the loop becomes overhead. OMH adds loopability_assessment/v1 so Hermes can preserve the ambition while converting it into a current loop goal that has a bounded arena, observable problem, next verification, and stop condition.

Task

Change the button color.

Route to a direct task or one-cycle delivery workflow. A persistent loop is unnecessary.

Project

Reduce install-to-first-value friction.

Start a loop because the next correct implementation must be discovered and verified.

Ambition

Make this a 100k-star OSS.

Keep this as the north star, then propose the first bounded loop goal before any tick advances.

When to use it

Open source ambition

Make this a 10k-star quality OSS.

Hermes treats the phrase as a north star, then proposes a first loop goal such as reducing first-run friction without pretending market response was observed.

Long refactor

Finish all major refactoring safely.

The loop records each risk, plan, handoff, feedback gate, and checkpoint so the work can continue across context limits.

Product development

Keep improving onboarding until it feels polished.

Hermes cycles discovery, plan, executor handoff, QA feedback, and user-facing status while preserving evidence boundaries.

What the tick means

loop_runtime/v1 is a prepared queue, not execution evidence. It can name a worktree plan, subagent plan, connector intent, or executor handoff, but observed work must be recorded separately.

loop_engineering/v1 shows task discovery, distribution, execution, verification, and next-task decision state.
Automation, worktree, skill, connector, and subagent blocks stay visible as separate states.
verification_plan names the expected signal, evidence required, failure action, and stop signal for the current queue item.
Inner-loop checks are cheap and frequent: syntax, compile or import, focused test, command smoke, and schema validation.
Outer-loop checks are slower and rarer: integration evidence, semantic review, adversarial verifier, release gate, or human review.
Workflow patterns such as fan-out, adversarial verification, tournament, and triage batch are orchestration metadata, not dispatch evidence.
failure_mode_summary warns about verification gaps, comprehension debt, and cognitive surrender before the loop keeps moving.
Cost policy keeps reads bounded, reuses stable schema scaffolds, and avoids extra verifier lanes unless evidence warrants them.
omh loop run-once --loop <id> can wake exactly one eligible tick for wrappers, cron, or future automation without running code, and its result distinguishes created_tick from pending_queue_exists.
Prepared queue item is not worktree creation.
Prepared subagent plan is not subagent dispatch; subagent results should return status, summary, evidence refs, and next actions.
Prepared connector plan is not connector I/O.
External outcomes stay waiting until observed.

Verification inside the loop

Verification is part of the loop design, not an afterthought. OMH keeps cheap inner-loop checks close to the current step and reserves outer-loop checks for higher-risk or more semantic judgment. If a check is missing or fails, Hermes should route back to research or plan instead of advancing the queue as if the work succeeded.

Small loop

Test as stop signal

Name the cheapest check that proves this step is done before the queue item is prepared.

Cycle shape

Plan -> execute -> verify

Keep one planned step, one handoff or execution lane, and one observed verification signal per cycle.

Failure modes

One task at a time

Warnings for verification gap, comprehension debt, and cognitive surrender keep the loop from looking safer than it is.

Start card to queue lifecycle

A chat wrapper can ask OMH for a loop_start_card/v1 before a loop artifact exists. After the user chooses a permission profile, Hermes can start the loop, tick it, show the prepared queue and loop engineering snapshot, render a handoff with verification intent for the actionable item, then close that item as observed or blocked.

1 Start card

Explain the goal boundary and ask for permission profile.

2 Runtime queue

Prepare the next worktree, subagent, connector, or handoff plan.

3 Handoff

Render the selected queue item as executor-ready instructions.

4 Observe or block

Record evidence refs, or stop the queue item with a reason.

How Hermes can present it

# founder-loop

user

./loop Make OMH a 10k-star quality Hermes-native project.

Hermes Agent BOT

workflow: loop phase: interview next: reframe north star

This is a north star; I will shape the first loop goal.

I will keep the ambition large, pick a bounded arena, name an observable problem, choose the next verification signal, and queue runtime steps only after the current loop goal is clear.

Observe only Handoff only Execute with gates

A loop tick is not execution, review, CI, merge, publication, or market response evidence.