witt3rd/oh-my-hermes/omh-v18-loop-taxonomy

OMH v18 Loop Taxonomy

The loop taxonomy lives in Saturate and OMH maps to it. Every OMH skill is an instance of a Saturate loop kind. The kind determines iteration semantics, turn result types, terminal states, and spec schema.

Source of truth: ~/src/witt3rd/saturate/ARCHITECTURE.md (Loop Taxonomy section) OMH architecture: ~/src/witt3rd/oh-my-hermes/docs/v18/architecture.md

The Six Loop Kinds

class MetricOptimizationKind(LoopKind): pass  # hypothesis → measure → keep/revert
class ConsensusKind(LoopKind):          pass  # agents deliberate until agreement
class TaskExecutionKind(LoopKind):      pass  # tasks executed until plan verified
class InformationSeekingKind(LoopKind): pass  # search until gaps close
class ClarificationKind(LoopKind):
    HUMAN_GATED: ClassVar[bool] = True        # ONLY human complete() ends this
class SelectionKind(LoopKind):          pass  # spawn N candidates, keep top-k, repeat

Plus structural (non-loop) kinds: BatchKind (fan-out/synthesize), OnceKind (single execution).

OMH Skill → Loop Kind Mapping

ClarificationKind.HUMAN_GATED is structural

The scheduler NEVER marks a ClarificationKind task terminal. There is no should_terminate overload for it. Only an explicit human complete() call ends it. This is a type property, not configuration.

omh-deep-interview does NOT post to the work queue. It manages state via plain file I/O to .omh/state/interview--{id}.json. The sentinel (.omh/specs/*-spec.md with status: confirmed) is the output.

Per-Kind Turn Result Algebras (key patterns)

MetricOptimizationKind:

Accepted(hypothesis, previous_value, new_value)  # keep + advance baseline
Discarded(hypothesis, reason: Regressed|Crashed|Unchanged)  # revert
# crashed ≠ regressed — crashed does NOT update baseline stats

ConsensusKind:

RoundComplete(round, verdicts: Dict[Role, Verdict])
ConsensusReached(proposal, rounds)
# Terminal: ConsensusTerminalReached | Deadlocked | Blocked | Cancelled

TaskExecutionKind:

TaskPassed(task_id, learning)
TaskFailed(task_id, error_fingerprint)
AllTasksPassed(summary)
# Terminal: PlanComplete | PlanBlocked | Exhausted | Cancelled

InformationSeekingKind:

FindingsAdded(new_findings, remaining_gaps)
Sufficient(findings)
# Terminal: InformationSufficient | InformationExhausted | Cancelled

ClarificationKind (HUMAN_GATED):

CoverageUpdated(dimension, new_level)
HumanConfirmed(spec)  # only terminal — no automatic predicate

The Four-Operation Queue Interface

OMH v18 plugin exposes omh_queue with these operations:

post(mode, instance_id, kind, name, ...)    → task_id
claim(mode, instance_id)                    → ClaimResult{status: claimed|not_found|running}
release(mode, instance_id)                  → None  # CRITICAL: resets RUNNING→PENDING on clean exit
write_state(mode, instance_id, state)       → None  # updates last_heartbeat
cancel(mode, instance_id)                   → None  # sets cancel_requested=1
complete(mode, instance_id, terminal_state) → None  # marks terminal
status(mode, instance_id)                   → dict|None  # read-only

`release()` is critical for ralph/autopilot: Each driver iteration exits cleanly after write_state(). Without release(), the next claim() waits 300s for crash-recovery timeout. With release(), reclaim is immediate.

Three backends, detected at runtime:

1. OMH internal queue (always — zero config, embedded SQLite)
2. Kanban (if omh_backend: "kanban" in config)
3. Saturate (if SATURATE_URL set — v18.2+)

When this fires

Load when:

• Working on any OMH v18+ skill and need to know which loop kind applies
• Designing a new OMH skill and need to declare its loop kind
• Working on Saturate integration and need the OMH side of the interface
• Debugging queue behavior — especially claim() returning running vs not_found

⚒️