Kybernetes loop architect
Runtime-adaptive steering layer and portable control contract for hard agentic work: durable goals, proportional control, bounded delegation, verification, and recovery.
npx -y skills add pariyar07/kybernetes --skill kybernetes-loop-architectAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 12 stars12 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Design a bounded execution contract using dynamically generated independent lenses when decision consequence and unresolved uncertainty justify architecture work.
SKILL.md
4.9 KB, as published. Nobody here has run it
Kybernetes Loop Architect
This is an advisory helper, not a governor. It designs or reframes loop shape and returns a proposal. It cannot mutate parent canonical lifecycle state, accept completion, activate detached machinery, or spend beyond its boundary.
Invoke When
Invoke when one or more apply:
- decision consequence is high and unresolved uncertainty could change action;
- high/extreme work needs activation, topology, feedback, or continuity design;
- a major reframe changes objective, evidence, side effects, or ownership;
- runtime capability loss requires a new substrate or migration architecture.
Do not invoke for clear, reversible, immediately verifiable routine work. Task length alone is neither necessary nor sufficient.
Required Inputs
- Objective, explicit program kind, and current evidence gaps.
- For finite work, measurable DONE and its completion verifier; for continuing work, the health invariant, review horizon, and cycle verifier.
- Constraints, permissions, external effects, reversibility, and HITL boundary.
- Duration, cadence, activation source, continuity, concurrency, and budget.
- Available verifier types and observed runtime capabilities.
- Parent control-record pointer and canonical revision when governed.
If a required input is absent, return the smallest question or explicit assumption that changes the architecture. Do not invent authority.
Generate Lenses
Generate lenses dynamically from material uncertainties in this task. Each lens must have a distinct question, competence needed, evidence to inspect, likely blind spot, and rejection criterion. Use the minimum set that exposes important tradeoffs. Do not select from a fixed personality or profession roster.
Run lenses independently before synthesis. Record for each:
- claim and recommended control move;
- evidence and source quality;
- assumptions and confidence;
- risks or conditions that would reverse it.
The synthesis must preserve disagreement rather than averaging it away. Lens output is comparator augmentation, never verification.
Design The Execution Contract
Return:
- activation posture: immediate, human, time, or event;
- loop topology: foreground, chain, route, parallel section, vote, orchestrator-workers, evaluator-optimizer, or bounded autonomous loop;
- feedback architecture: verifier, rejection path, iteration/no-progress cap;
- continuity architecture: foreground, durable, checkpointed, recurring, detached, migrated, or fresh-context;
- runtime needs as capabilities, not product names;
- canonical state, evidence, notification, permissions, and stop boundaries;
- selected modules and portable fallback;
- key alternatives, dissent, assumptions, and residual uncertainty.
For extreme detached work, the proposal must include program kind, strategy_id,
progress model and metric, measurement window, minimum delta, actionable capacity,
fallback coverage, cumulative no-progress cap, strategy envelope, escalation owner,
retirement, and either finite DONE verification or the continuing review horizon
and cycle verifier. The governor may produce this compact contract inline without
invoking this helper.
Prefer the simplest pattern that demonstrably absorbs task variety.
Temporary Review And Promotion
Expert lenses are temporary, read-only reviews by default. Apply evidence-gated promotion only when a lens needs recurring evidence gathering, a distinct objective, or ongoing ownership. A promoted lens becomes a dynamically named child workstream based on its objective, with owner, scope, sensor, boundary, budget, canonical revision, and return contract.
Promotion is a governor decision. This helper may recommend it but cannot create canonical child state or activate runtime machinery.
Standalone Use
When no governor is active, produce a standalone architecture proposal with
assumptions, evidence gaps, risk boundaries, and a portable next-step contract.
Do not claim it is active, scheduled, verified, or canonical. Recommend invoking
kybernetes:loop-governor for execution when durable control is needed.
Return To Governor
Return a concise packet:
architecture_status: proposed | needs_input | unsafe
canonical_revision: <revision or none>
trigger: <consequence and uncertainty>
selected_pattern: <smallest sufficient loop shape>
lenses: <independent findings and dissent>
execution_contract: <activation, topology, feedback, continuity, capabilities>
verification_needed: <rejection-capable sensor>
promotion_candidates: <none or evidence-backed objectives>
risks_and_assumptions: <material items>
The governor validates revision, selects or rejects the proposal, records any decision, and retains lifecycle and completion authority.