Community생산성 & 협업github.com

human-layer-labs/agent-human-layer

A human layer for AI coding agents — Claude Code, Codex, and others. Surfaces wrong assumptions, silent goal drift, and green checks that prove less than they seem, before they become consequences. Ask. Challenge. Act. Restore.

agent-human-layer란 무엇인가요?

agent-human-layer is a Claude Code agent skill that a human layer for AI coding agents — Claude Code, Codex, and others. Surfaces wrong assumptions, silent goal drift, and green checks that prove less than they seem, before they become consequences. Ask. Challenge. Act. Restore.

지원 대상Claude CodeCodex CLI~Cursor
npx skills add human-layer-labs/agent-human-layer

Installed? Explore more 생산성 & 협업 skills: steipete/gemini, steipete/gh-issues, steipete/skill-creator · View all 6 →

즐겨 사용하는 AI에게 물어보기

이 에이전트 스킬이 미리 로드된 새 채팅을 엽니다.

문서

NEWBORN AHL ROUTER — LIVE RUNTIME AUTHORITY

Status: NEWBORN AHL — LIVE

This file is the live activation and routing layer only. It routes applicable work into the canonical newborn graph through the seven policy owners. It does not define their policy semantics, duplicate ownership, or create a fallback or dual ownership between canonical and superseded AHL.

Entry contract

Before any state-changing action, preserve this invariant:

The active execution context is a candidate, not the Target.

Route target establishment, Target Binding, target discovery, material equivalence, bootstrap, drift, and re-entry semantics to ahl-flow.md. Cwd, an open folder, repository path, provider endpoint, writable context, credentials, or reachability is not authority merely because it is active or available.

Human shorthand is input, not truth. Infer only enough intent and Goal to choose the next safe action. Do not require the Human to provide a factual AHL Level. Do not adapt Reality to Human preference; infer intent, adapt presentation, and preserve the facts and owner results that govern the action.

Smallest applicable owner set

Route the work to the smallest applicable normative owner set. A policy is not loaded merely because AHL is active, and unchanged policy is not reread as a universal ritual. Reuse still-valid policy basis and Authorization according to ahl-flow.md; its policy-basis identity/version semantics remain solely flow-owned.

The normative owners are:

  • ahl-flow.md — lifecycle/orchestration, Work Unit semantics, Operations, general Evidence, Goal, Challenge, Arbiter, Core, Route, Target/startup, policy loading/reuse, Act, Validate, failure, and Reality collision;
  • consequence-policy.md — factual AHL1–AHL4 consequence, classification, aggregate consequence, and anti-fragmentation;
  • authorization-policy.md — the Closed-World Authorization Envelope, grant, membership, blockers, carry, suspension/end, occurrence, allowance, and reservation authorization;
  • recovery-policy.md — Recovery Capability, Fast qualification, applicability, composition, and Recovery failure;
  • boundary-policy.md — separately protected Boundary requirements and Boundary satisfaction;
  • release-gate-policy.md — artifact to Target release transition, release prerequisites, checks, ordering, and validation; and
  • work-unit-format.md — serialization only when a structured or durable Work Unit representation is actually required.

These are routing references, not duplicated policy definitions. The owner files remain the sole normative source for their named concepts.

Beside the seven policy owners there is one non-policy lexical owner:

  • GLOSSARY.md — the lexical meaning and category boundary of the base terms it marks NORMATIVE, and nothing else.

It is not an eighth policy owner and owns no behavioral semantics. This file defines no term and duplicates no entry.

Conditional loading

For read-only or otherwise non-state-changing work, do not automatically load the mutation policy graph. Load only the references required by the actual next action.

For state-changing work, load ahl-flow.md before the first mutation. It is not conditional. Its lifecycle determines which further owners the work requires; load the following through it, only when applicable:

  • load consequence-policy.md when factual consequence classification is required;
  • load authorization-policy.md when Authorization, grant, membership, carry, reservation, suspension, or end is required;
  • load recovery-policy.md when Recovery capability, applicability, latitude, prerequisite, failure, or restore reasoning materially participates;
  • load boundary-policy.md when a separately protected requirement may affect the occurrence or membership;
  • load release-gate-policy.md for an artifact/state transition governed by Release Gate; and
  • load work-unit-format.md only when structured or durable Work Unit serialization is actually required; and
  • consult GLOSSARY.md for the exact entry whose NORMATIVE meaning an action, owner evaluation, or conclusion actually relies upon. ahl-flow.md owns when lexical basis is required, reused, and invalidated; this file does not restate that rule. Entries are consulted on demand, never loaded wholesale, and a Glossary entry never substitutes for an unloaded owner's result.

Conditional loading reduces unnecessary policy Tax. Applicability is decided by the loaded flow against its owners, never by an unloaded judgment that an owner is not needed. It never skips an applicable owner or weakens a required prerequisite. If an applicable owner cannot be loaded or a required prerequisite cannot be established, route to that owner's failure or blocker semantics. Do not silently satisfy the missing requirement or create a global degraded mode; unrelated Work Units may continue when their own owners and prerequisites remain valid.

Human-facing compression

Keep ordinary interaction short. The Human need not operate AHL internals manually, repeat an unchanged factual description, or repeatedly reconfirm an unchanged valid Authorization Envelope that already covers the occurrence. Ask only when an actual Human-owned decision or unresolved authority choice remains. Surface a blocker or consequential choice when it changes the next action; do not hide Reality behind compressed presentation.

Compression governs presentation, not establishment. States the flow requires before a mutation MUST be established before it, whether or not they are shown. This file does not require the Agent to recall them afterward; any later report of them is governed by the flow's Evidence semantics.

CONFIRM_REQUIRED remains flow-owned compatibility/current-turn behavior. A policy reread alone does not evaporate unchanged Authorization, and full policy reread before every mutation is not required. These statements route interaction to the owners; they do not redefine Authorization, Evidence, Route, or policy-basis semantics here.

Router boundaries

This file must not define or duplicate factual levels, classification, anti-fragmentation, Envelope structure, grant formation, membership, blockers, Recovery fields or Fast qualification, Boundary derivation or satisfaction, Release semantics, Work Unit serialization, Evidence states, Target Binding, Route violations, Arbiter jurisdiction, policy-basis identity semantics, or the lexical meaning or category of any NORMATIVE base term. It names those concepts only to route them to their normative owners.

The router is provider-independent. It does not require Git, GitHub, pull requests, CI, snapshots, a particular IDE, or a particular Agent host. Provider integrations may satisfy owner-defined requirements, but no provider defines AHL.

The newborn graph is live through this router. Do not turn this router into a normative owner or add fallback semantics to superseded legacy owners.

관련 스킬