nthlayer-workers

Tier 2 of the NthLayer ecosystem. Consolidated runtime process containing five worker modules: observe, measure, correlate, respond, learn. Communicates with nthlayer-core (Tier 1) exclusively via HTTP API — never accesses the SQLite store directly.

pip install nthlayer-workers
nthlayer-workers serve --core-url http://localhost:8000

What it is

A single Python process running five cooperating modules behind a ModuleRunner. Each module has its own cycle interval, persists state to core's component_state for crash recovery, and submits verdicts and assessments to core via CoreAPIClient. Worker failure is degradation, not outage.

• Apache 2.0 licensed.
• No direct DB access. Everything goes through the core HTTP API.
• Single-process, single-instance per core in v1.5. Multi-instance HA is a v2 target.

The five modules

observe

Deterministic runtime infrastructure — SLO assessment, drift detection, topology discovery. No LLM. Three sub-modules registered separately with the runner:

Also ships a CLI-only deploy gate: nthlayer-workers gate --service X resolves the service tier, fetches the latest slo_status, runs check_deploy(), and submits a deploy_gate assessment. Exit codes: 0=APPROVED/WARNING, 1=eval error, 2=BLOCKED.

measure

LLM-aware AI decision quality evaluation. One-way safety ratchet. 60 s cycle.

• Per judgment SLO: queries Prometheus, submits judgment_slo_evaluation assessment.
• On HEALTHY → BREACH transition: emits quality_breach verdict with severity (low / high / critical).
• Severity-based deterministic governance reduces autonomy (fully_autonomous → autonomous → limited_autonomous → advisor → observer); emits autonomy_change verdict if reduced.
• Autonomy is a one-way ratchet — never restored without explicit human approval.
• Quality scoring of agent outputs uses Instructor-backed structured LLM calls (ModelEvaluator).

correlate

asyncio session-window correlation, topology drift, contract divergence. Three sub-modules:

Snapshots include a non-blocking 5 s NL summary generated via Instructor (structured_call). Failures are logged and counted, never raised.

respond

Multi-agent incident-response coordinator. 30 s cycle. Trigger ingestion is situation-shaped:

• Primary trigger: correlation_snapshot assessments (with cursor snapshot_after).
• Fallback: quality_breach verdicts older than fallback_threshold_seconds with no associated snapshot — used when correlate is degraded.

Pipeline: [TRIAGE] → [INVESTIGATION, COMMUNICATION] → [REMEDIATION] → [COMMUNICATION]. Each step is an Instructor-backed agent producing a verdict; AWAITING_APPROVAL is the worker-mode pause point (resumption is P3-E.3 work). Per-step asyncio.wait_for(step_timeout_seconds). The first incident verdict's parent_ids are the trigger verdict IDs — that's the cross-module lineage bridge.

State ({cursors, incidents}) persisted to component_state. Terminal incidents pruned after 24 h.

learn

LLM-powered retrospective + outcome resolution. Two sub-modules:

Calibration signals fire only on real resolution, never on expiry — absence is not a quality signal.

CLI

# Start the worker runtime (registers all modules)
nthlayer-workers serve \
  --core-url http://localhost:8000 \
  --instance-id worker-01 \
  --prometheus-url http://localhost:9090 \
  [--collect-interval 60] [--drift-interval 1800] [--topology-interval 86400] \
  [--correlate-interval 10] [--topology-drift-interval 3600] [--contract-interval 3600] \
  [--measure-interval 60] [--respond-interval 30] \
  [--outcome-interval 60] [--retrospective-interval 30] \
  [--expiry-threshold-days 7]

# Deploy gate (CLI-only in v1.5)
nthlayer-workers gate --service payment-api [--tier critical] [--commit-sha SHA] \
  [--core-url http://localhost:8000]

# Print version
nthlayer-workers -V

How it talks to core

Every module is a WorkerModule Protocol implementation: name, restore_state, process_cycle, get_state. ModuleRunner orchestrates:

1. On startup: restore_all_state() — each module loads its cursor / dedup / hysteresis from GET /component-state/{name}.
2. Each tick: any module whose cycle has elapsed runs process_cycle(). Output is submitted to core as verdicts (CloudEvents-wrapped) or assessments.
3. After each cycle: put_component_state(name, get_state()) — state survives restart.
4. Heartbeat emitted to core only when ≥1 module ran on this tick (avoids tight-loop heartbeat noise).
5. SIGTERM/SIGINT → graceful shutdown → final state persist + heartbeat.

Cycle failures are logged and counted, never raised. The runner keeps going.

Configuration

Worker-specific config lives under workers.{module}.* in the unified nthlayer.yaml consumed via nthlayer_common.config.Config. Examples:

workers:
  respond:
    cycle_interval_seconds: 30.0
    fallback_threshold_seconds: 60.0
    terminal_retention_seconds: 86400.0
    step_timeout_seconds: 90.0
  learn:
    expiry_threshold_days: 7

LLM models are resolved via NTHLAYER_MODEL (default anthropic/claude-sonnet-4-20250514).

NthLayer ecosystem

The v1.5 consolidation absorbed five previously-standalone repos (nthlayer-observe, nthlayer-measure, nthlayer-correlate, nthlayer-respond, nthlayer-learn) into this single process. Those repos are now archived with deprecation releases on PyPI pointing here.

Licence

Apache 2.0