@burtson-labs/stealth-core-runtime

Host-agnostic agent runtime that powers both the Bandit CLI and the VS Code extension.

Same tool-use loop, same skill resolution, same compaction logic — different hosts.

Install

pnpm add @burtson-labs/stealth-core-runtime @burtson-labs/agent-core

Quick start

stealth-core-runtime is the higher-level convenience layer on top of @burtson-labs/agent-core — it ships the provider implementations, the model capabilities + behavior profile catalogs, the task queue for subagents, and the system-prompt builders. Use this when you want CLI / extension parity without re-wiring every piece yourself.

import { createStealthRuntime, type StealthHostBindings } from "@burtson-labs/stealth-core-runtime";

const bindings: StealthHostBindings = {
  // Provider + model selection
  providerClient,                  // Ollama / OpenAI-compatible / Bandit cloud
  modelId: "gemma3:12b",

  // Host integration (filesystem, shell, telemetry, etc.)
  fsAdapter,
  shellAdapter,
  telemetry,
};

const runtime = createStealthRuntime(bindings);

const plan = await runtime.plan("Audit src/auth.ts for unhandled errors");
const result = await runtime.execute();

Model intelligence

import {
  getModelCapabilities,
  getModelBehaviorProfile
} from "@burtson-labs/stealth-core-runtime";

const caps = getModelCapabilities("gemma3:12b");
// → { contextWindow, supportsTools, supportsVision, tier, ... }

const behavior = getModelBehaviorProfile("gemma3:12b");
// → { preferredToolProtocol, textToolFallback, safeContextBudget, ... }

The catalog covers Bandit, Gemma 3 / 4, Qwen 2.5 / 3.6, Llama 3.x, GPT, Claude, and friends. Per-workspace overrides via parseModelBehaviorConfig + .bandit/model-profiles.json.

Typical pairings

| You need | Add | |---|---| | Memory loading / hooks / @-mentions | @burtson-labs/host-kit | | React UI for plan / chat / diff | @burtson-labs/agent-ui | | Sanitize raw model output for chat display | @burtson-labs/core-chat | | Run on Node host | @burtson-labs/agent-adapters-node | | Run in browser host | @burtson-labs/agent-adapters-web |

What's in the box

• Provider implementations — OllamaProvider, BanditEngineProvider (cloud), OpenAIProvider, plus shape adapters for OpenAI-compatible endpoints
• Model capabilities catalog — getModelCapabilities(modelId) returns context window, tool-calling support, vision support, tier, and runtime options for every known model (Bandit, Gemma 3/4, Qwen 2.5/3.6, Llama, GPT, Claude, etc.)
• Model behavior profiles — getModelBehaviorProfile(modelId) separates harness strategy from raw capability detection: preferred tool protocol, text-tool fallback, safe input/output budgets, compaction mode, prompting template, thinking default, parallel-tool limits, and known failure modes. parseModelBehaviorConfig / registerModelBehaviorConfig power workspace .bandit/model-profiles.json overrides in both hosts.
• Task queue — fair-scheduled subagent backgrounding so the parent turn doesn't block
• Rewrite generator — small-model-friendly streaming patch emitter, used when the model can't reliably produce apply_edit find/replace pairs
• System prompt builders — both CLI and extension variants, with capability-aware branches (vision-on, tool-calling-on, etc.)

Status

Stable. This is the most-tested package in the monorepo after @burtson-labs/agent-core. Breaking changes need a coordinated PR across both hosts.

Tests

pnpm --filter stealth-core-runtime test

The suite covers model capability/behavior resolution, provider behavior under each tool-calling mode, the task queue's fairness guarantees, and a small fleet of replay fixtures captured from real failure traces.

License

Apache License 2.0 — Copyright 2026 Burtson Labs.