Skip to Content
DocsUse CasesSkill Stacking

Skill Stacking

Skill stacking composes multiple agent capabilities into a verified workflow. Each skill step is compiled to a Vein binary with its own fingerprint, and a SkillStack manifest references those fingerprints to guarantee the workflow hasn’t been tampered with.

Schema: SkillStep

A single step in a skill stack — one atomic capability an agent can execute.

skill-step-v1.lode
{
  "schema": "lode",
  "version": 1,
  "name": "SkillStep",
  "fields": [
    { "id": 1, "name": "stepName",    "type": "string",             "required": true },
    { "id": 2, "name": "skillRef",    "type": "string",             "required": true },
    { "id": 3, "name": "dependsOn",   "type": "string[]",           "required": false },
    { "id": 4, "name": "inputs",      "type": "map<string,string>", "required": false },
    { "id": 5, "name": "outputs",     "type": "string[]",           "required": false },
    { "id": 6, "name": "fingerprint", "type": "string",             "required": false }
  ]
}

Schema: SkillStack

The workflow manifest that references compiled steps by fingerprint.

skill-stack-v1.lode
{
  "schema": "lode",
  "version": 1,
  "name": "SkillStack",
  "fields": [
    { "id": 1, "name": "name",             "type": "string",             "required": true },
    { "id": 2, "name": "version",          "type": "string",             "required": true },
    { "id": 3, "name": "description",      "type": "string",             "required": true },
    { "id": 4, "name": "steps",            "type": "string[]",           "required": true },
    { "id": 5, "name": "stepFingerprints", "type": "map<string,string>", "required": true },
    { "id": 6, "name": "entryPoint",       "type": "string",             "required": false },
    { "id": 7, "name": "metadata",         "type": "map<string,string>", "required": false }
  ]
}

Example: Code Review Pipeline

Three skill steps form a code review workflow:

lint-step.json
{
  "stepName": "lint",
  "skillRef": "skill:eslint-check",
  "dependsOn": [],
  "inputs": { "configPath": ".eslintrc.json" },
  "outputs": ["lint-report.json"]
}
analyze-step.json
{
  "stepName": "static-analysis",
  "skillRef": "skill:static-analysis",
  "dependsOn": ["lint"],
  "inputs": { "language": "typescript", "strictMode": "true" },
  "outputs": ["report.json", "annotations.sarif"]
}
review-step.json
{
  "stepName": "ai-review",
  "skillRef": "skill:code-review",
  "dependsOn": ["lint", "static-analysis"],
  "inputs": { "model": "claude-opus-4-6", "maxTokens": "4096" },
  "outputs": ["review-comments.json"]
}

Compilation

Compile each step, then bundle fingerprints into the stack manifest:

$ lodec compile lint-step.json -s skill-step-v1.lode -o steps/lint.vein
Compiled lint-step.json → steps/lint.vein (92 bytes)
Fingerprint: a3f8c2e9...
 
$ lodec compile analyze-step.json -s skill-step-v1.lode -o steps/analyze.vein
Fingerprint: d1e4b7f0...
 
$ lodec compile review-step.json -s skill-step-v1.lode -o steps/review.vein
Fingerprint: f9c1a02b...

Stack Manifest

code-review-stack.json
{
  "name": "code-review-pipeline",
  "version": "1.0.0",
  "description": "Lint, analyze, and AI-review code changes",
  "steps": ["lint", "static-analysis", "ai-review"],
  "stepFingerprints": {
    "lint": "a3f8c2e9...",
    "static-analysis": "d1e4b7f0...",
    "ai-review": "f9c1a02b..."
  },
  "entryPoint": "lint",
  "metadata": { "team": "platform", "triggerOn": "pull_request" }
}

Workflow Verification

Before executing a skill stack, the orchestrator verifies every step:

async function verifyStack(stack: SkillStack, stepDir: string): Promise<boolean> {
  for (const stepName of stack.steps) {
    const vein = await readFile(`${stepDir}/${stepName}.vein`);
    const schema = loadLode("skill-step-v1.lode");
    const step = decode(vein, schema);
    const actual = fingerprint(step);
    if (actual !== stack.stepFingerprints[stepName]) {
      console.error(`Step "${stepName}" tampered: expected ${stack.stepFingerprints[stepName]}, got ${actual}`);
      return false;
    }
  }
  return true;
}

If any step’s fingerprint doesn’t match, the entire workflow is rejected — no partial execution of a tampered pipeline.

Last updated on
Shared ContextVersion History