@agent-pattern-labs/iso-canon

@agent-pattern-labs/iso-canon

Deterministic canonicalization and entity matching for AI-agent workflows.

iso-canon answers: what stable key represents this URL, company, role, or company-role pair? It normalizes common workflow identifiers and compares them without model calls, MCP servers, or prompt-token-heavy duplicate rules.

It is:

• Deterministic: the same input and profile always produce the same key.
• Explainable: comparisons return a verdict, score, and reasons.
• Domain-neutral: JobForge examples are included, but the package only knows URL/text canonicalization profiles.

Install

npm install @agent-pattern-labs/iso-canon

CLI

iso-canon normalize url "https://www.example.com/jobs/123?utm_source=x"
iso-canon normalize company "OpenAI, Inc."
iso-canon key company-role --company "Anthropic, PBC" --role "Senior SWE - Remote US"
iso-canon compare company "OpenAI, Inc." "Open AI" --config examples/jobforge-canon.json --profile jobforge
iso-canon explain --config examples/jobforge-canon.json --profile jobforge

compare prints same, possible, or different. It exits 0 for all valid comparisons so callers can decide how strict to be.

JobForge-Style Example

iso-canon key company-role \
  --company "Anthropic, PBC" \
  --role "Senior SWE, AI Platform - Remote US" \
  --config examples/jobforge-canon.json \
  --profile jobforge

Output:

company-role:anthropic:senior-software-engineer-ai-platform

Config Shape

{
  "version": 1,
  "profiles": [
    {
      "name": "jobforge",
      "url": {
        "dropHash": true,
        "stripQueryParams": ["utm_*", "gh_src", "source"]
      },
      "company": {
        "aliases": {
          "open ai": "openai"
        },
        "suffixes": ["inc", "llc", "pbc"]
      },
      "role": {
        "aliases": {
          "swe": "software engineer"
        },
        "stopWords": ["remote", "us", "united states"]
      },
      "match": {
        "strong": 0.92,
        "possible": 0.78
      }
    }
  ]
}

Profiles extend built-in defaults, so you only need to declare domain-specific aliases, suffixes, stop words, and thresholds.

Library

import {
  canonicalizeCompanyRole,
  compareCanon,
  loadCanonConfig,
  resolveProfile,
} from "@agent-pattern-labs/iso-canon";

const config = loadCanonConfig(JSON.parse(configText));
const profile = resolveProfile(config, "jobforge");
const key = canonicalizeCompanyRole("OpenAI, Inc.", "Senior SWE - Remote US", profile).key;
const duplicate = compareCanon("company", "OpenAI, Inc.", "Open AI", profile);

Boundary

iso-canon does not decide which source of truth wins and does not mutate trackers, ledgers, indexes, or caches. Domain packages own source precedence and decide whether a possible match should block work, warn, or route to a human review step.