@agent-pattern-labs/iso-context

@agent-pattern-labs/iso-context

Deterministic context bundles for AI-agent workflows.

Agents often burn tokens because the question "which files should this role load?" lives in prose. iso-context moves context selection into local JSON policy: resolve bundle inheritance, read the declared files, estimate tokens, check per-file and per-bundle budgets, and render a compact context pack without model calls.

It is local-only, dependency-free, and MCP-free. Use it for mode runbooks, reference files, project facts, or any other context set where "load only these files for this task" should be executable policy instead of repeated prompt instructions.

Install

npm install -D @agent-pattern-labs/iso-context

CLI

iso-context list --policy context.json
iso-context explain apply --policy context.json

iso-context plan apply \
  --policy context.json \
  --root /path/to/project

iso-context check apply \
  --policy context.json \
  --root /path/to/project \
  --budget 12000

iso-context render apply \
  --policy context.json \
  --root /path/to/project \
  --target markdown

Every command accepts --json for machine-readable output.

Policy Shape

{
  "defaults": {
    "tokenBudget": 9000,
    "charsPerToken": 4
  },
  "bundles": [
    {
      "name": "base",
      "description": "Always-loaded workflow contract.",
      "files": [
        { "path": "iso/instructions.md", "maxTokens": 3500 }
      ]
    },
    {
      "name": "apply",
      "extends": "base",
      "description": "Application form-fill context.",
      "tokenBudget": 12000,
      "files": [
        "modes/apply.md",
        "modes/reference-geometra.md",
        { "path": "modes/reference-portals.md", "required": false }
      ]
    }
  ]
}

Accepted top-level input can be { "bundles": [...] }, an array of bundles, or one bundle object.

Semantics

• extends supports one parent or an array of parents.
• Parent files and notes are inherited before child values.
• Re-declaring the same file path in a child overrides scalar file fields while preserving the original file order.
• required defaults to true; missing optional files do not fail checks.
• tokenBudget may be set globally or per bundle.
• maxTokens may be set per file.
• Token estimates are deterministic and local: ceil(characters / charsPerToken), with charsPerToken defaulting to 4.

Library API

import {
  loadContextPolicy,
  planContext,
  renderContextPlan,
} from "@agent-pattern-labs/iso-context";

const policy = loadContextPolicy(JSON.parse(rawPolicy));
const plan = planContext(policy, "apply", {
  root: process.cwd(),
  includeContent: true,
});

if (!plan.ok) process.exit(1);
console.log(renderContextPlan(plan, "markdown"));

Fit With The iso Stack

• iso-context defines which files should enter context for a task.
• isolint makes the prose inside those files safer for smaller models.
• iso-harness emits the harness files where those contexts are referenced.
• iso-route defines which model a role should use.
• iso-capabilities defines what a role may do.
• iso-contract defines artifact shape.
• iso-ledger records domain events about those artifacts.
• iso-guard audits whether the actual run obeyed policy.

For JobForge, context bundles can represent the difference between apply, scan, batch, and tracker mode loads without repeatedly telling every agent to remember the full context-loading matrix.