decompose()

decompose() calls an LLM with a structured tool (emit_plan) to turn a free-form goal string into a Contract[]. Each contract represents one tenant's work.

Signature

function decompose(
  prompt: string,
  ctx: {
    adapter: ProviderAdapter;
    model: string;
    budget?: Budget;
  }
): Promise<Result<Contract[]>>

Parameters

Parameter	Type	Description
prompt	string	The high-level goal to decompose
ctx.adapter	ProviderAdapter	Adapter for the LLM call
ctx.model	string	Model to use for decomposition
ctx.budget	Budget (optional)	Budget to consume for this call

Basic usage

import { decompose } from '@flint/landlord';
import { anthropicAdapter } from '@flint/adapter-anthropic';

const adapter = anthropicAdapter({ apiKey: process.env.ANTHROPIC_API_KEY! });

const result = await decompose(
  'Build a REST API with user authentication, CRUD operations, and API documentation',
  { adapter, model: 'claude-opus-4-7' }
);

if (result.ok) {
  for (const contract of result.value) {
    console.log(`${contract.role}: ${contract.objective}`);
    console.log(`  depends on: ${contract.dependsOn.join(', ') || 'nothing'}`);
  }
}
// → researcher: Research best practices for REST API auth
//     depends on: nothing
// → implementer: Implement the API with auth and CRUD
//     depends on: researcher
// → documenter: Write API documentation
//     depends on: implementer

How it works

decompose() calls call() with:

1. A system prompt that instructs the model to act as the Landlord orchestrator
2. The user's goal as the user message
3. A single tool emit_plan that forces the model to return structured JSON

The model calls emit_plan({ contracts: [...] }). Each contract in the array is validated against ContractSchema (Zod). Malformed contracts are silently dropped. If the model doesn't call emit_plan, decompose() returns { ok: false }.

Writing effective decompose prompts

The quality of decomposition depends heavily on the prompt. Tips:

Be specific about output format:

await decompose(
  'Build a REST API. Each tenant should produce files in its work directory. ' +
  'The final tenant should produce an index.ts entry point.',
  { adapter, model }
);

Specify the number of tenants:

await decompose(
  'Create a 3-step pipeline: (1) research, (2) write, (3) review. No more tenants.',
  { adapter, model }
);

Describe dependencies explicitly:

await decompose(
  'Build a data pipeline where ingestion must complete before transformation, ' +
  'and transformation must complete before the report is generated.',
  { adapter, model }
);

Inspecting the plan before running

Use decompose() directly to preview the plan without running tenants:

const plan = await decompose(myGoal, { adapter, model });
if (plan.ok) {
  console.log(JSON.stringify(plan.value, null, 2));
  // Review the contracts, then pass them to runTenant() or build orchestrate() manually
}

Error cases

Condition	Result
LLM doesn't call emit_plan	{ ok: false, error: Error('LLM did not call emit_plan') }
All contracts are malformed	{ ok: true, value: [] } (empty array)
LLM call fails (network, budget)	{ ok: false, error: AdapterError | BudgetExhausted }

See also

• orchestrate() — runs decompose + execution together
• Contracts — Contract field reference
• runTenant() — run a single contract directly