workforce

Shared AgentWorkforce primitives for persona-driven orchestration.

Core frame

A persona is the runtime source of truth:

• prompt (systemPrompt)
• model
• harness
• harness settings
• optional skills array of { id, source, description } entries for reusable capability guidance (e.g. prpm.dev packages)

Each persona supports service tiers:

• best
• best-value
• minimum

Tiering controls depth, latency budget, and model cost envelope — not the quality bar. All tiers should enforce the same correctness/safety standards; lower tiers should be more concise, not lower-quality.

A routing profile is policy-only. It does not carry runtime fields; it only selects which persona tier to use per intent and explains why.

Packages

• packages/workload-router — TypeScript SDK for typed persona + routing profile resolution.

Personas

• personas/frontend-implementer.json
• personas/code-reviewer.json
• personas/architecture-planner.json
• personas/requirements-analyst.json
• personas/debugger.json
• personas/security-reviewer.json
• personas/technical-writer.json
• personas/verifier.json
• personas/test-strategist.json
• personas/tdd-guard.json
• personas/flake-hunter.json
• personas/opencode-workflow-specialist.json
• personas/npm-provenance-publisher.json

Routing profiles

• packages/workload-router/routing-profiles/default.json
• packages/workload-router/routing-profiles/schema.json

TypeScript SDK usage

import { resolvePersona, materializeSkillsFor } from '@agentworkforce/workload-router';

const selection = resolvePersona('npm-provenance');
// selection -> { personaId, tier, runtime, skills, rationale }
// selection.runtime.harness -> opencode | codex | claude
// selection.runtime.model   -> concrete model
// selection.skills          -> [{ id, source, description }, ...]

// Turn the persona's declared skills into a harness-correct install plan.
const plan = materializeSkillsFor(selection);
for (const install of plan.installs) {
  // e.g. ['npx', '-y', 'prpm', 'install', 'prpm/npm-trusted-publishing', '--as', 'codex']
  spawnSync(install.installCommand[0], install.installCommand.slice(1), { stdio: 'inherit' });
}

OpenClaw integration pattern

1. Map user request to intent:
   • implement-frontend
   • review
   • architecture-plan
   • requirements-analysis
   • debugging
   • security-review
   • documentation
   • verification
   • test-strategy
   • tdd-enforcement
   • flake-investigation
   • opencode-workflow-correctness
   • npm-provenance
2. Resolve profile policy + persona runtime via resolvePersona(intent).
3. Spawn subagent with returned harness/model/settings/prompt.

See runnable mapping example:

• examples/openclaw-routing.ts

This keeps runtime configuration in personas, while routing policy stays explicit, typed, and auditable.

Skills on personas

A persona can declare a skills array of reusable capability packages (e.g. from prpm.dev):

"skills": [
  {
    "id": "prpm/npm-trusted-publishing",
    "source": "https://prpm.dev/packages/prpm/npm-trusted-publishing",
    "description": "OIDC-based npm publish without long-lived tokens"
  }
]

Persona JSON is harness-agnostic — it declares what skill is needed, not how to install it. The SDK's materializeSkills(skills, harness) / materializeSkillsFor(selection) helper turns the declaration into a concrete install plan, routing each skill to the right on-disk convention per harness:

Harness	Install flag	Skill directory
claude	prpm install --as claude	.claude/skills/
codex	prpm install --as codex	.agents/skills/
opencode	prpm install --as opencode	.agents/skills/

Each returned SkillInstall carries an argv-style installCommand, installedDir, and installedManifest path. The helper is pure — it never shells out or touches disk — so callers (relay workflows, OpenClaw spawners, ad-hoc scripts) decide how to execute it. Once installed, Claude Code auto-discovers skills from .claude/skills/; for other harnesses, read the manifest off disk and inject it into the agent's task body.

Eval framework (scaffold direction)

Next step is a benchmark harness to score persona/tier combinations on:

• quality (task pass rate)
• cost
• latency

Then publish a versioned “recommended tier map” so default routing is data-backed.

Quick start

corepack enable
pnpm install
pnpm run check

This runs minimal guardrails across the workspace:

• lint (currently TypeScript-only)
• typecheck (package + examples)
• test (Node test runner)