eve-forge

name	eve-forge
category	orchestration
version	1.0.2
author	broomva
description	Forge a personalized eve agent for a business end-to-end — absorb the business's artifacts, author the eve agent/ dir, validate, deploy to Vercel (or a VPS), smoke-test against ground truth, register, and evolve. The deterministic core is three safety gates distilled from a driven benchmark (BRO-1677): deploy-safety (never ship auth:none), validate (eve info --json → 0 diagnostics + tools registered), and smoke (drive the deployed agent, assert vs a ground-truth example). USE WHEN building/onboarding an eve agent for a tenant, "forge an eve agent", "deploy an eve agent", "onboard <business> onto eve", or when the Claude-Code orchestrator/forge must turn absorption inputs into a running eve operator. NOT FOR benchmarking frameworks (that's a one-off), running the operator itself (the forge builds it; the operator runs cheap on eve), or non-eve agent frameworks.
tags	["eve","agent-forge","meta-agent","vercel","deployment","claude-agent-sdk","dogfood","agent-substrate"]
compounding	each forged tenant is a versioned agent/ dir + tenant-spec.json (authored-agents-as-data); re-onboard/evolve = a diff
provenance	["docs/reports/2026-07-04-life-vs-eve-benchmark.html","research/entities/concept/eve-agent-orchestrator.md","BRO-1677"]

eve-forge — turn a business into a deployed eve agent

The orchestrator/forge (a Claude Agent SDK program) produces a deployed, tenant-scoped eve agent from a business's absorption inputs. This skill encodes the real eve workflow + every trap learned dogfooding it, and gates the consequential steps so the forge cannot repeat the benchmark's mistakes.

Latent vs deterministic split:

• Latent (agent judgment, this SKILL.md): absorb (Word template + transcript
  • examples → tenant-spec.json) and author (write the eve agent/ files in the business's voice from the templates).
• Deterministic (scripts/, tested): preflight (Node ≥ 24), deploy-safety (never ship unlocked auth), validate (eve info clean), smoke (output vs ground truth). Precision work lives in code; the latent space invokes it.

The 8-stage pipeline

#	Stage	How	Gate
1	Absorb	read the template + 1–2 transcripts + 2–3 filled examples → write tenant-spec.json (see references/templates/tenant-spec.example.json) + a ground-truth truth.json (required substrings + case-scoped forbidden). Stage these OUTSIDE the tenant dir — eve init refuses a non-empty target ("has no package.json")	latent
2	Scaffold	python3 scripts/eve_forge.py preflight then nvm use 24 && npx eve@latest init <slug>, then move tenant-spec.json/truth.json in	preflight blocks if Node < 24 (the npx trap)
3	Author	fill from tenant-spec.json: copy references/templates/{fill_document,send_document}.ts → agent/tools/; write agent/instructions.md (business voice + "strip HTML comments from the output"); — remove → (never ). Do NOT hand-write ; the scaffold already ships