Spec writer

This agent holds the spec writer role for the Requirements and Design stages of the Agent Development Lifecycle.

It takes what somebody wants an agent to do, in their own words, and writes the agent's own files the package maker will take, the layout a package is built from (agent-descriptor.json, agent-runtime-requirements.json, plugin.json and a SKILL.md per skill): what the agent offers, what each job takes and produces, what it refuses and what it records. It writes no spec.json: the draft is its own working record, and the files are taken from what the builder makes of it. It never invents the three things only the person can decide (the name, the price and access to any system), and it hands back the questions it could not settle instead of guessing. The package maker's own check runs over every draft, so a draft that will not build is caught here, and no files are delivered for it.

What it needs

• Node 22 or later. The scripts use only node's own modules.
• A text model for the one judgment in the job, bound as the text-reasoning system: TEXT_API_KEY, and TEXT_MODEL to pick the model. The default is Claude on Google Vertex AI. A bare model name is a Gemini model.
• Python 3 and the package maker's build.py on the host to check each draft. It looks for /opt/fleet/agentbuilder-build.py; set BUILD_PY to point it at ../package-maker/build.py or your own builder. Without one the files still go out, written by scripts/files.mjs (the builder's own projection, held to it byte for byte by test/files.test.mjs), and the reply says the build check did not run.

Running it

This folder is a package source. Its offerings and steps are declared in agent-descriptor.json and agent-runtime-requirements.json, each skill's instructions are in skills/<slug>/SKILL.md, the scripts are in scripts/, and the standing rules the model works under are in rules.md and references/.

node --test test/*.test.mjs                                   # 118 tests
python3 ../package-maker/build.py . ./out                     # an Agent Package

Install the package on any host that runs Agent Packages, bind text-reasoning, and send it a request.

Replacing it

This is a reference implementation. Any agent that holds the spec writer role can take its place, and its output is checked the same way: files the package maker accepts, with the open questions named.