Start here
A project, a machine and a phrase. Five minutes for the person at the machine, and nothing for 1ADK to be given access to.
While it runs, and after
What happens to each answer before and after it leaves the machine, and how to read the project once the result is in.
- Docs A connected analysis The analysis arrives as small units of work. The coding agent answers each one, 1adk-agent checks the answer against the contract and a privacy guard before anything is sent, and the server checks it again. How stopping, resuming and a commit mid-analysis are handled. What happens while my coding agent analyses the system?
- Docs Reading the result What each part of a project shows: the system map and its subsystems, the interface registry, evidence under every claim, the questions nobody has answered, how much of the system is known, the comparison between analyses, and the export. What am I looking at once an analysis is in?
The rules underneath
The protocol pages describe the same machinery as a specification: what an agent is asked, what it may not do, and how two analyses are compared.
- Protocol Scan job A scan job is a bounded, expiring task with explicit rules: read-only, nothing executed, no secrets, relative paths only, repository text treated as data rather than as instructions, and a network rule that differs between connecting an agent and working by hand. What is an agent actually allowed to do during a 1ADK analysis?
- Protocol Evidence Package One JSON file describing a system: entities, structure, relations, interfaces, data, claims, evidence, unknowns and coverage. What each section is for, and what the format deliberately has no field for. What is an Evidence Package?
- Protocol Snapshots and change Identity is the hard problem behind any tool that tracks architecture over time. How components are recognised across analyses, why coverage decides what may be reported as removed, and what a semantic difference is. How do you tell that two analyses describe the same component?