Documentation Preview
Concepts, workflows, and operating principles.
A developer reference explaining how IRISH manages project context, scopes missions, executes sandboxed specialists, and anchors validation receipts.
Core Concepts
The IRISH Object Model
01 / Projects
A project binds a local directory, a canonical git remote (such as a GitHub repository), project-specific validation commands (e.g. npm test, cargo test), and an append-only mission log. Projects persist across daemon restarts.
02 / Missions
A mission is a single, atomic unit of engineering work. It is initiated by natural language intent or an issue description, compiles into a structured mission graph, and tracks state from planning to sandboxed execution, validation, and approval.
03 / Specialists
Specialists are domain-constrained execution agents. Rather than an omnipotent assistant, specialists operate with bounded tool surfaces and explicit file masks. The Developer Specialist focuses strictly on issue-to-PR code modifications and test verification.
04 / Evidence Packages
Every mission artifact is recorded as cryptographically verifiable evidence: git diffs, command-line execution logs, duration metrics, and exit codes. A mission cannot claim success without passing evidence.
Questions & Answers
Frequently Asked Questions
Does IRISH send my private code to cloud databases?
No. The IRISH daemon runs locally on your workstation. It operates directly against your local filesystem and git working tree. Code is only pushed to your designated GitHub remote when you explicitly approve a pull request publication.
Can IRISH accidentally push broken code to my main branch?
No. The IRISH supervisor enforces two hard boundaries: first, machine validation (compiler, linter, tests) must exit with code 0; second, publication is restricted to pull requests and requires human owner sign-off. Direct pushes to protected branches are blocked by policy.
Which LLM providers are supported?
IRISH features a decoupled adapter architecture. In development builds, reasoning and code generation can be routed to Anthropic Claude 3.7 Sonnet, OpenAI GPT-4o, or local offline models via Ollama. You are never locked into a single AI provider.
How can I test the system today?
You can explore our interactive, read-only Mission Command Center directly in your browser. It renders a full, realistic mission lifecycle with diff reviews, terminal validation evidence, and approval gates.