Skip to main content
An agent needs the repository standard, a concrete move and feedback on the result. Iris Code supplies those deterministically through MCP, CLI and project rules, while the agent or human remains responsible for editing the code.

​
MCP

File scope is free. Folder and workspace scope require Iris Code Pro. Both native and SDK MCP servers enforce the current agent governance policy before planning or verification. Agent-supplied before-source content is refused.

​
CLI scanning

CLI planning is read-only. A later standalone invocation verifies against Git HEAD because the planning process’s memory has ended. Generated project rules describe the same loop, honour disabled IDs and work without MCP: apply proved fixes, review suggestions in order, preserve behaviour and exports, stop when the file passes, then verify and test.

​
Finish verification

The CLI flag adds opt-in finish verification alongside the package guard. In either editor, Hooks has a separate Enable control with the same multi-select agent choice; it can be configured or removed independently. The hook verifies only a baseline ID already found in the current agent transcript and matched to a live MCP server for the same workspace. Source snapshots remain in server memory; the local registry contains connection metadata and IDs. Missing, null or compacted transcripts leave the hook inactive, with no guessed baseline or HEAD fallback. Only regressed requests a correction turn. not-improved is a warning, never a block. The shared counter permits at most maxCorrections blocks for a baseline, then asks the agent to stop and show the verdict and diff. Host loop signals provide a second backstop. Cursor warnings use stderr because its stop JSON has no advisory field. The hook does not call a model or change push/build enforcement. Codex also needs hooks enabled in its local configuration and project-hook approval. Vendor transcript formats are best-effort interfaces; explicit verification remains available when automatic matching is inactive. Overview, fixes and suggestions, conventions, verification, configuration, agent integration, and language coverage.