Give coding agents evidence before they call a change CI-ready.

Choose the command that matches the task

check: validate and measure

check analyzes the checkout and, when DiffCI can infer safe commands, executes the full and selected test paths sequentially. It reports paired runtime only when both commands pass and the checkout identity remains valid. Test commands may create generated files in the repository.

observe --no-send: analyze without test execution

Use observation mode when the agent should inspect selection evidence without executing repository scripts. It writes the report outside the checkout and sends nothing.

verify-workflow: inspect Action integration

This checks that the DiffCI observer remains separate, non-blocking, and outside the dependency chain of required jobs.

Install repository instructions

npx "@diffci.com/diffci@latest" init

init creates DiffCI guidance for supported agent instruction surfaces, including AGENTS.md, CLAUDE.md, Cursor rules, Copilot instructions, and diffci.config.json. Existing files are preserved unless replacement is explicitly requested.

Agent-specific setup:

• Codex — repository instructions, command interpretation, and PowerShell.
• Claude Code — CLAUDE.md workflow and guardrails.
• Cursor — project rules and optional MCP integration.
• GitHub Copilot — repository instructions and the observer Action.
• Grok — portable repository policy and CLI use.

Use DiffCI through MCP

Remote MCP clients can connect to the stateless, read-only guidance endpoint:

https://diffci.com/mcp

When tools must inspect the current checkout or run tests, configure the local stdio server:

{
  "mcpServers": {
    "diffci": {
      "command": "npx",
      "args": ["-p", "@diffci.com/diffci@latest", "diffci-mcp"],
      "cwd": "/path/to/repository"
    }
  }
}

Call diffci_check before a PR-ready answer. It has the same execution behavior as the CLI check command.

Interpret the result conservatively

• A selected command is evidence, not authority. Required project CI remains authoritative.
• REFUSED and ERROR are not passes. Run the repository's normal validation commands.
• Selection counts are not savings. Use paired execution before making a runtime claim.
• Fallback is a valid result. Lockfiles, workflows, root configuration, or incomplete dependency evidence can require the full suite.

For the measurement boundary, read the open evidence study and the affected-tests guide.