Skip to main content
The IronBee Action runs the same verification loop as the CLI, but unattended in CI. Instead of gating a developer’s task, it reviews what changed on a push or pull request, drives Claude Code through verification, and turns the result into a PR comment or a fix PR. This page explains what happens on each run.

The run, step by step

Every run is a single composite job that sets up the toolchain, verifies, and reports:
1

Install

Pins and installs @ironbee-ai/cli, @anthropic-ai/claude-code, @ironbee-ai/devtools, and Playwright Chromium (cached across runs).
2

Configure

Writes .ironbee/config.json with the collector URL and per-mode DevTools flags, then runs ironbee install --client claude to wire up hooks, skills, rules, and MCP config. Your API key is passed as an env var, never written to disk.
3

Prompt

Builds a verification prompt tailored to the trigger diff-based for pushes and PRs, full-application for manual and scheduled runs.
4

Verify

Runs Claude Code with /ironbee-verify, which exercises the change through each enabled DevTools mode and submits a verdict.
5

Report

Extracts the final verdict, uploads evidence, and posts a PR comment or opens a fix PR depending on the trigger.

The completion gate

Verification is enforced by the same completion gate the CLI installs, see How Verification Works for the full mechanism. In CI, the action instructs the agent to run /ironbee-verify, fix any issues it finds, and re-verify. The gate is what makes that meaningful: each cycle defines the tools that must appear on the wire before a pass counts, so the agent can’t declare success without actually exercising the change.

The verdict

To pass, the agent exercises the affected paths and submits a verdict: pass, fail, or unknown with the checks it ran, issues it found, and fixes it applied. The action surfaces the final verdict two ways: A run can go through several cycles, a failing verdict sends the agent back to fix and re-verify before the final result is recorded.

Verification scope

The prompt the action builds depends on what triggered it:
ScopeWhenWhat the agent verifies
Focusedpull_request, pushOnly the areas affected by the diff
Fullworkflow_dispatch, scheduleThe entire application, no diff required
See Running in CI for the full breakdown.

Fix behavior

When verification produces fixes, where they land depends on the trigger:
  • Pull request fixes are committed directly to the PR branch and the agent re-verifies.
  • Push, manual, scheduled fixes are moved to a new branch and opened as a fix PR; the original ref is left untouched.

Sessions and recording

Each run is recorded as a session to the IronBee collector and shows up in the Console, with the full timeline, recordings, and per-cycle verdicts. Locally, the same data is written under .ironbee/sessions/<id>/ and collected into the uploaded evidence artifact. The PR report links straight to the session and to each verification cycle in the Console.

What’s next?

Running in CI

How scope, fix behavior, and the PR report change per event.

Evidence

What the action records and uploads after a run.