getsentry/Warden

Event-driven agent that analyzes GitHub PRs using Claude Code SDK

Trust Signals

name	description	required	default
anthropic-api-key	Anthropic API key (sk-ant-...) or OAuth token. Can also be set via ANTHROPIC_API_KEY or CLAUDE_CODE_OAUTH_TOKEN env vars.	no	—
github-token	GitHub token for API access (defaults to GITHUB_TOKEN)	no	${{ github.token }}
mode	Action mode: analyze, report, or legacy run. analyze/report are pull request workflow only. Defaults to run for compatibility.	no	run
findings-file	Path to structured JSON findings file to read in report mode.	no	—
base-config-path	Path to the org-wide base warden.toml file to load before the repo config (relative to repo root)	no	—
base-skill-root	Path to the checked-out org repo root for local shared skills used by the base config (relative to repo root)	no	—
config-path	Path to the repo-local warden.toml config file that extends the base config in the same run (relative to repo root)	no	warden.toml
fail-on	Minimum severity level to fail the action (off, critical, high, medium, low, info). Use "off" to never fail.	no	high
report-on	Minimum severity level to show annotations in code review (off, critical, high, medium, low, info). Use "off" to disable reporting.	no	medium
max-findings	Maximum number of findings to report (0 for unlimited)	no	50
request-changes	Whether to use REQUEST_CHANGES review event when findings exceed fail-on threshold	no	false
fail-check	Whether to fail the check run when findings exceed fail-on threshold	no	false
parallel	Maximum number of concurrent trigger executions	no	5

name	description
findings-count	Total number of findings
high-count	Number of high severity findings
summary	Summary of the analysis
findings-file	Path to structured JSON findings file (always written, use for upload to GCS/S3/etc.)