billzhuang/Heimdall Kubernetes SRE

AI-powered, read-only Kubernetes SRE agent — diagnose cluster issues in CI/CD pipelines

Trust Signals

name	description	required	default
mode	Run mode: prompt — one-shot diagnosis (requires `prompt`) triage — whole-cluster health sweep schedule-once — run the configured triage schedule once and exit	no	prompt
prompt	Diagnostic question to ask Heimdall (required for prompt mode)	no	""
namespace	Kubernetes namespace to scope (triage mode only)	no	""
all-namespaces	Triage all namespaces (triage mode only; overrides namespace)	no	false
kubeconfig	Base64-encoded kubeconfig. Decoded and written to ~/.kube/config. Use secrets.KUBECONFIG_B64 to keep it out of the log.	no	""
anthropic-api-key	Anthropic API key. Falls back to the ANTHROPIC_API_KEY environment variable already present in the runner (e.g. from a repo/org secret).	no	""
model	LLM model override in provider/model format. Examples: anthropic/claude-opus-4-8, anthropic/claude-haiku-4-5-20251001 Defaults to the Heimdall built-in default (anthropic/claude-sonnet-4-6).	no	""
config	YAML content for heimdall.config.yaml. Written to a temporary file and passed to Heimdall via HEIMDALL_CONFIG. Use this to enable optional tools (Prometheus, Loki, etc.) or set a namespace lockdown.	no	""
fail-on-severity	Fail the action step when Heimdall finds issues at or above this level. Accepted values: critical \| warning \| info \| ok \| '' (never fail). Example: fail-on-severity: critical	no	""
post-summary	Write findings to the GitHub job summary (GITHUB_STEP_SUMMARY)	no	true

name	description
severity	Detected severity level: critical \| warning \| info \| ok
summary	High-level summary from the agent (prompt mode only)
answer	Full Markdown answer from the agent (prompt mode only)
remediation-steps	Newline-separated remediation steps (prompt mode only)
suggested-commands	Newline-separated suggested kubectl commands (prompt mode only)
summary-markdown	Rendered GitHub Markdown job summary (all modes)