sk-ilya/Flaky Test Analysis

Analyze CI failures and classify flakes using AI (Claude Agent SDK). See docs/USER_GUIDE.md for AI transparency details.

Trust Signals

name	description	required	default
repo	Repository to analyze, e.g. my-org/my-repo	yes	—
branch	Only analyze runs from this branch. Single name, comma-separated list, or * for all	—	main
lookback-days	How far back to search for failed runs, in days	—	7
run-ids	Workflow run IDs to analyze (comma-separated). Can be used alone or with branch filter	—	""
workflow	Workflow file(s) to analyze (from .github/workflows/). Single name, comma-separated list, or * for all	yes	—
skip-jobs	Job names to ignore (comma-separated). Useful for aggregation jobs that do not contain test output	—	""
context	Free-text hints for the classifier, e.g. test framework details, known infra quirks, job naming conventions	—	""
model	Claude model to use: sonnet, opus, or haiku	—	sonnet
stale-timeout	Safety limit: if no run finishes for this many minutes, cancel remaining work and report what was completed	—	60
max-turns-classify	Safety limit: max LLM round-trips each classifier agent can make before stopping	—	60
max-turns-correlate	Safety limit: max LLM round-trips the correlator agent can make before stopping	—	80
anthropic_api_key	Anthropic API key (for Claude)	yes	—
github_token	GitHub token with actions:read permission on the target repo	—	${{ github.token }}