madhan-karthikeyan/DecisionDrift Review

Detect violations of architecture decisions in PRs via deterministic rules and optional LLM semantic review.

Trust Signals

name	description	required	default
llm-api-key	API key for the LLM provider (OpenAI / Groq / OpenRouter, or "ollama" for local). Omit for deterministic enforce-only mode.	no	—
llm-model	LLM model name (e.g. gpt-4o, llama-3.3-70b-versatile, qwen2.5-coder:7b).	no	gpt-4o
llm-base-url	Base URL for the LLM API (e.g. http://localhost:11434/v1 for Ollama).	no	—
adr-dir	Path to the ADR directory within the repository.	no	docs/adr
fail-on-violation	Fail the CI check when violations are found.	no	false
min-confidence	Minimum confidence threshold (0.0–1.0) for bootstrap candidates. Lower values generate more ADRs.	no	0.3
max-pairs	Maximum number of (ADR, symbol) pairs to classify.	no	15
similarity-threshold	Minimum similarity score for a decision to be considered.	no	0.5
timeout-seconds	Maximum wall-clock time for the review in seconds.	no	300
sarif-output-path	Path to write SARIF output file. Leave empty to skip SARIF generation. Example: results/decisiondrift.sarif	no	""
review-mode	PR review submission mode. "comment" (default, just post a comment), "request-changes" (always request changes on violations), "auto" (approve on clean, request changes on violations).	no	comment

no outputs