ADR 0017: Use Tenacity for OpenAI retry/backoff

Date: 2026-01-14

Context: Loreley had multiple hand-rolled while + sleep retry loops around OpenAI calls (embeddings, commit summaries, trajectory chunk summaries), duplicating logic and making retry tuning inconsistent.

Decision: Use tenacity as the single retry/backoff mechanism for OpenAI calls and centralize configuration in loreley/core/openai_retry.py.

Decision: Preserve existing retry semantics by mapping *_max_retries to stop_after_attempt(...) and keeping linear backoff (backoff_seconds * attempt_number) via wait_incrementing(...), with a loguru-compatible before_sleep hook.

Consequences: Retry code is shorter and consistent across call sites; future changes (retry only transient errors, add jitter, add metrics/log hooks) require edits in one place; final failure errors may include Tenacity context in tracebacks.