v6.20.1

Auto-released from merge of PR #1796 (trigger: push).

• The thinking share covers GPT traffic too. Codex requests recorded 0 thinking tokens, so dario usage, --by-key and /analytics were blind on every ChatGPT-subscription lane. The engine's output_tokens_details.reasoning_tokens (counted inside output, like Claude's thinking) is now read on both GPT routes, the translated Messages route and the Responses passthrough, and lands in the same thinkingTokens the ledger keeps per day, model and consumer; a count above the output is capped at it. Rows written before read as before.

---

Built + tested + npm-published inline by cc-drift-auto-release.yml on this run — tokenless via OIDC trusted publishing, provenance-attested. See the workflow log for the publish output.