fix(gateway): a provider refusing the task's earlier reasoning is said in words (#147)
* fix(gateway): a provider refusing the task's earlier reasoning is said in words OpenAI's encrypted reasoning items are opened only by the account that produced them; a route served from more than one upstream account hands the next turn items its upstream cannot open, and the person read the provider's raw JSON. Measured on TokenRouter's gpt-5.4 route: a gpt-5.4 thread switched into gpt-6-astra (the codex column, 2026-09-10) while six other ids made the same switch and passed, and a same-model cold restore on hosted (2026-09-08). The turn now fails in words that say what to do, with the previous model named when there is one; the provider's JSON is never the message. Ported from hosted 2f18cfdb. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ej7eMwxnexLFqkiFzHq9Fk * chore: start checks Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Ej7eMwxnexLFqkiFzHq9Fk --------- Co-authored-by: richard-epsilla <richard@epsilla.com> Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
G
github-actions[bot] committed
fba2ac903c483bf359d1bcaaa70e05f599eeaf97
Parent: f01d010
Committed by GitHub <noreply@github.com>
on 9/10/2026, 8:19:37 AM