v1.89.6

huggingface/skillsv1.89.6Jun 12, 2026by autogame-17

AI Summary

This release focuses on fixing trace data quality issues to make Claude Code, cursor, and codex traces usable. It introduces fail-closed behavior for hub-decryptable uploads and improves streaming usage tracking capabilities.

Key Highlights

Proxy trace data-quality fixes for Claude Code, cursor, and codex traces.
Hub-decryptability fail-closed mechanism for profile analysis.
Valid-JSON truncation improvements with field caps increased to 1MB.
Session threading improvements with per-session extraction.
Streaming usage capture via passive SSE tee on both message endpoints.

New Features

Streaming usage capture (end-of-stream token usage)
OpenAI/codex shapes recognition
Per-session threading extraction
Hub-decryptable upload filtering

Full Release Notes

## @evomap/evolver v1.89.6

Proxy trace data-quality fixes so Claude Code, cursor, and codex traces are actually usable, plus hub-decryptability fail-closed (#251).

### Fixed
- **Valid-JSON truncation** — over-cap bodies are wrapped in a parseable envelope instead of hard-cut; field cap 64KB→1MB, upload 256KB→4MB.
- **Session threading** — per-session `session_id` extracted from `metadata.user_id` / `x-cursor-session-id`; rows now carry a trace id.
- **Streaming usage** — end-of-stream token usage / finish reason / provider response id captured by a passive SSE tee (bytes forwarded unchanged), on both `/v1/messages` and `/v1/responses`. Codex sessions thread by `previous_response_id` → `response.id`.
- **OpenAI/codex shapes** — `prompt_tokens`/`completion_tokens` and Responses `status`/`incomplete_details.reason` recognized.
- **Hub-decryptable uploads (fail-closed)** — when profile analysis is enabled, only rows carrying a hub-decryptable `hub_key_envelope` reach the warehouse; the hub trace public key can be distributed via a signature-verified config. No effect when analysis is disabled (the default).