Skip to content

Commit e3a2bf4

Browse files
committed
docs: update changelog for v2.2.7
1 parent a0ff9b9 commit e3a2bf4

1 file changed

Lines changed: 16 additions & 0 deletions

File tree

CHANGELOG.md

Lines changed: 16 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,5 +1,21 @@
11
# Changelog
22

3+
## v2.2.7 - 2026-06-05
4+
5+
### Features
6+
7+
- **WebSocket silent retry controls (#195).** Added Codex upstream WS settings for hiding upstream errors, silently retrying pre-first-token upstream failures, and capping silent retries (`codex_ws_hide_upstream_errors`, `codex_ws_silent_retry_enabled`, `codex_ws_silent_max_retries`). These settings are available in the admin UI and persist through both Postgres and SQLite.
8+
- **WHAM-only usage probe controls.** Added runtime/admin controls so usage probes can rely on the zero-cost WHAM endpoint without falling back to `/responses` probes when that is preferred.
9+
10+
### Fixes
11+
12+
- **WS upstream failure handling (#195).** Retryable upstream WS failures before the first token, including usage-limit, 429, 5xx, read errors, timeouts, and EOFs, now stay server-side while codex2api switches accounts and rebuilds the upstream WS connection. If retries are exhausted, clients receive a unified friendly message while the original upstream error remains in backend logs and usage records.
13+
- **Responses routing and model hardening (#198).** Hardened `/v1/responses` routing for recent Codex models including `gpt-5.3-codex-spark`, widened local plan gating so stale local `plan_type` records do not incorrectly block real upstream calls, and improved WS/TTFT handling around response payload content.
14+
- **OpenAI Responses compact routing.** OpenAI Responses API accounts added with `base_url` + `api_key` can now use `/v1/responses/compact`; compact bodies are normalized for that upstream path instead of being sent through the ChatGPT-only compact route.
15+
- **Usage-limit detection.** `usage_limit_reached` is now recognized even when wrapped inside `response.error`, `response.status_details`, or upstream 5xx-shaped payloads, so exhausted accounts are treated as quota-limited consistently.
16+
- **Accounts toolbar wrapping.** Improved the admin accounts toolbar layout so search/filter/action controls wrap cleanly on narrower viewports.
17+
- **Security scan recovery.** Updated React Router to a patched release and raised the Go toolchain directive to `1.26.4`, clearing the failing frontend npm audit and backend govulncheck jobs.
18+
319
## v2.2.6 - 2026-06-03
420

521
### Features

0 commit comments

Comments
 (0)