CliRelay

v0.5.9 - Gemini Images on /v1/images, config.yaml Saves That Stop Losing Settings, and Codex Models That Survive Restarts

Pair it with codeProxy v0.6.2, which adds the Gemini image pickers and shows why a config.yaml save was rejected. Either upgrade order works; see Upgrade Notes. PUT /v0/management/config.yaml now rejects saves it used to accept and then silently lose, so scripts that write config.yaml should read the Upgrade Notes first.

Highlights

1. Antigravity's Gemini image model on /v1/images and the console

• What was wrong: gemini-3.1-flash-image on Antigravity already returned pictures through the chat endpoints, but nothing on the images side could reach it. /v1/images/*, the console's image page and the model probe listed only gpt-image, Grok and MiniMax, and the Antigravity executor did not understand an images request.
• What's new:
  • /v1/images/generations and /v1/images/edits serve Gemini image models while an Antigravity credential offers them. A model Antigravity publishes later needs no code change.
  • aspect_ratio and image_size map onto Gemini's image settings: the fourteen ratios upstream accepts (1:1 to 21:9) and the 512 / 1K / 2K / 4K tiers. A WIDTHxHEIGHT size maps to the nearest ratio and tier, so existing clients keep working.
  • Ratios are sent in upstream's own spelling. 21:9 used to be reduced to 7:3, which upstream rejects (#1148).
  • Reference images can be uploaded or sent as base64 data URIs. Remote URLs are not fetched, and masks are refused.
  • Responses report the real output_format (Gemini returns JPEG), the pixel size and token usage. A request that returns no image (a safety block, a text-only answer) fails as an invalid request, so the account is not cooled down.
  • The chat endpoints are unchanged.

2. config.yaml saves no longer report success while losing settings

• What was wrong: PUT /v0/management/config.yaml answered 200 {"ok": true} in two cases where the change then existed nowhere:
  • The runtime settings a save spelled out (proxy-url, debug, request-retry, payload and others) were stored one key at a time, after the file was written. A key that failed to store was only logged. The reload then stripped it from the file and restored the old stored value.
  • The sections the database owns (routing, proxy-pool, api-keys, api-key-entries, api-key-permission-profiles) were written to the file, then dropped or overridden by the stored value. Hand-typed API keys even worked as keys of that one node until its next start.
• The fix:
  • Runtime settings are stored first, in one transaction. If that fails, nothing changes and the save is rejected with 409 config_version_conflict or with a 500 that gives the cause.
  • A save that would change a database-owned section is rejected with 422 config_yaml_database_owned_sections before anything is stored or written. The message names each section and the page where it is edited: Channel Groups, Outbound Proxies, API Keys or Portal Account Permissions. A copy of the stored value still passes, as do the published example keys and sections the database does not hold yet.
  • Once the database holds API keys, a reload drops api-keys / api-key-entries from config.yaml, so hand-typed keys no longer act as keys of one node only.
  • Only hand-typed YAML is affected. The panel's dedicated pages use their own endpoints.

3. Codex model lists survive restarts

• What was wrong: on 2026-10-11 a deploy restarted production, and for half an hour Codex clients got 502 unknown provider for model gpt-6.1-sol. The live model list was kept only in memory, and fetching it again failed. The fetch looked up each credential's proxy_id in the system tenant's proxy pool instead of the credential's own tenant, and the system pool still pointed at retired proxy endpoints.
• The fix:
  • Every model list fetched fr

v0.5.8 - Go Security Update, Sign-in Lockout Warnings, and No More Idle HTTP/2 Connection Pile-ups

This release fixes eleven Go standard library and golang.org/x/net vulnerabilities; upgrade promptly. Pair it with codeProxy v0.6.1, whose sign-in forms use the fields added here to show the attempts left and to count a lock down. Either upgrade order works; see Upgrade Notes.

Highlights

1. Go 1.26.9 security update

• What was wrong: v0.5.7 was built with Go 1.26.6 and golang.org/x/net v0.57.0. govulncheck now reports eleven vulnerabilities, GO-2026-6603 to GO-2026-6617, that CliRelay's code can reach:
  • HTTP/2 server: a crash from an HPACK encoder race, memory exhaustion from trailer headers, excessive CPU from repeated window changes, and a flow-control accounting error;
  • HTTP/2 client: malformed framing-related headers were accepted;
  • HTTP/1: connections could fall out of sync after a CONNECT response, on the server and on the client;
  • Header parsing: Range headers had no size limit, and MIME header parsing could bypass its memory limit;
  • crypto/tls: malformed ECH extension references were not rejected;
  • Windows only: os.Root directory creation could follow junctions out of the root.
• The fix:
  • Builds use Go 1.26.9 and golang.org/x/net v0.60.0, and govulncheck reports no vulnerabilities that the code reaches.
  • The Docker image built its updater binary with Go 1.26.1; it now uses 1.26.9 as well.
• Upgrade promptly: CliRelay is both an HTTP server and an HTTP/2 client.

2. Sign-in warns before it locks, and says how long a lock lasts

• What was wrong: users were locked out of sign-in without warning and then given the wrong explanation, so they kept guessing.
  • The admin sign-in locked on the third wrong password, without saying so.
  • During that lock, wrong passwords still got "invalid credentials". Only the correct password got account_locked, which the panel shows as "This account is disabled or locked. Contact your tenant administrator." In production, an administrator typed the correct password during a lock, read that message, and gave up for an hour and a half.
  • Neither the admin nor the portal sign-in said how many attempts were left. The failure that tripped a lock looked like one more wrong password.
  • Throttle rejections carried the wait only in the Retry-After header, which the panel does not read.
• The fix:

  • The admin panel and the portal now share one cooldown ladder:

    Failures	Lock
    5	1 minute
    10	5 minutes
    15	15 minutes
    20	30 minutes
    25, and every 5th after	60 minutes

    The count restarts after 15 minutes without a failure. There is no permanent lock. The admin panel's first lock moves from 3 failures to 5, matching the portal.

  • The admin sign-in checks the lock before it checks the password. A sign-in during a lock gets 429 login_cooldown, whatever password is typed. account_locked now means only a lock set by an administrator.

  • A wrong password that does not trip a lock gets 401 invalid_credentials with details.remaining_attempts. For an ordinary run of mistypes the number is the same whether or not the username exists, so it does not reveal which accounts exist.

  • The wrong password that trips a lock gets 429, with the longest of the per-IP, per-account and cooldown waits.

  • login_rate_limited and login_cooldown responses carry details.retry_after_seconds in the body, as well as the Retry-After header.

3. Idle HTTP/2 connections to chatgpt.com no longer pile up

• What was wrong: a user reported 373 connections from one relay to chatgpt.com:443. They were hours old, and each was exchanging only about 4.7 kB an hour in each direction. Production showed the same: 218 such connections 3.5 hours after a deploy. Two problems combined:
  • The Codex quota probe built

v0.5.7 - Monitor Center API, Credential Imports, and Faster Recovery From Dead Connections

Pair this release with codeProxy v0.6.0. The redesigned panel's monitor center, guided "Add AI account" dialog and credential imports use the endpoints added here. Upgrade the backend first; see Upgrade Notes.

Highlights

1. Monitor center endpoints

• What was missing: the monitor page could only read /usage/chart-data: request counts, a success rate and token totals. There were no latency percentiles, first-token times, failure sources, rankings or live traffic.
• Overview: GET /v0/management/usage/monitor/overview.
  • Ranges 1h, 6h, 24h, today, 7d, 14d and 30d, filterable by portal user or API key, model and channel.
  • Totals compared with the previous period of equal length, and continuous zero-filled series.
  • Latency and first-token P50 / P90 / P95 / P99 / max / avg with a histogram.
  • Model, channel and portal-user rankings with sparklines, a weekday × hour heatmap, the consumer → model → channel flow, and recent failures.
• Realtime: GET /v0/management/usage/monitor/realtime covers the last hour per minute, current and peak RPM / TPM, and a five-minute window. It is cheap enough to poll every 15 seconds.
• Built to stay cheap:
  • Totals, series, rankings, heatmap and flows read the rollup buckets, so they cover the whole history.
  • Percentiles read request details only inside their retention window, and the response says where coverage starts.
  • A section that fails is logged and returned empty instead of failing the page.
  • Both endpoints are tenant-scoped and need monitor.read.
• Timezone label: services started without TZ reported their zone as Local. The label is now the IANA name, or the UTC offset (UTC+08:00) when the zone has no name. Bucket edges are unchanged.

2. Add AI accounts from credentials you already hold

• What's new: besides the browser and device sign-ins, an account can now be added from a credential you hold. CliRelay exchanges it for the same tokens a browser login would mint and saves the account under the caller's tenant. Four imports are supported:
  • a claude.ai session cookie;
  • an OpenAI / Codex CLI refresh token;
  • a Google / Antigravity refresh token;
  • a Grok web SSO cookie.
• Errors that say what went wrong:
  • A bad credential returns 422. It is never 401, which the panel would read as its own session expiring.
  • A Cloudflare challenge returns 502 and is reported as an egress problem rather than a bad credential. A timeout returns 504.
  • Secrets an upstream echoes back are masked in the message.
• The permission is auth_files.oauth, the same as for OAuth sign-in.

3. OAuth sign-ins report how they finish

• What was wrong: the panel only received a URL and a state, so it showed one generic "paste the callback URL" box for every provider, including device-code providers where there is nothing to paste. A login that had expired, or that had been replaced by a newer one, also read as {"status":"ok"}, which the panel showed as success.
• The fix:
  • The *-auth-url endpoints now return flow (redirect, code or device), expires_at, and for device logins user_code, verification_uri and expires_in.
  • Claude accepts ?callback_mode=code. Anthropic's own code page is then used, so a remote panel never sends the browser to a localhost forwarder.
  • Expired and superseded logins are reported as such. Late callbacks get 409 with oauth_login_expired or oauth_login_superseded, on a single node and in cluster mode.

4. Dead HTTP/2 connections are detected in about 45 seconds instead of about 15 minutes

• What was wrong: the fingerprinted transports, for Anthropic and the TLS-fingerprinted providers, reuse one HTTP/2 connection per host. A connection that had silently died still looked usable. A request written onto it was retransmitted

v0.5.6 - 多节点集群、滚动发布与 Codex 新模型即时可用 / Multi-node Cluster, Rolling Deploys, and Live Codex Models

核心更新亮点 / Highlights

1. Codex 新模型不用再等一次发版 / New Codex models work without a release

• 请求带上当前官方 CLI 版本：ChatGPT 会按客户端版本拒绝模型。gpt-6.1-sol 在 codex_cli_rs/0.149.1 和清单里写的 0.153.0 上都会被拒，当前稳定版 0.159.1 可以通过。请求现在使用官方 Codex CLI 的最新稳定版本；版本源暂时不可用时，退回到已经验证能过门的版本，有更新的版本会自动跟上。
• 根目录渠道自动带上新模型：默认渠道组的允许列表不再挡住 Codex 刚刚发现的模型。模型广场不用再去根目录渠道手动勾选。已经排除的模型仍然排除，其它渠道组保持原来的固定名单。

2. 多节点集群 / Multi-node cluster

• 共享运行时状态：凭证、配置、路由、会话亲和、账号并发、API Key 与终端用户限额、冷却、登录节流可以跨节点共享。
• 节点身份与就绪：进程带上节点身份，/readyz 能区分本节点和集群。用量日志在数据库中断时仍会留下，配额在最近一次用量读数上继续放行。
• DNS 与出口：clirelay-dnswatch 按节点健康和出口探测调整记录，避免把流量切到出不了网的节点。
• 备份：PostgreSQL 基础备份、WAL 归档，以及每周恢复演练。

3. 发布改为单实例替换 / Deploys replace the one live process

• 不再另起第二个槽：同一台机器上的第二个进程会抢同一个集群节点身份。发布现在替换 nginx 正在转发的那一个进程，失败时把上一份二进制换回去。
• 多节点仍逐台滚动：一台健康之后才继续下一台，并支持经跳板机连接。

详细改动记录 / Detailed Changes

后端 CliRelay

• fix(codex)：请求使用当前官方 CLI 版本；默认渠道组自动接纳 Codex 发现的新模型（#1109）。
• feat(cluster)：PostgreSQL 协调器、Redis 共享状态、跨节点限额 / 会话 / 冷却 / 配置 / 凭证（#1092–#1099）。
• feat(deploy)：逐节点滚动发布、节点自己的资源限制、可选跳板机（#1094、#1098）。
• feat(egress)：/readyz/egress，DNS 优先选择出口正常的节点（#1100）。
• feat(cluster)：基础备份、WAL 与每周恢复演练（#1107）。
• fix(update)：更新器缺失时返回 503 而不是 502（#1106）。
• fix(cluster)：降级节点重新加入时不再拖住整个集群；蓝绿继任者保留原节点的成员行（#1099、#1105）。

升级提示 / Upgrade Notes

• 单节点部署会替换当前正在服务的进程，切换期间有短暂中断，不再并行启动第二个槽。
• 集群模式需要按 deploy/cluster 的说明准备 PostgreSQL、Redis 和节点身份。没打开集群配置的单节点行为与原来一致，只是发布方式改成单实例替换。
• 面板请搭配 codeProxy v0.5.4。

未进入本次发版、仍开在 dev 上的 PR：#1108、#1104、#1101、#1086、#1085。