Repository navigation
v0.5.9 - Gemini Images on /v1/images, config.yaml Saves That Stop Losing Settings, and Codex Models That Survive Restarts
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.yamlnow 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-imageon 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/generationsand/v1/images/editsserve Gemini image models while an Antigravity credential offers them. A model Antigravity publishes later needs no code change.aspect_ratioandimage_sizemap onto Gemini's image settings: the fourteen ratios upstream accepts (1:1to21:9) and the512/1K/2K/4Ktiers. AWIDTHxHEIGHTsize maps to the nearest ratio and tier, so existing clients keep working.- Ratios are sent in upstream's own spelling.
21:9used to be reduced to7:3, 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 pixelsizeand tokenusage. 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.yamlanswered200 {"ok": true}in two cases where the change then existed nowhere:- The runtime settings a save spelled out (
proxy-url,debug,request-retry,payloadand 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 runtime settings a save spelled out (
- The fix:
- Runtime settings are stored first, in one transaction. If that fails, nothing changes and the save is rejected with
409 config_version_conflictor with a500that gives the cause. - A save that would change a database-owned section is rejected with
422 config_yaml_database_owned_sectionsbefore 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-entriesfrom 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.
- Runtime settings are stored first, in one transaction. If that fails, nothing changes and the save is rejected with
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'sproxy_idin 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 from upstream is saved to
provider-model-lists.jsonin the state directory and restored at startup, before the first registration pass. A restored list keeps routing until a newer answer replaces it. - Model discovery (Codex manifests, and the per-credential lists for Antigravity, xAI and Kimi) uses the credential's tenant configuration, so
proxy_idresolves in that tenant's pool, as chat requests already did.
- Every model list fetched from upstream is saved to
4. Antigravity: long agent sessions and malformed function calls
Contributed by @testAccountDeltas (#1143, #1139) and @RUIIIOVO (#1145). Thank you both.
- Sessions past one million tokens: Antigravity accumulates a conversation on its side per session id. Once a long tool loop passed 1,048,576 tokens, every request in that session failed with
400 … exceeds the maximum number of tokens, and the conversation was stuck. The executor now retries on a new session id, at most twice per request, without using up a capacity retry. Sessions below the limit keep their prefix-cache hits. Each new turn starts on the original session again and switches after one quick 400. MALFORMED_FUNCTION_CALL: Gemini sometimes plans a tool call, fails to emit it, and ends the turn after only thinking. Clients took that as a finished turn, so agents stopped halfway through a task. The executor now asks again, at most twice, while nothing but thinking has been sent. A streaming client still sees one response with one ending. Once text or a tool call has gone out, the answer stands.
5. Panel probes work in tenants whose default group excludes the model
- What was wrong: the image console, the video console and the model test pick an account the same way API traffic does. A model excluded from the tenant's
defaultchannel group failed withmodel_not_allowed_by_channel_group, and accounts markedexclude-from-defaultcould not be probed at all. - The fix: these three probes may use any account of the tenant. API traffic is unchanged.
Detailed Changes
Backend CliRelay
- feat(images): serve Antigravity Gemini image models on
/v1/imagesand the console (#1147). - fix(images): send Gemini aspect ratios in the spelling upstream accepts (#1148).
- fix(management): let panel probes use any credential of the tenant (#1149).
- fix(model-discovery): keep upstream model lists across restarts and fetch them through the tenant's proxy (#1150).
- fix(management): reject config.yaml saves whose runtime settings fail to store (#1151).
- fix(management): reject config.yaml saves that change database-owned sections (#1152).
- test(management): pin the reason carried by
{"error": "<code>", "message": "<reason>"}bodies, which codeProxy v0.6.2 now shows (#1153). - fix(antigravity): rotate the session id at the per-session 1M-token limit (#1143 and #1139, by @testAccountDeltas) and ask again when Gemini ends a turn with
MALFORMED_FUNCTION_CALL(#1145, by @RUIIIOVO). Landed together in #1154, which keeps the re-request on the rotated session and adds executor tests.
Upgrade Notes
- Scripts that write config.yaml (
PUT /v0/management/config.yaml):- A save can now fail where it used to return 200:
409 config_version_conflict,422 config_yaml_database_owned_sections, or500with the cause. A failed save changes nothing, so retry it, or edit the named sections where the message says. - Change routing, proxy pools and API keys through their own management endpoints, not by writing them into config.yaml.
- A save can now fail where it used to return 200:
- Either upgrade order works:
- codeProxy v0.6.2 on an older backend shows no Gemini image pickers; everything else in it works.
- Older panels keep working on v0.5.9. Their size picker still drives Gemini image models, but some rejected saves show only an error code such as
invalid_config. - CliRelay pulls the latest codeProxy release on its own, roughly every three hours.
- New state file:
provider-model-lists.jsonin the state directory. It is written atomically, so blue-green slots on one host can share it. - No changes required: there are no configuration or database schema changes.
中文说明(点击展开)
v0.5.9 - /v1/images 支持 Gemini 生图、config.yaml 保存不再悄悄丢设置、Codex 模型清单重启不丢
请与 codeProxy v0.6.2 搭配:新面板增加了 Gemini 生图的比例和分辨率档位,并会显示 config.yaml 保存被拒绝的原因。前后端谁先升级都可以,详见升级提示。
PUT /v0/management/config.yaml现在会拒绝以前接受、随后却悄悄丢掉的保存,写 config.yaml 的脚本请先看升级提示。
核心更新
1. /v1/images 和控制台支持 Antigravity 的 Gemini 生图模型
- 原来的问题:Antigravity 上的
gemini-3.1-flash-image已经能通过对话接口返回图片,但生图这边都用不了它。/v1/images/*、控制台的生图页和模型测试只列出 gpt-image、Grok 和 MiniMax,Antigravity 执行器也不认识生图请求。 - 新增:
- Antigravity 账号提供 Gemini 生图模型时,
/v1/images/generations和/v1/images/edits就能调用它们。Antigravity 以后发布的新模型不需要改代码。 aspect_ratio和image_size对应 Gemini 的图片参数:上游接受的 14 种比例(1:1到21:9)和512/1K/2K/4K四档。传WIDTHxHEIGHT尺寸时会换算成最接近的比例和档位,原有客户端照常可用。- 比例按上游自己的写法发送。原来
21:9会被约分成7:3,上游不认(#1148)。 - 参考图可以上传,也可以用 base64 data URI。不抓取远程 URL,不支持蒙版。
- 响应里带真实的
output_format(Gemini 返回 JPEG)、像素size和 tokenusage。没有返回图片的请求(被安全策略拦截、只返回了文字)按无效请求处理,不会让账号进入冷却。 - 对话接口没有任何变化。
- Antigravity 账号提供 Gemini 生图模型时,
2. config.yaml 保存不再一边报成功、一边丢设置
- 原来的问题:
PUT /v0/management/config.yaml在两种情况下返回200 {"ok": true},改动却哪里都没留下:- YAML 里写出的运行时设置(
proxy-url、debug、request-retry、payload等)是在写完文件之后逐个存进数据库的。某个键存失败只记日志;随后的重载把它从文件里删掉,再套回数据库里的旧值。 - 由数据库管理的段(
routing、proxy-pool、api-keys、api-key-entries、api-key-permission-profiles)被写进文件后,又被删掉或被数据库里的值覆盖。手写进去的 API Key 甚至会在那一个节点上生效,直到它下次重启。
- YAML 里写出的运行时设置(
- 修复:
- 先存运行时设置,并且放在一个事务里。存失败时什么都不改,保存被拒绝,返回
409 config_version_conflict,或者附带原因的500。 - 会改动数据库管理段的保存,在存储和写文件之前就以
422 config_yaml_database_owned_sections拒绝。提示里会列出每个段,以及应该去哪个页面修改:渠道分组、出站代理、API Keys 或门户账号权限。与数据库中现有值相同的内容照常通过,官方示例 Key 和数据库里还没有的段也照常通过。 - 数据库里已有 API Key 后,重载时会从 config.yaml 中去掉
api-keys/api-key-entries,手写的 Key 不会再只在单个节点上生效。 - 只影响手写的 YAML,面板的专用页面走各自的接口。
- 先存运行时设置,并且放在一个事务里。存失败时什么都不改,保存被拒绝,返回
3. Codex 模型清单重启后不再丢失
- 原来的问题:2026-10-11 一次部署重启了生产服务,之后半小时 Codex 客户端一直收到
502 unknown provider for model gpt-6.1-sol。实时模型清单只存在内存里,重新拉取又失败了:拉取时按系统租户的代理池去解析账号的proxy_id,而不是账号所属租户的代理池,而系统租户的代理池仍指向已经下线的代理地址。 - 修复:
- 每次从上游拉到的模型清单都会存到状态目录下的
provider-model-lists.json,启动时在第一次注册前恢复。恢复的清单会一直参与路由,直到被更新的结果替换。 - 模型发现(Codex 清单,以及 Antigravity、xAI、Kimi 的按账号清单)改用账号所属租户的配置,
proxy_id在该租户自己的代理池里解析,与对话请求保持一致。
- 每次从上游拉到的模型清单都会存到状态目录下的
4. Antigravity:长 Agent 会话与 MALFORMED_FUNCTION_CALL
由 @testAccountDeltas(#1143、#1139)和 @RUIIIOVO(#1145)贡献,感谢两位。
- 会话超过一百万 token:Antigravity 会按 session id 在上游累积整段对话。长时间的工具循环一旦超过 1,048,576 token,这个会话的每个请求都会返回
400 … exceeds the maximum number of tokens,对话就卡住了。现在执行器会换一个新的 session id 重试,每个请求最多换两次,不占用容量重试的次数。没超过上限的会话照常命中前缀缓存。每一轮新的对话仍会先用原来的会话,碰到一次很快返回的 400 后再切换。 MALFORMED_FUNCTION_CALL:Gemini 有时计划调用工具却没有发出调用,只输出了思考就结束了这一轮。客户端会把它当作正常结束,Agent 就停在半途。现在只要还没发出思考以外的内容,执行器就会重新请求,最多两次。流式客户端看到的仍是一个响应、一次结束;一旦已经发出文字或工具调用,就保留这次回答。
5. 默认分组排除了某个模型时,面板测试照常可用
- 原来的问题:生图控制台、视频控制台和模型测试选账号的方式与 API 请求相同。被租户
default渠道分组排除的模型会报model_not_allowed_by_channel_group,标了exclude-from-default的账号则完全测不了。 - 修复:这三处面板测试可以使用租户里的任意账号,API 请求不受影响。
详细改动
后端 CliRelay
- feat(images):
/v1/images和控制台支持 Antigravity 的 Gemini 生图模型(#1147)。 - fix(images):Gemini 比例按上游接受的写法发送(#1148)。
- fix(management):面板测试可以使用租户里的任意账号(#1149)。
- fix(model-discovery):上游模型清单重启后不丢失,并通过账号所属租户的代理拉取(#1150)。
- fix(management):运行时设置存储失败时拒绝 config.yaml 保存(#1151)。
- fix(management):拒绝改动数据库管理段的 config.yaml 保存(#1152)。
- test(management):用测试固定
{"error": "<代码>", "message": "<原因>"}错误体里的原因,codeProxy v0.6.2 会显示这个原因(#1153)。 - fix(antigravity):会话达到 1M token 上限时轮换 session id(#1143、#1139,作者 @testAccountDeltas),Gemini 以
MALFORMED_FUNCTION_CALL结束时重新请求(#1145,作者 @RUIIIOVO)。两者在 #1154 中一起合入,#1154 还让重新请求保持在轮换后的会话上,并补充了执行器测试。
升级提示
- 写 config.yaml 的脚本(
PUT /v0/management/config.yaml):- 以前返回 200 的保存现在可能失败:
409 config_version_conflict、422 config_yaml_database_owned_sections,或附带原因的500。失败的保存不会改动任何内容,可以重试,或按提示去对应的地方修改。 - 路由、代理池和 API Key 请通过各自的管理接口修改,不要写进 config.yaml。
- 以前返回 200 的保存现在可能失败:
- 前后端谁先升级都可以:
- codeProxy v0.6.2 跑在旧后端上时不显示 Gemini 生图的比例和档位,其余功能都可以用。
- 旧面板在 v0.5.9 上照常可用。它的尺寸选择仍能驱动 Gemini 生图模型,只是部分被拒绝的保存只显示
invalid_config这样的错误代码。 - CliRelay 大约每三小时会自动拉取 codeProxy 的最新版本。
- 新的状态文件:状态目录下的
provider-model-lists.json。它以原子方式写入,同一台机器上的蓝绿两个实例可以共用。 - 无需改动:配置和数据库结构都没有变化。
Multi-arch binaries and the Docker image are published to GitHub Releases and GHCR. Released together with codeProxy v0.6.2.