English
If you have any questions or suggestions about the software, please reach us through the QQ group or GitHub Issues and we will get back to you as soon as possible.
Added
gpt-oss-120b-mediumin the bundled Codex catalog: the model is now part of the shipped catalog (revision 9), so it appears in the Codex model picker without a separate download.
Improved
- ACP model names are now published as-is: the ACP provider emitted an alias on top of an already renamed model, which the Claude Desktop alias table cannot express. Models are now published under their
acp/prefixed name with no alias indirection. - Claude Desktop no longer reports the managed configuration as needing attention:
behavesAsis no longer written into the managed model entries, because Desktop builds that do not recognise it flag the whole configuration. The family routes Desktop already knows (claude-opus-5,claude-sonnet-5,claude-haiku-4-5) are published directly and the upstream model stays inlabelOverride. Files written by an older version cannot resurrect the field. - ACP API streaming correctness: Chat Completions and Responses clients now receive a complete OpenAI-compatible stream — an initial assistant role delta, text or indexed tool calls, one terminal finish reason and
[DONE], with a stable completion ID, model and creation time across chunks. Output queues apply backpressure while staying cancellable instead of dropping frames under load, and usage is omitted rather than fabricated when the CLI reports no token counts.
Fixed
- Antigravity usage was inflated and lost: the usage scanner counted every mirror of a session (
transcript.jsonl,transcript_full.jsonl,chunks/**, IDE and backup copies), so the reported total was several times the real usage while some per-request models showed as "unknown model". Each session now contributes once from a single canonical transcript, models are read per request step from the session database with the settings label as a fallback, and identical samples are deduplicated while retries and compaction checkpoints are preserved. The first sync after upgrading rebuilds the Antigravity rows once, and the total drops to the real value. - Antigravity usage could disappear after an unreadable file: a session whose transcript could not be read had its cached rows deleted anyway, because any other readable file in the session authorised the clean-up. Cached rows now survive whenever a read fails, mirrored transcripts are compared by the requests they actually cover rather than by file name, and legacy keys are migrated instead of being counted alongside their replacements.
- ACP API write failures lost the underlying cause: saving a Claude Desktop alias mapping failed with
invalid_configbecause the ACP provider wroteproxy-urlat the provider level, which the core rejects foropenai-compatibilityentries. The field is now written inside the key entry, an existing provider-level value is normalised into the key entries, and the core error text is appended to the write-failure message. - ACP API could fail mid-completion: an upstream failure after the first chunk was reported as a truncated success. Failures before output now return an HTTP error, and failures after output emit a terminal SSE error without a success finish reason or
[DONE], so a failed turn can no longer look completed.
简体中文
关于软件,如果您有任何问题或建议,请通过 QQ群 或 GitHub Issues 反馈,我们会尽快处理。
新增
- 内置 Codex 模型目录新增
gpt-oss-120b-medium:该模型已加入随应用发布的模型目录(revision 9),在 Codex 模型选择器中直接可见,无需额外下载。
改进
- ACP 模型改为直接发布原始模型名:此前 ACP 服务在已改名的模型上再加一层别名,而 Claude Desktop 的别名表无法表达“重命名的重命名”。现在直接以
acp/前缀的模型名发布,不再写入别名字段。 - Claude Desktop 不再把托管配置标记为需要注意:不再向托管模型条目写入
behavesAs,因为无法识别该字段的 Desktop 版本会把整个配置判为异常。改为直接发布 Desktop 已认识的系列路由(claude-opus-5、claude-sonnet-5、claude-haiku-4-5),上游模型名保留在labelOverride中。旧版本写入的字段也不会再被恢复。 - ACP API 流式输出修正:Chat Completions 与 Responses 客户端现在能收到完整的 OpenAI 兼容流——首个 assistant 角色增量、文本或带索引的工具调用、唯一的结束原因与
[DONE],各分片共用同一个 completion ID、模型名和创建时间。输出队列在拥塞时施加背压同时保持可取消,不再丢帧;CLI 未提供 token 计数时不编造用量。
修复
- Antigravity 用量被重复计算且存在丢失:用量扫描此前会把同一个会话的每份镜像(
transcript.jsonl、transcript_full.jsonl、chunks/**、IDE 与备份副本)都统计一遍,总量达到真实用量的数倍,同时部分请求的模型显示为「unknown model」。现在每个会话只从一份规范 transcript 计入一次,模型按请求步骤从会话数据库读取并以设置标签作为回退,完全相同的样本会被去重,而重试与压缩检查点仍然保留。升级后首次同步会重建一次 Antigravity 数据,总量回落到真实值。 - 文件读不出来时 Antigravity 用量会消失:此前只要会话里有任意一个文件读取成功,其余文件的缓存行就会被删除,即使它们读不出来。现在读取失败一律保留缓存行;对照的 transcript 按实际覆盖的请求比较而不是按文件名判断;旧键会被迁移,不再与新键并存重复计数。
- ACP API 写入失败时丢失了根因:保存 Claude Desktop 别名映射会以
invalid_config失败,原因是 ACP 服务把proxy-url写在 provider 层,而核心不接受openai-compatibility条目的这种写法。现在写入 key 条目内部,已存在的 provider 层取值会被归一到 key 条目,并在写入失败的提示中附上核心的原始错误。 - ACP API 可能在补全中途失败却显示成功:首个分片之后的上游故障此前会被当作截断的成功结果。现在输出前的故障返回 HTTP 错误,输出后的故障发出终止性 SSE 错误且不带成功结束原因与
[DONE],失败的回合不会再看起来像已完成。