| 文件 | 最后提交记录 | 最后更新时间 |
|---|---|---|
feat: implement per-command process groups with registry for reliable cleanup Co-Authored-By: GLM-5 | 19 天前 | |
feat(llm-client): thread cached input tokens through Usage into trace spans Account for cache-read tokens end to end: - Usage carries cached_tokens (OpenAI prompt_tokens_details.cached_tokens; Anthropic cache_read_input_tokens), threaded into LLM_CALL / TURN trace spans. - Anthropic marks cache_control ephemeral on the stable system-prefix block ONLY, never the volatile Context block, so the breakpoint is read back across turns instead of being rewritten every turn. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> | 21 天前 | |
feat: initial commit with core features | 2 个月前 | |
feat(llm-client): thread cached input tokens through Usage into trace spans Account for cache-read tokens end to end: - Usage carries cached_tokens (OpenAI prompt_tokens_details.cached_tokens; Anthropic cache_read_input_tokens), threaded into LLM_CALL / TURN trace spans. - Anthropic marks cache_control ephemeral on the stable system-prefix block ONLY, never the volatile Context block, so the breakpoint is read back across turns instead of being rewritten every turn. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> | 21 天前 | |
fix(llm-client): keep tool-call id/name when stream continuation chunks send empty Some OpenAI-compatible gateways repeat a tool call's id/name as empty strings on stream continuation chunks. Don't overwrite the established id/name with the empty string — keep the first non-empty value — so the assembled tool call keeps a valid id/name instead of being silently blanked. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> | 21 天前 | |
敏感数据保险箱530阶段性实现 支持对API key的保护，白盒密钥和TEE内生国密两种方式 | 1 个月前 |