新增: 媒体引用化存储与消息溯源

- media_store 模块: sha256 去重落盘 + 阿里云 OSS 双写(签名手写), OSS 失败降级本机引用 - 飞书图片入站: image 消息下载装配 vision 片, 3 次退避+120s 超时+残件清理防线 - 出站回传: 报告 [交付] 路径行自动上传飞书原生附件, 凭据/密钥名拦截兜底 - 溯源 V52: ai_messages.source 列, 飞书 msg_id/chat_id/发送者随消息落库 - spawn offload: 桌面粘贴图 base64 统一转落盘引用, SQLite 不进大字符串 - 引用统一 url 字段(https/file://)不自造 scheme, image_ref 发送前按需展开
2026-10-11 07:33:18 +08:00
parent 8aefe3254c
commit 18e50c5af7
23 changed files with 1542 additions and 105 deletions
@@ -0,0 +1,122 @@
+//! 本地图片引用展开(#958 图片入站引用化)
+//!
+//! 消息持久化只存引用(ContentPart::Image.url = `file://<绝对路径>`,base64 空),
+//! 发送请求前在此统一展开为 base64 内嵌——存储面瘦 referencing(一张 5MB 图不再以
+//! 6.7MB base64 字符串进 SQLite 随历史全量重放),转换面按需读盘(SSD 毫秒级,
+//! 对比 base64 内嵌的每请求全量携带,磁盘读便宜得多)。
+//!
+//! 展开点是 provider 转换层的入口:主会话/子代理/标题/压缩等全部请求路径在此
+//! 单点覆盖,调用方(飞书入站/桌面粘贴图)只管落盘产引用,不关心 provider 差异。
+//!
+//! 失败语义:引用文件丢失/不可读时降级为 Text 片「[图片文件缺失: path]」——
+//! 诚实告知模型图不可用,不让请求 400(空 base64 会被两个协议端点拒绝)。
+
+use crate::provider::ContentPart;
+use base64::Engine as _;
+
+/// `file://` 引用前缀。内部引用标识(两端自洽即可,不进端点):落盘方拼
+/// `file://<原生绝对路径>`,展开方 strip 后按原生路径读(Windows 盘符路径直接可用)。
+pub const LOCAL_FILE_SCHEME: &str = "file://";
+
+/// 判定 Image 片是否为本地引用(base64 空 + url 带本地 scheme)
+fn is_local_ref(url: Option<&str>, base64: Option<&str>) -> bool {
+    base64.is_none() && url.is_some_and(|u| u.starts_with(LOCAL_FILE_SCHEME))
+}
+
+/// 把消息内的本地图片引用展开为 base64 内嵌片;非本地片原样透传。
+/// 文件缺失/不可读时该 Image 片降级为 Text 片(请求不炸,模型得到可读的缺失说明)。
+pub fn inline_local_image_parts(parts: Vec<ContentPart>) -> Vec<ContentPart> {
+    parts.into_iter().map(inline_local_image_part).collect()
+}
+
+/// 单片版(provider 转换臂逐片 map 用)
+pub fn inline_local_image_part(p: ContentPart) -> ContentPart {
+    match p {
+        ContentPart::Image { url, base64, media_type, alt }
+            if is_local_ref(url.as_deref(), base64.as_deref()) =>
+        {
+            let path = url.unwrap();
+            match std::fs::read(&path[LOCAL_FILE_SCHEME.len()..]) {
+                Ok(bytes) => ContentPart::Image {
+                    url: None,
+                    base64: Some(base64::engine::general_purpose::STANDARD.encode(bytes)),
+                    media_type,
+                    alt,
+                },
+                Err(_) => ContentPart::Text {
+                    text: format!("[图片文件缺失: {}]", &path[LOCAL_FILE_SCHEME.len()..]),
+                },
+            }
+        }
+        other => other,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn local_ref_expands_to_base64() {
+        let dir = std::env::temp_dir().join("df-ai-image-ref-test");
+        std::fs::create_dir_all(&dir).unwrap();
+        let path = dir.join("expands.png");
+        std::fs::write(&path, b"\x89PNG fake-bytes").unwrap();
+
+        let out = inline_local_image_parts(vec![ContentPart::Image {
+            url: Some(format!("file://{}", path.display())),
+            base64: None,
+            media_type: Some("image/png".into()),
+            alt: None,
+        }]);
+        match &out[0] {
+            ContentPart::Image { url, base64, media_type, .. } => {
+                assert!(url.is_none());
+                let b = base64.as_deref().expect("base64 应已填充");
+                assert_eq!(b, base64::engine::general_purpose::STANDARD.encode(b"\x89PNG fake-bytes"));
+                assert_eq!(media_type.as_deref(), Some("image/png"));
+            }
+            other => panic!("应保持 Image 片, 实际 {other:?}"),
+        }
+        std::fs::remove_file(&path).ok();
+    }
+
+    #[test]
+    fn missing_file_degrades_to_text_not_empty_base64() {
+        let out = inline_local_image_parts(vec![ContentPart::Image {
+            url: Some("file:///nonexistent/df-ai-no-such.png".into()),
+            base64: None,
+            media_type: Some("image/png".into()),
+            alt: None,
+        }]);
+        match &out[0] {
+            ContentPart::Text { text } => {
+                assert!(text.contains("[图片文件缺失"), "应降级为缺失说明, 实际 {text}");
+                assert!(text.contains("df-ai-no-such.png"));
+            }
+            other => panic!("缺失文件应降级 Text 片, 实际 {other:?}"),
+        }
+    }
+
+    #[test]
+    fn non_local_parts_pass_through() {
+        let passthrough = vec![
+            ContentPart::text("hello"),
+            ContentPart::image_base64("image/png", "aGVsbG8="),
+            ContentPart::image_url("https://x/a.png"),
+            // base64 非空的 file:// 不是引用(优先内嵌),原样透传
+            ContentPart::Image {
+                url: Some("file:///x.png".into()),
+                base64: Some("aGVsbG8=".into()),
+                media_type: Some("image/png".into()),
+                alt: None,
+            },
+        ];
+        let out = inline_local_image_parts(passthrough);
+        assert_eq!(out.len(), 4);
+        assert!(matches!(&out[0], ContentPart::Text { .. }));
+        assert!(matches!(&out[1], ContentPart::Image { base64: Some(_), .. }));
+        assert!(matches!(&out[2], ContentPart::Image { url: Some(u), .. } if u.starts_with("https://")));
+        assert!(matches!(&out[3], ContentPart::Image { base64: Some(_), .. }));
+    }
+}