文件名含非 ASCII 字符时 workspace API 返回 500

[baryhuang]

找到根因了。

问题：文件名含非 ASCII 字符时 workspace API 返回 500

测试结果（直接 curl GET /v1/files/{id}?network=...）：

文件名	HTTP
regulatory_cost_overview.txt	200 ✓
New_Recording_143-transcript.txt	200 ✓
04-23 Team Meeting_ Task Management... Product Rebranding-transcript.txt	200 ✓
All Mentor Matic Mentors (1).csv	200 ✓
Business Model Canvas - Strategyzer - v2.1.pdf	200 ✓
04-24 会议记录：QModel品牌...-transcript.txt	500 ✗
04-23 团队周会：任务管理优化...-transcript.txt	500 ✗
04-22 Coaching Session_ ... — CarMomo-transcript.txt	500 ✗（em-dash "—"）

规律：文件名里含 CJK 字符或 em-dash（—）的全部 500；纯 ASCII 文件名全部 200。 内容里有中文没问题（New_Recording_143-transcript.txt 是中文录音 52KB，读得出来）。

根因（服务端 bug）

https://workspace-endpoint.openagents.org/v1/files/{fileId} 在生成响应时拿原始 filename 去拼 header（大概率是 Content-Disposition: attachment; filename=...）。HTTP header 只允许 Latin-1，非 ASCII 字符直接让后端崩，返回裸 500 Internal Server Error（body 只有 21 字节，不是正常 JSON）。

正规做法是 RFC 5987：filename*=UTF-8''%E4%BC%9A%E8%AE%AE...

[linear]

OADEV-59

[zomux]

Check in

[dvirarad]

Heads-up: this looks like it was already fixed on main in commit e1bf8e59 (Apr 27 — fix(workspace-backend): RFC 6266-encode non-ASCII filenames in Content-Disposition).

workspace/backend/app/routers/files.py now uses the exact pattern this issue describes — RFC 5987 filename*=UTF-8''<percent-encoded> plus an ASCII fallback, with a comment explicitly attributing the prior 500s to Starlette failing during header encoding.

If your reproducer was against workspace-endpoint.openagents.org (hosted), it's likely the deployment hasn't picked up the fix yet. Worth a quick re-test with the same CJK / em-dash filenames once the hosted env is bumped — if 200s come back, this issue can be closed.

If the hosted env is already on a newer build and you're still seeing 500s, the next suspect would be the upload path (POST /v1/files) or a CDN/proxy in front of the API, not the download handler.