Repository navigation
Image/Audio format= emits image/jpg, audio/mp3, audio/m4a — unregistered MIME types that disagree with the path= suffix table #3585
Description
Activity
Hi maintainers,
I reproduced this locally:
Image(data=..., format="jpg")andAudio(data=..., format="mp3")emit unregistered MIME types (image/jpg,audio/mp3), differing from the suffix mapping used bypath=.I have a fix with tests ready that normalizes
format=through the same mapping table aspath=while preserving generic fallbacks. If this approach looks good, could you assign this issue to me? (Assigning will also automatically reopen PR #3593, which was closed by the bot).Disclosure: Prepared with AI assistance under human review.
Confirmed on
main. The asymmetry is exactly as described.src/mcp/server/mcpserver/utilities/types.py,Image._get_mime_type():if self._format: return f"image/{self._format.lower()}" if self.path: suffix = self.path.suffix.lower() return {".png": "image/png", ".jpg": "image/jpeg", ...}.get(suffix, "application/octet-stream")
The
format=branch is an unchecked interpolation — it never consults the suffix table two lines below it, soformat="jpg"emitsimage/jpgwhilepath="x.jpg"emitsimage/jpeg. Same forAudio(audio/mp3vsaudio/mpeg,audio/m4avsaudio/mp4).One thing worth flagging for whoever picks this up, since it affects the scope: the path in this report is the v1 location. On
main(v2) the file has moved tosrc/mcp/server/mcpserver/utilities/types.py—fastmcpno longer exists. So the fix needs landing on both lines (thev1.xmaintenance branch and v2), and a v2-only fix would leave v1 emitting the bad types.Suggested shape, if useful: a small module-level suffix→MIME map that both branches share, plus
mimetypes.guess_type()as the fallback for formats outside the table, soformat=andpath=cannot disagree by construction rather than by convention. Happy to also add the regression tests (format="jpg" == path=".jpg") if this gets assigned.I don't want to open a PR blind —
AGENTS.mdcloses PRs that aren't linked to an issue assigned to the author, so assigning this to me first is the efficient path.
Initial Checks
Release Line
2.x (current stable). The same code exists on
v1.x(src/mcp/server/fastmcp/utilities/types.py).Description
ImageandAudiocompute the MIME type two different ways. Withpath=they use a suffix table (.jpg -> image/jpeg,.mp3 -> audio/mpeg,.m4a -> audio/mp4). Withformat=, whichdocs/servers/media.mdsays to pass whenever you usedata=, they just dof"image/{format}"/f"audio/{format}". So the same name gives different results depending on which keyword it arrives through, andformat=puts non-registered types on the wire:Image(data=..., format="jpg")->mimeType: "image/jpg"(should beimage/jpeg)Audio(data=..., format="mp3")->audio/mp3(should beaudio/mpeg)Audio(data=..., format="m4a")->audio/m4a(should beaudio/mp4)This is easy to hit: switching a tool from
Image(path="chart.jpg")toImage(data=buf.getvalue(), format="jpg")silently changes the wire output fromimage/jpegtoimage/jpg, and a consumer that validates image MIME types rejects it.Expected:
format="jpg"andpath="x.jpg"give the same registered type.Proposed fix: route
format=through the same table thepath=branch already has, keepingimage/<format>/audio/<format>as the fallback for names outside it (soformat="bmp"etc. are unchanged). I have this fix with parametrized tests ready locally (100% branch coverage ontypes.py,ruff/pyrightclean) and would like to open the PR if this is accepted.Example Code
Python & MCP Python SDK
Disclosure: found, reproduced and the draft fix tested locally with the help of an AI coding agent (Claude Code); I reviewed the report before filing it.