Re-encodes WebP attachment images to PNG so alpha-channel images load on LLM backends that cannot decode WebP (e.g. llama.cpp).
Install
# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)
dsh plugin --profile web add github:MrGnUm/dsh-image-webp-to-png
Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).
README
A DeepSeek Harness (DSH) bundle that fixes the 400: Failed to load image or audio file error you get when sending alpha-channel images to a backend that
cannot decode WebP — most notably llama.cpp (its vision decoder is stb_image,
which has no WebP support).
It is an installable plugin: it lives in your profile, is re-applied on every DSH
start, and survives DSH updates without touching the root-owned
/usr/lib/node_modules/@deepseek-ai/dsh/node_modules/@deepseek-ai/dsh-attachment-local
file.
Why images fail
DSH normalizes every attached image in dsh-attachment-local. The codec choice is
hard-coded (encodingLadder):
const mediaType = hasAlpha ? "image/webp" : "image/jpeg";
So any image with an alpha channel is re-encoded to WebP. (A clean 8-bit sRGB image with no metadata that fits the size/limit budget passes through as-is; opaque images become JPEG.)
- Opaque → JPEG → llama.cpp decodes it fine.
- Alpha → WebP → llama.cpp's
stb_imagecannot decode WebP → the request is rejected with400 {"code":400,"message":"Failed to load image or audio file"}and the server logsE mtmd_helper_bitmap_init_from_buf: failed to decode webp buffer.
There is no configuration knob in dsh-attachment-local to change this codec,
and the server side (llama.cpp) is not going to add a WebP decoder, so the fix
belongs on the client.
How the bundle fixes it
The adapter (e.g. dsh-llm-pi-ai, used by your fedotov provider) builds the
outgoing request by calling the attachment seam:
attachments.readImageRequest(ref, target, signal) // -> { data, mediaType, width, height, ... }
This bundle wraps that one seam method at DSH start (idempotently, guarded by a
Symbol). Whenever the underlying store returns a WebP request-image, the
wrapper transparently re-encodes those bytes to PNG:
- PNG preserves the alpha channel (JPEG would flatten it).
- PNG is decodable by every backend (llama.cpp, OpenAI-compatible, …).
- Dimensions are unchanged (pure format conversion, no resize).
- The result is cached by
variantId, so a given image is re-encoded once.
Everything else passes through untouched (JPEG stays JPEG, PNG stays PNG).
The wrapper is fail-open: if sharp is missing, a decode fails, or the seam is
not ready at load time, the original WebP result is returned unchanged and DSH
keeps working exactly as before. This bundle can never block or fail a request.
Files
dsh-image-webp-to-png/
├── package.json # bundle manifest (dsh.bundle.patch)
├── cordis.patch.yml # self-declares the entry (id: image-webp-to-png)
├── lib/index.mjs # the bundle: wrap(attachments) + webpToPng()
├── install.sh # install into a profile (default: web)
└── uninstall.sh # remove from a profile
Install
Requirements: pnpm on PATH, and a DSH profile (default ~/.dsh/profiles/web).
cd dsh-image-webp-to-png
./install.sh # -> ~/.dsh/profiles/web
# or a different profile:
DSH_PROFILE_DIR=/path/to/profile ./install.sh
install.sh (idempotent, backs up package.json first):
- Adds
dsh-image-webp-to-pngas alink:dependency pointing at this directory. - Adds it to the profile's
dsh.profile.bundlesarray. - Runs
pnpm installin the profile to create the symlink + update the lockfile.
The fix takes effect on the next DSH web session (immediately if the profile
has patchReload: "live").
Install from GitHub
Repo: MrGnUm/dsh-image-webp-to-png
(private). From any machine with pnpm, git, and access to the repo — HTTPS
clone works with the gh credential helper (no token prompt):
git clone https://github.com/MrGnUm/dsh-image-webp-to-png.git
cd dsh-image-webp-to-png
./install.sh # -> ~/.dsh/profiles/web
Then restart the DSH web session (or it hot-reloads if the profile has
patchReload: "live"). Remove later with ./uninstall.sh.
Verify
Send an alpha-channel image (a PNG with transparency) to your fedotov model
and ask it to describe it. Before the bundle: 400 … Failed to load image or audio file. After: the model describes the image.
You can also watch the DSH logs for the one-time line:
dsh-image-webp-to-png: wrapped attachments.readImageRequest (WebP -> PNG).
Revert
./uninstall.sh # clean removal
# or, using the backup install.sh made:
./uninstall.sh ~/.dsh/profiles/web/package.json.bak-<timestamp>
For the admin (server side, optional)
The client-side fix above is sufficient. If you would rather fix it at the source,
the root cause is the hard-coded hasAlpha ? "image/webp" : "image/jpeg" in
dsh-attachment-local/lib/index.js (encodingLadder, and the matching
pipeline.webp(...) branch in encode()). Upstream has no option to emit PNG for
alpha images; a small patch adding a PNG branch (and bumping the
request-image-v6 transform version so stale WebP cache entries are regenerated)
would remove the need for this bundle.
Compatibility notes
- Works with any DSH profile that mounts the local attachment store
(
LocalAttachmentStore). If a future DSH changes the seam method name (readImageRequest) or its return shape, the wrapper degrades to a no-op (fail-open) and the original behaviour resumes. sharpis resolved from the DSH installation (/usr/lib/node_modules/@deepseek-ai/dsh/node_modules/sharp); this bundle has no runtime dependencies of its own.
Links
More in this category
yjh051108/dsh-routing-suite★ 6997
One repository, three parts: a runtime injector for DSH plugin packages (inject, hot-reload, unload, promote a dev staging tool to the front, route self-heal, plus a settings-page plugin manager that lists, unloads and drags folders in to internalize), a task-aware reasoning-mode router agent preset (router-standard / router-spec / router-react), and a graded two-level task protocol whose six tools (commit_star, lock_stage, revise_do, edit_plan, mark_task, redteam_verdict) pin task state to disk. The injector implementation ships in-tree, so the install carries its own behaviour rather than a dependency list.
strukto-ai/mirage#dsh★ 3696
Swaps the filesystem and bash providers for a mirage virtual workspace: file tools and shell commands run over mounted resources (RAM, S3, Redis, Slack, Gmail, Notion, Postgres) instead of the host disk, with per-mount read/write/exec modes, per-command sandbox routing (monty, pyodide, quickjs in process; docker, e2b, daytona remote), and installed CLIs (git, gh, slack, linear, ntn, gws, or one you register) as head words in the virtual terminal.
hust-open-atom-club/oh-dsh★ 322
Community distribution: TUI, desktop, and Web UI as one bundle with layered installation.
weijiafu14/pi2dsh★ 217
Pi Host ABI compatibility engine: after one install, unmodified Pi extensions from npm mount as native DSH plugins with `dsh plugin add <pi-package>`. Verified end to end on stock DSH with pi-mcp-adapter (full MCP manager: OAuth, resources, prompts, MCP Apps, elicitation, sampling), @tintinweb/pi-subagents, pi-code, pi-hermes-memory and pi-background-tasks; `pi2dsh inspect` reports a package's compatibility before installing.
Fishquito7/dsh-skill-mcp-panel★ 203
Manages DSH skills and MCP servers from the web settings: skill cards with hot enable/disable, workspace scopes, groups, batch migration and drag-and-drop import, plus stdio/HTTP MCP CRUD with connection tests, secret redaction and the unified dsh-panel CLI.
MichengAI/dsh-skills-manager★ 194
Manages local DeepSeek Harness skills in Settings and shows shared Agent skills as read-only.
Community comments
Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.