DeepSeek Harness Troubleshooting: 12 Common Errors

DeepSeek Harness (dsh) has no public issue tracker — GitHub Issues are disabled on the official repository, and the README points users to GitHub Discussions instead. That means the most reliable record of real-world problems and fixes lives in Discussions threads, not a curated FAQ. Below are twelve issues pulled directly from those threads, with what actually causes each one and how it was resolved.

Why this list looks the way it does

gh issue list against deepseek-ai/deepseek-harness returns "the repository has disabled issues" — the API confirms has_issues: false. The project's only official feedback channels are GitHub Discussions (across Q&A, General, Ideas, and Show Your Plugins! categories) and a Discord community linked from the README. If you hit something not covered here, Discussions is the place to search or ask, not Issues.

The 12 errors

#	Symptom	Root cause	Fix	Source
1	`npx @deepseek-ai/dsh web` fails on Arch Linux, plugin tree fails to load, error points to `node-pty`	npm doesn't trust `node-pty`'s install script to run its own compile step	Switch to `bun add --global @deepseek-ai/dsh` (bun trusts it by default), or manually `npm install -g node-pty` and clear `~/.npm/_npx` before retrying	Discussion #49
2	`dsh web --host 0.0.0.0` errors out	Not a bug — the docs state this is "intentionally not supported yet for safety," since it would expose remote code execution to the network	Use `127.0.0.1` (the default); if you need remote access, put it behind your own authenticated reverse proxy or tunnel rather than binding to all interfaces	Discussion #76
3	Global install (`pnpm add -g @deepseek-ai/dsh`) crashes on `dsh --profile headless --help` with `Cannot find package '@deepseek-ai/cordis-plugin-timer'`	Reproduced on Windows 11 + Node 24 + pnpm 11.9; points to a dependency-tree resolution gap	Community reports it as resolved in later installs, but no public patch details were confirmed at the time of this writing — if you hit this, try a clean reinstall with a current pnpm version first	Discussion #55
4	Windows: "add workspace" fails with `directory picker failed: win32 folder dialog worker exited before reporting a result`	The native directory picker depends on the `koffi` native binding, which fails to install	(1) Reinstall with `npm i -g @deepseek-ai/dsh --ignore-scripts --registry=https://registry.npmjs.org` to skip the script that crashes; (2) edit `~/.dsh/profiles/web/cordis.patch.yml` to disable the native/auto directory picker and insert `directory-picker-browse` plus its client UI package instead	Discussion #30
5	`npx @deepseek-ai/dsh web` fails with `The requested module 'node:zlib' does not provide an export named 'createZstdDecompress'`	Node version too old (`createZstdDecompress` needs Node ≥22.15; dsh targets 22.19+/24+)	Upgrade Node to a supported version	Discussion #100
6	dsh prints `failed to load .env: EISDIR` on every startup	A directory (not a file) named `.env` exists in the working directory; two separate `.env`-loading code paths read the path without first checking it's a regular file	Officially acknowledged as a bug with a known reproduction commit; rename or remove the `.env` directory as a workaround until it's patched	Discussion #71
7	`pnpm run build` from source fails with `[ELIFECYCLE] Command failed with exit code 1`	The `build` script internally mixes `npm run` and `pnpm` calls, which conflicts with the pnpm-only workflow the README describes	No other package manager is officially supported — stick to pnpm end to end rather than mixing npm and pnpm invocations	Discussion #86
8	A custom vision model (Claude/GPT-family) can't process images — "image not recognized"	The custom provider's model entry in `settings.yaml` doesn't declare `input: [text, image]`; undeclared models are treated as text-only and attachments are rejected before the request is even sent	Add the `input` field to the model definition (or use `modelOverrides` for a built-in-directory provider)	Discussion #112
9	"I want long-term memory"	dsh has no built-in memory capability — this is the most-discussed open Idea in the repo	There's no official fix; the community path is a third-party MCP memory server (see the official `examples/mcp-memory/` reference configs) or a memory-focused plugin	Discussion #14
10	First GitHub-sourced plugin install fails immediately with a build-permission error	pnpm 10+ blocks `prepare` scripts on git dependencies by default	Add the package to `allowBuilds: true` in the profile's `pnpm-workspace.yaml` — see our dedicated GitHub install guide for the full walkthrough	`apps/cli/reference/README.md`, "Plugin management"
11	An archived session can't be viewed or restored	No documented design for restoring archived sessions was found in the official docs at the time of this report	Treat archiving as effectively final until the docs say otherwise; don't rely on it as a recoverable state	Discussion #40
12	Chinese-language workspace paths get truncated or aren't supported when selecting a workspace on Windows	Two independent reports of Windows-specific path-handling bugs with non-ASCII (Chinese) path names	No official fix confirmed as of this writing; avoid non-ASCII characters in your workspace path on Windows if you hit this	Discussion #47, Discussion #107

The two you're most likely to hit

#2 (host 0.0.0.0) and #10 (allowBuilds) come up constantly because they're not really bugs — they're deliberate friction. The --host 0.0.0.0 restriction exists because dsh's Web UI, once reachable from the network, is remote code execution as a service; there's no config flag that safely bypasses this, only proper network isolation (a reverse proxy with its own auth, a VPN, an SSH tunnel) in front of it. The allowBuilds prompt exists because installing a plugin runs code on your machine, and pnpm 10+ won't do that silently. If you're running dsh in a team setting, both of these are worth reading about in more depth in our guide to running DeepSeek Harness in a team.

When the fix isn't here

This list draws specifically from dsh-tutorial-facts research current as of August 2026. Because dsh is in active developer preview with frequent releases and no formal changelog, new issues surface (and old ones get fixed) faster than any single article can track. If your error isn't above:

Search GitHub Discussions directly — it's the actual source these fixes came from.
Check whether it's specifically a plugin installation problem rather than a runtime one — see our narrower plugin install errors and fixes guide, which covers pnpm-not-found, allowBuilds, and missing build output in more depth.
Run dsh --profile <name> --dump-config to rule out a misconfigured patch layer before assuming it's a dsh bug.

FAQ

Where do I actually report a new bug?

GitHub Discussions, not Issues — Issues are disabled on the official repository (has_issues: false). The README also links a Discord community as a secondary channel.

Is dsh's version stable enough that these fixes stay accurate?

dsh is explicitly in developer preview with no SemVer promise and no GitHub Releases; the README states plainly that there will be compatibility-breaking changes. Treat every fix above as accurate as of August 2026 and verify against current Discussions if you're on a much newer release.

Why does `.env` as a directory break dsh (#6)?

Two separate code paths that load .env files don't check whether the path is a regular file before reading it, so a directory accidentally named .env triggers a Node EISDIR error. It's an acknowledged bug with a known reproduction, not user error.

For plugin-specific installation failures, see DeepSeek Harness plugin install errors and fixes. For the GitHub install path specifically, installing plugins from GitHub covers allowBuilds in depth. If you're setting dsh up for a team rather than yourself, running DeepSeek Harness in a team covers the --host restriction and sandbox/telemetry configuration that matters at that scale.