跪拜 Guibai
← All articles
Artificial Intelligence

DeepSeek Harness's 35k Plugin Count Is a Mirage: Only ~1k Actually Install

By 狒狒AI ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The gap between claimed and working plugins signals a governance vacuum that costs users real time and security: permissions are effectively unchecked, plugins conflict, and upgrades wipe configs. For anyone building atop an agent shell, the lesson is to audit the ecosystem you adopt and treat plugin counts as marketing, not engineering metrics.

Summary

DeepSeek Harness's plugin ecosystem claims 35,000 third-party plugins, but that number does not survive contact with reality. A manual inspection of 11,000 repositories found only about a thousand that can actually be installed per the project's own rules. The official announcement on November 5 said roughly 35k, dshmarket lists 10,778, GitHub's `dsh-plugin` tag shows 31,259, and a Leiphone screening found 1,883 public rejections, 93% due to "cannot install per dsh rules." The root cause is that there is no authoritative definition of a plugin: the only guidance is to add the tag to your repo, with no approval or actual connection to dsh. The result is a pile of mislabeled repositories, including a 2020 resume generator ranking fourth in best-match search and 505 repos named `deepseek-harness-desktop`.

The ecosystem's lack of governance extends to permissions and stability. A probe plugin succeeded in all nine tests even under the read-only tier, reading `~/.ssh` files and API keys from environment variables, writing to `/tmp`, and connecting to the internet. Plugins bypass permission checks because they are the shell itself, making the tier system effectively decorative. Real-world failures are common: the top ten plugins by downloads conflict when installed together, a version upgrade wiped 32 plugin entries and shrank configuration from 34 to 2 items requiring five hours of manual repair, and installing plugins invalidates the KV cache with no user-facing warning. The only practical advice is to check source, dependencies, permissions, and backups before installing anything.

Takeaways
— Official claim: ~35k third-party plugins; dshmarket lists 10,778; GitHub tag shows 31,259; a scan of 11k repos found 1,883 public rejections, 93% 'can't install per dsh rules'.
— Only 2 plugins actually replace the main loop; 69 plugins swap models (e.g., connecting Codex, Claude, Grok, Kimi subscriptions into DeepSeek's harness).
— dshmarket category counts: UI enhancements 2,150, tools and capabilities 1,478, Just for Fun 354, security and permissions 279.
— A probe plugin under read-only permissions successfully listed ~/.ssh files, read API keys from env, wrote to /tmp, and connected to the internet.
— A Windows upgrade from v0.2.1-rc.5 to v0.2.2-rc.1 lost 32 plugin entries, shrank configuration from 34 to 2 items, and required ~5 hours manual repair.
— Top ten plugins by downloads conflict when installed together, forcing two to be disabled.
— billion-context (2.7M weekly downloads) claims 1/5 context compression, but on a two-turn session recommends not compressing due to fixed overhead.
— dsh-cost-meter recorded a session costing 0.11 yuan, with ~90% spent on 'AI execute command' step and ~10% on direct model replies.
— Installation invalidates the KV cache; DeepSeek wrote this into 223 package READMEs but has no ecosystem-wide warning.
— dshmarket supports backup via JSON, WebDAV, or Gist; a package name typo (dsh-market) prevents installation.
— cc-safety-net blocks git reset --hard and git restore, but does not block rm -rf backend.
— archify takes 16 minutes and costs 0.8 yuan per diagram because it repeatedly screenshots to verify correctness.
Conclusions

The lack of a plugin definition means the ecosystem's size is mostly decorative; the genuinely functional subset is tiny and dominated by utilities that provide visibility or safety, not novel capabilities.

The permission model is fundamentally broken: plugins bypass the tier checks because they are the agent shell itself, making read-only mode meaningless and turning installation into unconditional consent.

Category data shows most plugin usage is cosmetic (UI enhancements outnumber capability extensions), suggesting users prioritize interface customization over agent intelligence.

The cost structure revealed by dsh-cost-meter indicates optimization should target tool call frequency and return payload size, not model pricing, which is counterintuitive but concrete.

Concepts & terms
dsh (DeepSeek Harness)
An open-source agent shell that claims 'everything is a plugin,' allowing users to swap out models, tools, sessions, sandbox, UI, and even the agent main loop. The command-line interface uses the command `dsh`.
KV cache
A key-value cache that speeds up inference by storing previously computed attention keys and values. In dsh, installing plugins invalidates the entire KV cache, degrading performance until rebuilt.
danger-full-access
A permission tier in dsh that grants full access to the local system. It is required on Windows because there is no official sandbox backend, and it disables most protective checks, effectively allowing plugins to do anything the user account can do.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗