dsh-enhance

简体中文

A toolkit of native plugins that fill the gaps of the DeepSeek Harness / DeepSeek models: capabilities the official stack does not have or does poorly, as one composition row each, configured entirely from the Web settings UI. Installation is one command (./install.sh); there is no npm publishing and no manual YAML editing.

What you get

Both plugins are host-side Cordis plugins: one fetch per tool call, per-call timeouts, caller-cancellation aware, API keys marked secret and redacted on the wire.

Quick start

1. Prerequisites — dsh on PATH, npm, pnpm, Linux with GNU coreutils (GNU sed is detected explicitly). Node >= 22.18 for the repo tooling (--verify / npm test); the plugins themselves run on Node 18+.
2. Install everything:

   ./install.sh

   Idempotent — safe to re-run. It packs both plugins, installs them into the web profile, mounts the two rows in the profile's patch layer (every session of that profile gets the tools), creates the enhance agent preset (shipped standard preset + built-in web_search disabled) and sets it as the default for new sessions. No restart is needed when dsh web is already running — the patch layer hot-reloads; otherwise just start dsh web. (./install.sh --restart performs a managed restart itself; see bin/restart-web.sh.)

3. Configure the tools in the Web UI: Settings → Plugins → dsh-vision / dsh-native-web (field-by-field guidance below). Until configured, the tools fail on call with a clear "not configured" error — expected.
4. Verify end to end: open any session and send the self-test prompt.
5. Check the regression suite (optional but recommended):

   ./install.sh --verify   # or: npm test

How install.sh works

Every step detects its own previous result and becomes a no-op (an installed profile short-circuits pack+install entirely):

1. hard-checks dsh / npm / pnpm / realpath / GNU sed — missing tools fail loudly, never skip;
2. packs both plugins and installs them with dsh plugin --profile <name> add — the real production install path; tarballs are kept under $DSH_HOME/enhance-pkgs so the profile's file: dependencies keep resolving across reinstalls (moving $DSH_HOME afterwards requires re-running install.sh);
3. appends the two mount rows to the profile's cordis.patch.yml — pure mount points; a half-mounted state self-heals (only the missing rows are added);
4. copies the installation's own shipped standard preset to the enhance agent preset with the built-in web_search disabled, and sets it as the default for future sessions. It deliberately does not copy the cordis preset: two cordis-family presets in one process collide on the host inspect provider;
5. optional --restart (web profile only, managed via bin/restart-web.sh) and --verify (runs npm test).

Configuration

Everything below is edited in the Web settings form (Settings → Plugins); changes apply on the next tool call, no restart.

dsh-vision

dsh-native-web

The web instance itself (docker compose, see firecrawl/firecrawl) must be running: the tools resolve it lazily and fail with a clear "no local web instance detected" error on the first call if it is down.

Mounting the rows (manual reference)

install.sh uses placement A automatically. The two placements are equivalent — pick one if you install by hand (the rows are pure mount points; copy from presets/vision.cordis.yml and presets/native-web.cordis.yml):

A. Profile patch layer — every session of that profile, from the start. Append an insert patch to the profile's cordis.patch.yml:

- insert:
    - id: vision
      name: '@vcxmug/dsh-vision'
    - id: native-web
      name: '@vcxmug/dsh-native-web'

The patch layer is hot-reloaded by a running dsh, survives reboots, and needs no preset picking or session restarts.

B. One agent preset — scoped to sessions on that preset. In the Web UI: Settings → Agent presets — create a preset (or duplicate an existing one) and add the rows:

- id: vision
  name: '@vcxmug/dsh-vision'
- id: native-web
  name: '@vcxmug/dsh-native-web'

Two cautions: config: keys in either placement act as the settings form's base layer; and do not duplicate a preset that registers first-party inspect providers (the cordis creative preset does) while sessions on the original run in the same process — both mounts register the same process-global providers and the second mount is rejected.

Updating & uninstalling

Updating the plugins: rebuild/pack a new tarball version and add it to the profile — dsh plugin --profile <name> add <new.tgz> updates the recorded file: spec. Then re-run ./install.sh to refresh the rest of the chain.

Uninstalling:

dsh plugin --profile web remove @vcxmug/dsh-vision @vcxmug/dsh-native-web
rm -rf "$DSH_HOME/.agent-presets/enhance"
# and remove the `agent-presets:` block from $DSH_HOME/settings.yaml if you
# no longer want the default-preset override

Requirements

• DeepSeek Harness 0.1.0-rc.6+ (the web profile; agent presets under $DSH_HOME/.agent-presets)
• Runtime (the plugins): Node 18+ (global fetch); repo tooling (tests/typecheck): Node >= 22.18
• Vision: any multimodal model endpoint that speaks the OpenAI Responses API (/responses) — official or third-party endpoints
• Native web: a reachable self-hosted Firecrawl-compatible instance (USE_DB_AUTHENTICATION=false instances accept any API key)

Repository layout

install.sh                  # one-command installer for the whole chain
bin/restart-web.sh          # managed `dsh web` restart with health check
packages/dsh-vision/        # vision plugin (TS in src/, built lib/ committed)
packages/dsh-native-web/    # native web plugin (TS in src/, built lib/ committed)
presets/                    # pure mount-point composition fragments
docs/                       # known limitations and self-test prompt
tests/                      # install-path regression suite + scripted mock LLM

Notes

• Why the plugins carry no runtime dependencies: the packages the Harness installation itself provides (@deepseek-ai/dsh-tools, schemastery, ...) are declared as optional peer dependencies. Installing them as regular dependencies would hoist a second copy into the profile's node_modules, shadow the Harness's own copy for composition rows, and break module-scoped identity (e.g. the dsh-tools tool-runtime Symbol) — every tool call then dies with Cannot read properties of undefined (reading 'prepare') after a cold boot. The regression suite pins this invariant; see tests/install-and-loop.test.ts.
• Native web vs MCP: MCP routes each call through the DSH MCP client and a firecrawl-mcp subprocess — extra hops and session turns. The native route is one direct HTTP call to your instance. (Prefer MCP? The upstream @deepseek-ai/dsh-mcp-client works for that.)
• Known limitations: see docs/known-limitations.md.
• End-to-end verification: docs/self-test-prompt.md.

Testing

npm test runs the install-path regression suite (no API key, no registry for the fixed packages): it packs both plugins with npm pack, installs the real tarballs into a throwaway profile with dsh plugin add (the production command), then asserts the fundamental invariant — the profile's node_modules never duplicates a package the Harness installation already provides — and finally drives a real headless agent tool loop through a scripted mock LLM. Requires Node >= 22.18 (native TypeScript execution, no build step), the dsh launcher on PATH (or the DSH env var), npm and pnpm. A missing tool fails the suite (it never skips): a skipped run would green-light exactly the regression this suite exists to catch.

npm run typecheck checks the test sources; npm run build rebuilds both plugins' lib/ from src/. Both need dev dependencies installed first (npm install at the repo root and inside each package directory).

License: MIT