dsh-ops-plugins

English | 中文

An ops plugin suite for DeepSeek Harness (dsh): it turns a dsh agent into a production-incident investigator — one that resolves kubectl / ceph / ssh credentials by name, runs read-only commands against your clusters, and organizes the investigation as a tree.

It installs as a single npm package, @elinpf/dsh-ops; the granular @elinpf/dsh-ops-* packages arrive as its dependencies, all published in lockstep versions.

Features

• Credentials register paths only; the agent sees profile names — secrets never enter model context
• Read-only by default; read-write needs a per-session grant with human approval, fully audit-logged
• Honest tool output — sensitive paths are scrubbed before anything reaches the model
• The trace tool structures investigations as trees, rendered git-graph-style in the web UI
• Environment inventory scanning, so the agent reasons over what actually exists
• A few methodology lines in the system prompt; full docs pulled on demand

Requirements

• DeepSeek Harness ≥ 0.1.0-rc (verified on 0.1.0-rc and 0.1.1-rc.2)
• pnpm ≥ 10
• kubectl on the host with cluster network reachability; ceph / ssh as needed

Installation

1. Install the package (dependencies and host-plane rows mount automatically):

   dsh plugin --profile ops add @elinpf/dsh-ops
   

2. For the web UI, edit ~/.dsh/profiles/ops/package.json and add the web host to the bundles:

   "dsh": {
     "profile": {
       "bundles": [
         "@deepseek-ai/dsh-base",
         "@deepseek-ai/dsh-web-app",
         "@elinpf/dsh-ops"
       ]
     }
   }
   

   @deepseek-ai/dsh-web-app resolves through the dsh installation; it cannot be installed via dsh plugin add.

If the install fails with a minimumReleaseAge error, add an exclusion to the profile's pnpm-workspace.yaml (matching the current version):

minimumReleaseAgeExclude:
  - '@elinpf/*@0.1.5'

Deployment

1. Materialize the ops preset:

   npx @elinpf/dsh-ops preset install --agents-home ~/.dsh
   

   The harness discovers user presets under ~/.dsh/.agent-presets/; without --agents-home the preset lands in ~/.agents and fails silently.

2. Edit ~/.dsh/profiles/ops/cordis.patch.yml, replacing the top-level array with:

   - id: agent-presets
     config:
       default: ops
   - id: session-reference
     disabled: true
   

3. Start (--no-open skips opening a browser; flags match dsh web):

   dsh --profile ops --no-open
   

4. Register credentials in ~/.dsh-ops/access.yaml — paths and connection parameters only, never secrets. list_access with help: true pulls the format docs; the web admin UI also covers registration.

Verify:

dsh --profile ops --dump-config | grep -A4 'id: agent-presets'      # default should be ops
dsh --profile ops --dump-config | grep -A2 'id: session-reference'  # should carry disabled: true

Then start an ops session in the web UI: list_access lists your profiles, kubectl resolves them, the trace panel renders, rw credential use raises an approval request.

Having an agent install it

Paste this into any dsh session and let the agent run the installation and deployment for you:

Read the README at https://github.com/Elinpf/dsh-ops-plugins,
install and deploy the @elinpf/dsh-ops plugin suite into the ops profile,
then confirm with the verification steps in the README.

Updating

dsh plugin --profile ops add @elinpf/dsh-ops@latest
npx @elinpf/dsh-ops@latest preset install --agents-home ~/.dsh   # the preset is a file on disk — re-copy it
dsh --profile ops --no-open                                       # restart

Use add @latest, not update — update does not cross minors. The preset does not refresh with the package; re-materialize it.

Uninstall

1. Remove the preset:

   npx @elinpf/dsh-ops preset remove --agents-home ~/.dsh
   

2. Remove the package:

   dsh plugin --profile ops remove @elinpf/dsh-ops
   

3. Restart the profile:

   dsh --profile ops --no-open
   

4. Optionally delete ~/.dsh-ops/ — the credential registry, environment inventory, and referenced credential files.

Uninstalling never touches your clusters: credentials are read-only references to files you own.

Security

• Secrets never enter services, logs, errors, or model context
• The access gate's threat model is "prevent mistakes, not malice"
• Design decisions live in docs/adr/; the domain glossary (Chinese) is CONTEXT.md