🧪 dsh-test-drive

• 1024 store channel: npm i -g dsh1024 once, then dsh1024 plugin --profile web add dsh-test-drive (counts toward the deepseek1024.com install ranking).

Isolated install-and-smoke test drives for DeepSeek Harness plugins.

Install, smoke, verify, and clean up in a throwaway profile — your real ~/.dsh stays untouched.

English · 简体中文 · Español · Português · हिन्दी

Compatibility

What you get

• test_drive tool — one target through the complete pipeline: dsh plugin add → --dump-config patch check → headless boot smoke (FAILED-marker scan + optional one-shot task) → optional capability assertion → dsh plugin remove → quarantined cleanup. Returns the structured record synchronously, or { kind: 'background', jobId } with background: true.
• /testdrive command — batch drive of a whitespace/comma-separated target list as a drive-batch background job over ctx.jobs, producing a matrix report (JSON + Markdown).
• drive_report tool — fetch any stored run (tdr_...), matrix (tdm_...), or the latest matrix; rendered as Markdown.
• Capability assertion — beyond "booted and exited": the optional capability stage drives one headless task that calls a named tool (or runs a /command) and verifies the durable session log recorded the invocation and the observed output contains expect. A clean boot is only a smoke test; observed proves a named capability really works.
• Structured results — every record carries the discriminator schema: "dsh-test-drive/v1" with first-class fields: stages.install.status (pass/fail), stages.smoke.status (pass/fail/boot-ok/skipped), stages.capability.status (observed/invoked/not-registered/skipped/failed), per-stage durationMs, sanitized summary/outputTail, and an overall verdict (pass/fail/partial/unknown). This is the machine-readable contract downstream scorers (dsh-score) consume.
• Safety by construction — every temp directory is created by this plugin under a dedicated dsh-test-drive- prefix, tracked in a live ownership registry, and removed only through a dry-run → quarantine-rename → delete ladder. The host profile is never read or written.

Quick start

Git channel

dsh plugin --profile web add github:PerryLink/dsh-test-drive#<commit-sha>

The first add fails because pnpm blocks the package's prepare build; copy the exact key pnpm printed into the profile's pnpm-workspace.yaml and re-run:

allowBuilds:
  'dsh-test-drive': true

npm channel

dsh plugin --profile web add dsh-test-drive

Prebuilt packages need no build allowance. Restart the profile, then use test_drive / /testdrive from a session.

Install & uninstall

dsh plugin --profile web add dsh-test-drive     # install (npm) — or the git form above
dsh plugin --profile web remove dsh-test-drive  # uninstall

Configuration

All keys are optional (defaults shown); invalid values fail loudly at load.

Tools & surfaces

test_drive

test_drive(target: string, headlessTask?: string, background?: boolean,
           capability?: { kind: 'tool' | 'command', name: string,
                          args: string, expect: string })

• target — git spec (github:owner/repo#sha, git+https://...), npm name, local path, or .tgz tarball.
• capability — assertion after the boot smoke: the agent calls name (tool) or runs /name (command) with args; the stage reads the durable session log and requires the observed output to contain expect. Needs DEEPSEEK_API_KEY (host env or forwardEnv); without it the stage is skipped, never failed.
• Returns the full structured record; see the sample below.
• background: true starts a drive-batch job and returns its id.

/testdrive <targets...>

Starts one background batch job; progress streams through the job output, and the final line names the matrix id for drive_report.

drive_report(id?)

Returns a run record (tdr_...), a matrix (tdm_...), or — with no id — the latest matrix.

Structured result sample

{
  "schema": "dsh-test-drive/v1",
  "run": { "runId": "tdr_9f2c...", "startedAt": "2026-08-16T00:00:00.000Z",
           "finishedAt": "2026-08-16T00:00:45.120Z", "durationMs": 45120,
           "harnessVersion": "0.1.1-rc.2", "pluginVersion": "0.1.0",
           "platform": "win32", "node": "v22.22.3" },
  "target": { "kind": "repo", "spec": "github:owner/dsh-click#abc123",
              "resolved": { "packageName": "dsh-click", "packageVersion": "0.1.0",
                            "hasBundleManifest": true } },
  "isolation": { "tempDshHome": true, "tempWorkspace": true, "tempStore": true,
                 "hostHomeTouched": false },
  "stages": {
    "install":   { "status": "pass", "exitCode": 0, "durationMs": 30412, "attempts": 2,
                   "summary": "install ok after allowBuilds allowance", "outputTail": "",
                   "allowBuildsNeeded": true },
    "config":    { "status": "pass", "exitCode": 0, "durationMs": 2310, "attempts": 1,
                   "summary": "dump ok (exit 0)", "outputTail": "",
                   "patchEffective": true, "layers": ["dsh-click"] },
    "smoke":     { "status": "boot-ok", "exitCode": 1, "durationMs": 4123, "attempts": 1,
                   "summary": "booted without loader failures; headless task did not complete (credentials/model unreachable)",
                   "outputTail": "", "bootFailed": false, "taskCompleted": false },
    "capability": { "status": "observed", "exitCode": 0, "durationMs": 8123, "attempts": 1,
                    "summary": "tool \"plugin_vet\" called and its result contains the expectation",
                    "outputTail": "", "capabilityKind": "tool", "name": "plugin_vet",
                    "expectMatched": true,
                    "detail": "tool \"plugin_vet\" called and its result contains the expectation" },
    "uninstall": { "status": "pass", "exitCode": 0, "durationMs": 5123, "attempts": 1,
                   "summary": "remove ok (exit 0)", "outputTail": "" },
    "cleanup":   { "status": "pass", "quarantined": true, "removed": true,
                   "summary": "owned temp root quarantined and removed" }
  },
  "verdict": "pass",
  "verdictReason": "install, patch, boot, and uninstall verified; headless task inconclusive (see smoke.summary)"
}

Verdict rules: install failure, boot failure (smoke.fail), or a capability stage that reached not-registered/failed ⇒ fail; install pass + patch effective + clean boot (pass/boot-ok) + uninstall pass ⇒ pass (with a capability note when observed); anything installed but missing a later assurance ⇒ partial; otherwise unknown.

CI (GitHub Actions)

The repository ships a composite action.yml that reuses dsh-test-drive in any plugin repo. It drives a target in an isolated throwaway profile and emits the report pair CI consumes: Markdown (PR comment) and JUnit XML (test reporter / status check).

# .github/workflows/test-drive.yml
name: test-drive
on: [pull_request, workflow_dispatch]
jobs:
  smoke:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Drive this plugin
        id: drive
        uses: PerryLink/dsh-test-drive@v0.2.4
        with:
          target: github:${{ github.repository }}#${{ github.sha }}
      - name: Publish JUnit
        uses: EnricoMi/publish-unit-test-result-action@v2
        with:
          files: ${{ steps.drive.outputs.junit }}
      - name: Comment the report
        run: cat "${{ steps.drive.outputs.markdown }}"

action.yml inputs: target (required), headless-task (optional smoke task), dsh-version (the dsh CLI spec). Outputs: markdown, junit (report paths), and verdict (pass/fail/partial/unknown). The drive itself stays keyless (install → patch check → boot smoke → uninstall → cleanup); a capability assertion is the only stage that needs DEEPSEEK_API_KEY, and it is skipped, never failed, without one.

The same report pair is available programmatically: renderDriveResult / renderMatrix (Markdown) and renderDriveJUnitXml / renderMatrixJUnitXml (JUnit) are exported from the package root over a settled DriveResult or MatrixRecord.

Permissions & data

• Only public services are consumed: ctx.subprocess, ctx.jobs, ctx.storageDomain, ctx.tools, ctx.commands.
• Reports are stored in the test_drive storage-domain (tables runs, matrices; latest-matrix pointer). When the composition has no storageDomain (e.g. the shipped headless profile), tools still work and report persistence is disabled with a logged reason.
• Child processes inherit a credential-scrubbed environment: host secrets never reach a tested profile unless you explicitly name them in forwardEnv. Values are never logged.
• All report/log strings pass through pure sanitizers: token literals, URL credentials, and bearer headers are redacted, temp-root paths are replaced with <testdrive-temp>, and tails are byte-capped.

Security boundaries

• Isolation. Each drive runs inside a fresh mkdtemp root under the OS temp dir: a throwaway DSH_HOME, a throwaway working directory, and a redirected pnpm store. The tested plugin's code only ever runs in that profile; your host profile is untouched.
• Ownership. A live registry records every root this plugin instance creates. Cleanup refuses anything that is not a registered direct child of the OS temp dir carrying the dsh-test-drive- prefix — no %TEMP% sweeps, no foreign prefixes, no real-home paths.
• Cleanup ladder. Before any mutation the full dry-run plan is logged (absolute paths). Removal renames the root into a dsh-test-drive-quarantine-<ts> directory first, verifies, then deletes; failures leave the directory quarantined and reported, never silently dropped. Cleanup runs in a finally on success, failure, timeout, and abort, and again on plugin teardown.
• allowBuilds is a real permission. Allowing a git package's prepare build executes that package's code at install time. The allowance is scoped to the throwaway profile only, but only test targets you trust, and pin commits.
• Headless smoke is keyless by default. The boot check needs no credentials; completing the one-shot task does. Forward credentials explicitly (forwardEnv) and never log them.

Known limitations

• Installing registry/git targets requires network access from the child dsh/pnpm processes.
• The smoke task needs model credentials to reach pass; without them it reports the honest boot-ok.
• In compositions without storageDomain, reports are not persisted (drive_report fails honestly).
• dsh must be locatable on PATH (or set dshBin); on Windows the npm .cmd/.bat shim is parsed automatically, a bare .ps1 resolution asks for dshBin.
• Batches default to serial execution; raising batchConcurrency shares the pnpm-store disk, not correctness.

Development

pnpm install
pnpm run typecheck && pnpm run typecheck:ci && pnpm test
pnpm run build && pnpm run verify:self-contained && pnpm run verify:artifacts && pnpm pack

• typecheck resolves @deepseek-ai/* through the local harness checkout; typecheck:ci checks against the published 0.1.2-alpha.5 types.
• Tests use the real Context/Session/ToolRuntime/LocalJobRegistry/storage stack with a scripted subprocess provider.
• Real-CLI end-to-end (requires network + dsh on PATH): DSH_TESTDRIVE_E2E=1 pnpm run test:e2e — drives this package's own checkout through the real install-smoke loop.
• Release: node scripts/release.mjs <x.y.z> (bumps, stamps CHANGELOG, re-runs the gate, commits + tags; never pushes).

Topics

dsh, dsh-plugin, deepseek-harness, deepseek, cordis, plugin-testing, install-smoke, compatibility-matrix, ci

Contributors

PerryLink — design and implementation.

PerryLink DSH Plugin Family

This project is one of the 33 DeepSeek Harness plugins maintained by PerryLink. If this one helps you, the others likely will too:

License

Apache-2.0