dsh-verify-gate

English | 中文

An open-source DeepSeek Harness plugin that requires the agent to run the workspace’s verification commands before it can conclude that work is done.

Pass/fail is decided only by process exit codes. A successful run writes an auditable receipt. After that, any successful mutating tool (for example edit, write, or bash) invalidates the receipt, so verification must pass again before conclude.

Install

Local path:

dsh plugin --profile web add /absolute/path/to/verify-gate

Or from GitHub (the repo includes the built lib/ entry points):

dsh plugin --profile web add github:blueWhalei/dsh-verify-gate

Restart Web, open the target workspace, then:

1. Run /verify run, or have the model call verify_run
2. Call verify_conclude only after every command exits 0

How commands are chosen

Resolution order:

1. Non-empty config.commands always wins
2. Otherwise auto-detect (on by default) requires exactly one common workspace marker:
   • package.json with scripts.test → pick the package manager from the lockfile / packageManager, then run <pm> test
   • go.mod → go test ./...
   • Cargo.toml → cargo test
3. If nothing matches, or more than one ecosystem matches → fail with an error that asks you to set commands, instead of guessing

For Python, Java, and other stacks with many local conventions, set commands explicitly.

Custom config

Change config only when the default is wrong. Put this in the profile’s cordis.patch.yml (it replaces that plugin’s entire config):

- id: dsh-verify-gate
  name: dsh-verify-gate
  config:
    commands:
      - id: test
        run: pytest -q
    autoDetect: true
    sandboxMode: danger-full-access

sandboxMode defaults to danger-full-access so package managers can use a global store. If you tighten the sandbox, confirm the verify commands still run for real.

Everyday entry points

Receipts are stored under <workspace>/.dsh/verify-receipts/. Most application repos should ignore that directory.

License

MIT