dsh-playwright-native

In DeepSeek Harness, the canonical path to expose a tool to an agent is ctx.tools.register(defineTool(...)). Many plugins choose to decompose a complex CLI into a set of semantic sub-tools, each with its own custom schema, so the agent ends up calling an interface wrapped by a third party. dsh-playwright-native takes the opposite approach: it registers a single passthrough tool named playwright that forwards the agent-supplied args array verbatim and in order to the local playwright binary. No install, test, or show-report splitting. No parameter remapping. What the agent types is exactly what executes.

The implementation rests on just two pillars: args are concatenated as-is onto the command line, and execution runs through ctx.shell. This seam is the same one used by the official bash and pwsh tools, so it inherits sandbox policy, DSH environment variables, output truncation, and timeout/abort classification automatically. A non-zero exit code is returned as an [exit code: N] marker, letting the model decide what to do next. When a playwright test worker fork or a playwright install browser download is blocked by the file sandbox, the plugin retains the sandbox_permissions escalation path, retrying with the narrowest broader scope and prompting for approval before re-execution.

This plugin suits DSH workflows where an agent needs full lifecycle control over Playwright: running E2E test suites, installing browsers, codegen recording, taking screenshots, and opening HTML reports. Every operation flows through native CLI arguments, unconstrained by a custom schema, making behaviour transparent and predictable.