ManagedCode.Prostir.Cli.win-x64 0.5.261027.4

Prostir CLI

dotnet prostir connects this computer to Prostir and runs coding tasks here. Install it once with the .NET 10 SDK, then start it in your project folder:

dotnet tool install --global ManagedCode.Prostir.Cli
cd my-project
dotnet prostir

The first run opens your browser: sign in, pick this computer (or add it), done. The terminal then shows a live dashboard and waits for tasks. Every later run is just dotnet prostir again. On a computer without a browser, copy the one-line command from the worker page instead: dotnet prostir wc1.… (it works once and expires in a few minutes). dotnet prostir logout disconnects the computer.

Ctrl+C and SIGTERM use the standard .NET ConsoleLifetime to stop the command and finish Worker cleanup without reporting cancellation as a CLI failure.

On Worker startup, a background check looks for a newer stable ManagedCode.Prostir.Cli release on NuGet. If found, the activity log shows the available version and update command. The check has a five-second timeout, does not delay connection, and installs nothing. Feed failures are ignored. Help, version and MCP commands do not run this check.

The folder you start in is the folder tasks work in; use dotnet prostir --workspace <path> for another one. Claude Code, Codex, Gemini and installed GitHub Copilot are found automatically on your PATH, including when Node.js/npm or Bun is installed. When Claude, Codex or Gemini is missing and Node.js/npm or Bun is available, turning it on in Prostir installs the qualified version privately. Every installed CLI version is allowed; missing features are reported by the actual CLI operation. Installed tools still undergo SDK readiness checks; Copilot reads your available account models through its official SDK. Turning a tool on is also the permission to use its models — there is no separate approval step.

The dashboard on stderr shows the connection status and uptime, the Worker name and your account, the folder, system and version, machine CPU and memory, what prostir itself uses, the processes it started, coding tools, the running task (model, elapsed time, steps, current activity and the latest output), recent results and an activity log. Redirected output switches to plain timestamped lines automatically. Missing Claude Code, Codex and Gemini installations show terminal shortcuts [1], [2] and [3]. Press the shown key to install the SDK-verified official package through locally detected npm or Bun. Progress stays in the tools section. Installation does not enable task execution: Worker provider permissions still apply. If no supported package manager is available, the dashboard shows the prerequisite and official setup link. [C], [X], [G] and [P] open installation/update guides for Claude Code, Codex, Gemini and Copilot, tailored to macOS, Windows or Linux. [O] opens the official instructions in the default browser and Escape returns to the dashboard. Guide commands are displayed for manual execution with the original installation method; they are never executed automatically.

Headless hosts can start with an exact Worker credential from redirected stdin (dotnet prostir --credential-stdin) or PROSTIR_WORKER_CREDENTIAL; dotnet prostir <code> --session-only connects without saving the connection.

Optional local settings for Ollama, llama.cpp, MLX, ONNX, Copilot or provider overrides are imported through stdin; an entry replaces the discovered one for the same provider:

cat <<'JSON' | dotnet prostir config import --stdin
{
  "providers": [
    {
      "provider": "Ollama",
      "models": ["llama3.2"],
      "enableExecution": true,
      "localModel": { "endpoint": "http://127.0.0.1:11434/v1/", "probeMaxOutputTokens": 8 }
    }
  ]
}
JSON

Each task runs through its own grain in this device's independent local Orleans silo and gets a fresh SDK session. Credentials live only in the current-user OS vault. There is no local command journal or restart replay; process loss leaves interrupted/unknown execution and never grants another inference invocation.

The composed tool is still being qualified. Application deploy and Manual deploy both restore CLI with Prostir.slnx, then build and verify framework-dependent .NET 10 CLI RID packages and the common tool pointer in the solution Build job. After successful API promotion, the shared release stage publishes those package artifacts and verifies public NuGet installation. Matching-OS Worker, vault and MCP behavior remains a separate qualification lane. Native AOT is optional research and does not block ordinary tool publication.

Ordinary solution builds and tests skip the Copilot SDK's native runtime download. CLI distribution commands explicitly pass --property:CopilotSkipCliDownload=false so published tool packages contain the pinned runtime. Running Copilot from a local build requires that opt-in or an explicitly configured compatible local executable.

CLI release versions match the application deployment's run number and exact commit. For example, base 0.4 from root Directory.Build.props and deployment run 20261007.3 produce 0.4.261007.3 for all four RID packages, the tool pointer and dotnet prostir --version. No CLI version or separate manual release is needed. Publication uses the existing Prostir variable group's secret nuget.

For system MCP over HTTP, run dotnet prostir mcp install --client codex, --client claude or --client gemini; each uses the provider's SDK-resolved native launch command and verifies the exact registration. Authenticate with the client's supported sign-in flow. dotnet prostir mcp login --token-stdin verifies and stores an explicit Prostir bearer for the optional stdio proxy. Proxy stdin and stdout are reserved for MCP framing. For a session-only proxy, provide PROSTIR_MCP_BEARER_TOKEN and run dotnet prostir mcp proxy --session-only.