mcpField
A
assign

assign

github.com/prathamdbyAgent workflows6367 installsUpdated 4d ago
Repository

About assign

Assign is a skill for running a supplied task through opencode, codex, or claude in a selected working directory. It handles prompt transport, process monitoring, cleanup, and verification before reporting the outcome.

Key features

  1. 01Select agent, model, and directory
  2. 02Validate prerequisites and registry entry
  3. 03Pass task through a temporary prompt file
  4. 04Track process output and status
  5. 05Clean up and verify results

Use cases

  1. 01Running a task with a selected external agent
  2. 02Monitoring a non-interactive agent process
  3. 03Checking code changes and requested test evidence

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/prathamdby/skills --skill assign

After installing, type/assignin Claude Code to confirm it loaded.

Bundled filesREFERENCE.md

Details

From the project README

Assign

Flags

FlagDefaultEffect
--agent <name>opencodeopencode, codex, or claude
--model <model>registry defaultOverride the selected agent's model
--dir <path>current directoryExternal agent working directory

If --agent names an unsupported agent, stop: "Unknown agent <name>. Supported agents: opencode, codex, claude. See ./REFERENCE.md."

1. Preflight

Require a non-empty positional task, an existing working directory, the agent executable, and the registry entry in ./REFERENCE.md. For Codex outside a git repo, use the registry's non-git branch.

Record: agent/model | dir | prompt path | process | last progress | exit | terminal.

Done when inputs and the exact non-interactive command are resolved, or a specific missing executable, directory, or auth blocker is reported.

2. Create transport

Create a collision-safe prompt file with restrictive permissions using the system temp directory. Write the task bytes unchanged. Never pass task text as a shell argument. Register cleanup before launching so success, error, signal, or interruption removes the prompt.

Done when the prompt can be read back byte-for-byte and parallel assignments cannot share its path.

3. Execute and monitor

Run the selected registry command through a tracked process and pipe the prompt on stdin. Capture stdout, stderr, process ID, and exit code. Monitor output, process state, and target-directory changes. Silence alone is not proof of a hang. Inspect the process after a silent interval; terminate only when it is waiting for input despite non-interactive flags, the user cancels, or the execution deadline is reached. Use the reference's termination sequence.

Done when the process exits or is terminated, with no orphaned child process.

4. Verify and report

Always run cleanup, then confirm the prompt file is absent. Non-zero exit, spawn failure, timeout, or empty output is not success. For code work, inspect the resulting diff and requested test evidence; do not trust the CLI summary.

Report the agent, model, directory, exit code, observed result, verification, and one terminal value: SUCCESS, FAILED, TIMED_OUT, or INTERRUPTED.

Delegate exactly the supplied task. Do not add commit, push, build, or scope unless the task itself asks for it.

Comments

Markdown supported