| name | assign |
| description | assign a task to an external coding agent and monitor it. Pipes the prompt via stdin to dodge quoting failures and auto-approves permissions to avoid silent hangs. Triggers: run a task with an external agent, hand off to OpenCode, Claude Code, or similar, execute a plan non-interactively. Flags: --agent <name> (default opencode), --model <model>, and --dir <path>.
|
Assign
/assign "Fix the null check in auth.ts" --agent opencode --model opencode-go/kimi-k2.6
/assign "Fix the null check in auth.ts" --agent claude
Flags
| Flag | Effect |
|---|
--agent <name> | Agent to delegate to. Default: opencode. |
--model <model> | Model to use. Agent default applies if omitted. |
--dir <path> | Agent working directory. Default: current directory. assign-prompt.tmp is always written to the current directory, not --dir. |
If --agent names an unsupported agent, stop: "Unknown agent <name>.
Supported agents: opencode, codex, claude. See ./REFERENCE.md."
Step 1: Write the prompt file
Write the full task prompt to ./assign-prompt.tmp in the current working
directory (not --dir). Never pass the prompt as a shell argument, quoting
breaks on multi-line or special-character prompts. Overwrite silently.
Step 2: Invoke the agent
Look up the agent's invocation template and required non-interactive flags in
./REFERENCE.md. Those flags are not optional: without them the agent blocks on
an invisible permission prompt and hangs silently. Substitute <model> and
<dir>, then pipe the prompt:
cat ./assign-prompt.tmp | <agent-command>
Step 3: Monitor output
Watch for progress (tool calls, file writes, step completions). If there is no
output for >60s after startup, the agent is hung on a permission prompt. Kill
it and check the agent's troubleshooting notes in ./REFERENCE.md.
Step 4: Verify and clean up
- Report the exit code and a brief summary of what the agent did.
- Delete
./assign-prompt.tmp.
- If the task was a code change, prompt the user to verify results.
Constraints
- Never infer or expand the task. Delegate exactly the prompt provided.
- Never commit, push, or build unless the prompt explicitly asked for it.
