rad-execute

name	rad-execute
description	Run a project through the rad orc execution pipeline. Use it when after /rad-plan has completed and ready to execute or when the user indicates they want to execute or run a project. It can also be used when resuming a rad-orc project that was previouisly executed but not completed.
user-invocable	true

You are an orchestrator. You will use the rad-orchestration skill to drive the execution pipeline. This skill is a thin relay: the execute resolve CLI command does all the classification and convention math and hands back a data envelope; you run only the human beats it flags, then the commands it returns. Do not re-derive run modes, branches, base branches, paths, or settings yourself — the CLI owns all of that.

${CLAUDE_PLUGIN_ROOT}/skills/rad-orchestration/scripts/radorch.mjs is the CLI for every call below. Run discovery calls silently — never narrate raw envelope output to the user.

Step 1: Resolve

Run execute resolve, adding --project <PROJECT> when a project name is available from the /rad-execute argument or the conversation (omit it otherwise):

node "${CLAUDE_PLUGIN_ROOT}/skills/rad-orchestration/scripts/radorch.mjs" execute resolve --project <PROJECT>

Read data from the envelope. The fields you act on: runMode, needsProject, candidates, ask, derived, projectDir, next, and — on unknown — reason.

Step 2: Resolve the project if asked

If data.needsProject is true, no project was determined yet:

• If a project name is obvious from the conversation, re-run Step 1 with --project <NAME>.
• Else if data.candidates is non-empty, ask the operator to pick one (show each candidate's name, tier, status), then re-run Step 1 with the chosen --project <NAME>.
• Else (data.candidates is empty) nothing is execute-ready: tell the operator that /rad-execute needs a project name or one in context, and suggest /rad-plan <PROJECT-NAME> (to plan an existing brainstorm) or /rad-brainstorm (to start a new one). Stop.

Step 3: Run the human beats for the run mode

Every confirm is driven by a data.ask.* flag — never decide one yourself. If a flag is absent, do not ask that question.

First, if data.ask.confirmDone is set, the project is already complete: ask "<data.project> is already marked done — run it again?" If the operator declines, stop. If they accept, continue.

Then branch on data.runMode:

• resume — already settled. Ask nothing. Go straight to Step 4.

• launch — you are at a main clone (or nowhere); a fresh session will be launched into a new worktree. Ask one combined askUserQuestion:

  • Launch flavor (always): Claude Code (auto-accept edits), Claude Code (yolo — bypass permissions), Copilot CLI, VS Code.
  • auto_commit — include the sub-question only if data.ask.autoCommit is set.
  • auto_pr — include the sub-question only if data.ask.autoPr is set.

  No separate plan-approval step: the execute prepare command in data.next confers approval (running /rad-execute is the approval), so the fresh session resumes without stopping at the plan gate.

• in-place — you will drive the pipeline in the current session. Ask one combined askUserQuestion containing only what data.ask flags (a side-project flags nothing → ask nothing):

  • if data.ask.reuseWorktree → confirm reuse: "you're in <data.derived.worktreeName>'s worktree on <data.derived.branch> — reuse it for <data.project>?" If data.derived.missingRepos is non-empty, fold into the same question: "it also needs <data.derived.missingRepos>, which aren't in this worktree — add them?"
  • else if data.ask.confirmHere → confirm location: "you're in <data.project>'s worktree on <data.derived.branch> — run here?"
  • auto_commit / auto_pr — include in the same question only when data.ask.autoCommit / data.ask.autoPr is set.

  If the operator declines a reuse / location confirm, stop. (No approval beat here either — execute prepare in data.next confers it.)

• unknown — stop. Relay data.reason to the operator verbatim and ask how to proceed. Do not guess a run mode or run any next command.

Step 4: Run the next commands

Run every command in data.next in order, each as:

node "${CLAUDE_PLUGIN_ROOT}/skills/rad-orchestration/scripts/radorch.mjs" <command from data.next>

The resolver already substituted everything it knows; substitute only the placeholders it left for your answers — change nothing else:

• {ac} / {ap} → always for a "yes" answer, never for "no" (present only when you asked).

• {flavor} / {pm} (launch only) → map the chosen launch flavor:

  Launch flavor	{flavor}	{pm}	--prompt
  Claude Code (auto-accept edits)	claude	acceptEdits	keep
  Claude Code (yolo — bypass permissions)	claude	bypassPermissions	keep
  Copilot CLI	copilot	drop --permission-mode	keep
  VS Code	vscode	drop --permission-mode	drop --prompt

  On Windows, prefix the Claude/Copilot launch line with MSYS_NO_PATHCONV=1 so the worktree path is not mangled.

After a launch you are done: the fresh session re-enters /rad-execute, resolves to resume against the now-settled state, and drives the pipeline. For in-place / resume, the final pipeline signal --event start returns the current pending action — continue driving the pipeline through the rad-orchestration skill. Plan approval is conferred for you before start (by execute prepare on launch/in-place, or a gate approve plan step on resume), so start advances past the plan gate without a prompt. Commit and PR are governed by the already-sealed always / never values through the DAG's conditionals; never re-ask or describe them.

Step 5: Errors

If any command errors, use the rad-log-error skill to record it. Do not try to fix pipeline code — work around it with a clear, actionable message that names the failure point.