[jflavan]

Lets each loop step (and chat's orchestrator and write-intent) choose its agent, subagents, teams of subagents, skills, tools and MCP tools. Setup, the console and the operator skill can all edit these settings. With no agents block, sessions start exactly as before, and tests pin that. Includes feat/mcp-tools (mcp_tools in every tool set, with MCP tools chosen during onboarding) and a merge of #87.

Spec: docs/superpowers/specs/2026-10-26-agent-setup-design.md. Plans 1–4 are in docs/superpowers/plans/, and the decisions are in docs/decisions.md.

Runtime

• New top-level agents block, with all plus one entry per step:

  • agent, subagents and teams (named groups run parallel or by pick);
  • skills, tools and mcp_tools;
  • sources: repo, which loads only what the repo declares.

  claude.mcp_tools moved to agents.all.mcp_tools.

• Enforced rules:

  • Read-only steps never get write tools.
  • No step gets git rules that commit, push or switch branches.
  • Fix, reconcile and revise run as the build's agent.
  • Built-in agents (Explore, Plan, …) can only be subagents.

• Repo agent definitions are read from main's commit, so a change can't rewrite the agent that judges it. Writers given --add-dir get deny rules for that snapshot.

• When a session starts, its init is checked against the setup. A missing agent or skill stops the step with a setup failure that says what's missing. loopstra start refuses while a step names an agent that doesn't exist.

Setup

• loopstra setup agents sets each step's agents, subagents, teams, skills and tools. It adds skills and agents from this machine, a plugin, a git repository or folder, or a short form:
  • Copies never follow links or escape the target.
  • .git is never copied.
  • Credentials are stripped from shown URLs.
  • ssh runs in batch mode unless you configured your own.
• The operator skill now has steps for choosing MCP tools and setting up agents.

Console

• Settings has an Agents section that edits the same values and shows where each one comes from (step, all or default). It can also copy, fetch, create, edit and remove agent and skill files.

• The console never writes the repo. The loop applies each agent-files request and commits it. A change is refused if:

  • the target has uncommitted, staged or untracked work;
  • the target is a link;
  • the file changed since the page read it.

  Request text can't carry control characters, so it can't forge a trailer.

• The session page shows what a step ran as. Each subagent's work is nested under the call that started it, and a subagent's prompt is never shown as a person's message.

Verified with the real CLI (2.1.311)

• Inline --agents with --agent keeps structured output only when the agent's tools include StructuredOutput.
• --disallowedTools also binds subagents.
• --setting-sources project keeps --settings hooks.
• Subagent events carry parent_tool_use_id.

Live runs: an intake ran as an inline agent, and a change went to done with build running as a custom agent. The fix sessions kept the build's session.

Known and not addressed here

A build session can still influence later sessions of the same step through its change's .claude/skills, .claude/settings.json, CLAUDE.md or Bash. This predates the PR. Agent definitions are protected; closing the rest needs a tighter setup for judges.

Verification

bun run check passes: typecheck, lint and deadcode are clean, and 1836 tests pass. Each task was reviewed, and each of the four parts had a final review and a re-review after its fixes. Browser checks covered phone and desktop widths, light and dark.

🤖 Generated with Claude Code