Delegate and resume repository tasks through Pi-Bolt MCP
Delegate a scoped repository task to Pi-Bolt over MCP, inspect structured results, and resume the same Pi session for follow-up work and verification.
npx skills add agentskillexchange/skills --skill delegate-and-resume-repository-tasks-through-pi-bolt-mcp
Use Pi-Bolt as a session-based repository worker for an MCP-capable parent agent. Start the installed executable as a stdio server with /full/path/to/pi-bolt --mcp. Give its prompt tool a bounded task and an explicit absolute cwd. Inspect the returned text, sessionId, stopReason, and usage before deciding whether to accept the result or continue the task.
What this skill actually does
Invoke this workflow when another agent needs to delegate repository inspection, an agreed code change, or follow-up verification to an existing Pi setup while retaining the same conversation. For example, ask it to investigate a failing test in a named checkout and report evidence; after reviewing the proposal, send an authorized follow-up using the returned sessionId. The optional model and thinking arguments control subsequent work in that session. Use list_sessions with the repository directory to recover a prior session rather than starting over.
The scope boundary is task delegation and continuation through two documented MCP tools. It excludes a general terminal-agent product tour, performance benchmarking, plugin compilation, editor setup, and fleet orchestration. Existing ASE Pi coverage describes interactive CLI operation, sandbox choices, and session sharing; this entry covers the built-in MCP call/result/continuation contract. The live Claude Code MCP bridge is framed as one-shot delegation, while this workflow retains Pi session files and the user’s Pi configuration across follow-ups.
Install an appropriate release asset from the exact upstream release page, check it against the published SHA256SUMS, and unpack it for the supported operating system and architecture. Run the executable in a terminal to configure a model provider. Review and trust only the intended project there before starting headless MCP work; upstream says MCP sessions cannot ask for project trust. Do not configure blanket project approval. Clients launched from a desktop may require the absolute executable path.
Configure your client’s stdio command and --mcp argument. Upstream documents Claude Code, Codex, Cursor, VS Code, and Claude Desktop configurations. MCP is the framework because this approved workflow requires MCP tool invocation, even though the underlying program has other modes. Pi can read and edit files and run commands in the selected directory using existing settings and extensions; the MCP wrapper supplies no additional sandbox guarantee. Cancellation stops the model and running tool while retaining a usable session. Calls to the same session serialize; distinct sessions can run concurrently.
Exact repository metadata, release information, npm identity, README, and dedicated MCP documentation were checked. No executable, provider login, project-trust change, model call, or delegated task was run. This entry is metadata verified and has not received a security review.