convoylisted

# convoy convoy runs a **governed, measurable multi-PR series**. You give it a series (PR-sized tasks with dependencies, a base branch, per-phase budgets, and a quality gate); it drives a coding agent to implement each PR in dependency order under budget, gates the result against the series' checks, repairs on a blocking failure, integrates the green branches, and records **per-spawn economy** (tokens, turns, cost, duration) as an append-only, versioned trace. It is headless — fire-and-walk-away, no human checkpoints. The plugin exposes four MCP tools: - **`convoy_run`** — run a series (or, with `dry_run`, pre-flight it for free; or, with `detach`, start it and get a handle back at once). - **`convoy_gate`** — run the series' `[[checks]]` against a workspace once, no spawn and no git mutation: the deterministic gate standalone, for verifying work produced outside convoy (see "The gate without the run" below). - **`convoy_init`** — scaffold a runnable starter series to adapt or smoke-test. - **`convoy_status`** — ask a run how it is doing, including one still in progress and one this server never started. Reads the ledger only: no spend, no state, no touch on the workspace, so polling is cheap and safe. A run **spends real model budget** and takes minutes to hours — it spawns a subprocess `claude -p` per PR. Always `dry_run` first (free, no side effects), then drop it for the real run. Because a real run is that long, do not hold a blocking `convoy_run` open for