MCP tools

run_status

The resolved run's registry row, re-read at call time: title, lifecycle state, progress line, `state_units` and `display_tasks` counts, worktree and branch, the recorded stop, and the owning process and its cleanup state.

Takes no arguments.

Any argument not listed above is rejected.

list_runs

The same row `run_status` returns, for every gantry run visible in the repo — not just the scoped one. Returns every run unless `scope` is "active", which returns the non-finished set ordinary `gantry list` shows.

Any argument not listed above is rejected.

read_ledger

The run's PROGRESS.md ledger, one row per projected task: its id, task path, title, latest ledger status, and whether the ledger has it DONE.

Takes no arguments.

Any argument not listed above is rejected.

read_plan

The run's `about.md` card as stored: the project title and description, a block per task, and the gate summary.

Takes no arguments.

Any argument not listed above is rejected.

list_tasks

The run's tasks as the pattern declares them, under the pattern's name: a tree of nodes each carrying its id, task path, title and status, nested to the depth the pattern runs at.

Takes no arguments.

Any argument not listed above is rejected.

read_report

A named diagnosis report under the run's reports/ directory.

Any argument not listed above is rejected.

tail_history

Recent records from the run's history.jsonl activity journal, with line/byte bounds.

Any argument not listed above is rejected.

tail_log

A named stage log with line/byte bounds. `task_path` addresses the log by a task's position in the run tree — its enclosing sub-orchestration ids, outermost first — so a log at any depth is reachable; omit it for a run-root log.

Any argument not listed above is rejected.

worktree_info

The run's worktree git state: branch, porcelain status, the 20 most recent commits, and the diff from the merge-base with the run's pinned home branch, bounded to 64 KiB. A run merged back and its worktree removed answers `live: false`.

Takes no arguments.

Any argument not listed above is rejected.

list_processes

Every process group the run recorded, from the same process model `gantry ps` and the build-run processes overlay render: the owning engine, and per group its state (live, stray, reaped, exited, stale), its process-group id, what it was spawned for, and its recorded and sampled fields. Read-only: it takes no arguments, signals nothing, and reports a stray group's remedy only as the `gantry stop <name>` command string.

Takes no arguments.

Any argument not listed above is rejected.

run_bash

Run a shell command in the run's worktree with the same wrapper state gantry uses for coding agents. The built-in process fence keeps a writable host view and binds the main checkout read-only with Git metadata and declared run-record outputs carved back writable; custom wrappers pass through verbatim, and disabled or degraded runs use no wrapper. The descriptor reports wrapper state, and each result reports whether the process fence, a custom wrapper, or no wrapper was active.

Any argument not listed above is rejected.

answer_prompt

Answer a pending prompt on the run's live engine, correlated by its prompt id — a run-control action over gantry's command channel. The answer's form must match the prompt's kind, or it is refused. Fails cleanly when no live engine is attached.

Any argument not listed above is rejected.