Supervise and reconcile persistent coding-agent crews with Firstmate
Dispatch scoped coding tasks into isolated worktrees, supervise durable task events, and reconcile the same crew after a primary-session restart.
npx skills add agentskillexchange/skills --skill supervise-and-reconcile-persistent-coding-agent-crews-with-firstmate
Use this workflow when several coding tasks must keep progressing across terminal sessions and supervisor restarts. Firstmate turns a supported coding harness into one operator-facing supervisor: it dispatches scoped tasks into isolated git worktrees, watches durable task events, escalates decisions, and reconciles the crew when the primary session resumes. The output is a task-by-task handoff with reviewable PRs, authorized local results, or investigation reports.
What this skill actually does
Inputs and scope
Collect the repository, independent task briefs, acceptance checks, selected harness, and project delivery mode. Upstream documents no-mistakes, direct-PR and local-only modes; select the intended mode explicitly with the operator and retain its documented merge authority. This entry covers a persistent local crew on the reference tmux backend, including supervision and restart reconciliation. Remote secondmates, Relay social posting, other backends, and maintenance of Firstmate itself require their own setup and authorization.
Set up the operational home
On macOS or Linux, install Git, the GitHub CLI, tmux and a supported primary harness. Authenticate the GitHub CLI through the operator’s normal account setup. Clone https://github.com/kunchenguid/firstmate into a dedicated directory, enter that checkout, start tmux new -s firstmate, and launch the selected harness there. For example, claude is an upstream-documented primary invocation. Read the checkout’s supervisor contract and follow its toolchain checks; obtain approval for missing-tool installation. The distro loads its instructions and bundled procedures from this home.
Dispatch, supervise and reconcile
Give the primary agent the selected repository, task briefs, checks and delivery mode. Confirm that each task has a distinct worktree and visible fm-<id> window before allowing parallel edits. Keep project changes in workers under the documented supervisor boundary. Firstmate’s Bash watcher observes durable status and wakes the supervisor for actionable changes rather than requiring continuous model polling. Treat watcher or state failures as unresolved supervision; a visible terminal alone does not prove that a worker is alive.
For a bounded task inspection, use bin/fm-peek.sh <id> from the Firstmate checkout. After a primary-session restart, resume in the same operational home and let the documented startup reconciliation inspect existing state and endpoints before creating replacement tasks. Preserve ambiguous or unreadable worker states for investigation; recover a worker only after the documented liveness checks establish that it is dead or missing. This prevents duplicate work and conflicting branches.
Return the handoff
Report each task’s endpoint and branch, observed verification results, deliverable link or report path, pending operator decisions, and unresolved supervision failures. Do not infer merge permission from a completed task. Follow the selected project’s delivery contract and request the required operator decision before landing work. Keep teardown behind confirmed completion so active worktrees remain available for recovery.
Source and framework
Reviewed exact upstream repository metadata, README and operational documentation; documentation snapshot: bd866530a07b23b57d933c6a745a9c81a2a94931. Firstmate explicitly supports multiple terminal harnesses, making Multi-Framework appropriate for this concrete supervision workflow. It is an instruction-and-tooling distro, not an MCP server. Installation, runtime behavior and security were not tested during discovery intake.