Oh My Subagents keeps Claude Code teams running

The open-source Oh My Subagents runtime packages local supervision for Codex and Claude teams with eight provider-neutral starter workflows, keeping assignments, waits and returned checkpoints outside an individual provider session, according to Github, and is designed for persistent, supervised parent-subagent delegation.

• The controller commits a manager’s delegated child assignments together with the parent wait, then gathers independently returned terminal checkpoints before continuing the parent with a complete wave.
• Each Task pins the published Workflow revision used at its start, while only the Task lead’s accepted completed or blocked checkpoint becomes the Result presented after integration.
• The package runs with Python 3.12 or newer on Linux, macOS 13 or later, and Windows 11 x64, using SQLite by default and PostgreSQL optionally.

As coding agents take on broader, multi-step work, coordination failure becomes a practical constraint: a useful investigation or review may be stranded in an interrupted session, while a parent may lack a reliable account of ownership and completion. Oh My Subagents appears to make that operational layer explicit, and it likely separates supervisory state from the conversational transcript that providers control.

Controller records preserve delegated work across interruptions.

When a Manager delegates a Wave, the runtime commits the child Assignments and the parent wait together in controller-owned state. This committed unit includes the parent wait. Children return terminal Checkpoints independently, and the controller resumes the parent after it has collected the complete Wave, rather than leaving the parent in a polling loop.

Managers can delegate recursively under the same direct-child rule: delegate, wait, inspect complete returns and integrate. The system records retries, replans, continuations and activity as runtime state; it treats child completion as evidence, while the Task lead’s accepted completed or blocked Checkpoint is the accountable Result. Browser closure, provider interruption, and controller restart do not discard accepted history or fabricate completion.

Eight starter workflows cover delivery, research, experiments and security.

Initialization publishes eight provider-neutral workflows: production-feature-delivery, incident-investigation-and-recovery, migration-and-modernisation, deep-research-and-decision-brief, decision-through-competing-prototypes, idea-to-validated-demo, experiment-and-replication-program, and security-audit-and-hardening. A one-Member Workflow is also valid when delegation would not add independent evidence, specialization, or integration. The catalog gives example missions, expected deliverables and guidance on when a simpler team is better.

Workflow definitions describe responsibility rather than a fixed schedule or DAG. At runtime, Managers may run dependent work sequentially, fan independent work out in parallel, iterate through implementation, review and repair, or replan a responsibility subtree without rewriting earlier revisions, completed work, or accepted history. Managers can combine these patterns within one Task. Its members retain reusable named responsibilities.

The visual Console and Operator use the same controller operations.

The visual Console includes a Workflow library for reusable teams, drafts and published revisions. Its Workflow Studio builds responsibility hierarchies on a horizontal canvas and publishes them, while Run Studio shows the live team, plans, activity, Human Requests, managed Actions, referenced files, and the precise completed or blocked Result. The library keeps drafts and revisions together.

Steering sends bounded context to a single active Member, and the separate Operator can draft or revise workflows, start and control Runs, answer Human Requests, and inspect managed Actions. Both interfaces call controller-owned operations, so conversation does not create a separate product record. The Console also presents the current plans of the live team.

Moving from Banksia Existing Banksia installations can migrate their configuration, database, provider credentials, Task history and service state with oms migrate-from-banksia before initialization. OMS says not to run oms init first. The project does not state a future release date; installations can run the controller as a native per-user service or through the foreground oms serve command.