Open Multi-Agent Files

Adaptive plan recovery

A run can now revise the not-yet-executed part of its task graph instead of carrying a plan that stopped fitting the work.

• Opt in with recovery.mode: 'repairable', then supply a Replanner or an onTaskOutcome callback that proposes an append-only PlanPatch.
• Patches are validated for agent eligibility, limits, task states, references, and the resulting DAG, then gated through the optional onPlanPatch approval.
• Application happens atomically at a task-outcome barrier, before downstream dispatch or failure cascade, so no original dependent starts against a stale plan.
• Revision history is exposed in results, progress events, and observability spans, and persisted through checkpoints with rollback when a checkpoint write fails.

See adaptive recovery.

Execution routing

• Added opt-in hybrid semantic routing through executionRouting: { strategy: 'hybrid' }. Deterministic Team decisions stay authoritative; deterministic Single decisions receive one structured semantic assessment.
• Added a provider-neutral TaskProfiler interface, a built-in LLMTaskProfiler, a strict task-profile schema, typed routing failures with timeout and fallback metadata, and semantic-routing observability.
• Automatic runTeam() routing remains deterministic, so existing runs are unchanged.

Provider support

• DeepSeek V4 Flash reasoning controls: AgentConfig.thinking.enabled maps to DeepSeek's native thinking.type, and thinking.effort accepts the DeepSeek-only value 'max' without forwarding it to OpenAI, Azure OpenAI, or GitHub Copilot.
• Refreshed DeepSeek model guidance, retired aliases, examples, CLI reference, and scaffold defaults.

Correctness and safety

• Task requirements are enforced as global hard constraints across every scheduling strategy. A task whose requiredTools, requiredCapabilities, requiredBackend, or requiredProvider no agent satisfies is rejected instead of being dispatched to an ineligible agent.
• Invalid task dependency graphs are rejected up front instead of executing a partially valid plan.
• The coordinator fails closed on an invalid plan rather than continuing with a plan it could not validate.
• OpenAI user aborts are classified as cancellation rather than as a retryable failure, and unsupported custom tool calls are rejected explicitly.

Runtime requirements

• Node.js 20 or newer is now required across @open-multi-agent/core, @open-multi-agent/otel, and create-oma-app. Node 18 reached end of life on 2025-04-30.
• The bundled openai dependency moved from v4 to v6.

Compatibility

• Every public export from 1.13.0 is still exported, and new result and configuration fields remain optional.
• Adaptive recovery adds a version 2 task-queue snapshot. fromSnapshot() still accepts version 1, so checkpoints written by earlier releases remain restorable.
• Runs that previously succeeded with an invalid task DAG or an unsatisfiable task requirement now fail at validation time. This surfaces a defect that was previously silent; correct graphs and rosters are unaffected.
• @open-multi-agent/otel@0.1.1 is republished to carry the Node 20 engines floor. Its @open-multi-agent/core@^1.11.0 range is unchanged.
• create-oma-app@0.7.0 pins generated starters to core 1.14.0.

Thanks

• HarperZ9 contributed the mid-task recovery conformance fixture in [#452], which pins the idempotency outcomes of four recovery cases against the current task-level checkpoint granularity.