Reference

Reference documents the runtime surface: the options, types, and behaviour of each subsystem. It assumes you already have a run working — if you do not, start with Quick start.

Most of these pages are vendored from the framework repository and refreshed by a scheduled sync, so they track the shipped release rather than being rewritten here by hand.

Configure models and tools

• Providers — hosted, cloud, and local models: built-in shortcuts, OpenAI-compatible endpoints, environment variables, and local tool-calling.
• Tool configuration — default-deny tool grants, presets and allowlists, the filesystem sandbox, custom tools, and MCP.
• Structured input — pass a complete LLMMessage array instead of a string: prior conversation turns, image blocks, runtime validation, and defensive copies.
• External agents — run local processes or ACP coding agents inside a task DAG, with explicit permission, usage, and lifecycle boundaries.

Control orchestration

• Execution routing — single-agent or team execution through explicit modes, governance policy, custom routers, and auditable decisions.
• Task scheduling — event-driven DAG execution, scheduling strategies, structured requirements and handoffs, priority, and approval modes.
• Durable approval gates — suspend a checkpointed run at a plan, round, dispatch, or tool-call gate, decide out of band, and resume from exactly the reviewed content.
• Consensus — runConsensus proposer-to-judge verification, the per-task verify hook, and the shared token-budget invariant.
• Model routing — opt-in deterministic policy that routes orchestration phases to different models by phase, agent, task role, priority, or leaf.
• Plan preview & replay — freeze a reviewed task DAG with createPlanArtifact, then execute it later with runFromPlan without calling the coordinator again.
• Shared memory — a namespaced key-value store shared across a team, in-process or through a custom MemoryStore backend.

Operate reliably

• Observability — TraceRecord v2 sinks, TraceStore implementations, optional OpenTelemetry export, and the offline post-run Run Viewer.
• Observability migration — move from onTrace to sinks, stores, and OpenTelemetry in reversible stages without changing runtime results.
• Observability performance — reproducible performance budgets, the benchmark method, and the current release snapshot.
• Run event journal — an opt-in append-only log of what entered the context, which tools ran, and how the plan and task states moved.
• Checkpoint & resume — opt-in per-run snapshots over any MemoryStore: persist task progress and resume after a crash, abort, or restart.
• Adaptive recovery — revise the not-yet-executed part of a task graph after an outcome, with validated, approvable, append-only plan patches.
• Context management — keep long runs under the token ceiling with context strategies, tool-result compression, and cross-provider reasoning.
• Evaluation — version EvalSets and scorers, persist results, gate CI, and sample completed production runs without changing business results.
• Egress policy — restrict which origins the built-in LLM adapters may reach, intersected across orchestrator, run, and agent scope.

Command line

• CLI — the JSON-first oma binary for shell and CI: commands, config files, output, and exit codes.

Elsewhere in the docs

• Introduction, Quick start, and Choose a run mode cover the first run.
• Orchestration controls, Control costs & budgets, and Production checklist cover what to decide before a run reaches production.