Optional staged pipeline runner¶
The core package includes a staged runner for projects that want AKMS to coordinate task planning, execution, review, checkpoints, and graph updates. This is an optional integration layer.
Stage model¶
The runner can:
- Build or load the graph
- Prepare role-specific context
- Dispatch an
AKMSAgentimplementation - Persist checkpoints and resume state
- Collect task/phase records
- Invoke explicit local graph updates
- Run code and physics reviewer roles
Runtime backends¶
The CLI currently maps friendly backend names to built-in adapters:
| Backend | Adapter model |
|---|---|
claude-sdk |
Claude Agent SDK |
claude-cli |
Headless claude command |
codex-sdk |
OpenAI Agents SDK |
codex-cli |
codex exec command |
local |
OpenAI-compatible local endpoint |
An explicit dotted --agent module.ClassName overrides the backend mapping.
Architectural rule¶
Do not make graph, route, loadout, failure-memory, or learning documentation assume this runner is present. Those services must remain independently usable from CLI, Python, MCP, or another harness.