AgentDispatch is split across several small repositories. Pick the repo closest to the change:
core— provider-neutral types, policy, routing, runtime service, and adapter contracts.mcp-server— agent-facing MCP tools and bootstrap.adapter-aws-agentcore— AWS Bedrock AgentCore provider implementation.worker-agentcore— cloud-side worker contract and reference runtime.cli— config bootstrap, diagnostics, task submission, and A2A follow-up.sdk-js— TypeScript client and MCP transport helpers.store-sqlite— local durable task, log, event, and artifact store.adapter-template— starting point for new providers.docs,website,.github— launch, docs, website, and organization defaults.
For package-local changes, run the package commands:
npm test
npm run typecheck
npm run buildFor cross-package changes from the local multi-repo workspace:
npm --prefix agentdispatch-docs run verify:local-e2eThat gate runs package tests, typechecks, builds, package-consumption smoke tests, a packaged MCP spawn_cloud_agent smoke, CLI init/doctor, MCP --check, website validation, and org-profile asset checks.
Live AWS AgentCore verification is opt-in because it touches external account state and can cost money:
AGENTDISPATCH_CONFIG=/absolute/path/agentdispatch.config.json \
npm --prefix agentdispatch-docs run verify:aws-liveTo submit a real cloud task after preflight:
AGENTDISPATCH_CONFIG=/absolute/path/agentdispatch.config.json \
AGENTDISPATCH_LIVE_DISPATCH=1 \
npm --prefix agentdispatch-docs run verify:aws-liveDo not claim live AWS dispatch has been verified unless you ran a live check with real credentials and a real AgentCore runtime ARN.
New cloud providers should implement the BackendAdapter contract. Keep provider-specific fields inside adapter config, target.details, or providerRefs; do not add provider-specific MCP tools unless the core provider model truly cannot express the use case.
The stable routing key is:
provider + capability + task_type + target.mode
Framework integrations such as LangChain, LangGraph, CrewAI, Strands, or custom agent loops belong inside worker-side framework adapters, not cloud backend adapters.
Keep PRs focused. Include:
- what changed
- which package or repo owns the change
- verification commands and results
- whether live cloud state was touched
- any compatibility impact on the MCP contract or adapter contract