coralos-run-remote-agents（Coralos 运行远程 Agent）

Coral

Use this skill to run a complete Coral Cloud session: discover available agents, build and submit a session payload, then monitor and close it.

Phase 1 — Discover (Registry)

Identify the agent you want to run before composing a session payload.

Fast commands:

• bash scripts/list_registry_agents.sh
• bash scripts/inspect_registry_agent.sh <source> <agent_name> <version>

Live endpoints:

• List: GET https://api.coralcloud.ai/api/v1/registry
• Inspect: GET https://api.coralcloud.ai/api/v1/registry/{source}/{agentName}/{version}

Before moving on, record:

• source
• agentName
• version
• notable option requirements and constraints

If you see {"message":"Not found"}, compare the URL to the current Coral Cloud API guide.

Phase 2 — Create Session

Build and submit a POST /api/v1/local/session payload using the agent identified in Phase 1.

Fast commands:

• bash scripts/create_session.sh examples/payloads/echo-session.json
• bash scripts/create_session_from_simple.sh examples/payloads/simple/echo-session.json

Payload notes:

• Payload shape is strict — start from examples/payloads/echo-session.json or expand from examples/payloads/simple/ using scripts/build_coral_session_payload.py.
• Required fields: agentGraphRequest, namespaceProvider, execution.
• Optional: session annotations, extendedEndReport.
• Custom tool/webhook flows are experimental until callback host policy is confirmed in your environment.

Success criteria:

• Non-error API response
• Real sessionId in the response

Phase 3 — Manage Session

Monitor progress and clean up after the session is running.

Fast command:

• bash scripts/get_session_state.sh <namespace> <session_id>

Live endpoint:

• State: GET /api/v1/local/session/{namespace}/{sessionId}
• Verify list and delete routes against your environment before automating cleanup loops.

Lifecycle checklist:

• Session exists
• Agents connect and begin execution
• Output appears in Coral Cloud thread when requested
• Session is closed when no longer needed

Success criteria:

• Real state payload for the target session
• Evidence of agent progress and/or thread activity