II.
AgentCoreImpl JSON
Structured · liveagent-core-impl:pi.core@current
Pi Core (current) json
Inspect the normalized record payload exactly as the atlas UI reads it.
{
"id": "agent-core-impl:pi.core@current",
"_kind": "AgentCoreImpl",
"_file": "agent-stack/core-impls/pi-core-current.yaml",
"_cluster": "agent-stack",
"attributes": {
"displayName": "Pi Core (current)",
"agentVersionId": "agentVersion:pi:ge-0-0-0",
"packageRef": "source-ref:pi-github",
"loopIteratorPolicy": "tool-use-loop",
"loopIteratorNotes": "Tool-use loop over Anthropic Messages or OpenAI Responses; tool calls are\nserialized (no parallel tool dispatch). Loop advances on the provider's native\nend-of-turn signal.\n",
"contextManagementStrategy": "harness-managed",
"compactionTriggerNotes": "Pi implements client-side compaction via its AgentHarness. The compaction\nsystem uses branch summarization — when context approaches limits, the\nharness summarizes conversation branches and replaces them with compact\nsummaries. Compaction settings are configurable (threshold, strategy).\n",
"subagentInvokerPolicy": "tool-call-handoff",
"subagentInvokerNotes": "Sub-tasks are modeled as ordinary tool calls whose handler may itself\nspawn a nested Pi session. The harness supports skill invocation via\nformatSkillInvocation for structured skill dispatch.\n",
"resultSynthesisPolicy": "model-direct",
"resultSynthesisNotes": "Final assistant text is streamed to stdout; tool-call deltas stream concurrently.\nSession ends when the provider end-of-turn fires.\n",
"stopDetectionStrategy": "structured-end-event",
"transportClientLibrary": "pi-internal",
"supportedTransportProtocols": [
"model-transport:anthropic-messages",
"model-transport:openai-responses",
"model-transport:openai-chat-completions"
],
"parallelToolCallHandling": "sequentialized",
"streamingFidelity": "full",
"thinkingChannelHandling": "not-supported",
"codingAgentLayer": "The coding-agent package (packages/coding-agent) layers on the agent\nharness to provide: CLI entry point and argument parsing, operational\nmodes (interactive, headless, scripted), configuration management with\nmigrations, package-manager integration for dependency operations, and\nthe core coding-agent loop that orchestrates tool use for software\nengineering tasks.\n"
},
"outgoingEdges": [
{
"from": "agent-core-impl:pi.core@current",
"to": "layer:4-agent-core",
"kind": "realizes",
"attributes": {}
},
{
"from": "agent-core-impl:pi.core@current",
"to": "agent:pi",
"kind": "composes",
"attributes": {
"role": "core"
}
},
{
"from": "agent-core-impl:pi.core@current",
"to": "model-transport:anthropic-messages",
"kind": "speaks",
"attributes": {}
},
{
"from": "agent-core-impl:pi.core@current",
"to": "model-transport:openai-responses",
"kind": "speaks",
"attributes": {}
},
{
"from": "agent-core-impl:pi.core@current",
"to": "capability:streaming",
"kind": "supports",
"attributes": {
"versionRange": "*",
"level": "full"
}
},
{
"from": "agent-core-impl:pi.core@current",
"to": "capability:supports-tool-use",
"kind": "supports",
"attributes": {
"versionRange": "*",
"level": "full"
}
}
],
"incomingEdges": [
{
"from": "agentVersion:pi:ge-0-0-0",
"to": "agent-core-impl:pi.core@current",
"kind": "composed_of",
"attributes": {
"role": "core"
}
}
]
}