displayName
Claude Code Core (1.x)
agentVersionId
packageRef
loopIteratorPolicy
tool-use-loop
loopIteratorNotes
Tool-use loop: each turn the model emits zero or more `tool_use` blocks; the runtime
executes them and appends `tool_result` blocks back into the message list. The loop
advances until the model returns an Anthropic Messages `stop_reason: "end_turn"`
(or `"tool_use"` after which tools are run and the next turn is fetched). The
PreCompact hook fires when context approaches the configured threshold and the
runtime injects a summarized history before the next turn.
contextManagementStrategy
platform-managed
compactionTriggerThresholdTokens
0
compactionTriggerNotes
PreCompact hook fires as the conversation approaches the active model's context
window (community-observed at ~70% fill). Vendor docs do not publish an exact
threshold; treat the integer field as TODO until Anthropic documents one.
subagentInvokerPolicy
tool-call-handoff
subagentInvokerNotes
Subagents are invoked via the built-in `Task` tool (a.k.a. the `Agent` tool). The
parent loop emits a `tool_use` for `Task` with a subagent name + prompt; the runtime
spawns a nested tool-use loop for the subagent and returns its final assistant text
as the `tool_result`. Up to 10 subagents can run in parallel.
resultSynthesisPolicy
model-direct
resultSynthesisNotes
The final assistant text is streamed verbatim to the host as it arrives; on
`stop_reason: "end_turn"` the runtime emits the `Stop` hook (and `SubagentStop`
for nested Task invocations) and closes the turn.
stopDetectionStrategy
structured-end-event
transportClientLibrary
@anthropic-ai/claude-agent-sdk
supportedTransportProtocols
parallelToolCallHandling
native
streamingFidelity
full
thinkingChannelHandling
passthrough
loopBudgetKnobs
compactionTrigger
both
streamingBlocks
effortLevelDelegation
native