Problem
Every sub-agent invocation currently spends two live tool calls (list_documents + list_relations) bootstrapping its view of the knowledge base via the preamble in plugins/archcore/agents/archcore-{assistant,auditor}.md (Option A, shipped and test-enforced). As the knowledge base grows, that per-invocation cost becomes measurable.
Expected result
Option B: precompute a knowledge-tree snapshot (the same payload as session-start) and prepend it at Task-dispatch time via a new archcore hooks <host> subagent-start subcommand, eliminating the per-invocation bootstrap calls.
Impact
Removes repeated bootstrap overhead for sub-agents, keeping delegated work fast as the documentation base scales.
Blocked
Claude Code exposes no documented sub-agent-start hook surface, so this needs host-side plumbing plus a new CLI subcommand. The adopted design (.archcore/plugin/subagent-knowledge-tree-bootstrap.adr.md) defers Option B until such a lifecycle hook exists on at least one host; revisit trigger is roughly 10x knowledge-base growth. Track as a blocked epic.
Proposed approach
References
Design rationale: .archcore/plugin/subagent-knowledge-tree-preload.idea.md (Option B) and subagent-knowledge-tree-bootstrap.adr.md. The empirical hook-coverage probe (separate issue) is what could re-open this.
Problem
Every sub-agent invocation currently spends two live tool calls (
list_documents+list_relations) bootstrapping its view of the knowledge base via the preamble inplugins/archcore/agents/archcore-{assistant,auditor}.md(Option A, shipped and test-enforced). As the knowledge base grows, that per-invocation cost becomes measurable.Expected result
Option B: precompute a knowledge-tree snapshot (the same payload as
session-start) and prepend it at Task-dispatch time via a newarchcore hooks <host> subagent-startsubcommand, eliminating the per-invocation bootstrap calls.Impact
Removes repeated bootstrap overhead for sub-agents, keeping delegated work fast as the documentation base scales.
Blocked
Claude Code exposes no documented sub-agent-start hook surface, so this needs host-side plumbing plus a new CLI subcommand. The adopted design (
.archcore/plugin/subagent-knowledge-tree-bootstrap.adr.md) defers Option B until such a lifecycle hook exists on at least one host; revisit trigger is roughly 10x knowledge-base growth. Track as a blocked epic.Proposed approach
subagent-startCLI subcommand and per-host injection plumbing.References
Design rationale:
.archcore/plugin/subagent-knowledge-tree-preload.idea.md(Option B) andsubagent-knowledge-tree-bootstrap.adr.md. The empirical hook-coverage probe (separate issue) is what could re-open this.