| feat(perf): performance tuning across ebpf collection and semantic persistence - perf(ebpf): decode kernel samples in the ring buffer callback, intern process observations behind Rc, drop redundant file-path re-normalization, reuse and parallelize preflight link teardown, and batch-read transport diagnostics counters - perf(semantic-action): parse LLM SSE streams incrementally per segment and eliminate deep clones / Vec builds in HTTP message link projection - perf(storage/sqlite): cache prepared statements with a fast-path action-id interning, skip unchanged membership updates, drop cold-field payload hashing, and make cold-field compression configurable - perf(llm): reuse live response HTTP parsing and frame responses incrementally - perf(daemon/ingest/tls/recording): light active-binding drain counting with per-phase slow-drain diagnostics, short-circuit the always-allow policy gate, reuse provider lookup per sync decision, and coalesce semantic persistence updates - perf(semantic): materialize LLM responses only at terminal, persist only final SSE actions, and gate assembled LLM/MCP payload retention behind config while keeping trace-close response links Co-Authored-By: gpt-5.6-sol | 4 天前 |