{"version":"1.0-before-execution","source":"CapyAgent development snapshot 77ac9b72409fc2b6afc5741fd2469c662758a1b4, same pinned baseline as earlier journal study; not latest product","question":"Can the recovered local journal distinguish a tool not yet called from a completed external effect whose result was not checkpointed?","matrix":{"boundaries":["before_start","before_request","after_effect","after_checkpoint"],"service_modes":["plain","idempotent"],"replicates":5,"total_cases":40},"procedure":"Run a new helper process using real FileStore, stop at an explicit barrier, SIGKILL only that child; reopen store in two new helper processes. A synthetic loopback HTTP service commits effects to SQLite. Research-only retry policy skips if recorded receipt exists and otherwise sends one request with same operation key.","metrics":["child termination signal","external effect count before recovery","recovered tool/result/unknown counts","history rows on second reopen","retry sent","external effects after retry"],"exclusions":["real agent loop or LLM","production retry policy","payment service","power loss","disk corruption","full network failure model","current product acceptance"],"safety":"Synthetic data, random loopback port, own child PID, isolated directories, no credentials or paid APIs; KotikTurbo daytime limited scope and shared tests.lock"}