--- fixture: 14-message transcript, 6 tool calls, scripted mock JevAsker (documented bring-your-own-transport path) --- --- negatives: missing key, malformed JSON, missing answers, bad probabilities, live endpoint with dummy key --- PASS first user message kept verbatim PASS final user message kept verbatim PASS kept result verbatim (t1) PASS truncated result keeps 300-char head + note (t2) PASS truncated call itself stays (t2) PASS dropped call removes call AND result (t3) PASS no orphaned tool results PASS pinned calls never dropped despite 0.1 answer PASS stats: before/after messages PASS stats: one request PASS reduction ratio in (0,1) PASS chars dropped as claimed PASS decision reasons tallied PASS state holds whole history PASS state omits tool outputs PASS goal defaults to last user prompts PASS tight request budget splits batches PASS merged answers apply across batches PASS short transcript makes no Jev request PASS missing key throws PASS malformed JSON rejected PASS missing answers rejected PASS non-numeric noul rejected PASS HTTP failure surfaces status PASS request hits documented System One endpoint PASS request carries bearer + JSON body PASS live endpoint rejects dummy key cleanly PASS token estimator positive and plausible FUNCTIONAL RESULT: 28 pass, 0 fail functional result: 28 pass, 0 fail (live Jev decision quality NOT exercised: no TypeSafe key; see review)