v0.19.0 — M5 context-engineering (transcript budget-fit)

Added — long, tool-heavy conversations stay coherent and stop ballooning token cost (M5)

When the assistant chains several tool calls (search laptop → read files → search again), the running transcript used to grow every step and get sent to the model in full each round — rising latency and cost with no ceiling. It's now budget-fitted: each round the transcript is trimmed to a token budget (memory.context_budget, default 24000), always keeping your question and the most recent turns, dropping the oldest middle history, and never splitting a tool call from its result. Normal chats sit far under budget so nothing changes; the fit only engages when a turn actually grows large. Set memory.context_budget to 0 to disable.

Pure context-engineering via litectx's assemble (0.25.0) wired through bare-agent's Loop (0.19.0) — the same libraries multis already depends on, no new dependency. Does not touch retrieved documents or saved notes. Full suite green, npm audit clean.