• Add a persistent IPython tool backed by a Jupyter kernel, with an auto-bootstrapped uv-managed Python runtime, so variables and imports survive across tool calls.
• Add the RLM harness and prime-agent-runtime bridge so IPython code can spawn recursive child agent sessions via rlm.run, with subagent tray, detail views, and chat-style child transcripts.
• Add /goal for long-running objectives and /usage for on-demand token, cost, and context usage.
• Add Prime Inference as a built-in OpenAI-compatible provider with a /login browser auth flow, plus a pi-style installer with R2-backed tarball releases.
• Make IPython the default tool set, restarting the kernel on compaction and using cached, compact cell rendering to cut TUI redraw lag.
• Fix RLM child usage aggregation, kernel port binding, child-run disposal races, and auto-compaction threshold checks.
• Remove legacy read/write/grep/find/ls tools, the separate RLM background API, and install/update telemetry pings.