v0.4.0 — install once, it just runs

The "make the promise literally true" release.

One-command setup, every platform. knitbrain setup --yes now writes the global MCP configs for Codex, Copilot CLI, and Zed (backed up, non-clobbering) instead of printing a snippet to paste — alongside the native Claude Code / Cursor / VS Code / Windsurf integration it already wrote.

The loop runs itself on hook-capable hosts. A SessionStart hook injects the operating protocol + your prior handoff + proven learnings into every session, so memory load no longer depends on the agent remembering to ask. PreToolUse hard-redirects large raw reads to the optimizer; PreCompact/Stop keep work resumable.

knitbrain wrap claude|codex|aider|copilot — launches your agent with the optimizer proxy wired in. No manual base-URL export.

Verified on Windows, not just Linux. CI now runs Ubuntu + Windows on Node 20 & 22; 233 tests green on both.

Numbers (run them yourself): ~48% of all tool-result tokens saved on real sessions, ~55% on sizable blocks, 60–70% on code/JSON/logs — npx knitbrain profile. Answer-preservation gated (errors, summaries, round-trip always 100%) — npx knitbrain evals. Lossless: every elision recovers byte-for-byte.

27 MCP tools · live stdio E2E · cold-start production audit (50 checks) · pure TypeScript, no native deps.