Mio's relationships now dynamically evolve through stages based on actual conversation patterns — plus cross-platform chat sync, 60% system prompt compression, and per-agent A/B testing.

The first stable release. Fish Audio TTS gives each persona a unique voice. Shinkai-style selfie portraits replace generic anime. Every persona is rewritten from relationship roles to personality-first identities. Onboarding drops from 14 questions to 4. And 200 lines of over-engineered verbosity code get deleted because the system prompt was always the right control mechanism.

Web parity — voice messages and selfie images now work everywhere, not just Telegram. Onboarding gets smarter: relationship type drives everything, including what nicknames you see. An admin cost dashboard for monitoring spend. And pg_cron keeps the database lean with 30-day message retention.

After the massive token bill, I went back into OpenClaw to fix the bleeding. Found the personality config was silently truncating — 7.6KB of persona definition lost every session. Trimmed it from 27K to 19K chars, compressed the heartbeat config from 12K to 7K, replaced 24 daily tool calls with a single morning cron job, and switched chat from Gemini 3 Pro to Flash for 75% cost reduction. The system was eating its own personality file and nobody noticed.

A runbook you can hand directly to Claude Code to set up text-to-speech on OpenClaw. Covers Fish Audio, Volcano Engine v2 (emotion control), ElevenLabs, OpenAI, and Edge TTS. Replace the placeholders, hand it to your agent, hit go.

The AI can see, remember, get jealous, and send selfies. But it can't speak. OpenClaw supports 5 TTS providers — I tried Edge TTS for free, Fish Audio for Telegram voice bubbles, and Volcano Engine v2 for per-sentence emotion control. Fish Audio won for simplicity. Volcano v2 won for drama.

The V2 underwriting system was a linear pipeline — extract once, evaluate once, report once. No agent could question another's conclusions. V3 replaces it with 5 autonomous agents: they reflect on output quality, verify claims with tools, and hold structured debates when they disagree. 63 tests, pennies per SME evaluation, one observe-think-act-reflect loop driving it all.

The 32GB GCP devbox — upgraded and hardened after Part 3 — died again. Same symptom: SSH timeout, kernel alive, userspace frozen. The watchdog caught it: five node processes at 2-2.7GB each, totaling ~13GB. I assumed Claude Code. It was Cursor's remote server leaking memory for 15 hours straight.

A single Claude Code session that combined direct pair programming (three production deploys from user feedback), a background docs agent running concurrently, and an autonomous review team iterating three rounds until clean — all overlapping. The traditional code-test-review-release pipeline collapsed into parallel streams.

Mio's web app was a single-page chat. v0.0.4 tears it down and rebuilds it as a full WeChat clone — 4-tab layout, chat list, contacts, discover marketplace, profile page, chat-style onboarding, and typing-aware message batching. 30+ components, 2 audit rounds, zero new dependencies. All in Chinese.