SolOS Architecture & Operation (rewrite)

Core model: SolOS = brain, OpenCode = muscles

• SolOS (orchestration) decides WHO works, WHAT context they hold, and HOW MUCH it costs.
• OpenCode (execution) makes the LLM calls, runs the tools, executes the work.

Verified components (live, 2026-09-22)

• A2A mesh — 8 agents on :8300–8307 (sol/ace/mira/nova/echo/alpha/atlas/min), run_<agent>.py.
• Central Registry — :8432 (tasks/agents/reports/board/memory search). ⚠️ :8433 is now only the A2A discovery registry — stale docs still point at it.
• Central Cron — :8436 (unified job board + scheduler).
• CostGov — :8430 (runs inside server.py; the solos-cost-governance.service unit was renamed).
• Session persistence — data/solos.db (sessions + messages).
• Memory plane — 3-layer (durable / semantic-facts / episodic) in solos.db, injected every turn via build_memory_plane.
• Personas — personas/<agent>.md, loaded + injected with the memory plane.
• Thin proxies — ~/.config/opencode/agents/<agent>.md dispatch OpenCode → A2A ("do NOT answer directly" — the anti-memory-loss fix).

Operation notes (this session's ground truth)

• Laptop agents (Tam :8447, Eva :8441) reach the mesh over tailscale: server socat forward 1830x→830x + reverse tunnel 8447/8443/9903.
• fail2ban jailed the laptop's tailnet IP (empty ignoreip) → whitelisted 100.64.0.0/10.
• Memory migration was partial (4/8 agents got durable memories) → rebuilt; default-agent ping noise + self-correction spam purged.
• opencode-web crash-loop = duplicate opencode serve processes fighting port 4090 → reconciled to one systemd instance.

Known gaps (fed to blueprint v2)

• Tam/Eva absent from central-cron AGENT_PORTS → can't receive prompt jobs yet.
• Eve on a separate TS runtime, memory-isolated (see 2026-09-22-vercel-eve-framework-analysis).
• Port/path drift (8432 vs 8433) across docs → single source of truth needed.