v0.10.3 — Performance & TUI

v0.10.3

Performance

• Tree cap: ctx.tree capped at 150 lines — eliminates unbounded token injection per request
• Memory cache: loadUnifiedMemory() now cached at module level, not called per-agent-turn
• Tool gating: spawn_task, web_fetch, web_search, memory, mcp now keyword-gated — ~750 tokens saved per plain coding task
• Telegram: reduced memory injection from 3,500 to 800 chars — faster bot responses

TUI

• Bracketed paste mode — eliminates 19-prompt-per-paste chaos
• Multi-column model selector — 116 models now readable
• formatProviderError — truncates HTML error floods at 400 chars

Providers

• ensureApiKeyForModel — prompts on model switch instead of silent fail
• GLM billing plan selector — pay-as-you-go vs Coding Plan routing
• OpenCode Go — fixed baseUrl and Anthropic→OpenAI provider routing
• .aura.json excluded from git (local config)

Security

• All credentials centralized to ~/.secrets/agents.env
• EnvironmentFile= in systemd services — no plaintext keys in unit files