Advisor Middleware

@xpert-ai/plugin-advisor adds a generic advisor tool to the agent runtime.

The tool lets the executor consult a separately configured model for:

• hard debugging
• architecture tradeoffs
• risky decisions
• second opinions on a plan

The plugin is provider-agnostic. It only depends on the configured Xpert model selection and does not assume Anthropic or any other vendor-specific protocol.

Key Behavior

• injects an advisor tool into the runtime
• appends a short executor prompt explaining when to use it
• enforces per-run and optional per-session usage limits
• forwards a curated slice of conversation history to the advisor model
• returns the advisor result as a normal ToolMessage

Configuration

• advisorModel: model selected with ai-model-select
• maxUsesPerRun: per-run quota, default 3
• maxUsesPerSession: optional session quota
• appendSystemPrompt: whether to guide the executor on advisor usage
• maxTokens / temperature: internal advisor call settings
• context.*: controls how much prior conversation is forwarded

Validation

Run:

npx nx build @xpert-ai/plugin-advisor
npx nx test @xpert-ai/plugin-advisor
fnm exec --using=20 -- node ../plugin-dev-harness/dist/index.js --workspace . --plugin ./middlewares/advisor/dist/index.js