Finder Reactor — Self-Guarded, Pause-Aware, Agentic Opportunity Engine

Core Mission: Turn raw X firehose + your CV + xAI into a reliable, low-intervention personal opportunity system. The finder decides what to pursue, prepares materials, tracks outcomes, and improves itself — while never acting on high-stakes things (CV changes to public profile, expensive runs, low-confidence fits) without explicit guard clearance and user pause points.

It is the "intelli" part that makes development (and usage) feel exponential: agents can drive large parts of the loop, the loop itself improves with use, and surplus from each cycle compounds the quality of future cycles.

Immutable Principles (Self-Guards Always Win)

1. Guard > Speed: No action crosses a threshold (cost, rate remaining, fit score, CV write) without a named guard check + audit log.
2. Pause for Human on Ambiguity or Stakes: Low/medium confidence, any CV promote, high token/cost estimates, X write side-effects → surface summary + explicit "proceed / tweak / ignore" path. User only intervenes when the guard fires.
3. Full Query + Context Control: The human/agent always owns the X query. Reactor never hides or hardcodes the "best" search.
4. Sidecar-First for External Mutation: CV promote always writes a sidecar proposal first. Master write to devprofile is a separate guarded step with diff + confirm.
5. Structured + Observable Decisions: Every xAI "decide" call returns machine-readable recommendation + confidence + guard triggers + rationale (zod/serde). UI + MCP + logs expose it.
6. Surplus After Every Cycle: After a search-analyze-prep-promote loop (or sub-part), produce at least one concrete improvement that makes the next similar loop cheaper or higher-value (better prompt, tighter guard, new skill, refined CV packet, etc.).
7. Reversible + Auditable: All state (leads, preps, decisions) is file-based JSON in app data. Easy to inspect, replay, or roll back. CV promotes produce .bak + git-friendly diffs.

Reactor Loop (Conceptual State Machine)