finder-reactor

name: finder-reactor description: Core autonomous, self-guarded decision loop for opportunity-finder apps. Handles discovery (e.g. X search), analysis (LLM + CV + platform context), prep generation, tracking, and guarded promote with cost/rate/fit/CV mutation guards, human pause points, structured decisions, and logging. Use when designing, implementing, or debugging the agentic heart of a finder platform. Fission for tight loops; fusion for reactor architecture and surplus.

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

Core Mission: Turn raw social/search firehose + user CV + LLM 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 a public profile, expensive runs, low-confidence fits) without explicit guard clearance and user pause points.

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, write side-effects → surface summary + explicit "proceed / tweak / ignore" path.
3. Full Query + Context Control: The human/agent always owns the search 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 the external profile repo is a separate guarded step with diff + confirm.
5. Structured + Observable Decisions: Every "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, produce at least one concrete improvement that makes the next similar loop cheaper or higher-value.
7. Reversible + Auditable: All state (leads, preps, decisions) is file-based JSON in app data. CV promotes produce .bak + git-friendly diffs.

Reactor Loop (Conceptual State Machine)

Search (tunable query + platform context)
  → Ingest + Classify
  → Analyze (pruned CV packet + platform skill/llms + post → fit, gaps, angle, confidence)
    Guard: cost, rate, basic validity
  → Decide (structured: pursue? score? prep now? pause for user? why?)
  → Prep (if cleared): letter, cv-delta, research, outreach (guarded)
  → Present + Track (UI/MCP)
  → Export / Apply (user or approved agent action)
  → Outcome → Learn
  → Promote (if pattern or explicit): guarded path to external CV repo
  → Surplus

Every arrow has a guard or pause hook.

Pause / Intervention Points (UI toasts/dialogs + MCP ask_user tool):

• Before expensive LLM batch.
• On fit 60–85 or "interesting but stretch".
• Before any promote.
• On rate limit approaching or cost budget hit.
• On low confidence decision.

Implementation Patterns

Fission: Prune CV to "CorePacket"; prune platform posts; token budget on every prompt; structured output schemas.

Fusion: Treat as FinderReactor state machine; cross-cutting guards (CVPromoteGuard, CostGuard, RateGuard, DecisionLogger, SurplusGenerator); persist "what worked".

MCP / Skill Exposure: Key entry points as MCP tools (search, analyze, prep, decide, promote). Document in root SKILL.md (agentskills.io style).

Guardrails (Non-Negotiable)

• Never send full raw CV JSON or unpruned threads to LLM unless user explicitly asks for "deep mode" with confirmation.
• Never write to external profile CV without sidecar + preview + two explicit confirms.
• Always surface rate limit headers and estimated spend.
• All autonomous "decide" outputs must include confidence (0–100) and intervention level.
• Reactor dev: follow bdd-strategizer for guard + pause logic.

Activation for Agents

• Architecture / new guard design: load finder-reactor + fusion-sage.
• Specific guard, prompt, or MCP tool: ai-optimization + cv-promote-guard, x-agent-resources, tauri-agentic.
• Parallel Rust vs React vs prompts: agent-orchestrator + git-worktrees.

Surplus Generation (Mandatory After Reactor Changes)

⚡ Finder Surplus (Q ≈ X.X)
This change would have [saved Y tokens / avoided Z bad decisions] on recent opportunities.
Concrete future win: [specific example].
Suggested follow-up: [one compounding improvement].

Track in .agents/skills/finder-reactor/surplus-log.md or fusion-state.

Related

x-agent-resources, cv-promote-guard, tauri-agentic, ai-optimization, fusion-sage, agentic-reactor.

Example provenance: battle-tested on collab-finder.