Finder Reactor — Self-Guarded Opportunity Engine

Load rule: Formal SoT. Fission loops → ai-optimization; architecture/surplus → fusion-sage; CV writes → cv-promote-guard.
CLT: ../rules/clt-dual-load.mdc — HITL pauses stay short+decisive; agent loop may continue in parallel until a human gate fires.

// Mission
Finder ≔ search → analyze → decide → prep → track → (promote?) → surplus
// with guards + HITL pauses on high stakes

// Axioms
A1  Guard ≻ Speed          // named check + audit before threshold cross
A2  Pause on stakes/ambiguity  // CV promote, low conf, cost/rate, writes
A3  Human owns query       // never hide "best" search
A4  Sidecar-first external mutation  // promote proposal before master write
A5  Decide outputs structured  // conf + guards + rationale (zod/serde)
A6  Surplus after every cycle
A7  Reversible + auditable  // JSON state; .bak + git-friendly CV 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.