Recon Dispute Agent — AI skill for Claude Code

recon-dispute-agent

[](https://github.com/jatin-gl/recon-dispute-agent/actions/workflows/ci.yml) [](https://www.python.org/) [-d97757)](https://www.anthropic.com/) [](LICENSE)

An **AI agent that investigates payment reconciliation discrepancies**. It takes the output of the [**payment-reconciliation-engine**](https://github.com/jatin-gl/payment-reconciliation-engine) — a list of transactions that don't line up between a PSP settlement and an internal ledger — and, for each one, runs a tool-using investigation with Claude: it pulls the PSP view, the ledger view, and the event trail, reasons to a root cause, and proposes a resolution. A second **verifier** agent then checks the conclusion against the evidence, and anything unproven is **escalated to a human**.

**The story.** The engine answers *"what doesn't reconcile?"* This agent answers *"why, and what should we do about it?"* — the judgment work a payments-ops analyst does by hand, framed as an agentic workflow with tool use, an adversarial verification pass, and a bias toward escalation over guessing.

It runs **fully offline with no API key** (a deterministic brain stands in for the model), which is how the whole thing is unit-tested end to end — then swaps to real Claude with a single flag.

---

What it demonstrates

• A real agentic loop over custom tools — evidence gathering, a schema-valid terminal tool call to finalize, bounded steps.
• A verifier / critic pass — an independent skeptical review of each proposed resolution, the standard guard against plausible-but-wrong LLM output.
• Escalation under uncertainty — UNKNOWN / MANUAL_REVIEW instead of a confident g