LLM Bidding System — AI skill for Claude Code

llm-bidding-system

[](https://github.com/Aidan2111/llm-bidding-system/actions/workflows/test.yml) [](https://github.com/Aidan2111/llm-bidding-system/actions/workflows/security.yml) [](https://github.com/Aidan2111/llm-bidding-system/releases/latest) [](LICENSE) [](pyproject.toml)

An auction router for LLM work. Multiple LLM agents (e.g. Claude Opus, Claude Sonnet, a GPT model, or a local Ollama model) **bid** on a piece of work, and a configurable utility function picks the winner. Each bid combines three ingredients:

1. A live self-assessment — each model is shown the task and asked for a structured bid: confidence (0–1), a brief approach, estimated tokens, and an effort class.
2. Deterministic risk context from agent-autonomy-score — the task's intent is scored 1–10 into a risk band (Low / Medium / High Risk) with named signals; the bidders see it, and it drives the risk-fit term.
3. Historical track record — every auction, bid, and reported outcome is stored in SQLite. Win rates, calibration-adjusted confidence, and per-band success rates feed back into future auctions.

This repo is an **auction router with supervised execution helpers**: it picks the winning model and tells you why. It can also ask a configured coding actor to propose a patch, but it does not apply code by itself. A human or supervising agent reviews, applies, edits, tests, and then reports the outcome back, which closes the feedback loop.

              ┌─────────────────────