AutoOPT

AutoOPT is a human-gated, repository-grounded workflow for optimization research. It coordinates numerical algorithm design through performance estimation, approved symbolic reasoning, and Lean verification while preserving researcher control over every consequential transition.

Installing this plugin once makes the complete eight-skill AutoOPT workflow available together. Installation does not provision external runtimes, commercial solvers, browser sessions, accounts, authentication, or host capabilities. Those prerequisites remain separate and are checked by the relevant skill before use.

AutoOPT pipeline

Included skills

auto-opt: orchestrates the human-gated AutoOPT pipeline.
research-repo-manager: grounds work in a versioned research repository and maintains evidence.
bnb-pep-skill: formalizes and implements Branch-and-Bound Performance Estimation Problems.
frontier-llm-consult: routes approved symbolic-discovery and proof-planning consultations.
chatgpt-pro-session: maintains an attended, reusable ChatGPT Pro research session.
chatgpt-pro-handoff: packages, previews, submits, monitors, and imports an attended consultation.
solve-with-highest-reasoning: runs a configurable-duration native reasoning campaign.
lean-verify: builds and checks Lean artifacts for optimization results.

The eight skills are one dependency closure. In particular, frontier-llm-consult preserves both the attended external route and the explicitly selected native reasoning route.

When AutoOPT is installed as a plugin, do not separately install bnb-pep-skill. The plugin already provides the matching packaged copy, and a standalone installation can mask it or drift to a different version.

Human-gated workflow

Ground the task and repository state before research work begins.
Construct and audit the numerical or performance-estimation problem.
Ask for approval before selecting a symbolic-reasoning route, exposing repository context, or uploading material externally.
Treat generated formulas and proofs as candidates until they are checked.
Ask for approval before formal verification and preserve the resulting evidence and qualifications.

The orchestrator never interprets plugin installation as permission to upload data, weaken a research gate, switch reasoning routes silently, or claim that a solver or Lean check succeeded.

Runtime prerequisites

Core skill scripts require Python 3.10 or newer.
BnB-PEP construction and execution require Julia 1.10 or newer, the project environment, JuMP-compatible packages, and an applicable solver. Commercial solvers require separate installation and licensing.
The attended external Stage 2 route requires a separately installed Chrome-control capability, the Codex Chrome extension connection, an authenticated ChatGPT Web session, and availability of the requested ChatGPT model and configuration.
The native Stage 2 route requires a host with the necessary reasoning, persistence, repository, subagent, and local-tool capabilities. The user must confirm the campaign duration before initialization.
Lean verification requires Lean, Lake, the required mathlib project, and any separately approved Comparator or Landrun tooling used by the verification profile.

When a prerequisite is unavailable, the applicable skill stops and explains what is missing. It does not silently select a different route.

Installing from GitHub with Codex

No download or manual clone is required. From any directory, run these two commands to install the plugin directly from the public GitHub repository:

codex plugin marketplace add https://github.com/Shuvomoy/AutoOPT
codex plugin add autoopt@autoopt

The first command asks Codex to fetch and register the marketplace stored in the repository. The second command installs the autoopt plugin from that marketplace.

Start a new Codex task after installation or reinstallation so that the task loads the current plugin contents. During development, invoke AutoOPT through the plugin-qualified entry shown by Codex to ensure that a separately installed standalone skill is not masking the packaged copy.

For local development in the canonical standalone AutoOPT-plugin distribution wrapper, register that wrapper's autoopt-local marketplace:

codex plugin marketplace add .
codex plugin add autoopt@autoopt-local

Plugin maintainers must update the portable and Codex manifest versions together. From the canonical distribution-wrapper root, use the atomic updater, validate the complete package, reinstall it through autoopt-local, and then start a new task:

python3 scripts/update_autoopt_plugin_cachebuster.py
python3 scripts/build_autoopt_plugin.py --write
python3 scripts/build_autoopt_plugin.py --check
python3 scripts/validate_autoopt_plugin.py
codex plugin add autoopt@autoopt-local

The Agent Plugins 1.0 specification defines the portable package layout but does not standardize one universal installation command. Other compatible clients may use a different local or marketplace installation flow.

Support tiers

Support claims are intentionally conservative and distinguish discovery from end-to-end behavioral validation.

Tier A covers Codex desktop and CLI, the primary targets for full structural and behavioral testing.
Tier B covers supported ChatGPT plugin surfaces. The package is compatible, subject to the tools and interaction capabilities exposed by the host.
Tier C covers VS Code and GitHub Copilot, Cursor, and Kiro. Only structural discovery and harmless entry-point tests are claimed until stage-level behavior is validated.

Claude Code native plugin installation is not claimed in version 0.1.0. The underlying skill material may be adapted separately in the future.

Development and provenance

The packaged skills/ tree is a generated release mirror of the authoritative AutoOPT skills. It is not a third synchronization authority. Before a release, the build reconciles the authoritative skill locations, copies an exact allowlist, applies narrowly defined portability overlays, and records source hashes and build provenance in SOURCE-MANIFEST.json.

The generated payload receives only declared, hash-checked packaging overlays: portable command paths for the two ChatGPT skills, plugin-aware BnB-PEP onboarding, punctuation normalization, and a bounded repository-manager subtitle. Any further packaged-source deviation must be documented and validated.

Local builds from a dirty source tree are suitable for smoke testing only and must record that state. Published releases must come from a clean committed source and reproduce byte for byte. The release validation command

python3 scripts/validate_autoopt_plugin.py --require-clean-source

requires both the authoritative Skills/ sources and the AutoOPT-plugin/ release inputs to satisfy the clean-release gate.

License

AutoOPT is licensed under the Apache License 2.0. See LICENSE and NOTICE.