Borda/AI-Rig

Practical agent workflows for Python, ML, and open-source maintenance. AI-Rig turns recurring work—scoping a change, reproducing a bug, reviewing a pull request, running an experiment, or checking release readiness—into explicit workflows with specialist ownership, evidence gates, and reviewable artifacts.
Seven packages serve two runtimes: six marketplace plugins for Claude Code, three for Codex, with codemap-py and bridge_CC-Codex shared by both.
Contents
- Start here
- What this setup enables
- Complete capability map
- Installed-all blueprint
- Install for Claude Code
- Install for Codex
- What each plugin solves
- Practical workflow sequences
- How the packages compose
- Optional integrations and add-ons
- Artifacts and evidence
- Current boundaries
- Repository checkout and synchronization
- Upgrade and remove
- Troubleshooting
- Documentation map
- Contributing
⚡ Start here
Choose the smallest plugin that solves the job in front of you. Every package installs independently; optional integrations add specialist depth or structural context without becoming hidden prerequisites.
| Need | Start with | What it gives you |
| ----------------------------------------------------------------------- | --------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------- |
| Keep Claude agent configuration, routing, and reusable guidance healthy | 🏭 foundry | 10 specialist agents, 11 configuration/content skills, rules, hooks, audit, and calibration |
| Triage GitHub work, review PRs, resolve feedback, and prepare releases | 🌱 oss | 4 agents and 5 maintainer skills with human-reviewed replies and no automatic publish step |
| Plan, build, fix, refactor, debug, or review Python changes | 🛠️ develop | 6 validate-first development workflows plus setup, with reproduction/demo/safety-net gates |
| Plan and run literature-grounded ML experiments | 🔬 research | 2 research agents and 10 skills for planning, judging, running, verifying, ablation, retrospectives, and Kaggle notebooks |
| Answer import, caller, coupling, rename, and affected-test questions | 🗂️ codemap-py | The same 6 structural-analysis skills for Claude Code and Codex, backed by a local static Python index |
| Use evidence-first workflows and specialist role cards in Codex | 🤖 Codex Rig | 13 workflow skills, 1 lifecycle manager, 15 role cards, shared gates, calibration, and artifact contracts |
| Hand bounded implement, advise, and review calls between the two hosts | 🌉 bridge_CC-Codex | 3 bridge verbs in both directions with explicit models, effort, budgets, compact envelopes, and recursion safety |
If you use Claude Code, read the Claude guide. If you use Codex, read the Codex guide.
🎯 What this setup enables
Show the evidence-first value of the full suite
- A bounded route from uncertainty to evidence: Foundry audits configuration and routing, Develop gates Python changes, OSS organizes maintainer work, Research keeps experiments reviewable, Codemap-py supplies static structure, and Codex Rig provides the parallel host-native workflow.
- Scope-selected specialist review: OSS and Codex Rig use narrow context packs and mandatory QA/challenge coverage where their contracts require it; smaller or lower-risk work can stay on the parent path.
- Review-to-remediation continuity: OSS and Codex Rig save findings as artifacts, ask which work to select, group related findings, and verify the selected closure rather than silently applying every comment.
- Validate-first development: Develop's feature path requires a failing demo, its fix path requires a failing regression test, and its refactor path protects current behavior before editing.
- Metric-guarded research: Research proposes one scoped change, measures the configured metric and guard, keeps improvements, and reverts regressions as auditable Git history.
- Calibration and recovery: Foundry and Codex Rig measure workflow/instruction quality, confidence, and known leaks; those results improve the process but do not guarantee model correctness.
These are workflow contracts, not promises that every run launches every specialist, finds every dynamic dependency, or publishes a correct change. The linked plugin READMEs define each route's prerequisites, fallback, and stop behavior.
🔧 Complete capability map
This is the literal shipped skill inventory. The host guides list every agent, role, rule, and hook; the plugin READMEs document arguments, outputs, prerequisites, recovery, and edge cases.