Open-source agentic coding plugins by Inferal. For Claude Code and Codex CLI.
Validate assumptions in implementation plans before execution.
# Claude Code
claude plugin install derisk@agentic-plugins
# Then use:
/derisk:derisk path/to/plan.mdThe skill walks you through:
- What makes you nervous about this plan? (Socratic exploration)
- Systematic assumption extraction
- Hypothesis formulation (you choose which to test)
- Experiment design with quality gates (in subagents)
- Execution (parallel where possible)
- Findings and proposed plan amendments with evidence
Experiments are stored in .inferal/derisk/{plan-name}/experiments/ as self-contained
executables (any language). Results go to .inferal/derisk/{plan-name}/findings.md.
Decide who writes what in an implementation plan.
# Claude Code
claude plugin install pair@agentic-plugins
# Then use:
/pair:pair path/to/plan.mdThe skill:
- Suggests an involvement level (hands-on / balanced / guided) based on plan characteristics
- Classifies each step as
aioryours - You override any classification with justification
- Annotates the plan with
<!-- pair:ai -->/<!-- pair:yours -->markers - During execution: AI scaffolds
yourssteps, you write the core logic
Custom involvement levels can be defined in .inferal/derisk/config.md.
Build an evidence chain for a claim. Not just for code: works for any factual statement where "sounds right" isn't enough.
# Claude Code
claude plugin install verify@agentic-plugins
# Then use:
/verify:verify "Stripe supports idempotency keys on POST in API version 2024-12"
/verify:verify # extract candidate claims from the last turn
/verify:verify --reverify <claim-slug> # re-run a prior packageThe skill walks you through:
- Pick the statement(s) to verify (from the last turn, a file, or supplied directly)
- Socratic decomposition: if this were true, what would we observe? How would we know if it's false?
- Source hunt with explicit credibility tiers (A: specs/RFCs, B: primary source code/papers, … E: community Q&A, F: hearsay)
- Hard-evidence push: if a runnable test can settle it, write one with a negative control
- Verdict (VERIFIED / LIKELY_TRUE / MIXED / UNCLEAR / LIKELY_FALSE / REFUTED / UNVERIFIABLE) with the full evidence chain
- Re-verification: the package stores machine-readable metadata and content hashes so future re-runs can detect when cited sources have changed
Packages live in .inferal/verification/{claim-slug}/ and include the claim,
ranked sources with excerpts, any experiments, and a dated verdict.
Add the marketplace, then install whichever plugins you want:
claude plugin marketplace add inferal-oss/agentic-plugins
claude plugin install derisk@agentic-plugins
claude plugin install pair@agentic-plugins
claude plugin install verify@agentic-pluginsgit clone https://github.com/inferal-oss/agentic-plugins.git ~/.agentic-plugins
mkdir -p ~/.agents/skills
ln -s ~/.agentic-plugins/plugins/derisk/skills ~/.agents/skills/derisk
ln -s ~/.agentic-plugins/plugins/pair/skills ~/.agents/skills/pair
ln -s ~/.agentic-plugins/plugins/verify/skills ~/.agents/skills/verify
# Restart CodexSee .codex/INSTALL.md for details.
Each plugin stores state under .inferal/ at the repo root:
.inferal/derisk/ # derisk + pair
├── config.md # Repo learnings, preferences, custom pair levels
└── {plan-name}/ # Per-plan derisking
├── hypotheses.md
├── experiments/
│ └── {name}/
│ ├── README.md
│ └── run.{ext}
├── findings.md
└── amendments.md
.inferal/verification/ # verify
├── config.md # Tooling notes for this repo
└── {claim-slug}/ # Per-claim evidence package
├── claim.md # Claim + Socratic decomposition
├── sources.md # Ranked source list with excerpts
├── experiments/ # Optional runnable checks
│ └── {name}/
│ ├── README.md
│ ├── run.{ext}
│ └── evidence/ # Raw captured output
├── evidence.json # Machine-readable metadata for re-verify
└── verdict.md # Verdict + reasoning (append-only)
Add .inferal/ to .gitignore or commit it: your choice. Experiments and
evidence packages are valuable to keep around.
These plugins were built using their own methodology. The .inferal/derisk/initial-implementation/
directory contains 9 derisking experiments from the creation of this repo.
MIT