Skip to contents

reproducibleai

reproducibleai creates, migrates, and validates durable repository context for reproducible AI-assisted development. Instead of copying chat instructions and preserving transcripts as working context, it creates a concise root AGENTS.md that routes work into maintained artifacts under dev/.

The standard separates standing instructions from goals, architecture, decisions, governance, workflows, schemas, feature specifications, and concise checkpoints. Git preserves superseded history. The package can also measure how effectively Codex follows those routes across repeated isolated runs.

Installation 

# install.packages("pak")
pak::pak("MVR-GIS/reproducibleai")

The core package is local-first. Scaffolding, migration planning and application, structural validation, competency-question authoring, offline scoring, summaries, and report rendering do not require a Codex installation, cloud SaaS account, or administrator rights.

Live agentic-routing evaluation is an optional connected capability. It requires:

  • a separately installed Codex CLI;
  • an authenticated Codex CLI session;
  • outbound connectivity to the Codex service; and
  • permission under the computer’s application, endpoint, identity, and network policies.

On Windows, OpenAI’s standalone installer installs into the user profile and does not require npm:

powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"

Restart the R IDE after installation, then run:

The diagnostic also discovers the standard user-local Windows installation when an IDE has a stale PATH. It does not install software, authenticate, contact a model, or consume model usage. On managed or disconnected equipment, live execution may remain prohibited even when package installation succeeds. reproducibleai reports that limitation and retains its offline functionality; it never attempts to bypass organizational controls.

Scaffold a new R package 

library(reproducibleai)

use_agentic_context(
  path = ".",
  profiles = c("base", "r-package")
)

validate_agentic_context(".")

Scaffolding is deterministic and idempotent. Existing repository-owned content is never overwritten.

Migrate an existing repository 

plan <- plan_agentic_context_migration(".")

# Review the structured operations before changing the repository.
plan$operations

result <- apply_agentic_context_migration(
  plan,
  approved = TRUE
)

Planning is read-only: legacy content remains in place. Application creates and validates replacements before removing superseded files. Destructive operations are refused when a source is untracked, has uncommitted changes, or requires manual review.

Profiles

Standard 0.1 includes:

  • base, which applies to every repository
  • r-package, which adds routes and workflow guidance for R package development

detect_agentic_context_profiles() reports evidence for candidate profiles but does not silently select or apply them.

Scope

The agentic-routing evaluation API defines private competency questions, runs fresh read-only codex exec sessions, scores evidence routing and expected answer content, and writes aggregate health reports.

The user workflow is presented as three connected articles:

  1. Review and freeze the benchmark
  2. Run the evaluation
  3. Interpret routing health
check_agentic_routing_prerequisites(path = ".")

benchmark <- derive_agentic_routing_questions(".")
print(benchmark) # every generated question begins as pending

review_dir <- file.path(tempdir(), "reproducibleai-routing-review")
write_agentic_routing_review(benchmark, review_dir)

# Edit questions.csv, inspect exclusions.csv, and then:
benchmark <- apply_agentic_routing_review(benchmark, review_dir)
benchmark_path <- file.path(tempdir(), "reproducibleai-benchmark.json")
write_agentic_routing_benchmark(benchmark, benchmark_path)
benchmark <- read_agentic_routing_benchmark(benchmark_path)

evaluation <- run_agentic_routing_evaluation(
  path = ".",
  questions = benchmark,
  repetitions = 5,
  approved = TRUE
)

health <- summarize_agentic_routing(evaluation)
write_agentic_routing_report(
  health,
  "dev/governance/agentic-routing-health.md"
)

Keep private fixtures and raw runs outside the evaluated repository. Freeze reviewed generated benchmarks outside the repository with write_agentic_routing_benchmark() so specification variants use the same criterion. Model responses remain stochastic; health reports summarize repeated observations rather than claiming deterministic model behavior. Live evaluation is intended as an explicitly initiated local development diagnostic, not as an automatic package-build or CI step. A committed aggregate report can communicate the observed health of repository context to maintainers and end-users.

Bug reports

Please open an issue at https://github.com/MVR-GIS/reproducibleai/issues.