Durable agentic context for a repository • reproducibleai

Why durable context?

An AI-assisted development session is most useful when the agent can quickly find current goals, architecture, decisions, contracts, and workflows. Full transcripts and copied instruction manuals consume context while allowing important facts to drift.

The reproducibleai standard uses:

a concise root AGENTS.md for standing rules and routes,
purpose-specific maintained artifacts under dev/,
checkpoints only for useful resumable state, and
Git for superseded history.

New repositories

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

The function preflights every target. It creates missing files, skips identical files, and refuses to overwrite differing repository-owned content.

Existing repositories

Migration is intentionally two-phase:

plan <- plan_agentic_context_migration(".")
plan$operations

result <- apply_agentic_context_migration(plan, approved = TRUE)

The plan is a durable, inspectable crosswalk. Application is blocked by any manual-review operation. When approved, it creates replacements, copies mapped legacy artifacts into their new routes, validates the structure, and removes superseded tracked sources last.

Validation

report <- validate_agentic_context(".")
report$valid
report$findings

# Suitable for CI:
validate_agentic_context(".", strict = TRUE)

Validation checks structure, required routes, the installed standard version, and changes relative to seeded file hashes. Local edits are warnings rather than errors because scaffolded files become repository-owned after creation.