Changelog • atlas

atlas (development version)

New $add_budget(steps, seconds): explicitly extend an exhausted session’s mechanical budget; the extension persists for resumes.
$tell() on an out-of-budget session now fails fast in R - nothing is sent to the LLM - with the remedy in the error message.
Fixed: a warning inside an agent code chunk (e.g. glm.fit: fitted probabilities numerically 0 or 1 occurred) aborted the rest of the chunk, so later statements - including the atlas_models assignment - silently never ran. Warnings are now collected into the output and execution continues, matching console behaviour.
New record_attempt agent tool: atlas keeps a live, mechanical tally of every model/tweak the agent evaluates - printed as a one-line scoreboard, compared against the best so far with stopping_tolerance, answered with a KEEP/DISCARD verdict, and counted toward stopping_rounds (the agent is told, in code-verified terms, when the stopping rule triggers). The tally persists to tally.csv and is returned as res$tally.
New atlas_message(dir, text): steer any running session - including fully autonomous ones - from another R session or terminal. The message is delivered with the agent’s next code execution as its highest-priority instruction.

Initial release.

atlas() builds up to n_models candidate models with an LLM agent: distribution-driven planning, plan approval, held-out comparison, feature refinement of the winner, and a written report. Stopping is controlled by patience and min_improve.
Domain knowledge as enforced constraints: constraint(), con_uses(), con_monotone(), natural-language extraction via extract_constraints(), automatic repair rounds, and a compliance table in the results.
Feature safety: exclude removes columns before the agent sees the data; atlas_leakage_screen() flags predictors that alone explain almost all of the outcome.
Persistent sessions: AtlasSession checkpoints every code chunk and conversation turn to a run directory; atlas_resume() reconstructs a session exactly by replaying the code log. Builds can be interrupted and steered mid-run (interject).
Optional extras: a point-and-click front-end (atlas_app(), requires ‘shiny’) and a validation workup for the winning model (requires ‘modelblueprint’).