CLAUDE.md — Claude Code project memory for finnts

@AGENTS.md

Claude-specific working agreements

Start by summarizing the plan (bullets) and the files you’ll touch.
Prefer running devtools::test() after code changes, and devtools::check() before finalizing.
If you change roxygen comments, run devtools::document() and include generated diffs.
Keep changes minimal: avoid drive-by refactors unless explicitly requested.
For multistep changes, test all supported date frequencies plus exact submodel routing, training rows, and lag-feature eligibility; the required daily multistep matrix runs in the standard suite.
Keep CRAN test runs deterministic and under the documented runtime budget. Preserve all existing skip_on_cran() markers, including non-network tests. Non-CRAN runs automatically execute all credentialed provider and agent stress tests.
Reuse directly fitted workflows or isolated artifact fixtures in downstream tests instead of repeating full forecast pipelines.
If a request is ambiguous, ask the smallest number of clarifying questions needed, then proceed conservatively.

Agent code must treat agent_info$llm as an immutable Chat template and create a fresh, empty-history session for every graph that can access the LLM, including EDA, forecast update, each time-series combo, and Q&A. Parallel code must clone inside the series workflow after foreach dispatch, preserve the ellmer 0.4+ serialization path, and avoid internal ellmer APIs or provider credential mutation.
PSOCK tests under devtools::test() may need a test-only .export for internal helpers because pkgload::load_all() differs from an installed namespace. Verify the installed package in a temporary library before adding .export to production code.
If this repo grows, move large topic rules into .claude/rules/*.md (testing, style, release).