v4.0.0

Cherries v4.0.0 records ordinary Python experiments in a local SHA-256 content-addressed collection, with foreground tools for review, analysis, and rclone archival.

Breaking changes

• Call cherries.input, cherries.output, cherries.log_output, and cherries.temp inside main; keep module imports and definitions passive. Inputs and declared outputs are staged in the run work folder.
• Retained payloads now live in the configured Cherries collection. Git owns experiment scripts, configuration, and authored reports. Previous run folders require explicit legacy import and do not become replay-verified native records.
• Comet is opt-in through CHERRIES_COMET=1; normal runs do not create Git commits. Execution failure payloads may be discarded according to the configured policy, while successful runs start unreviewed.
• The local recorder supports Linux and Python 3.12 or newer.

Records and evolving documentation

Each saved run retains its source and Git evidence, dirty patch, parameters, staged inputs, metrics, logs, outputs, receipt, and manifest. Source capture occurs at the main boundary after imports; the record does not claim verified replay or scientific validity.

Codex can maintain editable per-run Markdown during experiment discussions and attach each changed full version with cherries note --file. cherries --json show exposes current annotations and prior note versions alongside the frozen execution record. Ordinary notes create no dependency or retention hold; saved comparisons and figures use analysis records that protect their source runs. Analyses retain revision names, sequence numbers, and links to previous revisions.

Archive/restore and metadata sync run on demand through rclone. Checkpointed metadata events are verified against their saved SHA-256 proofs. Record reads hold a lease throughout access, and records with dependents cannot be discarded.

Install and Codex workflow

uv add liblaf-cherries==4.0.0

The run-cherries-experiments skill is maintained in this repository and installed separately from the Python package. Existing local skill installations need an update to gain the automatic run-memory workflow.

Validation

116 tests passed in each of six Linux configurations: Python 3.12, 3.13, and 3.14 with highest and lowest direct dependency resolutions. Documentation, wheel and source-distribution builds, package metadata checks, and isolated installed-CLI checks also passed.