v3.0.0

Token-Saver 3.0.0

Less repetition. More context to move forward.

Version 3.0.0 introduces Delta, strengthens SOLID architecture, and improves reliability across Linux, macOS, and Windows.

New features

• Delta for pytest and Ruff in Claude Code identifies new, changed, and unchanged diagnostics between consecutive runs.
• New and changed failures retain their full details; repeated failures are summarized when doing so reduces output.
• Explicitly passing tests are distinguished from failures that simply disappear from the latest output.
• delta show retrieves retained details without rerunning commands; delta clear removes the snapshot history.
• Sanitized snapshots are stored locally, with default retention limits of 24 hours and 100 runs.
• The portable compress command processes captured output, with JSON output and estimated token-budget checks.
• The replay command validates compression savings and diagnostic preservation against your own fixtures.

Architecture and maintainability

• Clearer separation between routing, compression, Delta comparison, storage, logging, statistics, and updates.
• Injectable dependencies, focused contracts, and pure comparison and formatting functions.
• Routing decisions and explanations now share a single evaluator.
• Existing public entry points remain compatible.
• Expanded README, architecture documentation, and contributor guidance.

Fixes

• Improved handling of pytest tracebacks, parametrized tests, and ambiguous results.
• Preserved secret redaction through processor chaining and compression fallbacks.
• Fixed routing for Windows Python .exe launchers and paths containing spaces.
• Ensured SQLite connections close after failures and indexes are restored during database recovery.
• Hardened snapshot handling against clock changes, invalid stored data, and concurrent cleanup.
• Prevented optional logging failures from exposing command text or raw exception details.

Validation

• 1,929 tests passed locally on both Python 3.12 and 3.14.
• All 9 CI checks passed, covering Linux, macOS, Windows, style, types, version consistency, and documentation.
• Reproducible benchmarks include the cost of retrieving diagnostic details.

Enabling Delta

Delta remains experimental and disabled by default. Enable it before launching Claude Code with export TOKEN_SAVER_DELTA_ENABLED=true.

It currently supports eligible single pytest and Ruff commands. Savings depend on the output; new or changed diagnostics may require more context to preserve their full details.