[ppgranger]

Repeated pytest and Ruff runs resend unchanged failure details to the agent. Add opt-in Delta comparisons: new and changed diagnostics retain their complete evidence, unchanged failures keep their identity and summary, and retained details can be retrieved without rerunning the command. Bump the source version, Claude/Antigravity manifests, benchmark header, and changelog to 3.0.0.

Behavior and boundaries

• Delta is experimental and disabled by default. It applies to supported single pytest/Ruff commands in the Claude wrapper with an explicit session; ordinary Antigravity and portable compression remain stateless.
• Keep parsing in processors, deterministic masking in its own module, bounded private SQLite storage in the store, and retrieval in the CLI. Classify Delta observations once through a pure comparison used by rendering and detail preservation.
• Move shell eligibility into an injectable host-independent policy with one evaluator for interception and explanation. Core, wrapper, and CLI explanation no longer import the Claude JSON hook; 545 fixture commands retain exactly the same decisions and explanations.
• Separate optional telemetry from compression with lazy UTF-8 auditing and a write-only persistence contract. Close writers and query connections after failures; avoid logging record/traceback dumps on audit write errors.
• Move update orchestration and its Git/archive/install operations into a dedicated adapter while retaining CLI entry points and update behavior. Tests use temporary profiles and simulated network/subprocess dependencies.
• Extract pure statistics formatting, pass CLI arguments explicitly without changing sys.argv, and share one SQLite schema across initialization and recovery so indexes and WAL remain intact. Document the extension contracts and remaining configuration couplings.
• Retain sanitized snapshots for 24 hours / 100 runs by default; only trusted global configuration or environment can opt in or change retention. Preserve custom processor redactions and reject ambiguous output.
• Harden pytest traceback/parameter handling, deduplicate proven repeated exception summaries, handle clock rollback and concurrent SQLite sidecar removal, and safely reject malformed stored JSON. Regression tests also cover extension failures, URL passwords with empty usernames, duplicate test identities, and installed-tree imports/execution.
• Recognize Windows Python .exe launchers and quoted paths while preserving interactive and recursive-wrapper exclusions; handle Windows directory access errors without masking genuine file permission failures. Real execution tests cover paths containing spaces and virtual environments.
• Promote Delta with runnable synthetic and real-command fixture benchmarks. The five-run real-tool fixtures save 32.5% of pytest output characters and 8.3% of Ruff output with one targeted read included; a full Ruff snapshot read costs more overall. These are fixture output measurements, not billing or agent-success claims.

Local validation

• Full suite on Python 3.12.8 and 3.14.7: 1,929 passed, 12 skipped on each, with EncodingWarning treated as an error.
• Ruff 0.15.15 lint/format, structural Python style, Pylint 4.0.5, and mypy pass.
• Version consistency check passes for 3.0.0.
• Jekyll build and rendered metadata, internal links, and sitemap checks pass.
• Copied-install smoke verifies default-off behavior, exactly-once command execution, comparison, and targeted retrieval outside the source checkout.

CI validation

All nine checks pass on commit 41680e8266d1ff63821ded42609b95a4f4c45a07: Linux with Python 3.10–3.13, macOS and Windows with Python 3.12, style/types, version consistency, and documentation/SEO. CI run.

This remains a draft for review; it does not publish a release or merge into main.