Grudges 0.3.0a1 — Regression tests for agent mistakes

Turn an agent mistake into a test, then check whether a new model or project rule fixes it.

uvx --from git+https://github.com/alexdevmotion/grudges@v0.3.0a1 grudges run

Already signed in to Codex or Claude? That runs one bundled example once. Choose Claude with --agent claude; test your rules with --rules ./AGENTS.md or --rules ./CLAUDE.md.

This release makes subscription replay the default, adds native verification for all 14 bundled tasks, simplifies task authoring and refreshes the README and animated walkthrough. API/Harbor replay remains an explicit alternative near the end of the docs.

Validated: 81 tests passed on Python 3.11 and 3.13 (one optional test skipped), all 28 native baseline/reference controls passed, and both live CLIs passed rules-delivery canaries. Release archives and installed-wheel tests passed.

Linux required for this alpha. Task verification runs on your host; use reviewed tasks. Subscription allowance applies. These examples demonstrate the workflow, not general model rankings.

Grudges 0.2.0a1 — Use your Codex or Claude subscription

Replay failures with your existing Codex or Claude subscription.

uv tool install https://github.com/alexdevmotion/grudges/releases/download/v0.2.0a1/grudges-0.2.0a1-py3-none-any.whl
grudges run path/to/task --runner local-codex --trust-local -k 1
grudges run path/to/task --runner local-claude --trust-local -m sonnet -k 1

• Native CLI login; no API key required, token copies, or automatic API fallback.
• Fresh sequential workspaces, frozen rules comparisons, and baseline/reference controls before model usage.
• Offline Docker verification and durable receipts for errors and missing attempts.
• Subscription usage labels, updated documentation, and simpler release instructions.

Linux tested. Requires Docker and a recent official CLI. Plan allowance still applies. Local mode is for trusted tasks: agents have host access, and Claude's restricted file tools differ from Codex's shell. These runs are not isolated cross-agent benchmarks. Harbor API replay remains available.

Validation: 78 tests passed on Python 3.11 and 3.13 (one optional test skipped); all 28 bundled controls passed; both live runners passed baseline/rules canaries. Wheel/sdist, installed-wheel suite, secret scan and desktop/mobile cards checked. See RELEASE-CHECKS.md for scope.

Grudges 0.1.0a1

First public alpha of Grudges: save coding-agent failures, turn them into replayable Harbor tasks, and compare models or project rules.

• Local capture for Claude Code and Codex, with private snapshots and transcripts.
• Frozen replay inputs, API-key-only native harness execution, and explicit error/missing-attempt accounting.
• Attempt-level HTML reports and 14 bundled example tasks with fairness notes.

Validation: 59 unit tests passed on Python 3.11 and 3.13; all 28 nop/oracle Docker controls passed their expected outcomes. Linux tested. The example packs are workflow examples, not a representative model leaderboard. No repaired-input claim is made for the historical private rules experiment.

Install the attached wheel with uv tool install ./grudges-0.1.0a1-py3-none-any.whl. Replay additionally requires Harbor 0.24.0 and Docker with Compose/Buildx; see the README. Not published to PyPI.

Artifacts include wheel, source distribution, source ZIP and SHA-256 checksums. Treat captures and raw replay outputs as private; review content before sharing.